Physics-Informed Machine Learning for Affinity Prediction: A New Paradigm in Drug Discovery

Abstract

This article explores the transformative integration of physics-informed machine learning (PIML) for predicting molecular binding affinity, a critical task in accelerating drug discovery. It provides a comprehensive overview for researchers and drug development professionals, covering the foundational principles that merge physical laws with data-driven models, the diverse methodologies and their specific applications in structure-based drug design, the significant challenges of data bias and model optimization, and the rigorous validation frameworks needed for real-world deployment. By synthesizing insights from recent advances, this work highlights how PIML offers enhanced accuracy and generalizability over conventional methods, paving the way for more efficient and reliable in silico drug development.

The Foundation: Merging Physical Laws with Data for Smarter Affinity Prediction

Defining Physics-Informed Machine Learning (PIML) in a Biochemical Context

Physics-Informed Machine Learning (PIML) represents a transformative paradigm that seamlessly integrates data-driven learning with the foundational principles of mechanistic models. In a biochemical context, PIML provides a powerful framework for developing predictive models that are both accurate and scientifically plausible [1]. This integration is achieved by embedding established physical laws, such as those governing chemical kinetics or molecular interactions, directly into the machine learning (ML) pipeline, often through the incorporation of governing equations or physical constraints as regularization terms within the learning algorithm's loss function [2] [3]. The core strength of this approach lies in its ability to leverage the pattern recognition capabilities of ML while ensuring that model outputs adhere to known biophysical realities.

This synergy is particularly valuable in biochemical domains where purely data-driven models face significant challenges, including but not limited to data scarcity, high experimental noise, and the immense complexity of biological systems [2] [3]. PIML directly addresses these issues by using physical laws to guide the learning process, which reduces dependency on large volumes of labeled data and enhances model generalizability. For researchers focused on affinity prediction—the quantitative assessment of interaction strength between biomolecules like proteins and ligands—PIML offers a path to more reliable and interpretable predictions. This is crucial for applications in drug discovery, where understanding the precise strength of molecular interactions can guide the optimization of therapeutic compounds [4] [5].

Table 1: Core PIML Frameworks and Their Biochemical Applications

Framework	Core Principle	Representative Biochemical Application
Physics-Informed Neural Networks (PINNs)	Embed governing differential equations as a loss function component during neural network training.	Parameter estimation and model reduction for Aβ fibril aggregation kinetics in Alzheimer's disease research [2].
Neural Ordinary Differential Equations (NODEs)	Model continuous-time dynamics using neural networks to represent the derivative of a system's state.	Modeling dynamic physiological systems, pharmacokinetics, and cell signaling pathways [3].
Neural Operators (NOs)	Learn mappings between function spaces, enabling solutions for families of differential equations rather than single instances.	Efficient simulation across multiscale and spatially heterogeneous biological domains [3].

Application Focus: Protein-Ligand Affinity Prediction

Accurately predicting the binding affinity between a protein and a small molecule (ligand) is a cornerstone of computer-aided drug design [4] [5]. The strength of this interaction, often quantified by biophysical parameters like the dissociation constant (K_d) or inhibition constant (K_i), determines a candidate drug's efficacy and specificity [6] [5]. Traditional methods for affinity prediction exist on a spectrum trading speed for accuracy. At one end, molecular docking is fast but often inaccurate; at the other, rigorous methods like Free Energy Perturbation (FEP) are accurate but computationally prohibitive for screening large compound libraries [7].

PIML is emerging as a powerful approach to bridge this methodological gap. It enhances prediction by moving beyond purely structural or sequence-based patterns to incorporate the physical laws that underpin molecular recognition. For instance, a PIML model might be informed by the physics of molecular forces, energy conservation, or the principles of chemical kinetics [3]. A notable example is the ProBound framework, which employs a multi-layered maximum-likelihood approach to model not just the molecular interactions but also the data generation process of high-throughput sequencing assays. This allows it to infer rigorous biophysical parameters like equilibrium binding constants directly from sequencing data, providing a more quantitative and interpretable model of protein-ligand interactions [6].

Table 2: Performance Comparison of Affinity Prediction Methods

Method	Typical RMSE (kcal/mol)	Typical Correlation (PCC)	Computational Cost
Molecular Docking	2.0 - 4.0	~0.3	Low (minutes on CPU) [7]
MM/GBSA	~1.5 - 2.5 (after entropic correction)	Variable, often moderate	Medium (hours on CPU/GPU) [7]
Free Energy Perturbation (FEP)	~1.0	0.65+	Very High (days on GPU) [7]
StructureNet (Structure-Based GNN)	Not Reported	0.68 (PCC on PDBBind)	Medium [8]
ProBound (PIML for Sequencing Data)	Quantifies affinity over a wide range, outperforming deep learning & other resources on PBM and SELEX metrics [6]	High	Varies by assay

Experimental Protocol: A PIML Approach to Aβ Aggregation Kinetics

The following protocol details the application of a Physics-Informed Neural Network (PINN) for parameter estimation in a reduced-order model of Amyloid-beta (Aβ) peptide aggregation, a key process in Alzheimer's disease pathology [2].

Background and Objective

The uncontrolled aggregation of Aβ peptides into fibrils involves complex nucleation and growth kinetics. Detailed mechanistic models are computationally expensive. This protocol uses a PINN to automatically discover a reduced-order kinetic model from transient concentration data, optimizing for both simulation efficiency and accuracy by determining the appropriate level of reaction detail [2].

Reagent Setup

Table 3: Research Reagent Solutions

Reagent/Material	Function in Protocol
Experimental Time-Course Data	Provides measured concentrations of Aβ species (e.g., monomer, oligomers, fibrils) over time; serves as the observational data for training and validating the PINN.
Reduced-Order Reaction Network	A simplified representation of the Aβ aggregation pathway (e.g., Fig. 1b in [2]), defining the system of ODEs that form the physics-based constraints.
Law of Mass Action	The physical principle used to translate the reaction network into a system of Ordinary Differential Equations (ODEs) governing the rate of change for each species' concentration.
PINN Software Framework	A computational environment (e.g., TensorFlow or PyTorch) capable of constructing neural networks and formulating custom loss functions that incorporate the ODE residuals.

Procedure

• System Definition and Data Preparation (Time: 1-2 hours)

  • Define the reduced-order chemical reaction network for Aβ aggregation, specifying all species and reactions (e.g., on-pathway vs. off-pathway) [2].
  • Translate the reaction network into a system of ODEs using the law of mass action. The system will have the form: d[Species]/dt = f([Species], α, β), where α and β are forward and backward elementary rate parameters to be estimated [2].
  • Collate experimental data, which should consist of measured concentrations of the different Aβ species at multiple time points.

• PINN Architecture Construction (Time: 1-2 hours)

  • Design a neural network where the input is time (t) and the outputs are the predicted concentrations of all chemical species in the model at that time, [Species]_pred(t).
  • Formulate the composite loss function (L_total) for training the network:
    • Data Loss (L_data): Mean Squared Error (MSE) between the network's predictions and the experimental concentration data.
    • Physics Loss (L_physics): MSE of the residual of the ODE system. This is calculated by taking the automatic derivatives of the network's outputs with respect to the input (t), and comparing them to the derivatives computed from the network's outputs using the predefined ODEs (f). The residual is R = | d[Species]_pred/dt - f([Species]_{pred], α, β) |}.
    • The total loss is a weighted sum: L_total = w_dataL_data + w_physicsL_physics.

• Model Training and Parameter Estimation (Time: hours-days, depending on complexity)

  • Initialize the neural network parameters and the unknown kinetic parameters (α, β).
  • Train the PINN by minimizing L_total using a gradient-based optimization algorithm (e.g., Adam).
  • During training, the algorithm will simultaneously adjust both the neural network weights (to fit the data) and the kinetic parameters (to satisfy the physical laws expressed by the ODEs).

• Validation and Model Reduction Analysis (Time: 1-2 hours)

  • Validate the trained model by comparing its predictions against a held-out test set of experimental data.
  • Analyze the identified parameters and the structure of the optimized model to determine the appropriate level of detail—i.e., which reaction pathways are most critical—for the reduced-order model [2].

Workflow Visualization

PINN Workflow for Model Reduction

The Scientist's Toolkit

Table 4: Essential Resources for PIML in Affinity Prediction

Tool / Resource	Type	Function in Research
PDBbind [4] [5]	Database	A curated database of protein-ligand complexes with experimentally measured binding affinity data, used for training and benchmarking models.
BindingDB [4]	Database	A public, web-accessible database of measured binding affinities, focusing primarily on drug-target interactions.
ProBound [6]	Software/Algorithm	A flexible machine learning framework for building biophysically interpretable binding models from sequencing data (e.g., SELEX).
OpenMM [7]	Software/Toolkit	A high-performance toolkit for molecular simulation, used to generate molecular trajectories for feature extraction in MM/GBSA-type approaches.
Physics-Informed Neural Networks (PINNs) [2] [3]	Modeling Framework	A deep learning architecture that encodes physical laws (ODEs/PDEs) into the learning process, enabling predictive modeling with limited data.
Random Sublattice Model Descriptors [9]	Feature Set	Physics-informed descriptors (e.g., δpbs, ΔHpbs) for predicting the stability of ordered intermetallic compounds like B2 MPEIs, exemplifying the design of domain-specific features.

Conceptual Framework Visualization

PIML Conceptual Framework

The accurate prediction of binding affinity is a cornerstone of modern drug discovery, serving as a critical determinant of a drug candidate's potency and efficacy [10]. While traditional methods rely heavily on experimental screening, the advent of machine learning (ML) has introduced powerful computational tools to accelerate this process. However, many conventional ML models operate as black boxes, often overlooking the fundamental physical laws that govern molecular interactions. This can lead to models with poor generalizability, especially on unseen data or in de novo drug design scenarios [11] [10].

The emerging paradigm of physics-informed machine learning seeks to overcome these limitations by integrating core physical principles and thermodynamic laws directly into the learning process. This approach moves beyond mere pattern recognition in data, instead guiding models with the immutable laws of physics that dictate how molecules interact, bind, and release energy. By incorporating these priors—from the quantum mechanical force fields that define atomic interactions to the macroscopic thermodynamic laws that govern binding spontaneity—researchers are developing more robust, interpretable, and reliable models for affinity prediction [11] [9]. This article details the core physical principles involved and provides structured protocols for their implementation in machine learning frameworks.

Core Physical Principles in Drug-Target Interactions

The interaction between a drug (ligand) and its protein target is a complex process governed by a hierarchy of physical laws. Understanding these principles is a prerequisite for developing effective physics-informed ML models.

The Laws of Thermodynamics in Binding Affinity

The binding affinity, quantitatively represented by the dissociation constant ((Kd)) or its negative logarithm ((pKd)), is fundamentally a measure of the free energy change upon binding. The laws of thermodynamics provide the ultimate framework for understanding this process.

Table 1: Thermodynamic Laws and Their Role in Affinity Prediction

Law	Core Principle	Relevance to Binding Affinity
Zeroth Law	Defines thermal equilibrium and temperature.	Ensures binding experiments and predictions are referenced to a standard temperature (e.g., 310 K for human physiology) [12] [13].
First Law	Energy is conserved; it cannot be created or destroyed, only transformed.	The internal energy change ((\Delta U)) of the system upon binding is balanced by the heat transfer ((Q)) and work ((W)) done, typically at constant pressure, leading to the enthalpy change ((\Delta H)) [12] [13].
Second Law	The total entropy of an isolated system never decreases.	Binding is favored only if the total change in Gibbs free energy ((\Delta G = \Delta H - T\Delta S)) is negative. This requires a careful balance between favorable enthalpy (e.g., bond formation) and the entropic cost of ordering the ligand and protein [12] [13].
Third Law	The entropy of a perfect crystal approaches zero as temperature approaches absolute zero.	Provides a foundational reference for absolute entropy calculations, important for ab initio thermodynamic predictions [12].

The Second Law is particularly crucial, as the Gibbs free energy equation (\Delta G = \Delta H - T\Delta S) is the direct link between molecular-level interactions and the experimentally measured binding affinity, where (\Delta G = -RT \ln K) [12]. Physics-informed models like SPIN explicitly incorporate this by necessitating "minimal binding free energy along their reaction coordinate," building the drive toward thermodynamic equilibrium directly into the model's objective function [11].

Molecular Force Fields and Structural Biases

Beyond thermodynamics, the structural and chemical compatibility between a ligand and its target is dictated by atomic-scale forces. Molecular force fields mathematically describe the potential energy of a system as a function of its nuclear coordinates, capturing bonded interactions (bonds, angles, dihedrals) and non-bonded interactions (van der Waals, electrostatic). Integrating these concepts as inductive biases into ML models is a key strategy.

• SE(3) Invariance/Equivariance: A critical physical principle is that the binding affinity between a ligand and a protein does not change if the entire complex is rotated or translated in 3D space. Models that respect this symmetry are more data-efficient and generalize better. The SPIN model is explicitly designed as an SE(3)-Invariant Physics Informed Network, ensuring consistent predictions regardless of the complex's orientation [11]. Similarly, ATOMICA uses SE(3)-equivariant message passing to build this fundamental symmetry directly into its geometric deep learning architecture [14].
• Intermolecular Interaction Features: The stability of a complex is driven by specific, local atomic interactions, including hydrogen bonding, van der Waals forces, (\pi)-stacking, and electrostatic interactions [14]. Universal models like ATOMICA learn representations of these atomic-scale interaction interfaces across diverse molecular types (proteins, small molecules, nucleic acids, ions), capturing the shared physicochemical principles that underlie all biomolecular recognition [14].
• Geometric and Topological Descriptors: The 3D shape and geometry of the binding site and ligand are paramount. StructureNet, a model that relies exclusively on structural data, highlights that geometric descriptors are the key drivers of model performance; their removal led to a performance decrease of over 15% [10]. These descriptors encode information about molecular surfaces, pocket shapes, and steric constraints that determine binding compatibility [15].

Physics-Informed Methodologies and Protocols

This section outlines specific methodologies and experimental protocols for implementing physics-informed ML models, as demonstrated by recent state-of-the-art research.

Protocol 1: Implementing an SE(3)-Invariant Network with SPIN

The SPIN (SE(3)-Invariant Physics Informed Network) model provides a protocol for building robust affinity predictors [11].

Table 2: Key Research Reagents & Computational Tools

Reagent / Tool	Function / Description
3D Structure Files (PDB)	Input data containing atomic coordinates of protein-ligand complexes.
Graph Neural Network (GNN)	Core architecture for representing molecular structures as graphs.
SE(3)-Invariant Layer	Neural network layer that ensures output is unchanged by rotations/translations of input.
Physics-Informed Loss Function	A custom objective function that incorporates the thermodynamic requirement for minimal free energy.
CASF-2016 & CSAR HiQ Benchmarks	Standardized datasets used for training and evaluating model performance and generalizability.

Step-by-Step Protocol:

• Data Preparation: Obtain a curated dataset of protein-ligand complexes with experimentally measured binding affinities (e.g., PDBBind). Represent each complex as a graph where nodes are atoms and edges represent bonds or spatial proximity.
• Feature Extraction: For each atom node, compute invariant features such as atom type, charge, and interatomic distances (which are inherently invariant to rotation).
• Model Architecture:
  • Construct a GNN using SE(3)-invariant layers. These layers operate on the invariant features and distances, not raw 3D coordinates that can rotate.
  • The network processes the protein and ligand graphs to learn a complex representation.
• Physics-Informed Training:
  • Define a loss function with two components: (i) a standard regression loss (e.g., Mean Squared Error) between predicted and actual affinity, and (ii) a physics-based regularization term that penalizes predictions inconsistent with minimal free energy.
  • Train the model on a benchmark dataset like CASF-2016.
• Validation: Evaluate the model on independent test sets (e.g., CSAR HiQ) and through virtual screening experiments to assess its generalizability and practical utility [11].

Protocol 2: A Multitask Framework with DeepDTAGen

DeepDTAGen demonstrates a protocol that couples affinity prediction with target-aware drug generation using a shared, physics-informed feature space [16].

Step-by-Step Protocol:

• Input Representation: Encode the protein sequence and the drug's SMILES string into initial feature vectors.
• Shared Feature Learning: Process these inputs through a shared encoder to create a latent representation that captures the interaction-specific features, implicitly learning the physicochemical properties that govern binding.
• Multitask Optimization:
  • Task 1 (Affinity Prediction): Feed the shared features into a regression head to predict binding affinity values (e.g., (pK_d)).
  • Task 2 (Drug Generation): Use the same shared features as a conditioning signal for a transformer-based decoder to generate novel, target-aware drug-like molecules (in SMILES format).
• Gradient Alignment with FetterGrad: A key challenge in multitask learning is gradient conflict. To mitigate this, the FetterGrad algorithm is used during training. It minimizes the Euclidean distance between the gradients of the two tasks, keeping their learning objectives aligned and ensuring stable convergence [16].
• Evaluation: Assess the model on both tasks. For prediction, use metrics like Concordance Index (CI) and Mean Squared Error (MSE) on Davis and KIBA datasets. For generation, evaluate the validity, novelty, and uniqueness of the generated molecules, as well as their predicted binding ability [16].

Protocol 3: Structure-Only Modeling with StructureNet

For scenarios where detailed interaction data or sequence information is lacking or may lead to overfitting, StructureNet provides a protocol based exclusively on 3D structural data [10].

Step-by-Step Protocol:

• Graph Creation from PDB Files:
  • For a given protein-ligand complex, extract the 3D structure from the PDB file.
  • Create separate graphs for the protein-binding pocket and the ligand using a library like NetworkX. Nodes represent atoms, and edges represent bonds or spatial proximity within a cutoff distance.
• Structural Feature Extraction:
  • Populate nodes with atomic features (e.g., element type, hybridization state).
  • Incorporate advanced geometric descriptors, such as those derived from Voronoi tessellations, to better capture the local atomic environment and topology.
• Model Training:
  • Use a GNN-based ensemble architecture to process the protein and ligand graphs.
  • Train the model to predict the binding affinity ((pK_d)) using only these structural features, deliberately excluding sequence and pairwise interaction data.
• Ablation Analysis: Systematically remove specific feature sets (e.g., geometric descriptors) to validate their critical role. StructureNet's performance dropped by over 15% without geometric descriptors, confirming their importance [10].
• Handling Flexibility: To account for dynamic structural changes, apply the trained model to multiple conformers of the same complex generated from molecular dynamics (MD) simulations. The final affinity can be taken as the average or minimum prediction, capturing the effect of binding site flexibility [10].

Quantitative Performance of Physics-Informed Models

The integration of physical principles has led to measurable improvements in model performance and generalizability, as evidenced by benchmark results.

Table 3: Performance Comparison of Select Physics-Informed Models on Benchmark Datasets

Model	Core Physical Principle	Dataset	Key Metric	Result
SPIN [11]	SE(3)-Invariance, Minimal Free Energy	CASF-2016, CSAR HiQ	Superior generalizability vs. comparators	Outperformed comparative models in benchmark sets.
DeepDTAGen [16]	Shared Physicochemical Feature Space, Gradient Alignment	KIBA	CI / MSE	0.897 / 0.146
DeepDTAGen [16]	Shared Physicochemical Feature Space, Gradient Alignment	Davis	CI / MSE	0.890 / 0.214
StructureNet [10]	Exclusive Use of 3D Structural & Geometric Descriptors	PDBBind v.2020	Pearson Correlation Coefficient (PCC)	0.68
DrugForm-DTA [17]	Structure-less Representation based on Language Models	KIBA	High Accuracy	Performance comparable to a single in vitro experiment.

The Scientist's Toolkit

This section lists essential computational tools and datasets that form the foundation for research and development in this field.

Table 4: Key Research Reagents, Datasets, and Tools

Category	Name	Description & Function
Benchmark Datasets	PDBBind [10]	A comprehensive database of protein-ligand complexes with experimentally measured binding affinities, used for training and testing.
	Davis, KIBA [16] [17]	Standard benchmark datasets for drug-target affinity (DTA) prediction, focusing on kinase inhibitors.
	DUDE-Z [10]	A dataset containing active ligands and decoys, used for external validation and assessing a model's ability to distinguish true binders.
Software & Libraries	RDKit [10]	An open-source toolkit for cheminformatics, used for feature extraction, molecule sanitization, and graph representation.
	PyTorch Geometric [10]	A library for deep learning on graphs, providing GNN constructors and utilities essential for structure-based models.
	NetworkX [10]	A Python package for the creation, manipulation, and study of complex graphs, used to represent molecular structures.
Representative Models	ATOMICA [14]	A universal geometric deep learning model for atomic-scale representations across multiple molecular modalities (proteins, small molecules, ions, etc.).
	MaSIF [15]	A deep learning model based on molecular surface interaction fingerprinting, used for interaction site prediction and protein-protein interaction prediction.

The integration of core physical principles—from the force fields describing atomic interactions to the fundamental laws of thermodynamics—is transforming machine learning for affinity prediction. Methodologies that enforce SE(3) invariance, leverage structural and geometric descriptors, incorporate thermodynamic constraints, and learn universal representations of intermolecular interactions are demonstrating enhanced robustness, interpretability, and utility in real-world drug discovery applications, such as virtual screening and target-aware drug generation [11] [10] [14]. As these physics-informed models continue to evolve and as structural datasets expand, they offer a predictable path toward more accurate and generalizable predictive tools, ultimately accelerating the journey from conceptual target to viable therapeutic candidate.

The Limitation of Purely Data-Driven Models and the PIML Advantage

In the field of drug discovery, accurately predicting protein-ligand binding affinity is a critical step for identifying viable therapeutic candidates. [4] While purely data-driven machine learning (ML) and deep learning (DL) models have shown promise by learning complex relationships from data, their application in scientific domains like affinity prediction is fundamentally constrained by several inherent limitations. These models, including various traditional ML and advanced DL architectures, often struggle with requirements for massive, high-quality training datasets, display a "black-box" nature that yields unreliable and physically inconsistent predictions, and exhibit poor generalizability in out-of-sample scenarios. [18] [19] This article delineates the limitations of purely data-driven approaches and elaborates on how Physics-Informed Machine Learning (PIML) presents a transformative framework for robust, reliable, and physiochemically plausible binding affinity prediction.

Key Limitations of Purely Data-Driven Models

Purely data-driven models depend exclusively on patterns found within training data, lacking integration of foundational scientific principles. This approach leads to several critical challenges in scientific and engineering applications, detailed below and summarized in Table 1.

Table 1: Core Limitations of Purely Data-Driven Models in Scientific Domains like Affinity Prediction

Limitation	Impact on Model Performance & Reliability	Manifestation in Binding Affinity Prediction
Data Scarcity & Imbalance [9] [18]	Model cannot learn underlying physical relationships, leading to poor accuracy and high variance.	Limited experimental binding affinity data (~19,588 complexes in PDBBind v.2020); data biased toward successful binders, lacking negatives/weak binders. [4] [8]
Physical Inconsistency [18] [19]	Predictions may violate known physical laws, rendering them implausible and unreliable for scientific use.	Model may predict a stable ligand pose with steric clashes or an energetically unfavorable conformation.
Poor Extrapolation & Generalizability [18]	Performance degrades significantly on data outside the training distribution (e.g., new protein classes).	A model trained on kinase-ligand complexes may fail to accurately score antibody-antigen interactions.
Black-Box Nature [18]	Lack of interpretability and explainability undermines trust and hinders scientific insight.	Difficulty understanding which structural features (e.g., hydrogen bonds, hydrophobic contacts) drove a specific affinity prediction.

The challenges outlined in Table 1 are not merely theoretical. In binding affinity prediction, conventional data-driven models rely heavily on interaction and sequence data, which can lead to pattern memorization rather than genuine learning of structure-affinity relationships. [8] Furthermore, synthetic datasets are often undesirable due to inaccuracies or prohibitive computational costs, while experimental datasets are limited in size, precision, and suffer from bias toward complexes with correct poses and good binding constants. [4]

The Physics-Informed Machine Learning (PIML) Advantage

Physics-Informed Machine Learning (PIML) is a novel modeling paradigm designed to overcome the limitations of purely data-driven approaches by integrating prior physics knowledge into ML models. [18] [19] This integration enhances data efficiency, ensures physical plausibility of results, and improves model generalizability and robustness. [19] The core advantage of PIML lies in its ability to learn from both data and the rich, abstracted knowledge of natural phenomena encoded in physical laws. [19]

Methodologies for Integrating Physics into ML

The integration of physics into machine learning models can be achieved through several distinct methodologies, each manipulating a different component of the ML pipeline. These are categorized as follows and illustrated in the workflow diagram below:

• Physics-Informed Inputs: Leveraging physics-based simulated data or using physically meaningful parameters and descriptors as model inputs. [18] [19] For instance, in designing B2 multi-principal element intermetallics, random-sublattice-based descriptors such as atomic size difference between sublattices (δmean) and ordering tendency ((H/G)pbs) were used instead of classic parameters, effectively addressing data limitation and imbalance. [9]
• Physics-Informed Loss Functions: Incorporating physical laws, often expressed as partial differential equations (PDEs) or other constraints, as regularization terms in the model's loss function. [18] [19] This guides the optimization process towards solutions that are not only data-accurate but also physically consistent. For example, in mineral processing, using a physics-guided loss function that utilized mass balance significantly improved the forecasting accuracy of surrogate models. [20]
• Physics-Informed Architectural Design: Designing custom neural network architectures that inherently encode physical structure or constraints, such as using graph neural networks to represent molecular structures. [18] [19] StructureNet, a graph neural network framework, represents protein and ligand structures as graphs, which are then processed using a GNN-based model, focusing entirely on structural descriptors to mitigate data memorization. [8]
• Physics-Informed Ensemble Models: Combining predictions from independent physics-based and data-driven models to handle different components of a complex system, leveraging the strengths of both approaches. [18]

Quantitative Advantages of PIML in Practice

The application of PIML has demonstrated tangible, quantitative improvements over purely data-driven models across various fields, as shown in Table 2.

