[Paper Review] EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

H. Stärk, Octavian-Eugen Ganea+3 more|arXiv (Cornell University)|Feb 7, 2022

Computational Drug Discovery MethodsComputer Science98 citations

TL;DR

EquiBind is an SE(3)-equivariant geometric deep learning model that directly predicts ligand binding location and bound pose, enabling ultra-fast blind docking with optional fast, torsion-angle fine-tuning using a closed-form von Mises solution.

ABSTRACT

Predicting how a drug-like molecule binds to a specific protein target is a core problem in drug discovery. An extremely fast computational binding method would enable key applications such as fast virtual screening or drug engineering. Existing methods are computationally expensive as they rely on heavy candidate sampling coupled with scoring, ranking, and fine-tuning steps. We challenge this paradigm with EquiBind, an SE(3)-equivariant geometric deep learning model performing direct-shot prediction of both i) the receptor binding location (blind docking) and ii) the ligand's bound pose and orientation. EquiBind achieves significant speed-ups and better quality compared to traditional and recent baselines. Further, we show extra improvements when coupling it with existing fine-tuning techniques at the cost of increased running time. Finally, we propose a novel and fast fine-tuning model that adjusts torsion angles of a ligand's rotatable bonds based on closed-form global minima of the von Mises angular distance to a given input atomic point cloud, avoiding previous expensive differential evolution strategies for energy minimization.

Motivation & Objective

• Address the high cost of conventional docking by predicting binding location and ligand pose directly without heavy sampling.
• Incorporate geometric inductive biases to ensure predictions are invariant to initial 3D placements.
• Model ligand flexibility through torsion-angle adjustments while keeping local bond geometry fixed.
• Provide a fast, differentiable framework that can complement energy-based fine-tuning methods.

Proposed method

• Use independent SE(3)-equivariant graph matching networks to jointly transform ligand and receptor features and coordinates.
• Represent ligand and receptor as spatial k-NN graphs and apply IEGMN to achieve pose prediction invariant to rigid-body transformations.
• Identify binding-site keypoints via SE(3)-equivariant attention and align them with an optimal transport loss to recover the docking transformation.
• Model ligand flexibility by (i) learning deformed atomic point clouds and (ii) enforcing chemically plausible conformers through a fast torsion-angle fitting step.
• Introduce a fast closed-form method to estimate rotatable-bond dihedrals by maximizing a von Mises likelihood, avoiding iterative differential evolution for torsion optimization.
• Optionally combine EquiBind with existing energy-based docking refinements (QuickVina2, SMINA, or GLIDE) for improved final poses.

Experimental results

Research questions

• RQ1 Can a geometry-aware, end-to-end model predict ligand binding locations and poses without heavy candidate sampling?
• RQ2 How can SE(3) equivariance be leveraged to ensure consistent docking predictions regardless of initial molecule orientations?
• RQ3 What is an efficient method to model ligand conformational flexibility while preserving chemical plausibility in docking?
• RQ4 To what extent can a post-prediction torsion-angle refinement improve pose accuracy without sacrificing speed?

Key findings

• EquiBind achieves significant speed-ups over conventional docking pipelines while delivering competitive or superior pose quality.
• Fine-tuning with energy-based refinements (e.g., Q, Q2, S) further improves results with manageable runtime costs.
• A fast, closed-form torsion-angle fitting method substantially reduces the time for conformer adjustment compared to differential evolution while maintaining chemical plausibility.
• Rigid-body variants (EquiBind-R) perform well for rapid rigid re-docking, enabling high-throughput screening pipelines.
• A hybrid approach that combines EquiBind with existing docking refinements yields a favorable trade-off between speed and accuracy.