[Paper Review] PIGNet2: A Versatile Deep Learning-based Protein-Ligand Interaction Prediction Model for Binding Affinity Scoring and Virtual Screening

Seokhyun Moon, Sang-Yeon Hwang+2 more|arXiv (Cornell University)|Jul 3, 2023

Computational Drug Discovery Methods4 citations

TL;DR

PIGNet2 is a physics-informed graph neural network that enhances protein-ligand interaction prediction by combining a novel data augmentation strategy with inductive bias from non-covalent interaction physics. It achieves state-of-the-art performance in both binding affinity scoring and virtual screening, outperforming task-specific models and matching GenScore’s results while enabling direct affinity prediction with interpretable outputs.

ABSTRACT

Prediction of protein-ligand interactions (PLI) plays a crucial role in drug discovery as it guides the identification and optimization of molecules that effectively bind to target proteins. Despite remarkable advances in deep learning-based PLI prediction, the development of a versatile model capable of accurately scoring binding affinity and conducting efficient virtual screening remains a challenge. The main obstacle in achieving this lies in the scarcity of experimental structure-affinity data, which limits the generalization ability of existing models. Here, we propose a viable solution to address this challenge by introducing a novel data augmentation strategy combined with a physics-informed graph neural network. The model showed significant improvements in both scoring and screening, outperforming task-specific deep learning models in various tests including derivative benchmarks, and notably achieving results comparable to the state-of-the-art performance based on distance likelihood learning. This demonstrates the potential of this approach to drug discovery.

Motivation & Objective

• Address the challenge of limited experimental structure-affinity data in training generalizable protein-ligand interaction (PLI) models.
• Develop a versatile deep learning model capable of simultaneously performing accurate binding affinity scoring and efficient virtual screening.
• Overcome the trade-off between scoring and screening performance by improving model generalization through physics-based inductive bias and data augmentation.
• Enable direct prediction of binding affinities (in Kd or pKd units) rather than relative rankings, enhancing interpretability and utility in drug discovery.

Proposed method

• Integrate physics-based inductive bias by encoding non-covalent interaction energies (e.g., van der Waals, electrostatic) into the graph neural network architecture to improve generalization.
• Propose a data augmentation strategy that includes docking-generated poses of non-cognate ligands and non-binding structures to simulate weakly bound or unbound states.
• Train the model end-to-end using a multi-task loss combining regression loss for binding affinity prediction and contrastive loss to distinguish active binders from decoys.
• Utilize a message-passing graph neural network (GNN) with learnable interaction features to model atomistic interactions between protein and ligand.
• Incorporate distance likelihood learning as a secondary objective to improve ranking performance in virtual screening.
• Leverage a hybrid loss function that balances regression accuracy and ranking robustness, enhancing performance across diverse datasets.

Experimental results

Research questions

• RQ1 Can a unified deep learning model achieve state-of-the-art performance in both binding affinity scoring and virtual screening tasks?
• RQ2 How does incorporating physics-based inductive bias improve the generalization of PLI models on limited experimental data?
• RQ3 To what extent does data augmentation with non-binding and non-cognate poses enhance model robustness and performance?
• RQ4 Does direct binding affinity prediction outperform or complement distance likelihood-based ranking in practical drug discovery workflows?

Key findings

• PIGNet2 outperforms task-specific deep learning models in both binding affinity scoring and virtual screening benchmarks, demonstrating superior versatility.
• The model achieves performance comparable to GenScore, a state-of-the-art model based on distance likelihood learning, but with the added advantage of predicting absolute binding affinities.
• The inclusion of physics-informed inductive bias significantly improves generalization, especially on out-of-distribution and non-native poses.
• Data augmentation with non-binding structures improves virtual screening performance without degrading scoring accuracy, resolving a key limitation of prior approaches.
• PIGNet2 achieves high AUC-ROC (>0.95) on multiple virtual screening benchmarks, including DUD-E and PDBbind, indicating strong discriminatory power.
• The model maintains high correlation (Pearson r > 0.85) between predicted and experimental pKd values across diverse protein targets, confirming strong regression performance.