Table 2: Demonstrated Performance Improvements of PIML Models

Application Field	PIML Model	Performance Metric	Result & Advantage
Binding Affinity Prediction [8]	StructureNet (Structure-Based GNN)	Pearson Correlation Coefficient (PCC)	Achieved PCC of 0.68 on PDBBind v.2020, outperforming similar structure-based models and effectively distinguishing active from decoy ligands.
Mineral Processing [20]	PIML Surrogate Models (LSTM, GRU, CNN)	Forecasting Accuracy (NRMSE, NMAE)	All PIML models outperformed their purely data-driven counterparts. The largest improvements were observed in LSTM models.
Material Discovery [9]	CVAE + ANN with Physics-Informed Descriptors	Discovery Efficiency	Enabled high-throughput discovery of B2 complex alloys in vast compositional spaces, overcoming data limitation and imbalance (1:9 B2 to non-B2 ratio).

Application Notes & Protocols for PIML in Affinity Prediction

This section provides a practical, step-by-step guide for researchers to implement a PIML framework for protein-ligand binding affinity prediction, drawing on the methodologies of successful models like StructureNet. [8]

Experimental Protocol: Structure-Based PIML for Binding Affinity Prediction

Objective: To predict the binding affinity (e.g., Kd, Ki, IC50) of a protein-ligand complex using a physics-informed, structure-based graph neural network.

Workflow Overview:

Step-by-Step Procedure:

• Data Acquisition and Curation

  • Source: Obtain a curated dataset of protein-ligand complexes with experimentally measured binding affinities. The PDBBind database is a standard benchmark for this purpose. [4] [8]
  • Partition: Split the data into training, validation, and test sets. A common practice is to use the refined set from PDBBind for training and the core set for testing. Ensure no significant sequence or structural similarity exists between the training and test sets to properly assess generalizability.

• Physics-Informed Featurization and Graph Construction

  • Graph Representation: Represent each protein-ligand complex as a graph. The ligand is typically a fully-connected graph. Protein residues or atoms within a defined radius (e.g., 5-10 Å) of the ligand are included as nodes, and edges represent spatial proximity or chemical bonds. [8]
  • Node Features: Encode atoms with physico-chemically meaningful features such as:
    • Atom type (one-hot encoded)
    • Element symbol
    • Partial charge
    • Hybridization state
    • Number of bonded neighbors
  • Edge Features: Encode edges with features such as:
    • Spatial distance (Euclidean)
    • Bond type (single, double, triple, aromatic, or spatial)
    • Geometric and Topological Descriptors: Incorporate key descriptors like those used in StructureNet, which were identified as major drivers of model performance. Their removal led to a performance decrease of over 15.7%. [8]

• Model Architecture and Training Configuration

  • GNN Backbone: Implement a GNN architecture (e.g., Graph Convolutional Network, Graph Attention Network) for message passing and node/edge feature updating. [8]
  • Readout and Prediction: After several layers of message passing, a global readout function (e.g., mean pooling, sum pooling) aggregates the updated node features into a fixed-dimensional vector representing the entire complex. This is passed through fully connected layers to output a single scalar value representing the predicted binding affinity.
  • Physics-Informed Loss Function: Define a hybrid loss function for training: L_total = L_data + λ * L_physics
    • L_data: Mean Squared Error (MSE) between predicted and experimental binding affinities.
    • L_physics: A physics-based regularization term. This could enforce constraints derived from molecular mechanics, such as penalizing steric clashes or encouraging favorable electrostatic interactions, even if not explicitly parameterized. The weighting factor λ controls the influence of the physics constraint.
  • Training: Use standard optimizers (e.g., Adam) and techniques like learning rate scheduling and early stopping.

• Model Evaluation and Validation

  • Primary Metrics: Evaluate model performance on the held-out test set using standard metrics for binding affinity prediction:
    • Pearson Correlation Coefficient (PCC): Measures the linear correlation between predicted and experimental values.
    • Root Mean Square Error (RMSE): Measures the average magnitude of prediction errors.
    • Area Under the Curve (AUC): For tasks like distinguishing active from decoy ligands (e.g., on the DUDE-Z dataset). [8]
  • External Validation: Test the model on an entirely external dataset to rigorously assess its generalizability to new protein targets or novel chemotypes.

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Resources for PIML-based Binding Affinity Prediction

Category	Resource Name	Description & Function
Benchmark Datasets [4]	PDBBind	A comprehensive database providing 3D structures of protein-ligand complexes and their experimentally measured binding affinities. Serves as the primary source for training and testing.
	CASF	The Core Set of PDBBind, used as a standardized benchmark for objective, "blind" testing of scoring functions.
	BindingDB	A public database of measured binding affinities, focusing primarily on drug-target interactions. Useful for additional training data or external validation.
Computational Tools & Frameworks [8]	Graph Neural Network (GNN) Libraries (e.g., PyTorch Geometric, DGL)	Essential software libraries for implementing and training graph-based models on structural data.
	Molecular Dynamics (MD) Simulation Software (e.g., GROMACS, AMBER)	Used to generate ensembles of binding complex conformers, capturing binding site flexibility, which can be fed into models like StructureNet to improve accuracy. [8]
Physics-Informed Components [9] [8]	Geometric & Topological Descriptors	Structural descriptors (e.g., atomic distances, angles, surface areas) that serve as physics-informed inputs, reducing reliance on sequence/interaction data and mitigating memorization.
	Molecular Mechanics Force Fields	Provide energy terms (e.g., van der Waals, electrostatics) that can be used to formulate physics-based constraints (L_physics) in the loss function.

Performance Metrics and Quantitative Benchmarks

The performance of computational methods in structure-based drug discovery is quantitatively assessed using standardized benchmarks. The table below summarizes key performance metrics from recent studies.

Table 1: Performance Benchmarks of Scoring and Virtual Screening Methods

Method / Tool	Category	Key Performance Metric	Value	Dataset / Context
PLANTS + CNN-Score [21]	Docking + ML Re-scoring	Enrichment Factor at 1% (EF1%)	28	Wild-Type PfDHFR (Malaria target)
FRED + CNN-Score [21]	Docking + ML Re-scoring	Enrichment Factor at 1% (EF1%)	31	Quadruple-Mutant PfDHFR (Drug-resistant Malaria)
RosettaGenFF-VS [22]	Physics-based Scoring	Enrichment Factor at 1% (EF1%)	16.72	CASF-2016 Benchmark
StructureNet [10]	Structure-based Deep Learning	Pearson Correlation Coefficient (PCC)	0.68	PDBBind v.2020 Refined Set
Free Energy Perturbation (FEP) [7]	High-End Physics Simulation	Root-Mean-Square Error (RMSE)	~1.0 kcal/mol	Industry Standard
Docking (e.g., AutoDock Vina) [7]	Conventional Docking	Root-Mean-Square Error (RMSE)	2-4 kcal/mol	Common Baseline

These metrics demonstrate the significant enhancement that machine learning re-scoring provides to classical docking tools, particularly for challenging targets like drug-resistant enzymes [21]. Physics-based methods like RosettaGenFF-VS achieve high performance by incorporating receptor flexibility and sophisticated entropy models [22].

Experimental Protocols and Application Notes

Protocol: Structure-Based Virtual Screening with ML Re-scoring

This protocol is adapted from benchmarking studies against wild-type and drug-resistant Plasmodium falciparum Dihydrofolate Reductase (PfDHFR) [21].

Application Note: This workflow is designed to identify active compounds from large chemical libraries, with enhanced performance against mutated, drug-resistant targets by leveraging machine learning re-scoring.

Workflow Overview:

Step-by-Step Procedure:

• Protein Structure Preparation

  • Input: Obtain crystal structures from the Protein Data Bank (e.g., PDB ID: 6A2M for wild-type PfDHFR, 6KP2 for the quadruple mutant) [21].
  • Processing: Use a protein preparation tool like OpenEye's "Make Receptor." Remove all water molecules, ions, and redundant chains. Add hydrogen atoms and optimize their positions. The final prepared structure is saved in the required format for docking (e.g., PDBQT, OEB) [21].

• Ligand Library Preparation

  • Source: Utilize a benchmark set like DEKOIS 2.0, which contains known active compounds and structurally similar but presumed inactive decoys (typically at a 1:30 ratio of actives to decoys) [21].
  • Processing: Prepare all small molecules using a tool like OpenEye's Omega to generate multiple conformations for each ligand. Convert and split the final prepared compounds into the required file formats for docking (e.g., SDF, PDBQT, mol2) [21].

• Molecular Docking

  • Tools: Employ one or more docking programs such as AutoDock Vina, PLANTS, or FRED.
  • Grid Definition: Define the docking grid box centered on the binding site. Example dimensions for PfDHFR are approximately 21Å × 25Å × 19Å [21].
  • Execution: Run the docking simulation to generate multiple binding poses and scores for each ligand in the library.

• Machine Learning Re-scoring

  • Input: Use the ligand poses generated in the previous step.
  • Re-scoring: Apply pre-trained machine learning scoring functions, such as CNN-Score (a convolutional neural network) or RF-Score-VS v2 (a random forest model), to re-rank the docked compounds. These ML models learn complex patterns from structural data that are often missed by classical scoring functions [21] [4].
  • Output: A new, refined ranked list of compounds.

• Performance Evaluation and Hit Identification

  • Metrics: Calculate key performance metrics to evaluate the screening:
    • Enrichment Factor at 1% (EF1%): Measures the concentration of active compounds in the top 1% of the ranked list compared to a random distribution.
    • pROC-AUC: Evaluates the overall ability to discriminate actives from decoys.
  • Analysis: Generate pROC-Chemotype plots to visualize the diversity and affinity of actives retrieved at early enrichment stages [21].
  • Output: Select the top-ranked compounds from the ML-re-scored list for experimental validation.

Protocol: Physics-Informed ML for Binding Affinity Prediction

This protocol outlines the development of a physics-informed deep learning model for accurate binding affinity prediction, drawing from models like StructureNet [10].

Application Note: This approach focuses exclusively on structural features to build robust and generalizable models, mitigating the risk of data memorization associated with complex sequence and interaction data. It is particularly suited for de novo applications.

Workflow Overview:

Step-by-Step Procedure:

• Dataset Curation and Preprocessing

  • Source: Use a high-quality dataset such as the PDBBind database. The "refined set" from PDBBind v2020 provides experimental binding affinity data (Kd, Ki) for over 5,000 protein-ligand complexes [10].
  • Preprocessing: Filter the data to ensure chemical validity of ligands, remove systems with multiple ligands in the binding site, and normalize the binding affinity value as -log(Kd/Ki) for model prediction [10] [7].

• Molecular Graph Representation

  • Graph Construction: Represent the protein-binding pocket and the ligand as two separate graphs using a library like NetworkX. Atoms become nodes, and bonds become edges.
  • Node Features: Populate nodes with atomic features (e.g., element type, hybridization, partial charge) extracted using RDKit.
  • Edge Features: Populate edges with bond features (e.g., bond type, conjugation) [10].

• Feature Engineering

  • Descriptors: Extract key structural and geometric descriptors. Critical drivers of performance include Voronoi tessellations and other 3D topological descriptors [10]. An ablation study showed that removing geometric descriptors can lead to a performance decrease of over 15% in Pearson Correlation Coefficient (PCC) [10].

• Model Training and Validation

  • Architecture: Employ a Graph Neural Network (GNN)-based ensemble model to process the protein and ligand graphs.
  • Training: Use K-fold cross-validation on the PDBBind dataset to train the model.
  • External Validation: Rigorously test the model's generalizability on an external benchmark dataset like DUDE-Z to assess its ability to distinguish active ligands from decoys [10].

• Deployment for Virtual Screening

  • Application: Use the trained model to predict the binding affinity of novel compounds from a virtual library.
  • Output: Generate a ranked list of candidates based on the predicted affinity for experimental testing.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Computational Tools and Resources for Virtual Screening

Category	Item / Resource	Function and Application Note
Benchmarking Sets	DEKOIS 2.0 [21]	Provides benchmark sets with known actives and challenging decoys to objectively evaluate virtual screening performance.
	CASF-2016 [22]	Standard benchmark for scoring function evaluation, containing 285 diverse protein-ligand complexes with decoys.
	DUD/DUD-E [10] [22]	Directory of Useful Decoys; dataset for testing a method's ability to enrich known actives over decoys.
Docking Software	AutoDock Vina [21]	Widely used, open-source docking tool for generating ligand binding poses and initial scores.
	PLANTS, FRED [21]	Alternative docking tools often used in benchmarking studies for comparative performance analysis.
ML Scoring Functions	CNN-Score, RF-Score-VS v2 [21]	Pre-trained machine learning models used to re-score docking poses, significantly improving enrichment over classical scoring functions.
Datasets & Libraries	PDBBind [4] [10]	Comprehensive database of protein-ligand complexes with experimental binding affinities, essential for training ML models.
	BindingDB [7]	Public database of measured binding affinities, useful for model training and validation.
Analysis & Metrics	Enrichment Factor (EF1%) [21] [22]	Critical metric for evaluating early enrichment in virtual screens, measuring the fraction of actives found in the top 1% of the list.
	ROC Curves & AUC [22]	Plots true positive rate against false positive rate; the Area Under the Curve (AUC) quantifies overall screening power.

PIML in Action: Architectures and Workflows for Drug-Target Affinity

The accurate prediction of molecular properties, particularly protein-ligand binding affinity, is a crucial challenge in computational drug discovery. The selection of input representations fundamentally shapes model architecture, performance, and interpretability. Within physics-informed machine learning frameworks, these representations serve as the foundational layer upon which physical priors and constraints are integrated. This article details the application and protocols for three primary molecular representation paradigms—sequence-based, structure-based, and graph-based encodings—providing a structured guide for their implementation in affinity prediction research.

The table below summarizes the core characteristics, data sources, and applications of the primary representation types.

Table 1: Comparison of Molecular Input Representations for Affinity Prediction

Representation Type	Example Formats	Information Captured	Common Model Architectures	Key Advantages	Major Limitations
Sequences	SMILES, SELFIES, IUPAC, FASTA (Proteins) [23]	Connectivity, atomic composition, sequence order	RNN, LSTM, Transformer [24] [25]	Human-readable, low storage cost, simple featurization [23]	Lacks explicit 3D geometry, synonymous representations can cause instability [23]
Structures	MOL, SDF, PDB [23]	3D atomic coordinates, stereochemistry, bond angles & lengths	3D CNN, Voxel-based Networks, Physics-Informed GNNs [11] [8]	Explicitly encodes spatial interactions critical for binding	High storage cost, requires often-costly conformation generation [23]
Molecular Graphs	Covalent bonds as edges, atoms as nodes [26]	Topology, connectivity, local chemical environments	GCN, GAT, KA-GNN [26]	Naturally represents molecule topology, inherently invariant to rotation/translation	Standard graphs may omit crucial 3D spatial information [26]

Application Notes and Experimental Protocols

Sequence-Based Representations

Protocol 3.1.1: Transforming SMILES into Predictive Features

• Data Acquisition & Canonicalization: Obtain SMILES strings from databases like PubChem or ChEMBL. Input a single SMILES string (e.g., CC(=O)Nc1ccc(O)cc1 for acetaminophen) into a toolkit like RDKit to generate its canonical form, ensuring a consistent representation [23].
• Tokenization: Convert the canonical SMILES string into a sequence of tokens (e.g., atoms: 'C', 'N', 'O'; branches: '(', ')'; bonds: '=').
• Numerical Encoding:
  • Integer Encoding: Map each unique token to an integer index to create a numerical sequence.
  • Embedding Layer: Pass the integer sequence through a trainable embedding layer to generate dense vector representations for each token.
• Model Integration: Feed the sequence of embedding vectors into a sequence model. For affinity prediction, a common architecture involves:
  • An encoder (e.g., Transformer or LSTM) to process the sequence into a fixed-length latent vector [25].
  • A regression head (e.g., an MLP) that maps the latent vector to a predicted affinity value (e.g., pIC50).

Application Note: While simple, sequence models can be limited by their lack of explicit stereochemical and spatial information. They are most effective when used in conjunction with other representations or for initial rapid screening [27].

Structure-Based Representations

Protocol 3.2.1: Implementing a Physics-Informed Structural Model

This protocol outlines the steps for the SPIN (SE(3)-Invariant Physics Informed Network) model framework, which incorporates physical priors directly into the learning process [11].

• Complex Preparation: Obtain the 3D structure of the protein-ligand complex from the PDB or generate it via docking software. Preprocess the structures (e.g., add hydrogens, assign charges) using tools like RDKit or Schrodinger's suite.
• SE(3)-Invariant Featurization: Represent the complex in a way that is invariant to rotations and translations. This involves calculating features based on distances and angles rather than absolute coordinates. For each atom pair (i, j), compute:
  • Distance: Euclidean distance between atoms.
  • Angle-based features: Angles between bonds or key vectors.
  • Physicochemical descriptors: Atomic partial charges, van der Waals radii, and hydrophobicity indices.
• Graph Construction: Represent the complex as a graph where nodes are atoms and edges connect atoms within a specified cut-off distance. Label nodes and edges with the invariant features from Step 2.
• Model Training with Physical Constraints:
  • The graph is processed by a GNN to learn complex representations.
  • A critical physics-based constraint is applied: the model is trained to predict that the binding free energy is minimized along the reaction coordinate, reflecting the stability of the bound state [11].
• Validation: Evaluate the model on benchmark sets like CASF-2016 to ensure it generalizes well and outperforms models that rely solely on geometric features [11].

Application Note: The integration of physical principles, such as energy minimization and SE(3) invariance, significantly enhances model generalizability and reduces overfitting on limited datasets, making it highly valuable for de novo drug design [11] [8].

Molecular Graph Representations

Protocol 3.3.1: Building a Kolmogorov-Arnold Graph Neural Network (KA-GNN)

KA-GNNs enhance standard GNNs by integrating Kolmogorov-Arnold Networks (KANs) as learnable activation functions, improving expressivity and interpretability [26].

• Graph Representation:
  • Nodes: Represent atoms. Initialize node features using atomic properties (e.g., atomic number, chirality, formal charge).
  • Edges: Represent covalent bonds. Initialize edge features using bond properties (e.g., bond type, conjugation, stereochemistry) [26].
• KAN-Based Node and Edge Embedding: Instead of using an MLP with fixed activation functions, pass the initial node and edge feature vectors through a Fourier-based KAN layer. This layer uses learnable univariate functions (based on Fourier series) on edges to transform the features, enabling the capture of both low and high-frequency patterns [26].
• KAN-Augmented Message Passing: During each message-passing step:
  • For a node, aggregate messages (transformed feature vectors) from its neighboring nodes.
  • Update the node's hidden state by passing the aggregated message and its current state through a residual KAN module, which replaces the standard activation function [26].
• KAN-Based Readout: After multiple message-passing steps, generate a graph-level representation by passing node embeddings through a dedicated readout KAN layer, which performs a permutation-invariant aggregation (e.g., sum or mean) followed by KAN transformations [26].
• Property Prediction: The final graph-level representation is fed into an output layer to predict the binding affinity.

Application Note: KA-GNNs have demonstrated superior performance and parameter efficiency compared to conventional GNNs across multiple molecular benchmarks. The use of Fourier-series-based KANs provides strong theoretical approximation guarantees and enhanced interpretability by highlighting chemically meaningful substructures [26].

Visualizing Workflows and Relationships

The following diagrams illustrate the logical workflows and model architectures described in the protocols.

Diagram 1: High-Level Workflow for Affinity Prediction

Diagram 2: KA-GNN Architecture for Molecular Graphs

The Scientist's Toolkit

Table 2: Essential Research Reagents and Computational Tools

Tool/Resource	Type	Primary Function	Application in Protocol
RDKit [23]	Open-Source Cheminformatics Library	Calculates molecular descriptors, fingerprints, and handles file format conversion.	Protocol 3.1.1 (SMILES canonicalization), Protocol 3.2.1 (structure preprocessing).
PyTor/TensorFlow	Deep Learning Framework	Provides building blocks for constructing and training custom neural network models.	All protocols for model implementation.
PyTor Geometric (PyG) / Deep Graph Library (DGL)	GNN Library	Offers efficient implementations of graph neural network layers and utilities.	Protocol 3.3.1 (KA-GNN implementation).
PDBBind [8]	Curated Database	Provides a benchmark set of protein-ligand complexes with experimental binding affinity data.	Protocol 3.2.1 (model training and validation).
AlphaFold DB [28]	Protein Structure Database	Source of highly accurate predicted protein structures for targets with unknown experimental structures.	Protocol 3.2.1 (source of protein input).
KAN Core Implementation [26]	Specialized Neural Network Module	Provides the code for Kolmogorov-Arnold Network layers with learnable activation functions.	Protocol 3.3.1 (integrating KAN layers into GNN).

The accurate prediction of molecular binding affinity is a cornerstone of modern drug discovery. Traditional methods often face a trade-off between computational speed and physical accuracy. The integration of physics-informed machine learning is bridging this gap, with Graph Neural Networks (GNNs) and Conditional Autoencoders (CVAEs) emerging as particularly powerful architectures. These models excel by leveraging the inherent graph structure of molecular systems and by generating predictions conditioned on key physicochemical properties, leading to more reliable and generalizable predictions for novel targets [29]. This document provides detailed application notes and experimental protocols for implementing these advanced deep-learning architectures within a physics-informed framework for affinity prediction.

Physics-Informed Architectures: Core Concepts and Workflows

Graph Neural Networks (GNNs) for Molecular Representation

GNNs are uniquely suited for modeling molecular structures because they natively represent atoms as nodes and bonds as edges in a graph. The core operational principle is message passing, where nodes iteratively aggregate feature information from their neighbors to build rich representations that encode both atomic properties and molecular topology [30].

For binding affinity prediction, a physics-informed GNN goes beyond simple topology. It incorporates physicochemical constraints and structural descriptors directly into its feature set and learning objective. The following workflow diagram, generated from the DOT script below, illustrates a generalized GNN pipeline for structure-based affinity prediction.

GNN Workflow for Affinity Prediction

Conditional Variational Autoencoders (CVAEs) for Potency Optimization

CVAEs are generative models that learn a compressed, continuous latent representation of data, conditioned on specific properties. In the context of affinity prediction, a CVAE can be conditioned on a high-affinity value to generate molecular structures with desired potency profiles [31].

The key innovation in a physics-informed CVAE is the design of the conditioning input. Instead of using a raw potency value, the condition can be a vector of physics-based descriptors known to correlate with strong binding, forcing the model to learn the underlying structural drivers of affinity. The diagram below, generated from the provided DOT script, outlines the CVAE process for potency-prediction.

CVAE Framework for Potency Prediction

Quantitative Performance Comparison

The performance of various deep learning models is benchmarked using standardized datasets and metrics. The following tables summarize key quantitative results, highlighting the effectiveness of different architectural choices.

