Source-linked AI summary
Development and evaluation of a deep learning model for protein-ligand binding affinity prediction
Marta M. Stepniewska-Dziubinska, Piotr Zielenkiewicz, Pawel Siedlecki
TL;DR
Binding-affinity prediction in structure-based ligand discovery remains limited by reliance on expert-engineered features. Pafnucy uses a 3D-grid deep neural network to predict exact binding affinity and outperformed all 20 state-of-the-art scoring functions tested on CASF.
Problem
Existing machine-learning scoring approaches still rely on expert-engineered features, motivating models that learn task-relevant representations directly from structural and affinity data.
Method
Pafnucy represents protein–ligand complexes as 3D grids and uses convolutional and dense layers to predict exact binding affinity while treating protein and ligand atoms uniformly.
Results
Pafnucy outperformed all 20 state-of-the-art scoring functions tested on the CASF scoring-power benchmark.
Takeaways & Limitations
Pafnucy can serve as a scoring function for virtual screening or an affinity predictor for novel molecules after complex generation.
Takeaways & Limitations
The method assumes that predictions remain similar regardless of the complex’s orientation.
Abstract
from arXiv · showhide
Structure based ligand discovery is one of the most successful approaches for augmenting the drug discovery process. Currently, there is a notable shift towards machine learning (ML) methodologies to aid such procedures. Deep learning has recently gained considerable attention as it allows the model to "learn" to extract features that are relevant for the task at hand. We have developed a novel deep neural network estimating the binding affinity of ligand-receptor complexes. The complex is represented with a 3D grid, and the model utilizes a 3D convolution to produce a feature map of this representation, treating the atoms of both proteins and ligands in the same manner. Our network was tested on the CASF "scoring power" benchmark and Astex Diverse Set and outperformed classical scoring functions. The model, together with usage instructions and examples, is available as a git repository at http://gitlab.com/cheminfIBB/pafnucy
1 Introduction
Structure-based screening predicts protein–ligand binding from 3D complexes, while scoring functions increasingly use machine learning to capture complex relationships. Pafnucy addresses feature-engineering and classification limitations with a deep neural network that predicts exact binding affinity from grid-based complex representations.
- Structure-based virtual screening augments drug discovery by predicting binding affinity or related scores between targets and candidate molecules.
- Machine-learning scoring functions are increasingly used because they can capture nonlinear and complex relationships in protein–ligand data.
- A key limitation of many machine-learning approaches is their continued reliance on expert-designed feature engineering.
- Deep learning reduces feature engineering by learning features through multiple layers of nonlinear transformations as models fit structural and affinity data.
- Earlier 3D-grid convolutional models used classification to identify active ligands, but classification loses information about interaction strength.
- Pafnucy combines a 3D-grid representation with convolutional and dense layers to predict the exact binding affinity value for structure-based applications.
2 Methods
Pafnucy represents protein-ligand complexes as 4D molecular grids and predicts binding affinity with a deep 3D convolutional neural network. The method uses structured dataset splits, external evaluation sets, and systematic rotations during training to improve generalization.
- Input representation: Complexes were cropped to 20-Å cubic boxes centered on the ligand and discretized into 1-Å grids of heavy-atom features.The resulting representation is a 4D tensor.
- Datasets and evaluation: PDBbind v. 2016 complexes were split into 11906 training, 1000 validation, and 290 external core-test complexes after removing overlaps and non-protein-ligand complexes.An additional external test set contained 73 complexes from the Astex Diverse Set.
- Network architecture: The architecture used three 3D convolutional layers with 64, 128, and 256 filters, followed by three dense layers with 1000, 500, and 200 neurons.Convolutional filters were 5-Å cubes, max-pooling patches were 2-Å cubes, and the network had a single output neuron.
- Training: Systematic rotations during training improved generalization, whereas models trained on single orientations showed much worse performance regardless of hyperparameters.Training used Adam with a 10^-5 learning rate, minibatches of 5 examples, and 20 epochs; the lowest-validation-error model was selected.
3 Results
Pafnucy was selected after validation error began increasing following 14 training epochs, and its predictions were evaluated using error and correlation-based metrics. It achieved accurate predictions on the two unseen test sets, outperforming other tested scoring functions on the v. 2013 core set and X-Score on the Astex Diverse Set.
