Source-linked AI summary
PIGNet: A physics-informed deep learning model toward generalized drug-target interaction predictions
Seokhyun Moon, Wonho Zhung, Soojung Yang, Jaechang Lim, Woo Youn Kim
TL;DR
Deep-learning DTI models face limited generalization in scarce, imbalanced data despite their accuracy and computational efficiency. PIGNet addresses this with physics-informed neural-network parameterization of atom-pair interactions and augmented training poses and ligands, achieving improved docking and screening power while enabling ligand-substructure interpretation.
Problem
Deep-learning DTI models have insufficient generalization under scarce and heavily imbalanced data, limiting reliable virtual screening.
Method
PIGNet predicts total binding affinity by summing neural-network-parameterized physics-informed atom-pair interactions and augments training with broader poses and ligands.
Results
PIGNet achieved better generalization and higher accuracy than other deep-learning models, with reported improvements in docking and screening power.
Takeaways & Limitations
Physically decomposed predictions support interpretation of ligand-substructure contributions for drug-candidate optimization and uncertainty-aware virtual screening.
Takeaways & Limitations
PIGNet does not explicitly include a solvent energy component, although water molecules affect ligand-binding thermodynamics and structure.
Abstract
from arXiv · showhide
Recently, deep neural network (DNN)-based drug-target interaction (DTI) models were highlighted for their high accuracy with affordable computational costs. Yet, the models' insufficient generalization remains a challenging problem in the practice of in-silico drug discovery. We propose two key strategies to enhance generalization in the DTI model. The first is to predict the atom-atom pairwise interactions via physics-informed equations parameterized with neural networks and provides the total binding affinity of a protein-ligand complex as their sum. We further improved the model generalization by augmenting a broader range of binding poses and ligands to training data. We validated our model, PIGNet, in the comparative assessment of scoring functions (CASF) 2016, demonstrating the outperforming docking and screening powers than previous methods. Our physics-informing strategy also enables the interpretation of predicted affinities by visualizing the contribution of ligand substructures, providing insights for further ligand optimization.
1 Introduction
DTI models must balance accuracy, computational cost, and generalization under scarce, imbalanced data. PIGNet addresses this by combining physics-informed pairwise interaction modeling with broader training augmentation and reports improved docking, screening, and interpretability.
- Motivation: Limited generalization under scarce, imbalanced data remains a challenge for deep-learning DTI models used in virtual screening.Virtual screening also requires high accuracy and low computational cost.
- Proposed approach: PIGNet models binding affinity as a sum of atom-pair interactions using neural-network-parameterized VDW, hydrogen-bond, metal-ligand, and hydrophobic energy components.This decomposition is intended to help the model recognize commonly observed local interactions in unseen complexes.
- Results: The model improves both docking power and screening power over traditional docking calculations and previous deep-learning models.The comparison is reported as a benchmark result without quantitative values in the supplied passage.
- Interpretability: Pairwise energy predictions allow ligand-substructure contributions to total binding free energy to be estimated for ligand optimization.The paper gives modifying less-contributing moieties into stronger-binding moieties as an example use.
2.1 Related works
Prior DTI models use 3D CNNs or graph-based representations, while physics-informed neural networks parameterize physically motivated functions to improve extrapolation beyond observed regimes.
- Previous DTI models: 3D CNNs improved DTI accuracy but may be limited by high-dimensional grids and the lack of explicit chemical-interaction and bond representations.Graph representations instead encode atoms and bonds as nodes and edges while incorporating 3D proximity.
- Previous DTI models: Graph representations can encode molecular structure with rotation- and translation-invariant relationships, supporting their use in DTI prediction.Atom pairs can be treated as neighbors when their Euclidean distance is within a threshold.
- Physics-informed learning: Hamiltonian neural networks were reported to improve generalization for systems governed by Hamiltonian mechanics.This work positions PIGNet within broader efforts to use physics-informed parameterization.
- Physics-informed learning: Physics-informed neural networks predict parameters of physically motivated energy functions rather than directly predicting total system energy.Prior work reported improved generalization outside the bonding region using this strategy.
2.2 Model architecture
PIGNet represents protein-ligand complexes as molecular graphs with separate covalent-bond and intermolecular-interaction structure, then updates atom features before computing interaction energies.