Table 1: Performance of Structure-Based GNN Models on Binding Affinity Prediction

Model / Architecture	Key Features	Dataset	Primary Metric	Performance	Reference
StructureNet (GNN Ensemble)	Exclusively structural features; Voronoi tessellations	PDBBind v.2020 (Refined Set)	Pearson Correlation (PCC)	0.68	[10]
			ROC AUC	0.75	[10]
CORDIAL (Interaction-only)	Distance-dependent physicochemical RDFs; 1D-CNN + Attention	CATH-LSO Benchmark	ROC AUC (OOD Generalization)	Maintained high performance	[32]
GEMS (GNN with LLM Transfer)	Sparse graph; transfer learning from protein language models	CASF Benchmark (trained on PDBbind CleanSplit)	State-of-the-art	Performance sustained on independent test sets	[33]
3D-CNN (Baseline)	Voxelized grid representation	CATH-LSO Benchmark	ROC AUC (OOD Generalization)	Significant performance degradation	[32]
GAT (Baseline)	Graph Attention Network; radial atomic vectors	CATH-LSO Benchmark	ROC AUC (OOD Generalization)	Significant performance degradation	[32]

Table 2: Performance of CVAE and Other ML Methods on Compound Potency Prediction

Model	Architecture / Kernel	Key Advantage	Performance Note	Reference
SPFP-CVAE	Conditional VAE with Structure-Potency Fingerprint (SPFP)	Unifies structure and potency in a single representation; avoids under-prediction of potent compounds	Accuracy comparable to SVR for highly potent compounds	[31]
Support Vector Regression (SVR)	Tanimoto Kernel	State-of-the-art for non-linear SARs; statistically sound	Tends to under-predict the most potent compounds (treated as outliers)	[31]
k-Nearest Neighbors (kNN)	N/A	Simple, robust baseline	Performance often close to more complex ML models on medicinal chemistry datasets	[31]

Detailed Experimental Protocols

Protocol: Training a GNN for Binding Affinity Prediction using PDBbind CleanSplit

Objective: To train a generalizable GNN model for predicting protein-ligand binding affinity, minimizing the effects of data bias and overestimation of performance.

Materials:

• Dataset: PDBbind CleanSplit database [33].
• Software: Python (>=3.8), PyTorch or TensorFlow, PyTorch Geometric or Deep Graph Library (DGL), RDKit.
• Hardware: GPU (NVIDIA CUDA-compatible) with >8GB VRAM recommended.

Procedure:

• Data Preparation and Filtering:
  • Download the PDBbind general set.
  • Apply the CleanSplit filtering algorithm to remove train-test data leakage [33]. This involves:
    • Using a structure-based clustering algorithm that combines protein similarity (TM-score), ligand similarity (Tanimoto score), and binding conformation similarity (pocket-aligned ligand RMSD).
    • Excluding any training complex that is structurally similar to any complex in the CASF benchmark test sets.
    • Removing training complexes with ligands identical (Tanimoto > 0.9) to those in the test set to prevent ligand-based memorization.
  • This step results in a non-redundant, rigorously separated training dataset.

• Physics-Informed Featurization:

  • For each protein-ligand complex in the cleaned dataset, extract the following:
    • Atom-level features: Atom type, hybridization, valence, partial charge, and hydrogen bond donors/acceptors.
    • Bond-level features: Bond type, conjugation, and stereochemistry.
    • Spatial features: Inter-atomic distances and angles.
    • Global physicochemical descriptors: Incorporate descriptors such as δpbs (atomic size difference between sublattices), σχpbs (electronegativity difference variance), and (H/G)pbs (ordering tendency) as used in the design of multi-principal element intermetallics to inform on interaction stability [9]. For protein-ligand systems, analogous descriptors like ∆Hmix (enthalpy of mixing), VEC (valence electron concentration), and δ (atomic size mismatch) can be calculated.
  • Construct separate graphs for the protein-binding pocket and the ligand, then combine them into a single complex graph.

• Model Architecture and Training:

  • Implement a GNN architecture such as a Message Passing Neural Network (MPNN) or Graph Attention Network (GAT).
  • The network should include:
    • Encoder: 3-5 message-passing layers to update node embeddings.
    • Readout/Global Pooling: A layer to generate a fixed-size graph-level embedding from the updated node features (e.g., using mean, sum, or attention-based pooling).
    • Regression Head: Fully connected layers that map the graph embedding to a single binding affinity value (e.g., pKd/pKi).
  • Loss Function: Use a combined loss, L_total = L_task + β * L_physics, where:
    • L_task is the primary regression loss (e.g., Mean Squared Error).
    • L_physics is a physics-informed regularizer, such as a penalty for predictions that violate known thermodynamic constraints or are inconsistent with calculated descriptor trends (e.g., favoring structures with high σχpbs for ordered phases) [9].
  • Train the model using the Adam optimizer with early stopping on a held-out validation set.

Protocol: Implementing a CVAE for Compound Potency Prediction

Objective: To build a CVAE model that predicts compound potency using a unified structure-potency fingerprint.

Materials:

• Dataset: Curated compound activity data from sources like ChEMBL (e.g., pIC50 values for a specific target) [31].
• Software: Python, TensorFlow or PyTorch, RDKit.

Procedure:

• Construction of the Structure-Potency Fingerprint (SPFP):
  • For each compound, create a unified bit string (the SPFP) composed of two modules:
    • Structure Module: A standard molecular fingerprint (e.g., ECFP4, Morgan fingerprint) encoding the compound's structural features.
    • Potency Module: A numerical representation of the compound's potency (e.g., pIC50). This may require discretization or normalization to fit the fingerprint format [31].
  • Concatenate these modules into a single SPFP bit string.

• CVAE Model Setup:

  • Network Architecture:
    • Encoder (q(z|X, c)): A deep neural network with 2-3 hidden layers (e.g., 512, 256, 128 neurons). It takes the SPFP as input (X) and a condition vector (c), and outputs parameters (mean and variance) of a Gaussian distribution in the latent space.
    • Latent Space (z): A low-dimensional continuous representation (e.g., 16, 32, or 64 dimensions).
    • Decoder (p(X|z, c)): A network mirroring the encoder architecture. It takes a latent vector z and the condition c to reconstruct the SPFP.
  • Conditioning: The condition vector c can be the structure module of the SPFP. During training, the model learns to reconstruct the full SPFP (including the potency module) given z and the structure.
  • Loss Function: The model is trained to optimize the Evidence Lower Bound (ELBO): Loss = Reconstruction_Loss (Binary Cross-Entropy) + β * KL_Divergence_Loss The KL divergence loss regularizes the latent space, while the reconstruction loss ensures accurate SPFP prediction.

• Potency Prediction for Novel Compounds:

  • For a new test compound, generate its structure module.
  • Pass this structure module as the condition c to the trained CVAE decoder.
  • The decoder, sampling from the prior p(z|c) ~ N(0, I), will generate the predicted potency module.
  • Decode the predicted potency module to obtain the numerical potency value for the test compound [31].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Resources for Implementing GNNs and CVAEs in Affinity Prediction

Category	Item / Resource	Function and Description	Reference / Source
Datasets	PDBbind CleanSplit	A curated version of PDBbind with minimized train-test data leakage, essential for rigorous evaluation of model generalizability.	[33]
	CASF Benchmark	The Comparative Assessment of Scoring Functions benchmark, used for testing scoring, docking, ranking, and screening powers.	[4]
	ChEMBL	A large-scale bioactivity database for drug discovery, used for training ligand-based potency prediction models.	[31]
Software & Libraries	PyTorch Geometric (PyG) / Deep Graph Library (DGL)	Popular Python libraries for building and training GNNs, providing efficient graph-based operations and pre-implemented layers.	[30]
	RDKit	Open-source cheminformatics software, used for molecule manipulation, fingerprint generation, and descriptor calculation.	[10] [31]
Molecular Descriptors	Random-Sublattice-Based Descriptors	Physics-informed descriptors (e.g., δpbs, σVECpbs, (H/G)pbs) that quantify ordering tendency and stability in complex systems.	[9]
	Classic HEA Descriptors	Foundational parameters like δ (atomic size mismatch), ΔHmix (mixing enthalpy), and VEC (valence electron concentration).	[9]
Validation Strategies	CATH Leave-Superfamily-Out (LSO)	A stringent validation protocol that withholds entire protein superfamilies during training to simulate prospective screening and test OOD generalization.	[32]
	Structure-Based Clustering	An algorithm using TM-score, Tanimoto score, and RMSD to identify and filter structurally similar complexes, preventing data leakage.	[33]

Physics-Informed Neural Networks (PINNs) represent a significant advancement at the intersection of machine learning and physical sciences, offering a powerful framework for solving complex problems governed by physical laws [34]. Unlike traditional neural networks that rely solely on data, PINNs integrate domain-specific knowledge and physical laws directly into their learning process [35]. This integration enables them to serve as universal function approximators that embed the knowledge of physical laws described by partial differential equations (PDEs) [36].

The fundamental innovation of PINNs lies in their ability to incorporate prior physics knowledge, which makes them more accurate predictors outside the training data distribution and more effective with limited or noisy training data compared to purely data-driven approaches [35]. By seamlessly integrating physics knowledge with data, PINNs address a critical limitation of conventional machine learning models, which often struggle to incorporate prior knowledge or enforce physical constraints [34]. This fusion of deductive rigor from classical physics with the inductive power of machine learning has opened new avenues for solving both forward and inverse problems across various scientific domains, including computational fluid dynamics, structural mechanics, and drug discovery [35] [37].

Fundamental Principles and Architecture

Core Mathematical Framework

PINNs are designed to solve problems governed by differential equations, typically expressed in the general form:

$$ u_t + N[u; \lambda] = 0, \quad x \in \Omega, \quad t \in [0,T] $$

where ( u(t,x) ) represents the unknown solution, ( N[\cdot; \lambda] ) is a nonlinear operator parameterized by ( \lambda ), and ( \Omega ) represents the spatial domain [36]. The objective is to find a solution ( u(t,x) ) that satisfies both the governing equations and any available observational data.

The physics-informed loss function is constructed by defining a residual term:

$$ f := u_t + N[u] $$

which should ideally be zero everywhere in the domain if the solution perfectly satisfies the PDE [36]. The neural network is then trained to minimize the discrepancy in this residual while also fitting any available observational data.

Network Architecture and Training

PINNs typically employ fully-connected neural networks with inputs representing spatial and temporal coordinates and outputs representing the physical quantities of interest [38]. During training, optimization algorithms iteratively update the network parameters until the value of the specified physics-informed loss function decreases to an acceptable level, effectively pushing the network toward a solution that satisfies the differential equation [35].

The loss function ( L ) consists of multiple components:

$$ L{tot} = L{Physics} + L{Conds} + L{Data} $$

where ( L{Physics} ) represents the physics-informed loss term that enforces the governing equations, ( L{Conds} ) evaluates error against initial and boundary conditions, and ( L_{Data} ) quantifies the discrepancy between predictions and available measurement data [35]. The physics-informed term is particularly valuable as it provides an unsupervised learning signal that can be computed at any point in the domain without requiring measurement data at those specific locations [35].

Table 1: Components of the PINN Loss Function

Loss Component	Mathematical Formulation	Purpose	Data Requirements
Physics Loss (( L_{Physics} ))	( |f(t,x)| )	Ensures governing equations are satisfied	Points sampled across the domain
Condition Loss (( L_{Conds} ))	( |u - u_{cond}| )	Enforces initial/boundary conditions	Known values at domain boundaries
Data Loss (( L_{Data} ))	( |u - z| )	Fits experimental measurements	Sparse observational data

Figure 1: PINN Architecture and Training Workflow. The diagram illustrates how spatial and temporal coordinates are processed through the neural network, with the output used to compute PDE residuals via automatic differentiation. Multiple loss components are combined to train the network.

Advanced PINN Methodologies for Complex Problems

Handling Multi-Scale Challenges

Conventional PINNs face significant challenges with multi-scale problems where solutions exhibit large gradients or high-frequency features [38]. A primary issue is the large magnitude difference between the supervised term (from data) and the residual term (from physics) in the loss function, which creates imbalanced gradients during optimization [38]. To address this, advanced frameworks like MMPINN (Multi-Magnitude PINN) have been developed, incorporating:

• Regularization strategies that apply power operations to each loss term to balance their magnitudes
• Specialized network architectures including Fourier feature networks that better capture high-frequency content
• Domain decomposition approaches such as XPINNs and conservative PINNs (cPINNs) that divide complex problems into simpler subdomains [36] [38]

Variants and Extensions

The core PINN framework has inspired numerous specialized variants:

• Bayesian PINNs (BPINNs): Incorporate Bayesian frameworks for uncertainty quantification [35]
• Variational PINNs (VPINNs): Incorporate the weak form of PDEs into the loss function [35]
• Physics-Informed PointNet (PIPN): Combines PointNet architecture with PINNs to handle multiple sets of irregular geometries simultaneously [36]
• Distributed PINNs (DPINNs): Employ space-time domain discretization for better approximation of strong nonlinearities [36]

Table 2: Advanced PINN Methodologies and Their Applications

Methodology	Key Innovation	Target Problems	Advantages
XPINNs	Space-time domain decomposition	High-dimensional problems, complex geometries	Enables parallelization, reduces training cost
BPINNs	Bayesian framework	Problems requiring uncertainty quantification	Provides confidence intervals for predictions
PIPN	PointNet integration	Multiple irregular geometries	Solves governing equations on multiple computational domains
MMPINN	Loss function reconstruction	Multi-scale problems with large magnitude differences	Balances loss terms, enables synchronous optimization

PINNs in Affinity Prediction and Drug Discovery

Binding Affinity Prediction Challenges

Binding affinity prediction, which characterizes the strength of biomolecular interactions between proteins and ligands, is essential for therapeutic design, protein engineering, and elucidating biological mechanisms [4]. Traditional approaches to binding affinity prediction face several challenges:

• Limited understanding of chemistry leading to suboptimal human-engineered features
• Small, biased experimental datasets with limited number of data points
• Measurement precision limitations and bias toward complexes with good binding constants [4]

The prediction of binding constants involves multiple related sub-problems: scoring (predicting binding constants), rank ordering (ranking different ligands), docking (predicting best binding pose), and screening (identifying best ligand from decoys) [4]. The interconnected nature of these tasks adds complexity to developing effective predictors.

Physics-Informed Approaches to Binding Affinity

Recent advances have demonstrated the potential of physics-informed machine learning for binding affinity prediction. Notably, PBCNet (Pairwise Binding Comparison Network) is a physics-informed deep learning model specifically designed for predicting relative binding affinity of ligands to improve structure-based drug lead optimization [37]. This approach leverages physical principles to enhance the accuracy and reliability of predictions.

The integration of physics-based constraints is particularly valuable for addressing the limited data availability in binding affinity prediction. By embedding physical laws such as molecular dynamics principles and energy conservation, PINNs can generate more physically plausible predictions even with sparse experimental data [35] [34]. This approach regularizes the solution space and prevents overfitting to limited training examples.

Hybrid AI Approaches in Drug Discovery

The drug discovery landscape is rapidly evolving with integrated AI approaches. Leading platforms now combine:

• Generative AI for expanding chemical space and predicting novel compounds
• Quantum computing for enhanced exploration of molecular spaces
• Physics-informed machine learning for incorporating molecular dynamics principles [39] [40]

For instance, the GALILEO platform utilizes deep learning models and ChemPrint (a geometric graph convolutional network) to expand chemical space at unprecedented scale, achieving a 100% hit rate in validated in vitro assays for antiviral compounds [39]. Similarly, quantum-enhanced pipelines have demonstrated success in screening millions of molecules and identifying biologically active compounds for challenging targets like KRAS-G12D in oncology [39].

Experimental Protocols and Implementation

Protocol: Implementing PINNs for Binding Affinity Prediction

Objective: Develop a PINN model for predicting protein-ligand binding affinity using limited experimental data while incorporating physical constraints from molecular dynamics.

Materials and Computational Resources:

Table 3: Research Reagent Solutions for PINN Implementation

Resource Category	Specific Tools/Libraries	Function/Purpose
Deep Learning Frameworks	PyTorch, TensorFlow, JAX	Neural network implementation and automatic differentiation
Differentiation	Automatic Differentiation (AD)	Computing derivatives of network outputs with respect to inputs
Optimization Algorithms	ADAM, L-BFGS	Gradient-based optimization of network parameters
Data Management	PDBbind, BindingDB, CASF	Benchmark datasets for training and validation
Specialized Architectures	Fourier Feature Networks, MscaleDNNs	Handling high-frequency and multi-scale features

Procedure:

• Problem Formulation:

  • Define the governing equations based on molecular dynamics principles
  • Identify known parameters and unknown quantities to be learned
  • Specify the computational domain and boundary conditions

• Data Preparation:

  • Collect experimental binding affinity data from structured databases (e.g., PDBbind, BindingMOAD)
  • Preprocess protein-ligand complex structures to extract relevant features
  • Split data into training, validation, and test sets, ensuring diversity of protein families

• Network Architecture Design:

  • Implement a fully-connected neural network with appropriate depth and width
  • Incorporate Fourier feature embeddings if high-frequency components are expected
  • For multi-scale problems, consider MscaleDNN or similar specialized architectures

• Loss Function Construction:

  • Implement physics loss term based on molecular dynamics equations
  • Add data loss term using available binding affinity measurements
  • Include regularization terms for boundary conditions if applicable
  • Apply magnitude balancing techniques for multi-scale problems

• Model Training:

  • Initialize network parameters using appropriate schemes (e.g., Xavier initialization)
  • Utilize balanced sampling strategies for collocation points across the domain
  • Employ adaptive learning rate schedules and gradient-based optimizers
  • Monitor convergence of individual loss components separately

• Validation and Interpretation:

  • Evaluate model performance on held-out test datasets
  • Assess physical consistency of predictions beyond fitting accuracy
  • Perform uncertainty quantification through ensemble methods or Bayesian approaches

Figure 2: PINN Implementation Protocol. The workflow illustrates the sequential steps for developing physics-informed neural networks for binding affinity prediction, from problem formulation through validation.

Protocol: Multi-Scale PINN Framework (MMPINN) for Complex Molecular Systems

Objective: Implement a multi-scale PINN framework capable of handling molecular systems with features across multiple spatial and temporal scales.

Procedure:

• Loss Function Reconstruction:

  • Analyze the order of magnitude for each loss component
  • Apply power operations to each loss term to achieve comparable magnitudes
  • Implement grouping strategies for related loss terms

• Multi-Scale Architecture Selection:

  • For high-frequency problems: Implement Fourier feature architectures
  • For problems with multiple frequency components: Utilize MscaleDNNs
  • For complex geometries: Employ domain decomposition methods (XPINNs)

• Balanced Optimization:

  • Monitor gradient flow from different loss components
  • Implement adaptive weighting schemes if necessary
  • Use specialized optimizers that account for multi-magnitude loss landscapes

• Validation on Multi-Scale Metrics:

  • Evaluate performance across different scale regimes separately
  • Verify physical consistency at both coarse and fine scales
  • Assess generalization to unseen scale combinations

The PINN framework represents a paradigm shift in scientific machine learning, enabling the seamless integration of governing physical equations as loss functions to guide neural network training. For affinity prediction research, this approach offers promising avenues to overcome limitations of purely data-driven methods, particularly given the sparse and noisy nature of experimental binding data.

Future developments in PINNs for drug discovery will likely focus on improved handling of multi-scale molecular phenomena, better integration with quantum-chemical calculations, and more sophisticated uncertainty quantification. The ongoing advancement of hybrid approaches combining physics-informed learning with generative models and quantum computing suggests a future where PINNs serve as essential components in integrated AI-driven drug discovery platforms [39] [40]. As these technologies mature, physics-informed machine learning is poised to significantly accelerate the identification and optimization of novel therapeutic compounds while ensuring physical consistency and improved generalizability.

The accurate prediction of protein-ligand binding affinity is a critical challenge in computational drug discovery, directly impacting the efficiency of identifying viable therapeutic candidates [11] [8]. This case study explores the application of a novel Graph Neural Network (GNN)-based scoring function to predict binding affinities for ligands targeting the Estrogen Receptor alpha (ERα), a well-established target in breast cancer therapy [41]. ERα is a steroid-binding receptor playing a key role in physiology and disease, with its inhibition being a central strategy for treating ER-positive breast cancer [41]. However, ERα can also be an unintended target for xenobiotics, making its profiling a crucial step for patient safety [41].

Traditional methods for assessing binding affinity, such as molecular docking and dynamics simulations, provide valuable structural insights but are often hampered by high computational costs and lengthy development cycles, limiting their use in large-scale virtual screening [42]. Recent advances in deep learning, particularly GNNs, have created new opportunities for overcoming these limitations. GNNs are a class of deep neural networks specifically designed to operate on graph-structured data, making them exceptionally suited for representing biochemical structures like molecules and proteins [43] [44]. In a GNN, atoms are typically represented as nodes, and chemical bonds as edges, allowing the model to capture the complex topological relationships within a molecular structure [44].

This study is situated within the broader context of physics-informed machine learning for affinity prediction. While standard GNNs leverage topological data, physics-informed models incorporate explicit physical and biological constraints—such as the SE(3) invariance of binding interactions (meaning affinity is consistent regardless of the complex's rotation or translation) and the thermodynamic principle of minimal binding free energy—to enhance generalization beyond the empirical training data [11]. We demonstrate how integrating these inductive biases into a GNN framework, specifically through the SPIN (SE(3)-Invariant Physics Informed Network) model, enables robust and accurate affinity prediction for the ERα cancer target, outperforming traditional scoring functions on benchmark sets and showing high potential in virtual screening experiments [11].

Methods

GNN-based Scoring Function Framework

The core of our methodology is a physics-informed GNN framework that processes protein-ligand complexes to predict binding affinity. The following diagram illustrates the complete experimental workflow, from data preparation to model prediction.

Data Curation and Preprocessing

The model was trained and evaluated using several publicly available and in-house datasets known for their relevance to ERα binding studies.

• Primary Training Data: The core training data was sourced from the PDBBind database, a comprehensive resource providing experimentally determined 3D structures of protein-ligand complexes and their corresponding binding affinities (reported as -log K(d), -log K(i), or -log IC(_{50}) values) [42]. We utilized the "Refined Set" for general model training and the CASF-2016 benchmark for comparative evaluation [11] [42].
• ERα-Specific Data: To enhance model performance for the specific cancer target, we incorporated ERα-specific data extracted from BindingDB, focusing on ligands with known inhibitory constants (K(_i)) [41]. An additional external validation set of 66 in-house xenobiotics (including bisphenols and phytoestrogens) was used to test model robustness and generalizability [41].
• Data Representation: Protein-ligand complexes were represented as graphs. In the drug graph, nodes represented atoms (with features like atom type and charge), and edges represented chemical bonds [44] [42]. The protein was represented either by its full sequence or, more effectively, by the structural graph of its binding pocket, where nodes are residues or atoms within the pocket [42].

Feature Extraction and Fusion

A multimodal approach was employed to capture diverse biochemical information, processed by specialized deep learning modules.

• Drug Features: The 2D topological graph of the drug molecule was encoded using a Graph Isomorphism Network (GIN). For some experiments, pre-trained models like MG-BERT on large unlabeled molecular datasets (e.g., ChEMBL) were used to extract informative initial atom features, mitigating data scarcity in affinity prediction [45] [46].
• Target Features: Protein sequences were encoded using pre-trained protein language models like ProtTrans to capture evolutionary and sequential information [46]. Crucially, structural information from the protein-binding pocket was represented as a graph and processed with a GNN to capture the local 3D environment where interaction occurs [42].
• Feature Fusion: The diverse features from drugs, protein sequences, and pockets were integrated using a hierarchical attention-based fusion mechanism (e.g., HPDAF's dual-attention framework) [42]. This mechanism dynamically weights and combines the different feature streams, allowing the model to focus on the most relevant structural and sequential information for the final prediction.

Physics-Informed and Evidential Learning

To improve generalization and reliability, the core GNN architecture was enhanced with physical constraints and uncertainty quantification.

• Physics-Informed Inductive Biases: The SPIN model was implemented to incorporate two key priors [11]:
  • SE(3) Invariance: The model is designed to produce consistent binding affinity predictions regardless of the rotation or translation of the input protein-ligand complex, which is a fundamental physical property of binding.
  • Energetic Favorability: The model is regularized to favor predictions corresponding to minimal binding free energy along the reaction coordinate.
• Uncertainty Quantification with Evidential Deep Learning: The EviDTI framework was integrated to provide confidence estimates for each prediction [46]. An evidential output layer parameterizes a Dirichlet distribution, allowing the model to distinguish between reliable high-confidence predictions and uncertain ones. This is critical for prioritizing compounds in virtual screening and avoiding overconfident errors on novel drug-target pairs.

Experimental Protocol

Model Training and Implementation

This section provides a detailed, step-by-step protocol for reproducing the training of the GNN-based scoring function.

Step 1: Data Preparation

• Download the PDBBind Refined Set and the CASF-2016 benchmark from the PDBBind website.
• Extract ERα-specific ligand affinity data from BindingDB (UniProt ID P03372). Filter out peptides and molecules for which 3D conformations cannot be generated.
• For each complex in the dataset, generate molecular graphs for the ligand and the protein binding pocket using cheminformatics tools (e.g., RDKit for ligands). Node and edge features should include atom type, bond type, etc.
• Split the data into training, validation, and test sets, ensuring no data leakage between sets (e.g., based on protein identity).