- Model selection: After 14 epochs, validation error began increasing, so the network weights from epoch 14 were saved as the final model.The increase indicated that the model had started to overfit.
- Evaluation metrics: Prediction performance was measured with RMSE and MAE, while score–measurement relationships were assessed using Pearson’s correlation coefficient (R) and regression standard deviation (SD).SD is the measure used in CASF.
- PDBbind test sets: Pafnucy produced accurate predictions on the two test sets that were unknown during training and validation, and outperformed every scoring function tested by Li et al. on the v. 2013 core set.Performance on the v. 2013 core set was substantially worse than on the other subsets, but still exceeded the tested scoring functions.
- Astex Diverse Set: Pafnucy outperformed X-Score on the Astex Diverse Set regardless of the evaluation measure, although correlation was lower for both methods.The Astex dataset contains only 73 complexes, making correlation more sensitive to small prediction changes.
4 Discussion
The discussion examines Pafnucy’s ability to handle rotated molecular complexes and interprets which input features and ligand-receptor regions drive its predictions. The model learned orientation-robust representations while emphasizing molecule identity and local interaction regions.
- Rotation robustness: Pafnucy was designed to produce similar predictions for differently oriented complexes despite lacking rotation invariance.The model had to learn equivalent information from inputs presented at different angles.
- Rotation robustness: PDE10A stability was assessed across 57 ligand complexes, each presented in 24 rotations.The complexes comprised 41 training, 6 validation, and 10 test examples.
- Feature interpretation: The moltype channel had the widest weight range, indicating that distinguishing protein from ligand was crucial for predicting affinity.The input contained 19 channels, including atom-type features expected to have low relevance.
- Feature interpretation: Across two orientations of complex 3WS8, the highest-impact region contained the ligand and its nearest neighborhood, including Gln726 and Phe729 interactions.Gln726 forms a hydrogen bond, while Phe729 forms a π −π interaction with the ligand.
- Feature interpretation: Among 15 corrupted complexes with the largest prediction drops, Tyr693 hydrogen bonding and Met713 hydrophobic contacts also affected predictions.The missing-data experiment deleted a sliding 5-˚A cubic box from 343 corrupted complexes.
- Rotation robustness: Activations differed markedly in the first hidden layer but became increasingly similar toward the output for two orientations.This shows that Pafnucy used different early filters yet extracted the same information from differently presented inputs.
5 Conclusions
The study presented Pafnucy, a deep neural network for structure-based ligand discovery that can score virtual-screening candidates or predict affinities for novel molecules after complex generation. Pafnucy outperformed all 20 state-of-the-art scoring functions tested on the CASF scoring-power benchmark and was released with code, notebooks, examples, and scripts.
- 5 Conclusions: Pafnucy is a deep neural network intended for structure-based ligand discovery campaigns.It can serve as a scoring function in virtual screening or as an affinity predictor for novel molecules after a complex is generated.
- 5 Conclusions: 20 state-of-the-art scoring functions were outperformed by Pafnucy on the CASF “scoring power” benchmark.The comparison involved all 20 scoring functions tested by the CASF authors.
- 5 Conclusions: Pafnucy and its source code are freely available at http://gitlab.com/cheminfIBB/pafnucy.The release includes Jupyter Notebooks for data preparation and result analysis.
- 5 Conclusions: Usage examples and scripts support preparing input data, predicting binding affinity, and training a new network.The materials are intended to facilitate common use cases.
Supplementary Figures
The supplementary figures illustrate the 3D-grid transformation, PDBbind dataset partitioning, training-error progression, and prediction stability under complex rotation.
- 3D-grid representation: The structure-to-grid transformation maps original atom positions onto grid coordinates, with the depicted flat molecule’s Z coordinates collapsed to Z=0.Original atom positions are shown in blue and grid positions in green.
- Dataset division: A supplementary scheme shows how the PDBbind database was divided into training, validation, and test sets.
- Training and model selection: The untrained model had RMSE values of 4.10 for training and 4.26 for validation, and the 14-epoch model was selected as final by validation error.
- Prediction stability: A supplementary figure assesses prediction stability with respect to rotation of the complex.