- Graph representation: PIGNet takes a molecular graph and atom-pair distances as input to predict the binding free energy of a protein-ligand complex.The graph contains node features and adjacency information.
- Graph representation: Two adjacency matrices separately encode covalent bonds within molecules and intermolecular interactions between protein and ligand atoms.This separation lets the architecture propagate the two types of structural information distinctly.
- Feature updates: Gated GAT units update node features using covalent-bond adjacency, while interaction networks update them using intermolecular-interaction adjacency.The gated GAT conveys covalent-bond information and the interaction network conveys protein-ligand interaction information.
- Affinity calculation: After feature updates, the architecture calculates four energy components from atom-pair distances and final node features and sums pairwise affinities into total binding affinity.The four components are VDW, hydrogen-bond, metal-ligand, and hydrophobic interactions.
2.3 Physics-informed parameterized function
PIGNet uses physically motivated interaction terms whose parameters are learned from atom features, distances, and corrected VDW radii, while a rotor penalty represents conformational entropy loss.
- Energy components: PIGNet combines VDW, hydrophobic, hydrogen-bond, and metal-ligand interactions with a rotor penalty to model binding affinity.The total affinity is a weighted sum whose learnable weights account for differences between calculated energies and true free energies.
- Parameterized functions: Each interaction energy depends on atom-pair features and interatomic geometry, including distance and a corrected sum of VDW radii.The correction term is generated by a fully connected layer from the two atom features.
- VDW interaction: The VDW term uses a 12-6 Lennard-Jones potential, with a learned coefficient representing the minimum VDW interaction energy.The total VDW energy sums pairwise contributions across eligible protein-ligand atom pairs.
- Other interactions: Hydrogen-bond, metal-ligand, and hydrophobic terms share a parameterized expression with interaction-specific coefficients and a learnable scalar.Hydrogen bonds and metal-ligand interactions use coefficients motivated by their electron donor-acceptor origin.
- Rotor penalty: The rotor penalty represents entropy loss from restricted ligand rotation and is proportional to the number of rotatable bonds.The number of rotatable bonds is computed with RDKit and scaled by a positive learnable parameter.
2.4 Monte Carlo dropout (MCDO) and epistemic uncertainty
PIGNet uses Monte Carlo dropout during testing to ensemble predictions and quantify epistemic uncertainty from prediction variance.
- 2.4 Monte Carlo dropout (MCDO) and epistemic uncertainty: 30 dropout models are ensembled during testing, and their averaged predictions provide the reported output.The dropout ratio remains 0.1, matching training.
- 2.4 Monte Carlo dropout (MCDO) and epistemic uncertainty: Prediction variance across the ensemble is interpreted as epistemic uncertainty, distinguishing ensemble and single-model PIGNet variants.
2.5 Loss functions
PIGNet combines energy accuracy, derivative constraints, and data-augmentation objectives to learn physically plausible affinities and distinguish stable binders from decoys.
- 2.5.1 Derivative loss: Minimizing energy loss fits experimental binding affinities, while derivative loss encourages stable poses through first- and second-position derivatives.The derivative objective makes validated structures local minima and controls potential-curve sharpness.
- 2.5.1 Derivative loss: Derivative regularization caps potential-curve sharpness at C_der2 = 20.0 to limit excessive energy changes from small positional shifts.
- 2.5.2 Data augmentation loss: Docking augmentation trains experimental poses to have lower energy than decoy poses by a margin of 1.
- 2.5.2 Data augmentation loss: Random- and cross-screening augmentation train non-binding pairs above the −6.8 binding-free-energy threshold to represent chemical diversity.Random screening uses synthetic compounds, while cross screening uses binders paired with other proteins.
- 2.5 Loss functions: PIGNet’s total loss is a weighted sum of energy, derivative, docking, random-screening, and cross-screening losses.The derivative, docking, random-screening, and cross-screening coefficients are 10.0, 10.0, 5.0, and 5.0, respectively.
2.6 Baseline models
The study compares PIGNet with 3D CNN and 3D GNN baselines designed to separate architecture effects from PIGNet’s physics-informed output.
- 2.6 Baseline models: The baseline set contains a reimplemented KDEEP-like 3D CNN and a 3D GNN derived from PIGNet but replacing its physics-informed output with fully connected layers.The CNN baseline uses PIGNet’s atom features while otherwise matching KDEEP’s architecture.