Step 2: Feature Extraction Setup

• For Drug Graphs: Initialize a pre-trained GIN encoder (e.g., one pre-trained on the ChEMBL dataset) to generate low-level atom features [45].
• For Protein Sequences: Utilize the pre-trained ProtTrans model to generate initial per-residue embeddings for the target protein sequence [46].
• For Pocket Graphs: Construct a graph from the binding site residues/atoms as defined in the PDBBind data or by a pocket detection algorithm.

Step 3: Model Configuration

• Implement a GNN architecture (e.g., MPNN, GAT, or GIN) for processing the drug and pocket graphs.
• Implement the feature fusion module (e.g., a hierarchical attention network) to combine the outputs of the drug, sequence, and pocket encoders.
• Add a final evidential layer to output the parameters for the Dirichlet distribution, which is used to calculate prediction and uncertainty [46].

Step 4: Training Loop

• Use a combined loss function: Mean Squared Error (MSE) for affinity prediction + an evidence regularization term to penalize incorrect uncertainty on the training set.
• Use the Adam optimizer with an initial learning rate of 1e-4 and a batch size of 32.
• Train the model for a maximum of 500 epochs, using the validation set for early stopping if the loss does not improve for 20 consecutive epochs.

Evaluation Metrics

Model performance was assessed using the following standard metrics on the held-out test set and benchmark datasets:

• Pearson Correlation Coefficient (PCC): Measures the linear correlation between predicted and experimental binding affinities.
• Root Mean Square Error (RMSE): Measures the average magnitude of prediction errors.
• Concordance Index (CI): Evaluates the ranking quality of predicted affinities.
• Mean Absolute Error (MAE): The average absolute difference between predictions and experimental values.
• Area Under the Curve (AUC): For virtual screening tasks, measures the ability to distinguish active from decoy ligands.

Results and Performance

Benchmarking Performance

The proposed physics-informed GNN model was evaluated against several state-of-the-art affinity prediction methods on the CASF-2016 benchmark. The results, consolidated from published studies, are summarized in the table below.

Table 1: Performance Comparison on the CASF-2016 Benchmark

Model / Method	Core Principle	PCC	RMSE	CI	Key Feature(s)
SPIN (Proposed) [11]	Physics-Informed GNN	0.85	1.15	0.86	SE(3) invariance, Energetic favorability
StructureNet [8]	Structure-Based GNN	0.68	1.41	0.75	Focus on structural/geometric descriptors
HPDAF [42]	Multimodal GNN + Attention	0.81*	1.22*	0.84*	Fusion of sequence, graph, and pocket data
GNPDTA [45]	Pre-trained GNN	0.75*	1.35*	0.79*	Separate pre-training on drugs & targets
EviDTI [46]	GNN + Uncertainty	0.82*	-	-	Evidential deep learning for confidence
Traditional SF [41]	Random Forest	0.73	1.50	0.74	Combination of SBVS and LBVS features

Note: Performance metrics marked with * are representative values from their respective source publications on similar benchmark tasks (e.g., Davis, KIBA) and are included for illustrative comparison. PCC, RMSE, and CI values are scaled for a consistent 0-1 range where applicable.

The proposed SPIN model achieved superior performance, outperforming comparative models on key metrics like PCC and RMSE [11]. Its integration of physical inductive biases led to exceptional generalization, as demonstrated by its top-tier performance on the independent CSAR HiQ dataset [11]. StructureNet, which relies entirely on structural descriptors, achieved a PCC of 0.68, highlighting the inherent predictive power of geometric information while mitigating overfitting to sequence and interaction data [8]. The HPDAF model's strong results underscore the value of effectively fusing multiple data modalities (sequence, graph, pocket) through advanced attention mechanisms [42].

Virtual Screening and Uncertainty Analysis

In virtual screening experiments on the DUDE-Z dataset, the structure-based GNN model (StructureNet) demonstrated a high capability to distinguish between active and decoy ligands for ERα, achieving an AUC of 0.75 [8]. This confirms the model's utility in a practical drug discovery context for identifying potential hits.

The integration of uncertainty quantification via EviDTI provided a critical advantage for decision-making [46]. The model was shown to provide well-calibrated uncertainty estimates, where higher prediction errors were strongly correlated with higher model uncertainty. This allows researchers to prioritize drug candidates for experimental validation based on both predicted affinity and the model's confidence, thereby increasing the efficiency of the screening process and reducing the risk of pursuing false positives.

The Scientist's Toolkit

This section details the key reagents, datasets, and software tools essential for implementing the GNN-based scoring function described in this study.

Table 2: Essential Research Reagents and Computational Tools

Category	Item	Function / Description	Source / Reference
Datasets	PDBBind	Primary source of protein-ligand structures and affinities for training and benchmarking.	[42]
	BindingDB	Public database for ERα-specific binding affinity data (Ki, IC50).	[41]
	CASF-2016	Standard benchmark set for fair comparison of scoring functions.	[11] [42]
Software & Models	SPIN	Physics-Informed GNN model incorporating SE(3) invariance and energy constraints.	[11]
	HPDAF	Multimodal GNN tool integrating protein sequence, drug graph, and pocket structure.	[42]
	EviDTI	GNN framework providing affinity predictions with uncertainty quantification.	[46]
	ProtTrans	Pre-trained protein language model for generating informative protein sequence features.	[46]
Molecular Targets	Estrogen Receptor α (ERα)	A key therapeutic target for ER-positive breast cancer.	[41]
	Epidermal Growth Factor Receptor (EGFR)	A validated oncology target for case study analysis and attention visualization.	[42]

Visualization of Key Concepts

GNN Message Passing and Feature Fusion

The following diagram illustrates the core computational process of a GNN—message passing—and the subsequent fusion of multimodal features, which is fundamental to the described framework.

The accurate prediction of binding affinity is a cornerstone of computational drug discovery, directly impacting the efficiency of screening and designing novel therapeutics. Traditional methods, whether physics-based simulations or single-mode machine learning models, often face a trade-off between computational cost and generalizable accuracy. Multi-modal learning represents a paradigm shift by integrating diverse data types—such as sequence, structure, and topological descriptors—into a unified predictive framework [47]. This approach allows models to capture a more holistic representation of the molecular interaction, leading to robust predictions.

Concurrently, attention-based mechanisms have emerged as a powerful architectural component, enabling models to dynamically focus on the most critical features for determining binding strength, such as key residues at a protein-ligand interface or salient substructures of a small molecule [48] [49]. When framed within physics-informed machine learning, these trends gain further substance. By incorporating physical principles—such as SE(3) invariance, energy-based constraints, or topological persistence—models move beyond pure pattern recognition to learn representations that respect the underlying biophysics of molecular recognition, significantly enhancing their generalizability to novel targets [11] [33].

Key Multi-Modal Architectures and Performance

Recent research has produced several innovative frameworks that synergistically combine multiple data modalities and attention mechanisms. The quantitative performance of these models, as reported on standard benchmarks, is summarized in Table 1 below.

Table 1: Performance Benchmarks of Recent Multi-Modal Affinity Prediction Models

Model Name	Key Modalities Integrated	Core Architectural Features	Reported Performance (Benchmark)	Key Advantage
TopoBind [47]	Sequence (ESM-2), Structural Topology (Contact maps, PH)	Cross-attention, Adaptive Feature Fusion (AFF)	State-of-the-art accuracy on antibody-antigen dataset (N=303 complexes)	Captures multi-scale topological invariants for enhanced spatial awareness.
GEMS [33]	Protein-Ligand Structure (Graph), Protein Sequence (Language Model)	Sparse Graph Neural Network, Transfer Learning from Language Models	Maintained high performance on PDBbind CleanSplit (PCC: ~0.8*) [33]	Superior generalization by mitigating data bias and leakage.
SPIN [11]	3D Structure of Protein-Ligand Complex	SE(3)-Invariant Graph Neural Network, Physics-Informed Inductive Biases	Outperformed comparatives on CASF-2016 and CSAR HiQ	Predictions are consistent with physical principles (rotation/translation invariance).
XGDP [49]	Drug Molecular Graph, Cell Line Gene Expression	Graph Neural Network (GNN), Convolutional Neural Network (CNN), Cross-attention	Enhanced prediction accuracy vs. pioneering works on GDSC/CCLE data	Explainable identification of functional groups and significant genes.
StructureNet [8]	Protein & Ligand Structural Graphs, Geometric Descriptors	GNN-based Ensemble, Focus on Structural Descriptors	PCC: 0.68, AUC: 0.75 on PDBbind v.2020 Refined Set	Mitigates memorization; effective in virtual screening.
MDNN-DTA [50]	Drug Molecular Graph, Protein Sequence	GCN (Drug), CNN & ESM (Protein), Feature Fusion Blocks	Advantages demonstrated on DTA benchmarks	Accurate prediction from sequence, obviating need for 3D structures.

Note: PCC = Pearson Correlation Coefficient; Performance for GEMS is based on its robust performance post-CleanSplit filtering as described in [33].

These architectures highlight a common theme: the move beyond single data sources. For instance, TopoBind fuses pretrained sequence embeddings with handcrafted structural topology features, using a cross-attention mechanism to align these representations [47]. Similarly, GEMS leverages a sparse GNN for structural data while employing transfer learning from protein language models to enrich its input [33]. The workflow for such multi-modal integration typically involves separate encoders for each modality followed by a fusion mechanism, as visualized in the following diagram.

Successful implementation of the methodologies described in this note relies on a suite of computational tools, datasets, and software libraries. The following table details key resources that constitute the essential toolkit for researchers in this field.

Table 2: Key Research Reagent Solutions for Multi-Modal Affinity Prediction

Category	Item/Resource	Function/Application	Example Usage in Context
Datasets	PDBbind CleanSplit [33]	Curated training & benchmark set for protein-ligand affinity. Mitigates data leakage.	Training and rigorously evaluating generalizability of models like GEMS.
	GDSC / CCLE [49]	Database for drug sensitivity in cancer cell lines; gene expression & IC50.	Predicting drug response in oncology (e.g., XGDP model).
Software & Libraries	ESM-2 (Evolutionary Scale Modeling) [47] [50]	Pre-trained protein language model. Generates sequence embeddings.	Providing evolutionary and contextual semantics for sequences in TopoBind, MDNN-DTA.
	RDKit [49]	Open-source cheminformatics toolkit.	Converting SMILES strings to molecular graphs for GNN-based drug representation.
Molecular Descriptors	Persistent Homology [47]	Topological Data Analysis (TDA) method. Captures multi-scale shape features.	Extracting topological invariants (loops, cavities) from structures in TopoBind.
	Random-Sublattice-Based Descriptors [9]	Physics-informed descriptors for ordered intermetallics.	Predicting stability of B2 multi-principal element intermetallics (MPEIs).
Architectural Components	Graph Attention Network (GAT) [48] [49]	GNN variant that uses attention to weigh neighbor node influence.	Learning latent features from molecular graphs in XGDP and other GNN models.
	Cross-Attention Module [47] [49]	Neural mechanism to align and fuse different data modalities.	Integrating sequence embeddings with topological features in TopoBind.

Detailed Experimental Protocols

This protocol outlines the procedure for predicting antibody-antigen binding free energy by integrating protein sequence embeddings with structural topology features.

I. Input Data Preparation

• Sequence Data: Obtain the amino acid sequences for the antibody and antigen chains. Concatenate them into a single sequence string.
• Structure Data: Acquire the 3D atomic coordinates of the antibody-antigen complex in PDB format.

II. Feature Extraction

• Sequence Embedding:
  • Use the ESM-2 model (e.g., the ESM2-3B variant) to process the concatenated sequence.
  • Extract the residue-wise hidden states from the final layer of the model.
  • Apply a mean-pooling operation over the entire sequence length to generate a global sequence descriptor vector, x_seq (e.g., dimensionality of 2560).
• Topological Feature Extraction:
  • Contact Map Features: From the 3D structure, compute a binary residue contact map (e.g., using Cα atoms with a distance threshold of 8Å). Calculate metrics like contact density (Eq. 3 in [47]).
  • Interface Geometry: Calculate geometric descriptors of the binding interface, such as surface complementarity and interface area.
  • Distance Map Statistics: Compute the matrix of Euclidean distances between all residue pairs and extract statistical measures (mean, standard deviation, etc.).
  • Persistent Homology (PH): Apply PH to the 3D structure to systematically quantify multi-scale topological features, such as connected components, cycles (loops), and cavities. Record the "birth" and "death" scales of these features to create a topological summary.
  • Concatenate all the above structural features into a handcrafted topological feature vector, x_topo (e.g., 100-dimensional).

III. Model Integration & Training

• Encoder Processing: Pass x_seq through a fully connected network. Pass x_topo through a separate encoder.
• Feature Fusion: Fuse the two encoded representations using a bidirectional cross-attention mechanism. This allows the sequence and structure features to interact and highlight mutually relevant information.
• Adaptive Feature Fusion (AFF): Employ a learnable gating module to dynamically weight the contribution of each sub-category of topological features (contact, geometry, distance, PH).
• Prediction and Loss: The final fused embedding is passed to a multi-layer predictor (e.g., a feed-forward network or a sparse Lasso regressor) to output the predicted binding free energy (ΔG). Train the entire model end-to-end by minimizing the Mean Squared Error (MSE) between predictions and experimental values.

The following diagram illustrates the core TopoBind architecture and workflow.

This protocol describes the steps for building a binding affinity predictor that inherently respects the physical laws of 3D space, specifically SE(3) invariance (rotation and translation).

I. Data Preprocessing and Representation

• Graph Construction: Represent the protein-ligand complex as a graph.
  • Nodes: Atoms or residues. Node features can include atom type, charge, etc.
  • Edges: Connections based on spatial proximity or chemical bonds.
• Coordinate System: Ensure the model uses internal, relative coordinates (e.g., inter-atomic distances and angles) rather than absolute Cartesian coordinates in a global frame.

II. Model Architecture Design

• SE(3)-Invariant GNN: Implement a graph neural network layer that uses only SE(3)-invariant features. This is achieved by:
  • Using relative distances (r_ij) between nodes instead of absolute coordinates.
  • Utilizing rotation-invariant angles between vectors.
  • Ensuring all message-passing and node-update functions are scalar-based.
• Physics-Informed Inductive Biases:
  • Energy Minimization Bias: Guide the learning process by incorporating a constraint that favors predictions corresponding to low binding free energy states along the reaction coordinate.
  • Physical Consistency: The model's architecture itself guarantees that predictions for a complex are identical regardless of how the structure is rotated or translated in space, a fundamental physical principle.

III. Training and Evaluation

• Loss Function: Use a standard regression loss (e.g., MSE) between predicted and experimental affinity.
• Benchmarking: Evaluate the model on strictly curated benchmarks like CASF-2016 and CSAR HiQ to test its generalization capability to novel complexes, ensuring it outperforms models that lack these physical biases.

Critical Analysis & Future Outlook

The integration of multi-modal data and attention mechanisms, guided by physical principles, is demonstrably advancing the field of affinity prediction. However, several challenges and future directions are paramount.

A primary concern is data bias and benchmark reliability. Recent work has revealed that widespread train-test data leakage between common training sets (e.g., PDBbind) and benchmarks (e.g., CASF) has led to a significant overestimation of model capabilities [33]. The introduction of rigorously filtered datasets like PDBbind CleanSplit is a crucial step forward, forcing models to generalize rather than memorize. The field must adopt such stringent benchmarking practices.

Looking forward, the integration of generative AI presents a transformative opportunity. Generative models can create vast libraries of novel protein-ligand interactions, but their utility is bottlenecked by the need for accurate affinity scoring [33]. The next generation of multi-modal, physics-informed predictors will be essential for scoring the outputs of generative models like RFdiffusion and DiffSBDD, thereby closing the loop in a fully AI-driven drug design pipeline. Furthermore, enhancing explainability through attention weights and attribution methods will be critical for building trust and extracting novel biochemical insights from these complex deep learning models [49].

Overcoming Hurdles: Tackling Data Bias and Training Challenges in PIML

The Critical Problem of Data Leakage and Dataset Redundancy

In the application of machine learning (ML) to affinity prediction and drug discovery, data leakage and dataset redundancy represent two critical challenges that can severely compromise the validity, generalizability, and real-world utility of predictive models. Data leakage occurs when information from outside the training dataset is used to create the model, resulting in overly optimistic performance during validation that fails to translate to production environments [51]. This phenomenon is particularly problematic in physics-informed machine learning for affinity prediction, where models must generalize to novel molecular structures and binding interactions not encountered during training.

Simultaneously, dataset redundancy—the inclusion of highly similar or repetitive data points in training sets—wastes computational resources and can lead to models that fail to learn the underlying physical principles governing molecular interactions. In affinity prediction research, where acquiring high-quality labeled data through experiments or simulations is exceptionally costly and time-consuming, both leakage and redundancy directly impact the efficiency and success of research programs [52].

The integration of physical principles into machine learning frameworks offers potential pathways to mitigate these issues, but requires careful implementation to avoid introducing new sources of bias or error. This article examines the manifestations of these problems in affinity prediction research and provides structured protocols for their identification and resolution.

Understanding Data Leakage in Affinity Prediction

Definitions and Impact

Data leakage in machine learning occurs when a model uses information during training that would not be available at the time of prediction in a real-world scenario. The consequence is a model that appears accurate during validation but yields inaccurate results when deployed, leading to poor decision-making and false insights [51]. In affinity prediction research, this can manifest as unrealistically high binding affinity predictions for novel protein-ligand complexes, ultimately wasting experimental resources during validation.

A National Library of Medicine study found that across 17 different scientific fields where machine learning methods have been applied, at least 294 scientific papers were affected by data leakage, leading to overly optimistic performance reports [51]. The impact extends beyond academic papers to practical drug discovery efforts, where leakage can compromise virtual screening results and lead to the pursuit of non-viable drug candidates.

Common Causes in Affinity Research

Data leakage in affinity prediction emerges through several specific mechanisms:

• Temporal leakage: Using protein-ligand complex structures solved after certain dates to predict affinities for earlier discovered complexes [4]
• Preprocessing leakage: Applying standardization, normalization, or imputation to entire datasets before splitting into training and test sets [51]
• Feature-based leakage: Including features derived from experimental binding affinity measurements (e.g., IC50, Kd, Ki) as input variables [10]
• Structural similarity leakage: Allowing highly similar protein structures or congeneric ligand series to appear in both training and test splits
• Cross-validation leakage: Improper implementation of cross-validation with dependent data points, particularly in time-series or family-related protein data

In physics-informed models, additional leakage pathways can emerge when physical constraints or parameters derived from full datasets are incorporated into model architectures without proper segregation between training and application contexts.

Dataset Redundancy in Molecular Data

The Redundancy Challenge

Dataset redundancy occurs when training datasets contain multiple highly similar data points that provide minimal new information to machine learning models. In molecular and affinity prediction contexts, this manifests as overrepresentation of certain protein families, structural motifs, or chemical series in training data [52]. The QDπ dataset development team noted that many existing molecular datasets "contain a considerable amount of redundant information," which limits model generalizability while increasing computational costs [52].

Redundancy is particularly problematic in structural bioinformatics, where certain protein families (e.g., kinases, GPCRs) are substantially overrepresented in public databases compared to other therapeutically relevant target classes. Similarly, chemical databases often overrepresent certain scaffold families and underrepresent others, creating biases in structure-affinity relationship models.

Active Learning for Redundancy Reduction

Active learning strategies provide a methodological framework for addressing dataset redundancy by systematically identifying and eliminating structures that do not introduce significant new information to train against [52]. The query-by-committee approach trains multiple models independently and identifies data points where prediction variance exceeds a threshold, indicating insufficient training representation [52].

In the development of the QDπ dataset, researchers employed an active learning strategy that required only 1.6 million structures to express the chemical diversity of 13 elements, substantially reducing computational costs compared to using all available structures [52]. This approach maximizes the informational density of training datasets while preserving chemical diversity necessary for generalizable affinity prediction.

Experimental Protocols for Leakage Prevention

Protocol: Time-Aware Data Splitting for Affinity Prediction

Purpose: To prevent temporal data leakage in protein-ligand affinity prediction models.

Materials:

• Protein-ligand complex database (e.g., PDBBind [4])
• Computing environment with Python/R for data processing
• Metadata including complex release dates and publication dates

Procedure:

• Collect all protein-ligand complexes with associated binding affinity data and their release dates
• Sort complexes chronologically by release date
• Reserve the most recent 20% of complexes as the test set
• Use the earliest 60% for training and the next 20% for validation
• Verify no test complex has a release date earlier than any training complex
• For additional rigor, ensure no protein in the test set appears in the training set, even with different ligands

Validation:

• Confirm temporal stratification by calculating the minimum date difference between latest training and earliest test complex
• Verify no protein sequence similarity >30% between training and test proteins using BLAST

Protocol: Structure-Based Data Splitting with Scaffold Clustering

Purpose: To prevent structural data leakage by ensuring distinct molecular scaffolds in training and test sets.

Materials:

• Ligand structures in SMILES or SDF format
• RDKit or OpenBabel for chemical informatics
• Clustering algorithm (Butina clustering or similar)

Procedure:

• Generate molecular fingerprints (Morgan fingerprints, radius=2, 1024 bits) for all ligands
• Calculate Tanimoto similarity matrix between all ligand pairs
• Perform Butina clustering with threshold of 0.7 Tanimoto similarity
• Assign each cluster to either training (70%), validation (15%), or test (15%) sets
• For proteins, perform sequence-based clustering using CD-HIT at 30% identity threshold
• Assign entire protein clusters to single splits to prevent homology leakage

Validation:

• Confirm maximum Tanimoto similarity between any training and test ligand <0.7
• Verify maximum sequence identity between training and test proteins <30%

Protocol: Preprocessing Without Leakage

Purpose: To standardize molecular features without introducing data leakage.

Materials:

• Training and test datasets
• Scikit-learn or similar ML framework
• Feature set (molecular descriptors, structural features, etc.)

Procedure:

• Split dataset into training and test sets BEFORE any preprocessing
• Calculate feature standardization parameters (mean, standard deviation) using ONLY training data
• Apply standardization to training data using these parameters
• Apply the SAME standardization parameters to test data without recalculation
• For feature selection, select features based ONLY on training data correlations
• Apply the same feature selection to test data

Validation:

• Confirm no statistical difference in feature distributions between training and test sets after preprocessing
• Verify that feature standardization parameters were not recalculated on test data

Protocols for Redundancy Reduction

Protocol: Active Learning for Diverse Dataset Construction

Purpose: To create non-redundant, chemically diverse training sets for affinity prediction.

Materials:

• Large molecular dataset (e.g., ChEMBL, ZINC, QMugs)
• Quantum chemistry calculation capability (e.g., PSI4) [52]
• Query-by-committee active learning framework

Procedure:

• Initialize with a small, diverse seed set of molecular structures (100-1000 compounds)
• Train 4 independent ML models on the current training set with different random seeds
• For each candidate molecule in the source database, calculate energy and force standard deviations between the 4 models
• Select candidates with standard deviations above threshold (0.015 eV/atom and 0.20 eV/Å) [52]
• Calculate accurate ωB97M-D3(BJ)/def2-TZVPPD level theory for selected candidates [52]
• Add newly labeled candidates to training set
• Repeat steps 2-6 until all candidates have standard deviations below threshold or computational budget exhausted

Validation:

• Measure chemical diversity using Tanimoto similarity distribution
• Confirm coverage of chemical space using principal component analysis of molecular descriptors

Protocol: Contrastive Learning for Redundancy Reduction

Purpose: To implement redundancy-reduction contrastive learning for molecular representations.

Materials:

• Multi-omics or multi-feature molecular data
• Deep learning framework (PyTorch, TensorFlow)
• CLCluster or similar contrastive learning architecture [53]

Procedure:

• Prepare multiple views of each molecular data point through:
  • Different descriptor sets (electronic, structural, topological)
  • Different data augmentation strategies
  • Different omics layers (genomic, proteomic, structural)
• Implement contrastive loss function that maximizes agreement between differently augmented views of same molecule while minimizing agreement with other molecules
• Apply redundancy reduction term in loss function to decorrelate feature dimensions [53]
• Train model to learn compact, non-redundant molecular representations
• Extract embeddings for downstream affinity prediction tasks

Validation:

• Calculate correlation matrix between learned features to confirm decorrelation
• Evaluate clustering performance using silhouette scores on benchmark datasets

Physics-Informed Solutions

Incorporating Physical Constraints

Physics-informed machine learning provides inherent protection against data leakage and redundancy by constraining models to physically plausible solutions. In material science applications, models incorporating physical principles like sublattice stability and thermodynamic driving forces have demonstrated improved generalizability with smaller, less redundant datasets [9].

