EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

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.

• The paper introduces a novel framework using SE(3)-equivariant GNNs that directly predicts binding sites and ligand poses, bypassing exhaustive candidate sampling.
• The paper incorporates a fast conformer fitting algorithm that optimizes torsion angles to model ligand flexibility without iterative refinement.
• The paper validates EquiBind’s performance with superior ligand RMSDs and centroid distances, highlighting its potential for efficient high-throughput screening.

EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

The field of computational drug discovery is marked by significant challenges, primarily due to the immense experimental space that arises from the combinatorial complexity of possible drug-like molecules and their interactions with proteins. This paper introduces EquiBind, a novel geometric deep learning model designed to predict drug binding structures with greater efficiency and accuracy compared to traditional methods. EquiBind addresses key limitations in current paradigms by leveraging SE(3)-equivariant graph neural networks (GNNs) to perform direct-shot predictions of receptor binding locations and ligand poses.

Key Contributions

1. Equivariant Graph Neural Network Architecture: EquiBind employs an Independent E(3)-Equivariant Graph Matching Network (IEGMN), combining intra and inter molecular graph message passing, which guarantees that predictions remain consistent regardless of initial molecular placements in three-dimensional space. This property is essential for ensuring robustness in applications where ground truth conformations are unknown or uncertain.
2. Fast and Direct Predictions: Unlike traditional docking approaches that rely on exhaustive candidate sampling, scoring, and fine-tuning, EquiBind performs direct-shot predictions, substantially reducing computational costs. This efficiency is particularly beneficial for virtual screening applications, where millions of potential drug candidates must be evaluated swiftly.
3. Modeling Ligand Flexibility: EquiBind incorporates ligand flexibility through the adjustment of torsion angles in rotatable bonds, achieved via a novel fast conformer fitting algorithm. This approach avoids expensively iterative processes by leveraging closed-form solutions for optimizing dihedral angles according to a von Mises distribution fitted to predicted atomic point clouds.
4. Empirical Validation and Performance: The efficacy of EquiBind is empirically validated against established methods like GNINA, SMINA, and commercial solutions such as GLIDE. EquiBind demonstrates superior performance in predicting ligand RMSDs and centroid distances, often with a significant reduction in inference time. Notably, EquiBind achieves high accuracy in the flexible blind self-docking scenario, exposing its ability to identify accurate binding sites with minimal prior conformational information.

Implications and Future Directions

The practical implications of EquiBind are substantial, particularly for high-throughput virtual screening tasks, where its efficiency and accuracy can greatly enhance the drug discovery pipeline. Moreover, the methodology employed in EquiBind underscores the potential of integrating geometric deep learning with drug discovery, suggesting avenues for further enhancements in terms of protein-ligand interaction modeling.

From a theoretical standpoint, EquiBind's success illustrates the merit of incorporating physical and geometric constraints into deep learning models, paving the way for similar applications in other molecular and structural biology domains. Future developments may involve expanding the scope of EquiBind to include protein flexibility or integrating the model with multi-objective tasks such as predicting biochemical properties alongside structural predictions.

In summary, EquiBind represents a significant step forward in computational methods for drug binding prediction. By combining geometric deep learning with chemically informed model adaptations, it offers a promising tool for accelerating and improving the accuracy of drug discovery efforts.