[Paper Review] ETDock: A Novel Equivariant Transformer for Protein-Ligand Docking

Yiqiang Yi, Xu Wan+3 more|arXiv (Cornell University)|Oct 12, 2023

Computational Drug Discovery MethodsComputer Science3 citations

TL;DR

ETDock proposes a novel equivariant transformer model for protein-ligand docking that integrates ligand graph-level features with 3D spatial representations using a TAMformer module, achieving state-of-the-art performance with 80+ ligands predicted within 2 Å RMSD and superior ECDF curves compared to TankBind and EquiBind.

ABSTRACT

Predicting the docking between proteins and ligands is a crucial and challenging task for drug discovery. However, traditional docking methods mainly rely on scoring functions, and deep learning-based docking approaches usually neglect the 3D spatial information of proteins and ligands, as well as the graph-level features of ligands, which limits their performance. To address these limitations, we propose an equivariant transformer neural network for protein-ligand docking pose prediction. Our approach involves the fusion of ligand graph-level features by feature processing, followed by the learning of ligand and protein representations using our proposed TAMformer module. Additionally, we employ an iterative optimization approach based on the predicted distance matrix to generate refined ligand poses. The experimental results on real datasets show that our model can achieve state-of-the-art performance.

Motivation & Objective

• Address the limitations of traditional docking methods that rely on scoring functions and fail to capture 3D spatial and graph-level ligand features.
• Overcome the shortcomings of deep learning-based docking models that neglect 3D geometric constraints and ligand graph-level information.
• Develop a model that jointly learns invariant and equivariant representations of proteins and ligands to improve docking pose prediction accuracy.
• Achieve state-of-the-art performance on the PDBbind v2020 benchmark dataset for protein-ligand docking pose prediction.

Proposed method

• Employ graph isomorphism networks (GIN) and graph vector prediction (GVP) to extract atomic-level features from ligands and proteins.
• Integrate functional-class fingerprints (FCFPs) with GIN-learned atom-level features to capture ligand graph-level information.
• Use learnable outer products to model ligand-protein interaction features.
• Design the TAMformer module with three layers: a triangle layer to encode geometric and spatial constraints, an attention layer to focus on key interaction regions, and a message passing layer to exchange scalar and vector information.
• Apply iterative optimization based on the predicted distance matrix to refine ligand poses.
• Train the model end-to-end to predict accurate 3D coordinates of ligands in the binding pocket using equivariant representations.

Experimental results

Research questions

• RQ1 Can an equivariant transformer model that integrates graph-level and 3D spatial features improve protein-ligand docking pose prediction accuracy?
• RQ2 How does the inclusion of ligand graph-level features (via FCFP) affect the performance of docking models compared to models relying only on atomic-level features?
• RQ3 To what extent does the TAMformer module’s triangle, attention, and message-passing layers enhance the model’s ability to capture physical constraints and key interactions?
• RQ4 How does ETDock compare to state-of-the-art methods like TankBind and EquiBind in terms of RMSD and centroid distance below the 2 Å threshold?
• RQ5 Can iterative optimization based on the predicted distance matrix further refine ligand pose predictions?

Key findings

• ETDock achieved a prediction accuracy where over 80 out of 363 ligands had a root-mean-square deviation (RMSD) below 2 Å, indicating high conformational accuracy.
• More than 120 ligands achieved a centroid distance below 2 Å, demonstrating accurate prediction of ligand position within the binding site.
• The cumulative distribution function (ECDF) curves for RMSD and centroid distance showed that ETDock outperformed both TankBind and EquiBind, particularly below the 2 Å threshold.
• In a case study on PDB 6JMF, ETDock achieved an RMSD of 0.78 Å, significantly lower than both TankBind and EquiBind, confirming superior conformational prediction accuracy.
• The model’s performance was validated on the PDBbind v2020 dataset, where it achieved state-of-the-art results in both RMSD and centroid distance metrics.
• The integration of FCFP with GIN-learned features and the use of learnable outer products enhanced the model’s ability to capture ligand-specific properties and interactions.