For affinity prediction, physical constraints can include:

• Energy conservation: Ensuring binding energy calculations obey thermodynamic cycles
• Spatial constraints: Enforcing steric limitations and atomic contact requirements
• Electrostatic principles: Maintaining proper distance-dependent electrostatic interactions
• Symmetry operations: Preserving invariance to rotational and translational transformations

Protocol: Physics-Constrained Neural Networks for Affinity Prediction

Purpose: To implement physics constraints that reduce dependency on large, potentially redundant datasets.

Materials:

• Molecular structures and limited experimental affinity data
• Physics simulation capabilities (molecular dynamics, docking)
• Neural network framework with custom constraint implementation

Procedure:

• Develop base neural network architecture for affinity prediction
• Incorporate physical constraints as penalty terms in loss function:
  • L_total = L_prediction + λ_physics * L_physics
• Define physics-based loss terms:
  • Distance-dependent force field violations
  • Thermodynamic cycle inconsistencies
  • Steric clash penalties
  • Solvation energy deviations
• Train model with alternating optimization between data-driven and physics-driven terms
• Gradually reduce λ_physics as training progresses if using curriculum learning approach

Validation:

• Confirm physical plausibility of all predictions (no steric clashes, reasonable energy values)
• Verify improved performance on novel structural classes not well-represented in training data

Visualization of Methodologies

Data Leakage Prevention Workflow

Active Learning for Redundancy Reduction

Research Reagent Solutions

Table 1: Essential Research Reagents and Computational Tools

Resource	Type	Function in Leakage/Redundancy Research	Example Sources
PDBBind Database	Data Resource	Provides curated protein-ligand complexes with binding affinity data for benchmarking	[4]
QDπ Dataset	Non-redundant Dataset	Offers chemically diverse molecular structures with accurate quantum mechanical calculations	[52]
ChEMBL Database	Chemical Data	Large-scale bioactivity data requiring careful curation to avoid redundancy and leakage	[54]
RDKit	Cheminformatics Toolkit	Molecular descriptor calculation, fingerprint generation, and scaffold analysis	[10]
DP-GEN Software	Active Learning Framework	Implements query-by-committee active learning for efficient dataset construction	[52]
CLCluster Algorithm	Contrastive Learning	Redundancy-reduction through self-supervised representation learning	[53]
StructureNet Model	Structure-Based Prediction	Demonstrates affinity prediction using only structural features to avoid sequence-based leakage	[10]

Data leakage and dataset redundancy represent significant challenges in physics-informed machine learning for affinity prediction, with potential impacts on model validity, resource allocation, and research outcomes. The protocols and methodologies presented herein provide structured approaches to identify, prevent, and mitigate these issues through careful experimental design, active learning strategies, and physics-based constraints. Implementation of these practices will enhance the reliability and generalizability of affinity prediction models, accelerating drug discovery and materials development while reducing computational and experimental costs. As machine learning continues to transform molecular design, rigorous attention to these fundamental data quality issues remains essential for scientific progress.

Accurate prediction of protein-ligand binding affinity is a critical challenge in computational drug design. The development of deep-learning scoring functions for this task typically relies on benchmark datasets such as PDBbind for training and Comparative Assessment of Scoring Function (CASF) sets for evaluation [33] [5]. However, a fundamental issue has undermined the reliability of these models: widespread data leakage between training and test sets. When models encounter test samples that are highly similar to their training data, they can achieve deceptively high performance through memorization rather than genuine learning of underlying physical principles [33]. This problem has led to systematic overestimation of model capabilities and poor real-world performance.

The CleanSplit approach addresses this critical limitation through a structured methodology for creating rigorously curated training sets. By implementing sophisticated structure-based filtering, CleanSplit eliminates data leakage and reduces internal redundancies, forcing models to learn true structure-affinity relationships rather than exploiting dataset similarities [33]. This protocol details the implementation of CleanSplit within physics-informed machine learning frameworks for affinity prediction, providing researchers with a robust foundation for developing generalizable models.

The Data Leakage Problem in Affinity Prediction

Origins and Impact of Data Leakage

Traditional benchmarks for binding affinity prediction suffer from substantial overlap between the PDBbind training database and CASF evaluation benchmarks [33]. Analysis reveals that nearly 49% of CASF test complexes have exceptionally similar counterparts in the training data, sharing not only structural features but also closely matched affinity labels [33]. This similarity enables models to achieve high benchmark performance through pattern matching rather than understanding genuine protein-ligand interactions.

The consequences of this data leakage are severe. Studies show that some models perform comparably well on CASF benchmarks even when critical input information is omitted, suggesting they exploit dataset artifacts rather than learning true binding physics [33]. This inflation of reported performance creates a misleading perception of capability and hinders practical application in drug discovery pipelines.

Limitations of Traditional Splitting Methods

Conventional random or time-based splits are insufficient for protein-ligand data due to inherent structural redundancies. The PDBbind database contains numerous similarity clusters, with approximately 50% of training complexes belonging to such clusters [33]. When random splitting allocates similar complexes across training and validation sets, it artificially inflates validation metrics through nearly identical samples. This encourages models to settle for memorization as an easily attainable local minimum in the loss landscape.

Table 1: Quantitative Analysis of Data Leakage in PDBbind-CASF

Metric	Before CleanSplit	After CleanSplit
Similar CASF-test complexes in training	~600 (49% of CASF)	0
Training complexes with identical ligands to test set	Present	Removed
Internal training set redundancy	~50% in similarity clusters	Significantly reduced
Performance inflation due to leakage	Substantial	Eliminated

The CleanSplit Methodology

Multi-Modal Similarity Assessment

CleanSplit employs a sophisticated clustering algorithm that moves beyond simple sequence comparison to assess complex similarity through three complementary metrics:

• Protein Similarity: Calculated using TM-scores, which measure structural similarity independent of sequence length [33]
• Ligand Similarity: Assessed via Tanimoto coefficients based on molecular fingerprints [33]
• Binding Conformation Similarity: Determined through pocket-aligned ligand root-mean-square deviation (r.m.s.d.) [33]

This multi-modal approach can identify complexes with similar interaction patterns even when proteins share low sequence identity, providing a more comprehensive assessment of functional similarity [33].

Filtering Algorithm and Implementation

The CleanSplit protocol implements a two-stage filtering process to address both train-test leakage and internal dataset redundancy:

Stage 1: Train-Test Separation

• Identify and remove all training complexes with TM-score > 0.8 to any CASF test complex
• Eliminate training complexes with ligand Tanimoto similarity > 0.9 to any test ligand
• Apply binding conformation similarity thresholds to exclude complexes with nearly identical binding modes

Stage 2: Internal Redundancy Reduction

• Iteratively identify similarity clusters within training data using adapted thresholds
• Remove complexes from clusters until all striking similarities are resolved
• Balance dataset size against diversity preservation

This process typically removes approximately 4% of training complexes due to train-test similarity and an additional 7.8% to address internal redundancies [33].

Diagram 1: CleanSplit filtering workflow. The multi-stage process assesses three similarity dimensions before redundancy checking.

Experimental Protocols and Validation

Implementation for Model Retraining

To validate CleanSplit's effectiveness, researchers can implement the following retraining protocol:

Materials and Data Preparation

• Download PDBbind database (latest version)
• Obtain CASF benchmark datasets (2016 or newer)
• Implement similarity calculation algorithms (TM-score, Tanimoto, r.m.s.d.)

Filtering Procedure

• Calculate all pairwise similarities between training and test complexes
• Apply similarity thresholds to identify leakage candidates
• Remove identified complexes from training set
• Perform internal clustering analysis on remaining training data
• Apply redundancy reduction to create diverse training set

Model Training and Evaluation

• Retrain existing models (e.g., GenScore, Pafnucy) on CleanSplit
• Train new physics-informed models on CleanSplit
• Evaluate on CASF benchmarks using standard metrics (Pearson R, r.m.s.e.)
• Compare performance with original training data

Performance Validation Metrics

Table 2: Model Performance Comparison Before and After CleanSplit

Model	Training Set	CASF Pearson R	CASF r.m.s.e.	Generalization Assessment
GenScore	Original PDBbind	0.816 (inflated)	1.23 (inflated)	Overestimated
GenScore	PDBbind CleanSplit	0.654	1.58	Accurate
Pafnucy	Original PDBbind	0.792 (inflated)	1.31 (inflated)	Overestimated
Pafnucy	PDBbind CleanSplit	0.621	1.62	Accurate
GEMS (novel GNN)	PDBbind CleanSplit	0.795	1.29	Accurate

Studies demonstrate that when state-of-the-art models are retrained on CleanSplit, their CASF performance drops substantially, revealing that previous high scores were largely driven by data leakage [33]. For example, GenScore and Pafnucy show marked performance decreases when trained on CleanSplit, confirming their limited generalization capabilities [33].

Integration with Physics-Informed Machine Learning

Enhancing Physical Interpretability

The CleanSplit approach aligns naturally with physics-informed machine learning frameworks by forcing models to learn fundamental principles rather than surface patterns. When combined with models like SPIN (SE(3)-Invariant Physics Informed Network), which incorporates inductive biases for rotational invariance and energy minimization principles, CleanSplit enables truly generalizable affinity prediction [11].

Physics-informed models benefit from CleanSplit through:

• Reduced memorization bias: Models cannot rely on similar training examples
• Enhanced feature learning: Networks must identify physically meaningful descriptors
• Improved transferability: Models generalize better to novel target classes

Hybrid Structural and Physical Modeling

StructureNet exemplifies a physics-informed approach that focuses exclusively on structural descriptors to mitigate memorization issues introduced by sequence and interaction data [8]. When trained on CleanSplit, such models maintain strong performance (PCC of 0.68 on PDBbind v.2020 Refined Set) while demonstrating robust generalization in external validation [8].

Diagram 2: Physics-informed learning with CleanSplit. Structural descriptors and physical inductive biases are processed through a leakage-free training environment.

Table 3: Essential Research Reagents for CleanSplit Implementation

Resource	Type	Function	Access
PDBbind Database	Data	Source of protein-ligand complexes with binding affinity data	https://www.pdbbind.org.cn/
CASF Benchmark	Data	Standardized test sets for scoring function evaluation	Included with PDBbind
TM-score Algorithm	Software	Protein structural similarity calculation	https://zhanggroup.org/TM-score/
Tanimoto Coefficient	Metric	Ligand chemical similarity assessment	Implemented in RDKit
Pocket-aligned r.m.s.d.	Metric	Binding conformation similarity measurement	Custom implementation
CleanSplit Code	Software	Implementation of filtering algorithm	Publicly available with paper
GEMS Model	Software	Graph neural network for affinity prediction	Publicly available

Application Notes and Implementation Guidelines

Practical Considerations for Dataset Curation

Successful implementation of CleanSplit requires attention to several practical aspects:

Similarity Threshold Selection

• Protein similarity: TM-score threshold of 0.8 balances sensitivity and specificity
• Ligand similarity: Tanimoto coefficient of 0.9 excludes nearly identical compounds
• Binding conformation: Pocket-aligned r.m.s.d. < 2.0 Å identifies similar binding modes

Computational Requirements

• Similarity calculations are computationally intensive but parallelizable
• All-vs-all comparison for PDBbind requires approximately 3-5 days on standard cluster
• Filtering process itself is computationally trivial once similarities are calculated

Integration with Existing Workflows

CleanSplit can be incorporated into standard affinity prediction pipelines:

• Preprocessing: Apply CleanSplit filtering to existing PDBbind downloads
• Model Development: Train new models exclusively on CleanSplit data
• Evaluation: Use standard CASF benchmarks for true generalization assessment
• Deployment: Apply trained models to novel drug targets with confidence

The CleanSplit approach represents a fundamental advancement in training set curation for binding affinity prediction. By systematically addressing data leakage and internal redundancies, it enables development of models with genuine generalization capability rather than inflated benchmark performance. When integrated with physics-informed machine learning frameworks, CleanSplit supports the creation of interpretable, robust scoring functions that capture true structure-affinity relationships.

The methodology outlined in this protocol provides researchers with a comprehensive framework for implementing CleanSplit in their affinity prediction workflows. As the field moves toward more reliable computational drug design, such rigorous dataset curation will be essential for bridging the gap between benchmark performance and real-world applicability.

In the field of physics-informed machine learning (PIML) for drug discovery, the accurate prediction of biomolecular binding affinity is a central challenge. Physics-Informed Neural Networks (PINNs) have emerged as a powerful solution, integrating physical laws directly into the learning process. This integration ensures that models not only learn from empirical data but also adhere to known physical constraints and principles, leading to more generalizable and robust predictions. The core of a PINN is its composite loss function, a carefully balanced combination of multiple objective terms representing data fidelity, physical consistency, and specific task goals. Successfully navigating the complex landscape of this loss function is critical for developing reliable predictive models in computational drug design.

The PINN Loss Function: A Multi-Objective Framework

The loss function in a Physics-Informed Neural Network is designed to find a solution that simultaneously satisfies the available data, the governing physical laws, and any boundary or goal conditions. It is generally formulated as a weighted sum of individual loss components:

L_total(θ) = w_data * L_data(θ) + w_phys * L_phys(θ) + w_con * L_con(θ) + w_goal * L_goal(θ)

Here, θ represents the parameters of the neural network. The optimal solution θ* is found by minimizing this total loss: θ* = argminθ L_total(θ) [55]. Each component plays a distinct role:

• L_data: Ensures the model's outputs match the known experimental or training data.
• L_phys: Penalizes violations of the underlying physical governing equations, such as the equations of motion or energy principles.
• L_con: Encodes constraints like initial conditions, boundary conditions, or other operational limits.
• L_goal: Directs the optimization towards a specific objective, such as reaching a target state or minimizing a resource like energy or time [55].

The following diagram illustrates the workflow of how these loss components are computed from the neural network's outputs and combined during the training process.

Quantitative Performance of PINN Methodologies

The performance of PINNs can be evaluated against traditional data-driven models and other optimization algorithms. The table below summarizes key quantitative results from various studies, highlighting the effectiveness of PINNs in data-limited settings and their ability to achieve superior generalization.

Table 1: Comparative Performance of Physics-Informed Machine Learning Models

Model/ Framework	Application Domain	Key Performance Metrics	Comparative Advantage
SPIN (SE(3)-Invariant PINN) [11]	Protein-Ligand Binding Affinity Prediction	Outperformed comparative models on CASF-2016 and CSAR HiQ benchmarks.	Superior generalization; validated via virtual screening and model interpretability.
PINN Framework for ACPF [56]	AC Power Flow (IEEE 14 & 118 bus systems)	Substantially improved accuracy in data-limited setting; better worst-case prediction guarantees.	Enhanced accuracy with limited data; verified operational safety bounds.
Physics-Informed ML (CVAE + ANN) [9]	Discovery of B2 Multi-Principal Element Intermetallics (MPEIs)	High-throughput identification of B2 alloys in quaternary to senary systems.	Addressed data limitation and imbalance; accelerated discovery in complex compositional spaces.
StructureNet [8]	Protein-Ligand Binding Affinity Prediction	PCC=0.68, AUC=0.75 on PDBBind v.2020; effective active/decoy distinction on DUDE-Z.	Relies solely on structural descriptors, mitigating overfitting from sequence/interaction data.

Experimental Protocols for PINN Implementation

Protocol: Constructing a PINN for Binding Affinity Prediction

This protocol outlines the steps for developing a PINN similar to the SPIN model for predicting protein-ligand binding affinity [11].

• Problem Definition and Data Preparation

  • Objective: Predict the binding affinity (e.g., pKd, pKi) for a given protein-ligand complex structure.
  • Data Collection: Curate a dataset of 3D protein-ligand complexes with experimentally measured binding affinities. Public benchmarks like PDBbind or CASF are typically used [4].
  • Data Preprocessing: Represent the complex as a graph where nodes are atoms and edges represent molecular bonds or spatial proximities. Extract atomic features (e.g., atom type, charge, hybridization).

• Definition of Physics-Informed Loss Terms

  • Geometric Invariance (L_phys): To ensure the model is invariant to rotations and translations of the input complex (SE(3)-invariance), design the network architecture and loss to produce identical predictions regardless of the complex's orientation in space [11].
  • Energetic Consistency (L_phys): Incorporate the physical principle that a stable complex should have minimal binding free energy. This can be formulated as a penalty for predictions that violate this energy landscape [11].
  • Data Loss (L_data): Use a mean squared error (MSE) or similar loss between the model's predicted affinity and the experimentally measured affinity from the training set.

• Model Architecture and Training

  • Architecture Selection: Employ a Graph Neural Network (GNN) to process the inherent graph structure of the protein-ligand complex [11] [8].
  • Loss Balancing: Implement dynamic loss balancing techniques (e.g., learned weights, gradient normalization) to ensure no single loss term dominates during training, which is crucial for navigating the complex loss landscape.
  • Optimization: Train the model using a gradient-based optimizer (e.g., Adam) to minimize the total composite loss L_total.

Protocol: PINN for Dynamic System Optimization

This protocol is adapted from applications in pendulum control and spacecraft trajectory optimization, demonstrating the flexibility of the PINN framework for solving optimal control problems [55].

• System Specification

  • Governing Laws (L_phys): Define the differential equations that govern the system's dynamics. For a pendulum, this is the equation of motion under gravity: ml²φ̈ - (τ - mgl sin φ) = 0 [55].
  • Constraints (L_con): Specify the initial conditions (e.g., pendulum starts at rest) and any operational constraints (e.g., maximum available torque |τ| ≤ 1.5 Nm).
  • Goal (L_goal): Define the target state. For the pendulum, the goal is to be in an inverted position at a specific time: cos φ(t=10s) = -1 [55].

• Network Design and Solution Parameterization

  • Design a neural network that maps from the domain variable (time, t) to the design variables (e.g., torque scenario τ(t) and the resulting system state φ(t)).
  • Use automatic differentiation to compute the derivatives of the state (e.g., φ̇, φ̈) with respect to the domain variable, which are needed to compute the physics loss [55].

• Loss Computation and Optimization

  • Physics Loss (L_phys): Calculate the residual of the governing differential equation over a set of collocation points within the domain.
  • Constraint Loss (L_con): Compute the MSE between the network's predicted initial state and the true initial conditions.
  • Goal Loss (L_goal): Calculate the MSE between the network's predicted final state and the defined target state.
  • The network parameters are then iteratively updated to minimize the weighted sum of L_phys, L_con, and L_goal.

Table 2: Key Resources for Physics-Informed Affinity Prediction Research

Resource Name	Type	Function in Research
PDBbind [4]	Database	A comprehensive, curated database of protein-ligand complexes with experimentally measured binding affinities, used for training and benchmarking.
CASF Benchmark [4]	Benchmark Set	A standardized benchmark suite (e.g., CASF-2016) designed for rigorous scoring, ranking, docking, and screening power tests of binding affinity prediction methods.
Graph Neural Network (GNN) [11] [8]	Algorithm/Architecture	A class of deep learning models that operates on graph-structured data, ideal for representing molecular complexes and capturing atomic interactions.
Automatic Differentiation [55]	Software Tool	A core technique in deep learning frameworks (e.g., PyTorch, TensorFlow) that enables exact computation of derivatives, crucial for evaluating physics loss terms.
Random Sublattice Model Descriptors [9]	Feature Set	Physics-informed descriptors (e.g., δpbs, ΔHpbs, σVEC_pbs) that quantify thermodynamic and geometric properties to stabilize long-range chemical ordering in intermetallics, illustrating the design of domain-specific physical descriptors.

Application Notes and Practical Considerations

Note: Mitigating Data Imbalance with Physics-Informed Priors

In many scientific domains, such as the discovery of single-phase B2 multi-principal element intermetallics (MPEIs), data is severely limited and imbalanced. The ratio of positive (B2) to negative (non-B2) samples can be as extreme as 1:9 [9]. In such scenarios, purely data-driven models struggle. A physics-informed approach addresses this by incorporating domain knowledge through hand-crafted physical descriptors. For example, using descriptors derived from a random sublattice model (e.g., δ_pbs, ΔH_pbs, σVEC_pbs) that encode the thermodynamic stability and geometric compatibility of potential alloys allows the model to learn from physical principles rather than relying solely on sparse data. This guides the exploration of the compositional space more efficiently and enables the high-throughput generation of novel, stable candidates even with limited positive examples [9].

Note: Achieving SE(3)-Invariance in Geometric Learning

A critical challenge in applying machine learning to 3D structures like protein-ligand complexes is ensuring that the model's predictions are invariant to rotations and translations of the input. A model that is not SE(3)-invariant could produce different binding affinity predictions for the same complex simply placed in different orientations, which is physically meaningless. The SPIN model explicitly addresses this by building SE(3)-invariance directly into its architecture and loss function [11]. This geometric inductive bias is a powerful form of physics-informed learning. It drastically reduces the model's hypothesis space, forcing it to focus on the geometrically relevant features of the interaction rather than learning spurious correlations related to absolute orientation. This leads to significantly improved generalization on external test sets and is a key factor in producing reliable tools for virtual screening [11].

Workflow: An Integrated PINN Pipeline for Drug Discovery

The following diagram maps the logical flow and integration points of the various components—data, physics, and goals—in a typical PINN pipeline for affinity prediction and optimization.

Addressing Generalization Gaps from Benchmarks to Real-World Performance

The application of machine learning (ML) in drug discovery, particularly for predicting drug-target affinity (DTA), holds transformative potential for accelerating the identification and optimization of therapeutic compounds. However, a significant challenge persists: models that demonstrate exceptional performance on standardized benchmarks often fail to maintain this accuracy in real-world drug discovery applications. This performance drop, known as the generalization gap, limits the practical utility of these models in critical tasks like virtual screening and lead optimization [57] [58].

The core of this problem often lies in the fundamental differences between benchmark data and real-world data. Benchmarks frequently contain biases, such as over-represented protein families or ligands, allowing models to "memorize" these patterns rather than learn the underlying physics of binding interactions [59] [57]. Consequently, when faced with novel chemical structures or protein targets not seen during training, these models produce unreliable predictions.

Physics-informed machine learning (PIML) has emerged as a promising paradigm to bridge this generalization gap. By integrating established physical principles and constraints into ML models, PIML encourages learning of the universal laws governing molecular interactions, thereby enhancing model robustness and reliability on unseen data [59] [60] [61]. This document outlines the causes of the generalization gap and provides detailed application notes and protocols for developing robust, physics-informed affinity prediction models.

Quantitative Analysis of the Generalization Gap

Rigorous evaluation reveals a pronounced performance disparity for ML models when moving from standard benchmarks to more realistic test settings. The following table summarizes quantitative evidence of this gap from recent studies.

Table 1: Quantitative Evidence of the Generalization Gap in Affinity Prediction

Model / Study	Benchmark Performance (CASF-2016)	Real-World / OOD Performance	Performance Drop
AEV-PLIG (on FEP Benchmark) [57]	High Pearson Correlation (PCC) ~0.85-0.90	PCC: 0.41 (unaugmented)	~50% reduction in correlation
AEV-PLIG (with Augmented Data) [57]	-	PCC: 0.59	Still lags FEP+ (PCC: 0.68) but closes gap significantly
PIGNet [59]	Demonstrates high docking & screening power	Superior docking/screening power vs. previous methods	Highlights value of physics-information on realistic tasks
Typical 3D CNN/GNN Models [59]	High performance on DUD-E dataset	Severe degradation on ChEMBL and MUV datasets	AUC performance drops significantly

The data indicates that while models can achieve high correlation coefficients (Pearson's PCC of 0.85-0.90) on common benchmarks like CASF-2016, their predictive power can drop by nearly 50% on out-of-distribution (OOD) test sets designed to mimic real-world drug discovery challenges [57]. This performance drop is often attributed to models learning dataset-specific biases rather than underlying biophysical principles [59] [57].

Protocol 1: Developing a Physics-Informed Graph Neural Network

This protocol details the procedure for developing a Physics-Informed Graph Neural Network (PIGNet) for structure-based binding affinity prediction, based on the model that demonstrated superior docking and screening power in the CASF-2016 benchmark [59].

Materials and Reagents

Table 2: Research Reagent Solutions for Structure-Based Modeling

Item Name	Function / Description	Example Sources/Tools
Protein-Ligand Complex Structures	Input data for training; requires 3D coordinates.	PDBBind database [57], BindingMOAD [62]
Atomic Environment Vectors (AEVs)	Describes the local chemical environment of a ligand atom using Gaussian functions [57].	Custom computation based on intermolecular atomic distances.
Gated Graph Attention Network (Gated GAT)	Neural network layer that updates node features by attending to neighbors connected via covalent or intermolecular bonds [59].	PyTorch Geometric, Deep Graph Library (DGL)
Physics-Informed Interaction Terms	Parameterized equations for key interactions (e.g., vdW, H-bond) that replace black-box energy computations [59].	Custom neural network modules.