2.7 Dataset
The experiments train on processed PDBbind data with three augmentations and evaluate on CASF-2016 and non-overlapping CSAR NRC-HiQ samples.
- 2.7 Dataset: The processed PDBbind 2019 refined set contains 4,514 training samples, while the held-out PDBbind 2016 core set contains 265 test samples.Redundant core-set samples were removed from training.
- 2.7 Dataset: Training augmentation adds docking, random-screening, and cross-screening complexes generated with Smina and IBS molecules.The passage reports 292,518 docking decoys and 831,885 random-screening complexes, with the cross-screening count truncated in the supplied text.
- 2.7 Dataset: CASF-2016 evaluation uses 283 scoring-and-ranking, 22,340 docking, and 1,612,867 screening samples after processing.
- 2.7 Dataset: The CSAR NRC-HiQ evaluation retains 48 samples from dataset 1 and 37 from dataset 2 after removing overlaps with PDBbind.
3 Results and Discussions
PIGNet was evaluated on CASF-2016 and CSAR NRC-HiQ benchmarks, where it combined strong generalization across virtual-screening tasks with interpretable physics-based interaction estimates and uncertainty quantification.
- 3.1 Assessment of the model performance and the generalization ability: PIGNet single and ensemble models outperformed previous methods in CASF-2016 docking and screening, with best rates of 87% top-1 docking, 55.4% top-1% screening, and 19.6% top-1% average EF.Its scoring and ranking powers exceeded docking methods and remained competitive with other deep learning approaches; CSAR NRC-HiQ results were consistent with CASF-2016 scoring results.
- 3.2.1 Effect of the physics-informed parametrized functions: The physics-informed model significantly increased docking and screening powers while maintaining comparable scoring power to the 3D GNN-based model.PIGNet without data augmentation still outperformed the augmented 3D GNN-based model in docking power, indicating complementary contributions from physical modeling and augmentation.
- 3.2.2 Effect of the DTI-adapted data augmentation strategy: Adding 1,652,085 augmented samples improved docking and screening power for both PIGNet and the 3D GNN-based model.The augmentation strategy did not improve benchmarks containing only true binding complexes, including scoring, ranking, and CSAR NRC-HiQ assessments.
- 3.3 Interpretation of the physically modeled outputs: PIGNet decomposed binding affinity into atom-pair energy contributions, enabling ligand-substructure analysis for PTPN1 and PAF-AH.The model uses neural networks to differentiate atom pairs rather than relying only on handcrafted interaction categories.
- 3.3 Interpretation of the physically modeled outputs: Learnable parameters produced broader carbon-carbon van der Waals distance-energy interactions and pair-type-dependent corrected radii, including an increasing trend from Csp2-Csp2 to Csp3-Csp3 pairs.The authors associate these deviations with learning sufficient physics information and speculate that corrected radii depend on atom radii.
- 3.4 Epistemic uncertainty quantification of PIGNet: Uncertainty estimates ordered samples by prediction quality, with lower-uncertainty groups showing higher Pearson correlation, while uncertainty filtering did not harm overall performance.The comparison also indicates that uncertainty quantification can help filter false positives.
4 Conclusion
PIGNet addresses poor generalization in DTI models trained on scarce and imbalanced data by combining physics-informed equations with broader training-data augmentation. The authors report improved generalization and accuracy, while noting missing explicit solvent-energy modeling as a limitation.
- 4 Conclusion: PIGNet achieved better generalization and higher accuracy than other deep learning models for DTI prediction.The conclusion attributes this outcome to physics-informed parameterized equations and broader chemical and structural diversity in the training data.
- 4 Conclusion: Physics-informed equations act as an inductive bias that guides the neural model toward the underlying physics of chemical interactions.
- 4 Conclusion: Augmenting training data with more chemically and structurally diverse protein-ligand complexes further improved model performance.An ablation study found that both physics-informed modeling and data augmentation contributed to generalization.
- 4 Conclusion: PIGNet does not explicitly include a solvent-energy component, leaving room to improve representation of water-mediated binding effects.The authors suggest introducing solvent energy into PIGNet as a future improvement.
Conflicts of interest
The authors report no conflicts of interest.
- Conflicts of interest: The paper declares that there are no conflicts of interest.
- Conflicts of interest: No competing interests are reported by the authors.
- Conflicts of interest: The conflicts-of-interest statement contains no declared conflicts.