Accurate and generalizable protein-ligand binding affinity prediction with geometric deep learning

TL;DR

The paper addresses the generalization challenge in protein-ligand binding affinity prediction for unseen proteins by introducing IPBind, a SE(3)-invariant geometric deep learning method that models interatomic potentials through energy differences between bound and unbound states. The approach uses frame averaging to achieve robust SE(3) invariance, pocket-focused graph representations, and a graph neural encoder to compute atomic contributions, with a training regime that combines Balanced MSE and ranking losses for both accuracy and relative binding strength. Key contributions include the SE(3)-invariant energy-difference formulation, robustness to predicted input structures, and atom-level interpretability demonstrated across CASF16 and Atom3D LBA benchmarks, achieving state-of-the-art or competitive results. The findings have practical implications for drug discovery, offering a scalable, efficient, and explainable tool for predicting binding affinities under realistic conditions where high-quality structures may be unavailable.

Abstract

Protein-ligand binding complexes are ubiquitous and essential to life. Protein-ligand binding affinity prediction (PLA) quantifies the binding strength between ligands and proteins, providing crucial insights for discovering and designing potential candidate ligands. While recent advances have been made in predicting protein-ligand complex structures, existing algorithms for interaction and affinity prediction suffer from a sharp decline in performance when handling ligands bound with novel unseen proteins. We propose IPBind, a geometric deep learning-based computational method, enabling robust predictions by leveraging interatomic potential between complex's bound and unbound status. Experimental results on widely used binding affinity prediction benchmarks demonstrate the effectiveness and universality of IPBind. Meanwhile, it provides atom-level insights into prediction. This work highlights the advantage of leveraging machine learning interatomic potential for predicting protein-ligand binding affinity.

Figures (5)

• Figure 1: Overview of IPBind A. IPbind leverages the difference between bound and unbound status of protein-ligand to predict their binding affinity B. Workflow of IPBind. The initial protein-ligand complex were used to construct three atom-level graphs, after encoder module, atomic contribution of each graph will be summed up to predict the binding affinity between input protein and ligand. n: number of atoms in each molecule, d: number of representation dimensions. C. Radius edges creation D. Message passing layer. E. The prediction module of IPBind.
• Figure 2: Performance comparison between model performance under different level of test protein sequence identity.
• Figure 3: Model performance unbder different structure input. A. Difference between crystal structure vs redocking and colding structures in particular binding state with conformation PDB ID: 3DP4 (cyan: crystal protein, pink: crystal ligand, red: redocking ligand, beige: cofolding protein, green: cofolding ligand). B. & C. Model performance by different structure input in RMSE for LBA60 and 30. D. & E. Model performance by different structure input in Pearson R for LBA60 and 30. Red horizontal line indicated the performance of PSICHIC, as 1.561, 1.290, 0.478, 0.763 in B., C., D., E., respectively.
• Figure 4: IPBind predictions vs PLIP interaction visualization.In the left-side, blue indicates a lower contribution of the atom to binding affinity, while red represents a higher contribution.
• Figure 5: Performance vs average training time over 5 runs on LBA-60 benchmark