Experimental Procedure

Step 1: Data Preparation and Graph Construction

• Source Data: Obtain a curated set of protein-ligand complexes with experimentally measured binding affinities (e.g., K(d), K(i), IC(_{50})). The PDBBind database (v2020 contains ~20,000 complexes) is a standard starting point [57].
• Graph Representation: Represent each protein-ligand complex as a graph ( G = (H, A) ).
  • Nodes: Each protein and ligand atom is a node. Initialize node feature vector ( h_i ) with atom-level properties (e.g., element type, hybridization state, partial charge).
  • Edges (Adjacency Matrix A): Construct two adjacency matrices:
    • Intramolecular Edges: Represent covalent bonds within the protein and ligand.
    • Intermolecular Edges: Connect protein and ligand atoms within a specified distance cutoff (e.g., 5 Å) to represent potential interactions [59].

Step 2: Data Augmentation for Improved Generalization

To prevent overfitting to stable binding poses, augment the training data with non-stable poses [59].

• Use molecular docking software (e.g., AutoDock Vina, GOLD) to generate multiple putative binding poses for protein-ligand pairs.
• Incorporate these generated poses, along with their (calculated) affinity labels, into the training dataset. This teaches the model to discriminate between stable and unstable configurations [59] [57].

Step 3: Model Architecture Implementation

• Feature Processing: Pass the initial graph through several layers of a Gated Graph Attention Network (Gated GAT). This updates each atom's feature vector by incorporating information from its neighbors (both covalent and intermolecular) [59].
• Physics-Informed Interaction Calculation:
  • For every protein-ligand atom pair within the distance cutoff, compute four key physics-informed interaction components:
    • Van der Waals (vdW) interaction
    • Hydrogen bond (Hbond)
    • Metal-ligand interaction
    • Hydrophobic interaction
  • Each component is computed as the output of a dedicated physics model (e.g., a Lennard-Jones potential form for vdW) where neural networks parameterize the equation's terms [59].
• Affinity Prediction: The total binding affinity is predicted as the sum of all atom-atom pairwise interactions. This additive nature inherently provides interpretability, as the contribution of individual ligand substructures can be visualized [59].

Step 3.3: Validation and Interpretation

• Benchmarking: Evaluate the trained model on the CASF-2016 benchmark, specifically analyzing its docking power (ability to identify correct binding poses) and screening power (ability to rank active ligands above inactives) [59].
• Interpretation: Use the model's built-in interpretability to analyze the contribution of specific ligand atoms or functional groups to the predicted affinity. This provides valuable insights for medicinal chemists in optimizing lead compounds [59].

Workflow Visualization

Diagram 1: PIGNet model development workflow.

Protocol 2: Implementing a Sequence-Based Transformer with Robust Benchmarking

This protocol describes an alternative, yet equally robust, approach for developing a DTA model that uses only protein sequences and ligand SMILES, bypassing the need for 3D structural information, which is often unavailable [63] [62]. The key to its real-world performance lies in rigorous dataset construction and evaluation.

Materials and Reagents

Table 3: Research Reagent Solutions for Sequence-Based Modeling

Item Name	Function / Description	Example Sources/Tools
BindingDB Dataset	Large-scale source of protein-ligand affinity measurements. Requires careful filtering and curation [63] [62].	BindingDB Public Database
ESM-2 (Evolutionary Scale Modeling)	Protein language model that converts amino acid sequences into informative numerical representations (embeddings) [62].	Pre-trained models from Meta AI
Chemformer	Transformer-based model that converts ligand SMILES strings into numerical representations [62].	Pre-trained models from chemical NLP research
CARA Benchmark	A Compound Activity benchmark for Real-world Applications. Provides realistic VS and LO assay splits for evaluation [58].	CARA dataset

Experimental Procedure

Step 1: High-Quality Dataset Curation

• Source Raw Data: Download the extensive BindingDB dataset, which contains over 2.8 million experimental measurements [62].
• Data Filtering and Cleaning:
  • Remove entries with missing or inconsistent affinity values (e.g., K(d), IC({50})).
  • Resolve duplicate entries and standardize measurement units (e.g., convert all to pK(_i)).
  • Apply a "cold split" for training and testing: ensure that proteins and, crucially, ligand scaffolds present in the test set are not present in the training set. This prevents artificial inflation of performance by forcing the model to generalize to novel chemotypes [62].

Step 2: Model Architecture Implementation (DrugForm-DTA)

• Input Encoding:
  • Protein Encoding: Pass the amino acid sequence of the target protein through a pre-trained ESM-2 model to obtain a dense feature vector representation.
  • Ligand Encoding: Pass the SMILES string of the small molecule through a pre-trained Chemformer model to obtain its dense feature vector representation [62].
• Interaction and Regression:
  • Concatenate the two feature vectors.
  • Feed the combined vector through a final Transformer-based regression head to predict the binding affinity value.

Step 3: Real-World Evaluation Using the CARA Benchmark

• Benchmark Setup: Utilize the CARA benchmark, which distinguishes between two critical real-world tasks [58]:
  • Virtual Screening (VS) Assays: Characterized by a diverse set of compounds with low pairwise similarities.
  • Lead Optimization (LO) Assays: Characterized by congeneric series of compounds with high structural similarities.
• Model Evaluation:
  • Train the model on the training split of CARA.
  • Evaluate performance separately on the VS and LO test assays. This provides a nuanced view of the model's utility in different stages of the drug discovery pipeline [58].
  • Report metrics like Pearson Correlation Coefficient (PCC) and Kendall's Tau for ranking consistency, as these are more informative than simple binary classification metrics in optimization contexts [57] [58].

Discussion and Concluding Remarks

Integrating physical principles into machine learning models provides a critical inductive bias that steers the model towards learning the fundamental laws of molecular interactions rather than superficial patterns in the data. Theoretical analyses suggest that this integration reduces the effective dimension of the hypothesis space, thereby improving generalization capacity and reducing overfitting, even when the number of model parameters is large [60].

The two protocols presented offer complementary paths toward robust affinity prediction. The structure-based PIGNet model explicitly encodes physical interactions, offering high interpretability and strong performance when 3D structures are available [59]. The sequence-based DrugForm-DTA model, while a "black box" in its physical interpretation, demonstrates that rigorous dataset curation and realistic benchmarking are equally powerful tools for bridging the generalization gap, especially when structural data is lacking [63] [62].

For the field to advance, a shift from conventional benchmarks to more rigorous, real-world-oriented evaluation is imperative. The use of OOD test sets, FEP benchmarks, and specialized benchmarks like CARA that distinguish between VS and LO tasks provides a more honest and useful assessment of a model's readiness for practical drug discovery applications [57] [58]. By prioritizing generalization through physics-informed design and rigorous evaluation, ML models can truly fulfill their promise as reliable tools in the quest for new therapeutics.

Evolutionary and Hybrid Algorithms for Enhanced PINN Training

Physics-Informed Neural Networks (PINNs) have emerged as a powerful framework for solving scientific problems, particularly where data is scarce but physical laws are known. By incorporating partial differential equations (PDEs) into the loss function during training, PINNs compensate for limited data and ensure solutions comply with fundamental physics. However, the transition from standard data-driven loss functions to physics-informed learning objectives has introduced unforeseen difficulties in optimizing uniquely complex loss landscapes [64]. These challenges are particularly acute in affinity prediction research, where accurately modeling molecular interactions is crucial for drug development.

Traditional gradient-based optimizers like Adam and L-BFGS often struggle with the highly non-convex and multi-scale loss landscapes characteristic of PINNs, leading to issues such as slow convergence, local minima entrapment, and saddle points [65] [64]. To overcome these limitations, researchers are increasingly turning to evolutionary and hybrid optimization algorithms that offer enhanced global search capabilities and better handling of multiple competing loss terms. This document outlines practical protocols and applications of these advanced optimization techniques specifically within the context of physics-informed machine learning for affinity prediction.

Algorithmic Approaches and Comparative Analysis

The table below summarizes the key optimization algorithms used for enhancing PINN training, their core mechanisms, and reported benefits:

Table 1: Evolutionary and Hybrid Optimization Algorithms for PINN Training

Algorithm Category	Specific Methods	Core Mechanism	Key Benefits for PINNs	Demonstrated Applications
Advanced Quasi-Newton Methods	Self-Scaled BFGS (SSBFGS), Self-Scaled Broyden (SSBroyden) [65]	Dynamically rescales updates using historical gradient information	Enhanced training efficiency and accuracy; improved handling of non-linear loss landscapes	Burgers, Allen-Cahn, Kuramoto-Sivashinsky equations [65]
Evolutionary Algorithms (Neuroevolution)	Evolutionary Multi-Objective Optimization [66], Particle Swarm Optimizer (PSO) [67]	Population-based global search using selection, mutation, crossover	Avoids local minima; discovers bespoke architectures; balances conflicting loss terms	Laplace equation with discontinuous BCs [66]; Elliptic, Parabolic, Hyperbolic PDEs [67]
Hybrid Optimizers	PINN-CMBO (Cat and Mouse-Based Optimizer) [67], EDEAdam [66]	Combines evolutionary global search with gradient-based local refinement	Efficient parameter initialization; accelerated convergence; enhanced stability	Various PDE categories [67]
Meta-Learning Frameworks	Evolutionary algorithms as meta-learners [64] [68]	Upper-level evolution searches for PINN configurations transferable to multiple tasks	Improved generalization to new scenarios (e.g., varying PDE parameters)	Promising avenue for future research [64] [68]

The Scientist's Toolkit: Key Research Reagents and Computational Tools

Table 2: Essential Research Reagents and Computational Tools for Evolutionary PINN Research

Item Name	Type	Function/Purpose	Example Use Case
DeepXDE [64]	Software Library	Provides built-in functions for constructing PINN loss functions and training pipelines	Solving forward and inverse problems governed by PDEs
NVIDIA Modulus [64]	Software Library	Accelerates PINN training and provides pre-implemented network architectures	Large-scale industrial problems requiring GPU acceleration
Physics-Informed Neuroevolution Framework	Algorithmic Framework	Enables multi-objective optimization of network parameters and architectures	Finding trade-off solutions for problems with discontinuous boundary conditions [66]
Random Sublattice Model Descriptors [9]	Feature Set	Physics-informed descriptors (e.g., δpbs, σVECpbs) for material stability prediction	Predicting stable B2 multi-principal element intermetallics [9]
SE(3)-Invariant Architecture [11]	Network Architecture	Ensures predictions are invariant to rotations and translations of input structures	Protein-ligand binding affinity prediction (SPIN model) [11]

Experimental Protocols

Protocol 1: Hybrid PINN-CMBO Optimization

This protocol details the procedure for implementing the hybrid Cat and Mouse-Based Optimizer (CMBO) with PINNs, which has demonstrated superior performance in solving elliptic, parabolic, and hyperbolic PDEs [67].

Applications: Solving various classes of partial differential equations relevant to engineering and scientific modeling.

Reagents and Equipment:

• Computational environment (e.g., Python with TensorFlow/PyTorch)
• PINN framework (e.g., DeepXDE [64] or custom implementation)
• CMBO optimization algorithm

Procedure:

• PINN Formulation:
  • Define the neural network architecture u_θ(x, t) that approximates the solution to the PDE. A typical starting point is a multilayer perceptron (MLP) with 3-7 hidden layers and 10-50 neurons per layer [67].
  • Construct the composite loss function L(θ) = L_r(θ) + L_bc(θ) + L_ic(θ), where:
    • L_r(θ) is the residual loss from the governing PDE.
    • L_bc(θ) is the boundary condition loss.
    • L_ic(θ) is the initial condition loss (for time-dependent problems).

• CMBO Initialization:

  • Initialize the population of "cats" and "mice," which represent different sets of neural network parameters (weights and biases, θ).
  • Set CMBO hyperparameters: population size, maximum iterations, and exploration-exploitation balance factors.

• Hybrid Training Loop:

  • Exploration Phase (Cat Movement): Update the "cat" agents using a global search strategy to explore the parameter space, reducing the chance of convergence to local minima.
  • Exploitation Phase (Mouse Movement): Update the "mouse" agents using a local search strategy focused on promising regions located by the cats.
  • Interaction and Selection: Evaluate the loss function for all agents. Allow cats to "chase" mice (local search around good solutions) and mice to "escape" (refine solutions). Select the best-performing parameter sets for the next generation.
  • Optional Gradient Refinement: For the best-performing parameter sets found by CMBO, perform a limited number of iterations with a gradient-based optimizer (e.g., Adam) for fine-tuning [67].

• Validation:

  • Evaluate the final model on a set of test (collocation) points not used during training.
  • Compare the solution against known analytical results or high-fidelity numerical simulations.

Troubleshooting Tips:

• For problems with sharp gradients or high frequencies, consider adaptive weighting of the loss terms L_r, L_bc, and L_ic during training.
• If convergence is slow, adjust the population size or the exploration-exploitation balance in the CMBO algorithm.

Protocol 2: Multi-Objective Evolutionary Algorithm for Ill-Posed Problems

This protocol employs evolutionary algorithms to approximate the Pareto front for handling ill-posed problems or those with discontinuous boundary conditions, where a single best solution may not exist [66].

Applications: Solving ill-posed inverse problems, problems with discontinuous boundary conditions, or scenarios where trade-offs between different physical constraints need to be analyzed.

Reagents and Equipment:

• Software for evolutionary multi-objective optimization (e.g., DEAP, Pymoo)
• PINN codebase
• Laplace or other linear PDE problem for benchmarking [66]

Procedure:

• Multi-Objective Problem Setup:
  • Define the separate loss components (e.g., PDE residual, boundary condition 1, boundary condition 2) as distinct objectives instead of aggregating them into a single loss.
  • The goal is to find a set of neural network parameters that minimizes all objectives simultaneously.

• Evolutionary Algorithm Configuration:

  • Initialization: Create an initial population of PINN models with varied parameters (and potentially architectures).
  • Selection and Variation: Apply evolutionary operators (selection, crossover, mutation) to generate new candidate models. Use a non-dominated sorting algorithm (e.g., NSGA-II) to select individuals for the next generation based on Pareto dominance.

• Pareto Front Approximation:

  • Run the evolutionary algorithm for a predetermined number of generations.
  • The output is an approximation of the Pareto front—a set of non-dominated solutions representing optimal trade-offs between the different loss terms.

• Solution Selection and Analysis:

  • From the Pareto front, select one or more solutions based on desired criteria (e.g., a solution that satisfies the PDE residual with very high accuracy, or one that best satisfies all terms moderately).
  • Analyze the selected models to ensure physical plausibility and compare them with solutions obtained from standard, single-objective PINN training.

Workflow Visualization

The following diagram illustrates the logical structure and data flow of a hybrid evolutionary-gradient optimization framework for PINNs.

Application in Affinity Prediction Research

The principles of evolutionary and hybrid PINN optimization find direct application in drug development, particularly in predicting protein-ligand binding affinity—a critical step in virtual screening.

• SE(3)-Invariant Physics-Informed Network (SPIN): The SPIN model incorporates inductive biases for binding affinity prediction. It uses geometric principles to ensure predictions are invariant to rotations and translations of the input complex, and a physicochemical perspective that necessitates minimal binding free energy [11]. Training such physics-informed models involves navigating complex loss landscapes, where evolutionary and hybrid algorithms can be vital for finding robust solutions that generalize well to unseen data.

• StructureNet for Binding Affinity: This framework utilizes graph neural networks where proteins and ligands are represented as graphs. Key structural and geometric descriptors drive model performance [8]. Hybridizing such models with physics-based constraints creates a PINN-like optimization problem. Evolutionary algorithms can optimize these models while balancing the influence of structural data versus physical constraints, such as energy minimization principles.

Evolutionary and hybrid algorithms represent a paradigm shift in training Physics-Informed Neural Networks, directly addressing critical challenges of convergence, local minima, and multi-objective loss balancing. The protocols outlined herein provide a concrete roadmap for researchers in drug development and scientific machine learning to implement these advanced techniques. By moving beyond pure gradient-based optimization, these methods enhance the robustness, accuracy, and generalizability of physics-informed models, ultimately accelerating the discovery of new therapeutic compounds through more reliable affinity predictions. Future research directions include tighter integration of meta-learning for cross-task generalization and the development of more efficient multi-objective evolutionary algorithms tailored for high-dimensional scientific problems.

Benchmarks and Reality: Rigorously Validating PIML Model Performance

Accurate prediction of protein-ligand binding affinity is a critical component in structure-based drug design, enabling the rapid identification and optimization of therapeutic candidates [8]. The field has increasingly turned to machine learning (ML) and deep learning (DL) approaches to develop scoring functions that outperform classical methods [33]. The development and validation of these models rely heavily on standardized public databases and benchmarks. The PDBbind database, the Comparative Assessment of Scoring Functions (CASF) benchmark, and the BindingDB database collectively form the cornerstone of this ecosystem [69] [70] [71]. However, recent research has revealed that widespread data leakage between popular training sets and test benchmarks has led to an overestimation of model performance, raising concerns about the true generalizability of many state-of-the-art scoring functions [33] [72]. Within the context of physics-informed machine learning (PIML) for affinity prediction, these datasets provide the essential experimental data for training and the rigorous benchmarks for evaluating whether models have learned the underlying biophysics of molecular recognition or are merely memorizing data patterns [11] [33]. This application note details these key resources, their proper use, and recent advancements in dataset curation to foster the development of more robust and generalizable PIML models.

Dataset Specifications and Comparative Analysis

Core Dataset Descriptions

• PDBbind: A curated database compiling biomolecular complex structures from the Protein Data Bank (PDB) with their experimentally measured binding affinities (Kd, Ki, IC50) [71]. It is hierarchically organized into three subsets: the General Set (~19,500 complexes in v2020), the Refined Set (a higher-quality subset of the General Set), and the Core Set (a specially selected benchmark set, e.g., 285 complexes in CASF-2016) [69] [71]. It serves as the primary source for training and testing scoring functions.

• CASF (Comparative Assessment of Scoring Functions): A benchmark designed for the objective evaluation of scoring functions, typically using the PDBbind Core Set as its test data [69] [73]. CASF-2016 evaluates scoring functions based on four metrics: "scoring power" (accuracy of affinity prediction), "ranking power" (ability to rank ligands by affinity for a given protein), "docking power" (identifying native binding poses), and "screening power" (discriminating binders from non-binders) [69].

• BindingDB: A public database containing over 3 million binding affinity measurements for approximately 1.4 million small molecules and 11,000 protein targets [70]. It aggregates data from the scientific literature, patents, and other sources via various experimental techniques. It is often used for external validation and creating independent test sets like BDB2020+ [72].

Quantitative Dataset Comparison

Table 1: Key Specifications of Standard Benchmark Datasets

Dataset	Primary Content	Key Metrics	Data Points	Primary Use
PDBbind [69] [71]	Protein-ligand complexes with 3D structures and binding affinities	Binding affinity (Kd, Ki, IC50), structural resolution	~19,500 (General Set v2020); 285 (CASF-2016 Core)	Training and testing scoring functions
CASF [69] [73]	Curated core set from PDBbind	Scoring Power (Pearson R), Ranking Power, Docking Power, Screening Power	285 (CASF-2016)	Benchmarking and comparative assessment
BindingDB [70]	Binding affinity measurements	Ki, Kd, IC50, EC50	~3.2 million measurements	External validation, independent testing

Table 2: Experimental Uncertainty in Binding Affinity Measurements [74]

Affinity Measure	Estimated Experimental Uncertainty (MAE in log units)	Notes
Ki, Kd, IC50 (Combined)	0.78	Characterized from bioactivity data in ChEMBL
Ki, Kd, IC50 (Combined)	RMSE: 1.04, Pearson R: 0.76	Serves as a reference for model performance upper limit

Critical Considerations and Recent Advancements

The Data Leakage Challenge

A significant challenge identified in recent years is the data leakage between the training set (PDBbind General/Refined sets) and the test benchmark (CASF Core Set) [33] [72]. This leakage arises from high structural similarity between complexes in these sets, meaning models can achieve high benchmark performance by memorizing similar training examples rather than learning generalizable principles of binding. One study found that nearly 49% of CASF test complexes have a highly similar counterpart in the training set, and a simple similarity-based algorithm could achieve competitive performance on CASF by exploiting this leakage [33]. This inflates performance metrics and reduces the real-world utility of models in drug discovery on novel targets.

Next-Generation Dataset Curation and Filtering

To address data leakage and quality issues, new datasets and curation workflows have been developed:

• HiQBind-WF: A semi-automated, open-source workflow that corrects common structural artifacts in PDBbind, such as incorrect bond orders, protonation states, and steric clashes [71]. It applies filters to exclude covalent binders, ligands with rare elements, and small inorganic molecules to create a higher-quality dataset for scoring function development.

• LP-PDBbind (Leak Proof PDBbind): A reorganized version of PDBbind that creates new training, validation, and test datasets by minimizing sequence and chemical similarity of both proteins and ligands between the splits [72]. This approach controls for data leakage more rigorously than random or time-based splits.

• PDBbind CleanSplit: A filtered training dataset created using a structure-based clustering algorithm that combines protein similarity (TM-score), ligand similarity (Tanimoto score), and binding conformation similarity (pocket-aligned ligand RMSD) [33]. It removes training complexes that are structurally similar to any CASF test complex, ensuring a more genuine evaluation of model generalization. Retraining existing models on CleanSplit caused a marked drop in their benchmark performance, revealing that their previous high performance was largely driven by data leakage [33].

Experimental Protocols

Protocol 1: Benchmarking a Scoring Function using CASF-2016

Objective: To evaluate the performance of a new or existing scoring function using the standard CASF-2016 benchmark.

• Data Acquisition:

  • Download the CASF-2016 benchmark set from the PDBbind-CN web server (http://www.pdbbind-cn.org/casf.asp) [69]. This set contains 285 high-quality protein-ligand complexes with reliable binding constants.

• Input Preparation:

  • For each of the 285 complexes, prepare the protein and ligand structure files in the required format for your scoring function (e.g., PDBQT, MOL2).
  • The native crystallographic poses are typically used for evaluating "scoring power" and "ranking power."

• Affinity Prediction:

  • Process each complex through your scoring function to obtain a predicted binding affinity (e.g., in pKd units).

• Performance Evaluation [69] [73]:

  • Scoring Power: Calculate the Pearson Correlation Coefficient (PCC or Rp) between the experimental and predicted binding affinities across all 285 complexes. A higher PCC indicates better predictive accuracy.
  • Ranking Power: For each protein with multiple ligands, calculate the Spearman rank correlation coefficient between the experimental and predicted affinity ranks. The overall ranking power is the average success rate across all such protein clusters.
  • (Optional) Docking Power & Screening Power: Follow the specific protocols outlined in the CASF publications to evaluate these additional metrics [73].

• Result Interpretation:

  • Compare your obtained PCC and other metrics with those of established scoring functions as reported in the CASF-2016 benchmark study [69].
  • A PCC significantly higher than the estimated experimental uncertainty ceiling (Rp ~0.76) may warrant investigation into potential overfitting or data leakage [74].

Protocol 2: Training a Model with a Leak-Proof Data Split

Objective: To train a physics-informed machine learning model for affinity prediction using a data split that minimizes leakage and maximizes generalizability.

• Data Selection:

  • Obtain the PDBbind General Set.
  • Apply a rigorous data splitting strategy. It is strongly recommended to use a pre-existing leak-proof split such as LP-PDBbind [72] or PDBbind CleanSplit [33] instead of a random split. These are available from their respective publications.

• Data Preprocessing & Featurization:

  • Implement physics-informed featurization. For a SE(3)-invariant model like SPIN, this includes:
    • Geometric Features: Calculate interatomic distances, angles, and use them to construct a graph representation of the protein-ligand complex [11] [8].
    • Physicochemical Features: Encode atom types, partial charges, and interaction types (e.g., hydrogen bonds, hydrophobic contacts) as edge and node features [11].

• Model Training:

  • Design a model architecture that incorporates physical inductive biases. For example:
    • Use a Graph Neural Network (GNN) to process the featurized complex [11] [33].
    • Enforce SE(3)-invariance by relying only on interatomic distances and not on global coordinates, ensuring predictions are rotationally and translationally invariant [11].
    • Introduce a regularization term based on the minimum binding free energy principle to guide the learning process [11].
  • Train the model on the training split of your chosen leak-proof dataset.

• Validation and Testing:

  • Use the validation split for hyperparameter tuning and early stopping.
  • Evaluate the final model's performance on the held-out test split of the leak-proof dataset.
  • For a final assessment of generalizability, test the model on a truly external dataset like BDB2020+ [72] or the CSAR HiQ sets [11] [75].

Workflow and Data Relationships

Table 3: Key Computational Tools and Datasets for Binding Affinity Prediction

Resource Name	Type	Function in Research
PDBbind Database [71]	Dataset	Primary source of protein-ligand complexes with 3D structures and binding affinities for model training.
CASF Benchmark [69] [73]	Benchmarking Tool	Standardized benchmark for objectively evaluating scoring power, ranking power, docking power, and screening power.
BindingDB [70]	Database	Source of extensive binding affinity data for external validation and creating independent test sets.
HiQBind-WF [71]	Curation Workflow	Open-source tool for correcting structural artifacts in protein-ligand complexes to create high-quality datasets.
Leak-Proof Splits (LP-PDBbind, CleanSplit) [33] [72]	Dataset Split	Reorganized data splits that minimize protein and ligand similarity between training and test sets to prevent data leakage and enable realistic evaluation of model generalizability.
Graph Neural Networks (GNNs) [11] [33]	Model Architecture	Deep learning framework well-suited for representing the inherent graph structure of protein-ligand complexes.
SE(3)-Invariant Networks [11]	Model Architecture	Neural networks that produce predictions invariant to 3D rotations and translations, a crucial inductive bias for structural data.

The development of robust scoring functions, particularly within the emerging paradigm of physics-informed machine learning (ML), is a cornerstone of modern computational drug discovery. The accuracy of these functions is not monolithic but is evaluated against three distinct, critical capabilities collectively known as the "evaluation powers": scoring power, docking power, and ranking power [4]. Scoring power assesses the model's ability to predict the absolute binding affinity value of a protein-ligand complex. Docking power evaluates the model's capability to identify the native binding pose among a set of decoy conformations. Finally, ranking power measures the model's proficiency in correctly ranking different ligands by their binding affinity for a given protein target [4]. These metrics are indispensable for validating the real-world utility of scoring functions in virtual screening and lead optimization, ensuring that they are not only statistically sound but also operationally effective in a drug discovery pipeline. This document outlines standardized protocols and application notes for the rigorous evaluation of these powers, with an emphasis on benchmarks and methodologies relevant for physics-informed ML approaches.

Quantitative Metrics and Benchmarking at a Glance

The performance of scoring functions across the three evaluation powers is quantified using a standardized set of metrics and benchmarks. The table below summarizes the core metrics and the most widely used benchmark datasets for this purpose.

Table 1: Core Evaluation Metrics and Standard Benchmarks for Scoring Function Validation

Evaluation Power	Key Quantitative Metrics	Primary Benchmark Datasets	Typical Performance Target
Scoring Power	Pearson Correlation Coefficient (PCC/Pearson's R), Root-Mean-Square Error (RMSE) [4]	PDBbind Core Set, CASF Benchmark [4] [76]	High PCC (e.g., >0.8) and low RMSE between predicted and experimental binding affinities [76].
Docking Power	Success Rate of identifying native pose (e.g., RMSD < 2.0 Å) as top rank [22] [77]	CASF (e.g., CASF-2016) [22]	High success rate across a diverse set of protein-ligand complexes.
Ranking Power	Spearman Rank Correlation Coefficient, Enrichment Factor (EF) [4] [22]	DUD-E, DUDE-Z [4] [10]	High Spearman correlation and high early enrichment (e.g., EF1% > 10) [22].

Protocols for Assessing Scoring Power

Objective and Definition

Scoring power measures the ability of a scoring function to accurately predict the absolute binding affinity of a protein-ligand complex, yielding a quantitative value such as pK_d (pK_d = -log₁₀K_d) or pK_i [78]. A model with high scoring power will show a strong linear correlation between its predictions and experimentally determined values, which is crucial for predicting binding constants during lead optimization.

Experimental Protocol and Workflow

The following protocol leverages the curated PDBbind database to ensure a standardized evaluation [76].

• Dataset Curation: Utilize the PDBbind "Core Set" as the benchmark. This set contains a diverse and non-redundant collection of high-quality protein-ligand complexes with experimentally measured binding affinities (K_d, K_i) [76]. Prepare the structures by adding hydrogen atoms, assigning correct protonation states, and performing energy minimization to optimize hydrogen bonding networks [76].
• Affinity Prediction: For each complex in the Core Set, use the scoring function to predict the binding affinity.
• Metric Calculation:
  • Calculate the Pearson Correlation Coefficient (PCC) between the predicted and experimental binding affinities (e.g., predicted pK_d vs. experimental pK_d). The PCC assesses the linear relationship, where a value of 1.0 indicates perfect correlation [76].
  • Calculate the Root-Mean-Square Error (RMSE) to measure the average magnitude of prediction errors.

Figure 1: Scoring power assessment workflow.

Protocols for Assessing Docking Power

Objective and Definition

Docking power evaluates a scoring function's ability to identify the correct, native binding pose of a ligand from a set of computationally generated decoy poses [4] [22]. This is a critical test of the function's accuracy in capturing the physical chemistry of the protein-ligand interaction.

Experimental Protocol and Workflow

The standard benchmark for this task is the Comparative Assessment of Scoring Functions (CASF) dataset, which provides pre-generated decoy poses [22].

• Pose Generation and Preparation: Use a benchmark dataset like CASF-2016, which provides multiple decoy poses (typically with Root-Mean-Square Deviation (RMSD) > 2.0 Å from the native crystal structure) for each protein-ligand complex [22].
• Pose Scoring and Ranking: Score all poses (native and decoys) for a given complex using the scoring function under evaluation.
• Success Rate Calculation: A "success" is counted if the pose with the best (lowest) predicted binding energy is the one closest to the native structure, typically defined as having a heavy-atom RMSD of less than 2.0 Å. The docking power is reported as the success rate across all complexes in the benchmark set [22] [77].

Figure 2: Docking power assessment workflow.

Protocols for Assessing Ranking Power

Objective and Definition

Ranking power, also referred to as "screening power," evaluates a scoring function's ability to prioritize active ligands over inactive ones for a specific protein target [4] [22]. This is directly relevant to the virtual screening task in drug discovery.

Experimental Protocol and Workflow

This protocol uses the DUD-E (Directory of Useful Decoys: Enhanced) dataset, which contains known active ligands and structurally similar but physiologically inactive decoys for multiple targets [22].

• Dataset Preparation: For a specific target protein from DUD-E, prepare the 3D structures of all active ligands and decoy molecules. Dock each molecule into the protein's binding site.
• Compound Scoring and Ranking: Score each protein-ligand complex (actives and decoys) using the scoring function. Rank all compounds from best (most favorable score) to worst.
• Performance Metric Calculation:
  • Enrichment Factor (EF): Calculate the EF at a given percentage (e.g., EF_1%), which measures the concentration of active compounds found in the top X% of the ranked list compared to a random distribution. For example, an EF_1% of 16.72 means the method found 16.72 times more actives in the top 1% than expected by chance [22].
  • Spearman Rank Correlation: This metric assesses how well the ranking of a series of known active ligands by predicted score matches their ranking by experimental affinity [4].

Figure 3: Ranking power assessment workflow.

The Scientist's Toolkit: Essential Research Reagents and Databases

The rigorous evaluation of scoring functions relies on access to high-quality, curated data and specialized software. The following table details key resources that constitute the essential toolkit for researchers in this field.

Table 2: Key Research Reagent Solutions for Evaluation Power Benchmarking

Resource Name	Type	Primary Function in Evaluation	Key Features
PDBbind [4] [76]	Comprehensive Database	Provides data for training and testing scoring power.	A curated collection of protein-ligand complexes with experimentally measured binding affinity data, including a refined set and a core set for benchmarking.
CASF Benchmark [4] [22]	Standardized Benchmark	Designed for the comparative assessment of scoring functions across all three evaluation powers.	Provides pre-processed datasets and decoy structures for standardized tests on scoring, docking, and ranking power.
DUD-E / DUDE-Z [22] [10]	Benchmark Dataset	Used primarily for evaluating ranking/screening power and virtual screening performance.	Contains active ligands and structurally similar but physiologically inactive decoys for multiple protein targets, minimizing false enrichment.
RosettaGenFF-VS [22]	Physics-Based Scoring Function	An example of an advanced scoring function used for high-performance docking and virtual screening.	A physics-based force field that combines enthalpy calculations with an entropy model, demonstrating state-of-the-art performance in benchmarks.
GOLD / AutoDock Vina [78] [22]	Molecular Docking Engine	Used to generate binding poses for ligands as input for scoring function evaluation.	Docking programs that generate multiple plausible binding conformations (poses) which can then be scored and ranked.

Critical Considerations for Physics-Informed ML

For physics-informed machine learning models, adherence to these standardized protocols is paramount. It is critical to perform vertical tests (where the test set contains proteins not seen during training) rather than just horizontal tests (where the same protein may appear in training and test sets bound to different ligands) to ensure generalizability and avoid overfitting [78]. Furthermore, the integration of physics-based terms, such as those accounting for solvation, lipophilic interactions, and torsional entropy, has been shown to be a key driver of performance in ML-based scoring functions, improving their physical realism and predictive accuracy on unseen targets [76].

Physics-Informed Machine Learning (PIML) represents a paradigm shift in computational science, strategically integrating physical laws with data-driven algorithms to overcome limitations of purely data-driven or traditional physics-based models. In the critical field of affinity prediction for drug discovery, this hybrid approach enables more accurate, interpretable, and generalizable predictions of biomolecular interactions. Traditional machine learning models often struggle with limited training data and fail to incorporate fundamental biochemical constraints, while conventional physics-based methods like molecular docking achieve speed but sacrifice accuracy, and rigorous methods like thermodynamic integration are computationally prohibitive [4] [59]. PIML elegantly bridges this divide by embedding physical principles—such as energy conservation, molecular force fields, and thermodynamic constraints—directly into the learning process [61] [79]. This synthesis creates models that learn from available data while maintaining consistency with established physical laws, offering enhanced robustness particularly valuable in data-scarce regimes common early in drug discovery campaigns.

Theoretical Foundations and Comparative Framework

Defining the Modeling Paradigms

Pure Machine Learning Models rely exclusively on patterns discovered from data without explicit physical constraints. In affinity prediction, these typically utilize structural or sequence data to predict binding constants through architectures including graph neural networks, 3D convolutional neural networks, and transformers [4]. While capable of achieving high accuracy with sufficient data, they often suffer from poor generalization outside their training distribution and can produce physically implausible predictions [59].

Traditional Physics-Based Models include molecular docking programs and scoring functions derived from empirical observations or simplified physical equations. These methods are computationally efficient but often insufficiently accurate due to necessary approximations and simplifications of complex molecular interactions [59]. Their rigidity limits application across diverse protein families and binding scenarios [4].

Physics-Informed Machine Learning seamlessly integrates components from both approaches. PIML incorporates physical knowledge through multiple mechanisms: embedding physical equations as regularization terms in loss functions, designing network architectures that inherently obey conservation laws, using physics-based features as model inputs, and incorporating physical simulations directly into training pipelines [61] [80] [79]. This hybrid strategy ensures predictions remain consistent with fundamental principles while maintaining the flexibility to learn complex patterns from data.

Comparative Analysis of Model Characteristics

Table 1: Fundamental characteristics across modeling paradigms

Characteristic	Pure ML Models	Traditional Physics-Based Models	Physics-Informed ML Models
Physical Consistency	Not guaranteed; can violate physical laws	Explicitly enforced through equations	Explicitly enforced through architectural constraints and loss functions
Data Efficiency	Requires large datasets; prone to overfitting with limited data	Highly data-efficient; can work without training data	Improved efficiency through physical priors; can generalize from limited data
Interpretability	Typically "black box"; limited mechanistic insight	High interpretability; direct physical meaning of parameters	Enhanced interpretability through physically meaningful intermediate variables
Computational Cost	Moderate to high inference cost; extensive training required	Fast inference; minimal to no training required	Moderate training cost; efficient inference similar to pure ML
Generalization Ability	Limited to training distribution; poor out-of-domain performance	Good transfer across systems sharing similar physics	Enhanced generalization through physical principles
Implementation Complexity	Moderate (standard ML pipelines)	Low (established software packages)	High (requires domain knowledge and ML expertise)

PIML Methodologies in Affinity Prediction

Architectural Strategies for Physics Integration

PIML implementations employ diverse architectural strategies to incorporate physical knowledge. Physics-constrained loss functions incorporate physical equations as regularization terms, directly penalizing predictions that deviate from established physical laws [79]. Hybrid architecture designs, such as dual-branch parallel frameworks, maintain separate processing streams for physical principles and data-driven patterns, later combining their outputs [80]. Physics-parameterized networks use neural networks to predict parameters within physical equations rather than directly predicting target values [59]. Graph-based physical representations model molecular structures as graphs with nodes and edges representing atoms and bonds, respectively, enabling direct computation of physics-based interactions like van der Waals forces and hydrogen bonding [8] [59].

Case Studies in Drug Discovery Applications

StructureNet: A Physics-Informed Graph Neural Network StructureNet exemplifies the structure-based PIML approach for protein-ligand binding affinity prediction. This framework represents protein and ligand structures as graphs, processed using a GNN-based ensemble deep learning model that focuses exclusively on structural descriptors [8]. By emphasizing geometric and topological descriptors over sequence and interaction data, StructureNet mitigates pattern memorization issues and demonstrates robust performance with a Pearson Correlation Coefficient (PCC) of 0.68 and AUC of 0.75 on the PDBBind v.2020 Refined Set [8]. Ablation studies confirmed geometric descriptors as crucial drivers of model performance, with their removal causing a PCC decrease of over 15.7% [8].

PIGNet: Physics-Informed Generalization for Drug-Target Interactions PIGNet enhances generalization in drug-target interaction prediction by incorporating atom-atom pairwise interactions parameterized with neural networks [59]. The model computes binding affinity as the sum of four physically meaningful energy components: van der Waals interactions, hydrogen bonds, metal-ligand interactions, and hydrophobic interactions [59]. This physics-informed strategy is coupled with comprehensive data augmentation using computationally generated random binding poses, substantially improving both docking power (identifying correct binding poses) and screening power (ranking potential ligands) on the CASF-2016 benchmark compared to previous approaches [59].

Generative AI with Physics-Based Active Learning This innovative approach combines a variational autoencoder (VAE) with nested active learning cycles that iteratively refine molecule generation using physics-based oracles [81]. The workflow integrates chemoinformatics predictors for drug-likeness and synthetic accessibility with molecular mechanics simulations for affinity assessment [81]. When applied to CDK2 and KRAS targets, the system generated novel, synthesizable scaffolds with high predicted affinity, with experimental validation confirming 8 of 9 synthesized molecules showing in vitro activity against CDK2, including one with nanomolar potency [81].

Experimental Protocols and Implementation

Protocol 1: Implementing a PIML Framework for Binding Affinity Prediction

Objective: Establish a standardized protocol for developing and validating physics-informed machine learning models for protein-ligand binding affinity prediction.

Materials and Data Requirements:

• Protein-ligand complex structures from PDBBind database (approximately 19,000 complexes with experimental binding affinities) [4]
• Molecular graph representation tools (RDKit or OpenBabel for graph construction)
• Physics-based feature calculators (van der Waals radii, partial charges, hydrophobicity indices)
• Computational environment (Python with deep learning frameworks PyTorch/TensorFlow, GPU acceleration recommended)

Procedure:

• Data Preparation and Preprocessing (Duration: 2-3 days)
  • Curate protein-ligand complexes from PDBBind refined set, ensuring structural integrity and binding affinity data quality
  • Represent each complex as a molecular graph with atoms as nodes and interactions as edges
  • Compute physics-based features including interatomic distances, interaction types, and energy components
  • Partition data into training/validation/test sets (70/15/15%) with careful attention to structural similarity to avoid data leakage

• Model Architecture Design (Duration: 3-5 days)

  • Implement graph neural network backbone using gated graph attention networks (gated GATs)
  • Design physics-informed interaction networks that explicitly compute pairwise atomic interactions
  • Incorporate physical equations (Lennard-Jones potential, hydrogen bonding potentials) as differentiable operations within the network
  • Establish loss function combining mean squared error for affinity prediction with physics-based regularization terms

• Training and Optimization (Duration: 5-7 days)

  • Initialize model with pre-trained weights if available
  • Employ three-step training strategy: (1) train data-driven branch, (2) train physics-informed branch with physical consistency alignment, (3) fine-tune both branches simultaneously [80]
  • Use Adam optimizer with learning rate scheduling and early stopping based on validation performance
  • Monitor both accuracy metrics (PCC, RMSE) and physical consistency measures

• Validation and Interpretation (Duration: 2-3 days)

  • Evaluate model on CASF-2016 benchmark for docking power, screening power, and scoring power [59]
  • Perform ablation studies to quantify contribution of physics-based components
  • Visualize atomic-level contribution maps to interpret predicted affinities [59]
  • Compare performance against traditional scoring functions (AutoDock Vina, Gold) and pure ML baselines

Troubleshooting Tips:

• If model shows poor generalization, increase diversity of augmented binding poses in training data
• If physical constraints are violated, strengthen regularization hyperparameters
• If training instability occurs, implement gradient clipping or adjust learning rate schedule

Protocol 2: Active Learning with Physics-Based Oracles for Molecular Optimization

Objective: Implement an active learning framework with physics-based oracles for generative molecular design with optimized binding affinity.

Materials:

• Initial compound library (ZINC database or ChEMBL compounds for target of interest)
• Cheminformatics toolkit (RDKit for molecular representation and property calculation)
• Molecular docking software (AutoDock Vina or similar for physics-based affinity estimation)
• Generative model architecture (Variational Autoencoder or Graph-based generative model)

Procedure:

• Initial Model Setup (Duration: 2-3 days)
  • Represent molecules as SMILES strings or molecular graphs
  • Pre-train variational autoencoder on general compound library to learn fundamental chemical space
  • Fine-tune on target-specific compounds if available
  • Establish criteria for drug-likeness (Lipinski's Rule of Five) and synthetic accessibility

• Nested Active Learning Cycles (Duration: 3-4 weeks, iterative)

  • Inner Cycle (Chemical Space Exploration):

    • Sample latent space to generate novel molecular structures
    • Filter generated molecules using chemoinformatic oracles (drug-likeness, SA, diversity)
    • Fine-tune VAE on molecules meeting threshold criteria
    • Repeat for predetermined iterations (e.g., 5-10 cycles) [81]

  • Outer Cycle (Affinity Optimization):

    • Evaluate accumulated molecules with physics-based affinity oracles (molecular docking)
    • Transfer high-scoring compounds to permanent-specific set
    • Fine-tune VAE on high-affinity candidates
    • Implement pose refinement through Monte Carlo simulations (e.g., PELE) [81]
    • Repeat for predetermined cycles (e.g., 3-5 outer cycles)

• Candidate Selection and Validation (Duration: 1-2 weeks)

  • Apply stringent filtration based on binding free energy calculations (MM/GBSA or FEP)
  • Select top candidates for synthesis consideration
  • Validate top candidates through experimental binding assays

Key Considerations:

• Balance exploration (novel chemical space) and exploitation (known high-affinity regions)
• Adjust oracle stringency based on iteration progress
• Implement diversity metrics to prevent mode collapse in generated compounds

Performance Analysis and Benchmarking

Quantitative Performance Comparison

Table 2: Performance comparison across model architectures on standardized benchmarks

Model Architecture	Benchmark Dataset	Performance Metrics	Key Advantages	Limitations
StructureNet (PIML) [8]	PDBBind v.2020 Refined Set	PCC: 0.68, AUC: 0.75	Focus on structural descriptors reduces data memorization; enhanced generalization	Limited to structural information; may miss sequence-based patterns
PIGNet (PIML) [59]	CASF-2016	Superior docking and screening power vs. traditional methods	Explicit atom-atom pairwise interactions; interpretable energy decomposition	Computationally intensive; complex implementation
Generative AI + Active Learning (PIML) [81]	CDK2 and KRAS targets	8/9 synthesized molecules showed in vitro activity; 1 with nanomolar potency	Successfully explores novel chemical spaces; high experimental validation rate	Resource-intensive process; requires multiple optimization cycles
Traditional Docking [59]	CASF-2016	Fast but less accurate	Computational efficiency; well-established workflows	Limited accuracy; poor generalization across protein families
Pure 3D CNN Models [59]	DUD-E, PDBBind	High correlation but poor screening power	Strong pattern recognition with sufficient data	Susceptible to data bias; limited out-of-domain generalization

Qualitative Comparative Analysis

The quantitative performance advantages of PIML approaches manifest across several critical dimensions. Data efficiency is markedly improved, with PIML models achieving superior generalization even with limited training data by leveraging physical principles as inductive biases [61] [80]. Interpretability is significantly enhanced through physically meaningful intermediate representations, such as PIGNet's decomposition into specific interaction types, providing actionable insights for lead optimization [59]. Generalization capability represents perhaps the most significant advantage, with PIML models maintaining robust performance across diverse protein families and scaffold types, substantially reducing false-positive rates in virtual screening scenarios [59].

Table 3: Key resources for implementing PIML in affinity prediction

Resource Category	Specific Tools/Databases	Primary Function	Application Notes
Structural Datasets	PDBBind [4], Binding MOAD [4]	Provide curated protein-ligand complexes with experimental binding affinities	Essential for training and benchmarking; PDBBind contains ~19,000 complexes
Benchmarking Suites	CASF-2016 [59]	Standardized assessment of scoring, docking, and screening power	Critical for comparative model evaluation
Molecular Representation	RDKit, OpenBabel	Cheminformatics toolkit for molecular graph construction and feature calculation	Enable conversion from structural data to graph representations
Physics-Based Simulation	AutoDock Vina, PELE [81]	Molecular docking and pose optimization	Serve as physics-based oracles in active learning cycles
Deep Learning Frameworks	PyTorch, TensorFlow	Implementation of graph neural networks and custom physics-informed layers	Support automatic differentiation for physics-based loss functions
Specialized PIML Tools	PiML Toolbox [82]	Interpretable model development and diagnostics	Provides specialized algorithms for physics-informed modeling
Data Augmentation Tools	Molecular dynamics simulations [8]	Generate additional conformations for training	Captures binding site flexibility; improves model robustness

Visualizing PIML Workflows

PIML Workflow for Drug Discovery

Model Architecture Comparison

The integration of physical principles with machine learning represents a fundamental advancement in binding affinity prediction, addressing critical limitations of both pure data-driven and traditional physics-based approaches. PIML frameworks demonstrate superior performance in key metrics including generalization ability, data efficiency, and interpretability while maintaining physical consistency—attributes particularly valuable in drug discovery where experimental data is often limited and physical realism is paramount. As the field evolves, several promising directions emerge: increased integration with multiscale modeling to capture cellular context, development of more sophisticated physics-informed generative models for molecular design, and adaptation to emerging structural biology data sources such as cryo-EM maps. With regulatory shifts toward reduced animal testing, including the FDA's phasing out of animal studies, sophisticated PIML approaches for in silico prediction are poised to play an increasingly central role in accelerating therapeutic development while reducing costs. The continued refinement of these hybrid methodologies promises to bridge the gap between computational prediction and experimental reality, ultimately enabling more efficient exploration of chemical space and more reliable identification of promising therapeutic candidates.

The Impact of Proper Data Splitting on Reported Performance

In the field of physics-informed machine learning (PIML) for affinity prediction, the accuracy and generalizability of models are paramount for successful drug design. However, a critical, often-overlooked factor that significantly influences reported performance metrics is the strategy used to split data into training, validation, and test sets. Inappropriate data partitioning can lead to data leakage and overestimation of model capabilities, rendering a model that performs excellently in benchmarks useless in real-world applications like virtual screening. This application note examines the profound impact of data splitting strategies, provides protocols for robust evaluation, and integrates these concepts within a PIML framework to enhance the reliability of binding affinity prediction.

The Data Splitting Problem: Evidence from Affinity Prediction

Evidence from recent literature consistently shows that conventional, naive data splitting methods inflate performance metrics, creating a significant gap between benchmark results and real-world predictive power.

Table 1: Documented Impacts of Data Splitting Strategies on Model Performance

Splitting Strategy	Reported Performance (Typical Context)	Performance on Independent Test	Key Findings
Random Splitting	High (e.g., Pearson R up to 0.97 on autocorrelated data) [83]	Poor (Negative R² in stratified split) [83]	Leads to data leakage; models memorize data instead of learning underlying physics. [83]
UniProt-Based Splitting	Lower than random splits [84]	More realistic, but can still lack high accuracy [84]	Preserves data independence but may not fully address structural similarities in complexes. [84]
Temporal Splitting	Lower than random splits [85]	Better reflects real-world deployment [85]	Addresses the inconsistency between offline evaluation and real-world, time-ordered data. [85]
Structure-Based (CleanSplit)	Lower than with leaked data (e.g., performance drop in top models) [33]	Genuinely reflects generalization [33]	Removing training complexes similar to test set causes performance drop, revealing previous overestimation. [33]

A seminal study on predicting protein-ligand binding free energy changes found that while machine learning models showed high predictive correlations (Pearson coefficients up to 0.70) under random partitioning, their performance declined significantly with UniProt-based partitioning, which better preserves data independence [84]. This highlights how conventional random splitting can lead to an overestimation of model accuracy.

Similarly, in click-through rate (CTR) prediction, a domain with similar evaluation challenges, models evaluated with random splits showed a poor correlation with actual online performance compared to those evaluated with temporal splits that mimic real-world data flow [85]. The core issue is autocorrelation in data, where similar data points are present in both training and test sets, allowing the model to "cheat" by effectively interpolating rather than truly learning the underlying function [83].

The PDBbind Case Study and Data Leakage

The PDBbind database is a standard benchmark for training and evaluating structure-based binding affinity prediction models. A critical analysis revealed substantial train-test data leakage between PDBbind and the commonly used Comparative Assessment of Scoring Functions (CASF) benchmark [33]. Alarmingly, some models performed well on the CASF benchmark even when critical information (e.g., the protein structure) was omitted, suggesting they were memorizing biases rather than learning protein-ligand interactions [86] [33].

To address this, the PDBbind CleanSplit was proposed, which uses a structure-based clustering algorithm to remove training complexes that are highly similar to any in the test set. When top-performing models were retrained on CleanSplit, their benchmark performance dropped substantially, confirming that prior high performance was largely driven by data leakage [33].

Experimental Protocols for Robust Data Splitting

Adopting rigorous data splitting protocols is essential for developing reliable PIML models for affinity prediction. The following workflows provide a template for robust experimental design.

Protocol: Structure-Based Data Splitting for Binding Affinity Prediction

Objective: To create training and test sets for protein-ligand binding affinity prediction that minimize data leakage and provide a genuine assessment of model generalization.

Materials:

• Dataset: PDBbind database (general or refined set).
• Software: Clustering algorithm capable of calculating protein (TM-score), ligand (Tanimoto similarity), and binding pose (pocket-aligned ligand RMSD) metrics [33].
• Benchmark: CASF core sets.

Procedure:

• Data Preprocessing: Curate the dataset to remove duplicates and resolve any inconsistencies in affinity measurements.
• Similarity Analysis: For every potential protein-ligand complex pair, calculate:
  • Protein structure similarity (e.g., TM-score).
  • Ligand chemical similarity (e.g., Tanimoto coefficient based on molecular fingerprints).
  • Binding conformation similarity (e.g., pocket-aligned ligand root-mean-square deviation).
• Define Filtering Thresholds: Establish thresholds for the similarity metrics to define "unacceptably similar" complexes. For example, a complex may be flagged if both its protein and ligand similarities exceed set thresholds [33].
• Identify & Remove Test Analogs: Compare all potential training complexes against the intended test set (e.g., CASF-2016). Remove any training complex that exceeds the similarity thresholds with any test complex.
• Reduce Internal Redundancy (Optional but Recommended): Apply the same similarity analysis within the training set. Iteratively remove complexes to break up large similarity clusters, creating a more diverse and less redundant training set [33].
• Final Split Generation: The remaining training complexes and the pristine test set constitute the final, leakage-free split (e.g., PDBbind CleanSplit).

Protocol: Anchor-Query Partitioning Framework

Objective: To leverage limited reference data to improve the prediction of mutation-induced changes in binding free energy.

Materials:

• Dataset: MdrDB or similar database containing protein mutants and their binding free energy changes.
• Embedding Model: Protein language model (e.g., ESM-2) for generating sequence embeddings.

Procedure:

• Data Partitioning: Split the dataset such that no protein sequence (or mutant thereof) appears in both the training and test sets (e.g., UniProt-based split).
• Define Anchor and Query Sets: From the partitioned data, designate a subset of known protein-ligand states as the "anchor set." The remaining data, representing unknown states, forms the "query set."
• Feature Integration: Generate features for both wild-type and mutant proteins using the protein language model.
• Model Training: Train a machine learning model not to directly predict the binding affinity of a query, but to learn the relationship between a query and the anchor points. The model leverages the known states as references to make predictions for the unknown query states.
• Validation: The proposed method has been validated across multiple systems, showing that even a small amount of reference data can significantly enhance prediction accuracy for unknown queries compared to standard splitting and modeling [84].

Diagram 1: Anchor-Query partitioning framework workflow.

Integration with Physics-Informed Machine Learning

Physics-Informed Machine Learning (PIML) presents a powerful solution to the generalization problem by incorporating physical laws as inductive biases, which can reduce over-reliance on potentially biased training data [59] [87] [88].

PIML Strategies for Enhanced Generalization

• Physics-Informed Architectures: Using models that inherently respect physical principles. For example, PIGNet is a physics-informed graph neural network that predicts binding affinity as a sum of atom-atom pairwise interactions (van der Waals, hydrogen bonds, etc.) [59]. This approach encourages the model to learn universal physics of molecular interactions rather than memorizing specific data points.
• Physics-Based Loss Functions: Physics-Informed Neural Networks (PINNs) incorporate governing equations (e.g., differential equations describing physical processes) directly into the loss function during training [87] [88]. This acts as a strong regularizer, guiding the model towards physically plausible solutions even for unseen data.
• Data Augmentation: Generating broader ranges of binding poses and ligands, including non-stable poses, for training can improve a model's ability to distinguish between true binders and decoys, thereby enhancing screening power [59].

Diagram 2: Logical relationship between data leakage, PIML and robust data splitting.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Materials and Tools for Robust Affinity Prediction Research

Item Name	Function / Application	Relevant Protocol / Context
PDBbind Database	A curated database of protein-ligand complexes with binding affinity data for training and benchmarking.	Structure-Based Splitting, General Model Training [86] [33] [4]
CASF Benchmark	A benchmark set specifically designed for the comparative assessment of scoring functions.	Final Model Evaluation (when used with a clean split) [33] [59]
ESM-2 Protein Language Model	Generates contextual, vector-based embeddings from protein sequences for feature extraction.	Anchor-Query Partitioning Framework [84]
Structure-Based Clustering Algorithm	Algorithm to compute multi-modal similarity (protein, ligand, pose) for identifying data leakage.	Generating PDBbind CleanSplit [33]
PIGNet Model	A physics-informed graph neural network that decomposes binding affinity into fundamental physical interactions.	Implementing PIML for improved generalization [59]
AutoDock Vina	A widely used molecular docking program for predicting binding poses; often used for comparison.	Benchmarking against conventional methods [33]

The strategy for splitting data is not a mere preliminary step but a fundamental determinant of the real-world value of a machine learning model in affinity prediction. The pervasive issue of data leakage, as evidenced in standard benchmarks, has led to an over-optimistic assessment of the field's progress. By adopting rigorous, structure-aware data splitting protocols such as CleanSplit and leveraging the generalization power of physics-informed machine learning, researchers can build more reliable and trustworthy models. This combined approach ensures that predictive performance is grounded in a genuine understanding of protein-ligand interactions, ultimately accelerating robust and effective drug discovery.

The accurate prediction of biomolecular binding affinity is a cornerstone of modern drug discovery, serving as a critical filter for identifying viable therapeutic candidates [4]. However, the true utility of any predictive model is not its performance on internal validation sets, but its generalization capability—its ability to make accurate predictions on novel, previously unseen data that reflects real-world application scenarios [89] [4]. This application note examines the framework for achieving and demonstrating true generalization in physics-informed machine learning (PIML) models for affinity prediction, with a specific focus on performance evaluation under strictly independent test conditions.

The challenge of generalization is particularly acute in therapeutic development, where models must perform reliably on distinct protein targets or novel chemical scaffolds not represented in training data [4]. Traditional machine learning approaches often struggle with this challenge, as they may learn dataset-specific biases rather than underlying physical principles [89] [61]. Physics-informed machine learning addresses this limitation by incorporating immutable physical laws and domain knowledge as inductive biases, constraining the hypothesis space and promoting learning of fundamental relationships rather than statistical artifacts [89] [60] [37].

Performance Analysis on Independent Benchmark Sets

Rigorous benchmarking against strictly independent test sets provides the most credible evidence of a model's generalization capability. The following analysis examines performance metrics across established benchmarks that are completely separate from model training data.

Table 1: Model Performance on Independent Binding Affinity Benchmarks

Model	Benchmark Set	Pearson's r	RMSE	Key Characteristic
SPIN [89]	CASF-2016	0.824	1.280	SE(3)-invariant + minimal free energy principles
	CSAR-HiQ	0.816	1.305	SE(3)-invariant + minimal free energy principles
PBCNet [37]	CASF-2016	0.807	1.350	Pairwise binding comparison
Hybrid FEP++ML [90]	16-target benchmark	0.790	1.410	Combined physics and machine learning
Traditional GNN [89]	CASF-2016	0.801	1.420	Geometric features only
Grid-based CNN [89]	CASF-2016	0.756	1.580	Voxelized representation

The superior performance of physics-informed models, particularly SPIN, on independent benchmarks demonstrates the value of incorporating physical principles. SPIN's integration of SE(3)-invariance (ensuring predictions are consistent regardless of molecular orientation) and the principle of minimal binding free energy provides inductive biases that generalize effectively to novel complexes [89]. Theoretical work suggests this improvement stems from physical constraints reducing the effective dimension of the hypothesis space, thereby preventing overfitting and enhancing performance on new data distributions [60].

Table 2: Key Benchmark Datasets for Strict Independence Testing

Dataset	Complexes	Use Case	Independence Principle
CASF-2016 [4]	285	Scoring power	Different complexes from training
CSAR-HiQ [89]	1,117	Ranking power	Novel targets & ligands
PDBbind core sets [4]	Varies (e.g., 290)	Virtual screening	Temporal hold-out
DUD-E [4]	22,886	Enrichment power	Distinct chemical scaffolds

Experimental Protocols for Generalization Assessment

Protocol: Strictly Independent Test Set Construction

Purpose: To create benchmark sets that provide unbiased estimates of real-world performance by ensuring no data leakage between training and evaluation phases.

Materials:

• PDBbind database (general set)
• Clustering software (e.g., BLAST, MMseqs2)
• Chemical similarity tools (e.g., RDKit, OpenBabel)

Procedure:

• Sequence-based clustering: Cluster proteins at 30% sequence identity threshold using BLAST or similar tools [4]
• Chemical scaffold separation: Cluster ligands using Bemis-Murcko scaffold analysis or Taylor-Butina clustering
• Temporal splitting: Reserve recently published complexes as test cases
• Cross-validation strategy: Implement leave-one-cluster-out cross-validation where entire protein families or scaffold classes are held out
• Similarity analysis: Verify that maximum Tanimoto similarity between training and test ligands is <0.7

Validation:

• Confirm no significant overlap in sequence space between training and test proteins
• Verify chemical diversity between training and test ligands
• Ensure binding affinity distributions are similar to prevent dataset bias

Protocol: Physics-Informed Model Training with Inductive Biases

Purpose: To train binding affinity prediction models that incorporate physical principles as inductive biases, enhancing generalization to novel complexes.

Materials:

• Graph neural network framework (e.g., PyTorch Geometric, DGL)
• Molecular structure processing tools (e.g., RDKit, OpenBabel)
• 3D complex coordinates from PDBbind or similar databases

Procedure:

• Graph construction:
  • Represent protein-ligand complex as a graph with atoms as nodes
  • Connect atom pairs within distance threshold (typically 4-5Å) [89]
  • Encode atom features (type, charge, hybridization) and edge features (distance, bond type)

• SE(3)-invariance implementation:

  • Use invariant coordinate representations (distances, angles, dihedrals) rather than absolute coordinates
  • Implement message passing that depends solely on relative distances and orientations
  • Apply random rotations and translations during training as data augmentation

• Energy minimization constraint:

  • Incorporate binding free energy minimization as a regularization term in the loss function: L = L_data + λ·L_physics
  • Define L_physics to penalize predictions that violate the principle of minimal free energy along reaction coordinates [89]

• Training regimen:

  • Pre-train on large unlabeled structural data when available (semi-supervised approach)
  • Fine-tune on labeled affinity data with multi-task learning
  • Utilize early stopping based on independent validation set performance

Validation:

• Test rotational and translational invariance of predictions
• Verify energy minimization trends along binding pathways
• Evaluate on independent benchmark sets following Protocol 3.1

Visualization of Methodologies and Relationships

Workflow for Generalization-Focused Model Development

Physics-Informed Neural Network Architecture

Resource	Type	Function in Generalization Research	Access
PDBbind [4]	Database	Comprehensive collection of protein-ligand structures with binding affinity data	Public
CASF-2016 [4]	Benchmark	Curated test set for scoring power evaluation with strict independence	Public
CSAR-HiQ [89]	Benchmark	High-quality test set for ranking power assessment	Public
BindingDB [4]	Database	Binding affinity data for protein-ligand and other biomolecular interactions	Public
SE(3)-invariant GNN [89]	Algorithm	Base architecture for rotation and translation invariant predictions	Open source
Physics-Informed Loss [89]	Method	Incorporates energy minimization principles as regularization	Implementation dependent
FEP+ [90]	Software	Physics-based simulation for hybrid machine learning approaches	Commercial
QuanSA [90]	Algorithm	Focused machine learning for ligand-based affinity prediction	Commercial/Academic

The integration of these resources enables a comprehensive approach to generalization research. Public benchmarks like CASF-2016 and CSAR-HiQ provide standardized evaluation frameworks, while SE(3)-invariant architectures and physics-informed loss functions incorporate domain knowledge that transfers effectively to novel targets [89] [90]. Hybrid approaches that combine physics-based simulation with machine learning have demonstrated particular strength in generalization, leveraging the complementary strengths of both methodologies [90].

Conclusion

Physics-informed machine learning represents a paradigm shift in binding affinity prediction, moving beyond black-box models to create solutions that are both accurate and physically plausible. The synthesis of foundational physics with advanced deep learning architectures like GNNs addresses critical data scarcity issues and enhances model interpretability. However, the field's maturity hinges on overcoming persistent challenges, particularly concerning data bias, optimization difficulties, and rigorous validation. Future progress will depend on developing more robust and generalizable models, the creation of cleaner and larger benchmark datasets, and the seamless integration of these predictors into broader AI-driven frameworks like AI Virtual Cells (AIVCs). As these advancements converge, PIML is poised to dramatically accelerate the discovery of novel therapeutics, ushering in a new era of efficient and rational drug design.

References

• 1. Physics-informed Machine Learning [https://www.pnnl.gov/explainer-articles/physics-informed-machine-learning]
• 2. Physics-informed machine learning for automatic model ... [https://www.nature.com/articles/s41598-025-92680-8]
• 3. Physics-Informed Machine Learning in Biomedical Science ... [https://arxiv.org/abs/2510.05433]
• 4. Binding Affinity Prediction: From Conventional to Machine ... [https://arxiv.org/html/2410.00709v2]
• 5. Prediction of protein–ligand binding affinity via deep learning ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10939342/]
• 6. Prediction of protein–ligand binding affinity from ... [https://www.nature.com/articles/s41587-022-01307-0]
• 7. Exploring Protein–Ligand Binding-Affinity Prediction | Rowan [https://rowansci.com/blog/exploring-protein-ligand-binding-affinity-prediction]
• 8. StructureNet: Physics-Informed Hybridized Deep Learning ... [https://pubmed.ncbi.nlm.nih.gov/40428123/]
• 9. A physics-informed machine learning framework for ... [https://www.nature.com/articles/s41524-025-01775-3]
• 10. StructureNet: Physics-Informed Hybridized Deep Learning ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12109334/]
• 11. Invariant Physics Informed Network for Binding Affinity ... [https://arxiv.org/abs/2407.11057]
• 12. Laws of thermodynamics [https://en.wikipedia.org/wiki/Laws_of_thermodynamics]
• 13. Thermodynamics | Laws, Definition, & Equations [https://www.britannica.com/science/thermodynamics]
• 14. Learning Universal Representations of Intermolecular ... [https://zitniklab.hms.harvard.edu/projects/ATOMICA/]
• 15. Advancements in deep learning for molecular interactions ... [https://www.sciencedirect.com/science/article/pii/S0006291X25005133]
• 16. DeepDTAGen: a multitask deep learning framework for ... [https://www.nature.com/articles/s41467-025-59917-6]
• 17. DrugForm-DTA: Towards real-world drug-target binding ... [https://pubmed.ncbi.nlm.nih.gov/41050467/]
• 18. A review of physics-informed machine learning for building ... [https://www.sciencedirect.com/science/article/abs/pii/S0306261924025534]
• 19. When physics meets machine learning: a survey of ... [https://link.springer.com/article/10.1007/s44379-025-00016-0]
• 20. Physics-informed machine learning surrogate models [https://www.sciencedirect.com/science/article/pii/S0892687525002523]
• 21. Benchmarking the Structure-Based Virtual Screening ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12363558/]
• 22. An artificial intelligence accelerated virtual screening ... [https://www.nature.com/articles/s41467-024-52061-7]
• 23. evolution of small molecule representations in drug discovery [https://pmc.ncbi.nlm.nih.gov/articles/PMC10689004/]
• 24. 9.1. Working with Sequences [http://d2l.ai/chapter_recurrent-neural-networks/sequence.html]
• 25. Biological Sequence Representation Methods and Recent ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12467266/]
• 26. Kolmogorov–Arnold graph neural networks for molecular ... [https://www.nature.com/articles/s42256-025-01087-7]
• 27. a comprehensive investigation via machine learning methods [https://www.nature.com/articles/s41598-024-84711-7]
• 28. Integrating artificial intelligence in drug discovery and early ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11909971/]
• 29. Recent advances in AI-driven protein-ligand interaction ... [https://www.sciencedirect.com/science/article/pii/S0959440X25000387]
• 30. A Survey of Graph Neural Networks for Drug Discovery [https://arxiv.org/html/2509.07887v1]
• 31. Predicting Potent Compounds Using a Conditional ... [https://www.mdpi.com/2218-273X/13/2/393]
• 32. A generalizable deep learning framework for structure ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12557521/]
• 33. Resolving data bias improves generalization in binding ... [https://www.nature.com/articles/s42256-025-01124-5]
• 34. Understanding Physics-Informed Neural Networks [https://www.mdpi.com/2673-2688/5/3/74]
• 35. What Are Physics-Informed Neural Networks (PINNs)? - MATLAB & Simulink [https://www.mathworks.com/discovery/physics-informed-neural-networks.html]
• 36. Physics-informed neural networks - Wikipedia [https://en.wikipedia.org/wiki/Physics-informed_neural_networks]
• 37. Physics-Informed Machine Learning [https://www.nature.com/collections/jaihfcabgi]
• 38. A practical PINN framework for multi-scale problems with multi-magnitude loss terms - ScienceDirect [https://www.sciencedirect.com/science/article/abs/pii/S0021999124003619]
• 39. The Future of Drug Discovery: 2025 as the Inflection Year ... [https://modelmedicines.com/newsroom/the-future-of-drug-discovery-2025-as-the-inflection-year-for-hybrid-ai-and-quantum-computing]
• 40. Leading AI-Driven Drug Discovery Platforms [https://www.sciencedirect.com/science/article/abs/pii/S0031699725075118]
• 41. Towards accurate high-throughput ligand affinity prediction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6956784/]
• 42. HPDAF: A Practical Tool for Predicting Drug-Target Binding ... [https://www.sciencedirect.com/science/article/abs/pii/S0223523425009183]
• 43. A Gentle Introduction to Graph Neural Networks [https://distill.pub/2021/gnn-intro]
• 44. What is a GNN (graph neural network)? [https://www.ibm.com/think/topics/graph-neural-network]
• 45. Graph neural pre-training based drug-target affinity prediction [https://pmc.ncbi.nlm.nih.gov/articles/PMC11439641/]
• 46. Evidential deep learning-based drug-target interaction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12297561/]
• 47. Multi-Modal Prediction of Antibody-Antigen Binding Free ... [https://arxiv.org/html/2508.19632v1]
• 48. Advancing drug discovery with deep attention neural ... [https://www.sciencedirect.com/science/article/pii/S1359644624001922]
• 49. Drug discovery and mechanism prediction with explainable ... [https://www.nature.com/articles/s41598-024-83090-3]
• 50. MDNN-DTA: a multimodal deep neural network for drug- ... [https://pubmed.ncbi.nlm.nih.gov/40182923/]
• 51. What is Data Leakage in Machine Learning? [https://www.ibm.com/think/topics/data-leakage-machine-learning]
• 52. The QDπ dataset, training data for drug-like molecules and ... [https://www.nature.com/articles/s41597-025-04972-3]
• 53. CLCluster: A redundancy-reduction contrastive learning ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12181778/]
• 54. Exploiting Machine Learning for End-To-End Drug Discovery ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6594828/]
• 55. Solving real-world optimization tasks using physics- ... [https://www.nature.com/articles/s41598-023-49977-3]
• 56. Physics-informed machine learning with optimization ... [https://www.sciencedirect.com/science/article/pii/S0142061523007986]
• 57. Narrowing the gap between machine learning scoring ... [https://www.nature.com/articles/s42004-025-01428-y]
• 58. Benchmarking compound activity prediction for real-world ... [https://www.nature.com/articles/s42004-024-01204-4]
• 59. PIGNet: a physics-informed deep learning model toward ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8966633/]
• 60. Understanding Generalization in Physics Informed Models ... [https://arxiv.org/html/2501.18879v1]
• 61. Physics-informed machine learning: A comprehensive ... [https://www.sciencedirect.com/science/article/pii/S0957417424015458]
• 62. DrugForm-DTA: Towards real-world drug-target binding ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12495441/]
• 63. DrugForm-DTA: Towards real-world drug-target binding ... [https://www.sciencedirect.com/science/article/pii/S2001037025003824]
• 64. Evolutionary Optimization of Physics-Informed Neural ... [https://arxiv.org/html/2501.06572v2]
• 65. Optimizing the Optimizer for Physics-Informed Neural ... [https://arxiv.org/abs/2501.16371]
• 66. Evolutionary PINN Learning Algorithms Inspired by ... [https://www.mdpi.com/2079-3197/11/8/166]
• 67. A new hybrid approach for solving partial differential ... [https://www.sciencedirect.com/science/article/pii/S2590037425000032]
• 68. Evo-PINN Frontiers and Opportunities - ADS [https://ui.adsabs.harvard.edu/abs/arXiv:2501.06572]
• 69. Comparative Assessment of Scoring Functions: The CASF ... [https://pubmed.ncbi.nlm.nih.gov/30481020/]
• 70. Info [https://www.bindingdb.org/rwd/bind/info.jsp]
• 71. A workflow to create a high-quality protein–ligand binding ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11967698/]
• 72. Leak Proof PDBBind: A Reorganized Dataset of Protein- ... [https://arxiv.org/html/2308.09639v2]
• 73. Assessing protein-ligand interaction scoring functions with ... [https://pubmed.ncbi.nlm.nih.gov/29517771/]
• 74. Experimental Uncertainty in Training Data for Protein ... [https://www.sciencedirect.com/science/article/pii/S2667318523000314]
• 75. Benchmark datasets for evaluation [https://bio-protocol.org/exchange/minidetail?id=9225162&type=30]
• 76. New machine learning and physics-based scoring ... [https://www.nature.com/articles/s41598-021-82410-1]
• 77. Comparing sixteen scoring functions for predicting ... [https://www.sciencedirect.com/science/article/abs/pii/S1093326315000285]
• 78. Machine Learning Scoring Functions for Drug Discovery ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9966217/]
• 79. Physics-Informed and Physics-Guided Machine Learning ... [https://www.mechademy.com/resources/physics-informed-and-physics-guided-machine-learning]
• 80. A generic physics-informed machine learning framework ... [https://www.sciencedirect.com/science/article/pii/S0306261925000443]
• 81. Optimizing drug design by merging generative AI with a ... [https://www.nature.com/articles/s42004-025-01635-7]
• 82. API Reference - PiML Toolbox [https://selfexplainml.github.io/PiML-Toolbox/_build/html/modules/classes.html]
• 83. Splitting data randomly can ruin your model [https://datascience.stanford.edu/news/splitting-data-randomly-can-ruin-your-model]
• 84. Evaluating Data Partitioning Strategies for Accurate ... [https://www.sciencedirect.com/science/article/pii/S2001037025004222]
• 85. The Effects of Data Split Strategies on the Offline ... [https://arxiv.org/abs/2406.18320]
• 86. The Impact of Data on Structure-Based Binding Affinity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10671244/]
• 87. Physics-Informed Machine Learning — PIML | by Joris C. [https://medium.com/@joris.c/physics-informed-machine-learning-piml-debe8f856c10]
• 88. Physics-Informed Machine Learning in Biomedical Science ... [https://arxiv.org/html/2510.05433v1]
• 89. SPIN: SE(3)-Invariant Physics Informed Network for Binding Affinity Prediction [https://arxiv.org/html/2407.11057]
• 90. Synergy and Complementarity between Focused Machine Learning and Physics-Based Simulation in Affinity Prediction - PubMed [https://pubmed.ncbi.nlm.nih.gov/34890185/]