PaperlayerView exact revision on arXiv

Source-linked AI summary

EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

Hannes Stärk, Octavian-Eugen Ganea, Lagnajit Pattanaik, Regina Barzilay, Tommi Jaakkola

arXiv:2202.05146v4q-bio.BMcs.LG

TL;DR

Drug binding prediction is costly because existing methods rely on heavy sampling, scoring, ranking, and fine-tuning, while structural data are scarce. EquiBind uses equivariant geometric deep learning for direct-shot protein–ligand structure prediction and shows strong performance, speed, and compatibility with fine-tuning.

  • Problem

    Drug binding prediction is computationally expensive, and available 3D structural data are scarce for learning reliable models.

  • Method

    EquiBind uses E(3)-equivariant graph neural networks and graph matching to directly predict ligand–receptor complexes, incorporating deterministic placement and ligand flexibility through rotatable-bond torsions.

  • Results

    EquiBind shows strong empirical performance against state-of-the-art baselines, is much faster, and supports improved final predictions when combined with fine-tuning.

  • Takeaways & Limitations

    EquiBind demonstrates the potential of single-shot geometric deep learning and hybrid fine-tuning workflows for advancing deep learning adoption in drug discovery.

  • Takeaways & Limitations

    EQUIBIND only implicitly models side-chain atom positions, and explicitly representing them might improve precise docking but adds considerable computational complexity.

Abstract

from arXiv · show

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.

1. Introduction

Drug binding is a costly but essential drug-discovery problem, while existing accurate docking methods are slow because they sample and refine many candidate complexes. EQUIBIND instead predicts binding structures directly with geometric deep learning, while modeling blind docking and ligand flexibility.

  • Motivation: $1 billion and 10 years are typical reported costs and timelines for developing and testing a single drug before possible FDA approval.The discovery process can also fail because of side effects or experimental disproof of therapeutic efficacy.
  • Motivation: Drug binding is a prerequisite for virtual screening but involves binding sites, orientations, conformational changes, and varied atomic interactions.Relevant interactions include hydrophobic, hydrogen-bonding, and π-stacking contacts.
  • Computational bottleneck: 146 seconds per ligand-receptor pair is the average GNINA runtime, while Glide is up to 9 times slower.Existing methods generate large candidate sets, then score, rank, and fine-tune them.
  • Approach: EQUIBIND uses geometric and graph deep learning to directly predict ligand-receptor complex structures without heavy candidate sampling.The model uses graph matching networks and E(3)-equivariant graph neural networks.
  • Approach: EQUIBIND is independent of initial molecular placements and orientations, producing the same complex for the same unbound structures.This property supports deterministic direct-shot prediction rather than dependence on starting coordinates.
  • Approach: Ligand flexibility is modeled by changing only rotatable-bond torsion angles while keeping bond lengths and angles fixed.The method investigates blind docking under re-docking and flexible self-docking settings with a rigid receptor.
  • Approach: A non-intersection loss is used to prevent steric clashes and unrealistic van der Waals interactions.The method focuses on blind docking but can be adapted when an approximate binding location is known.
  • Approach: A von Mises-based closed-form torsion optimization provides fast fine-tuning instead of expensive differential-evolution RMSD minimization.The procedure fits torsion angles to a predicted atomic point cloud and derives global optima in closed form.

2. Related Work

Related work spans computational structure prediction, protein representations, binding-site identification, docking pipelines, and learned scoring functions. EQUIBIND differs by combining direct geometric prediction with ligand conformational flexibility and closed-form torsion optimization.

  • Protein and molecular structure prediction: Deep learning has accelerated computational protein folding and generation of low-energy small-molecule conformations to complement expensive experimental structural data.These methods support applications including drug discovery.
  • Protein representations: Protein representations for molecular-interaction prediction include backbones and side chains, surfaces, atomic point clouds, amino-acid sequences, and hierarchical structures.Geometric and mesh-based methods use protein surface patches for fast structural representations.
  • Drug binding: Drug-binding prediction includes binding-site identification, molecular placements and orientations, and internal conformational deformations.These components correspond to distinct structural aspects of forming protein-ligand complexes.
  • Drug binding models: Docking software such as AutoDock Vina uses heavy sampling, scoring, ranking, and fine-tuning in a multistage pipeline.Subsequent methods improve individual components of this pipeline, including speed, scoring, and blind docking.
  • Learned scoring functions: 3D CNNs and interaction-graph GNNs improve traditional scoring functions, enabling pose fine-tuning and direct binding-affinity prediction.These models operate on voxelized complexes or molecular interaction graphs.
  • Fine-tuning: EQUIBIND instead optimizes torsion angles in closed form to match a predicted atomic point cloud in less than 1 second.This approach targets the computational cost of iterative torsion optimization.
  • Equivariant geometric learning: Geometric deep-learning methods incorporate Euclidean symmetries through equivariant message passing, with EQUIBIND extending rigid docking to ligand conformational flexibility.The approach follows prior SE(3)-equivariant work while modeling internal ligand changes.

3. EQUIBIND Model

EQUIBIND uses equivariant graph matching to transform ligand and receptor representations, identify binding geometry, and model ligand flexibility while preserving local chemical structure. Its pipeline combines soft distance-geometry projection with a closed-form torsion-angle fitting procedure, while acknowledging that unconstrained projection can yield implausible conformers.

  • Model inputs and representations: EQUIBIND takes a ligand molecular graph with a random unbound 3D conformer and a receptor-bound structure, assuming a rigid protein conformation.Both molecules are represented as spatial k-nearest-neighbor graphs; ligand nodes are atoms, while receptor nodes are residues.
  • Equivariant graph matching: The IEGMN jointly transforms ligand and receptor features and coordinates through intra- and inter-graph message passing using E(3)-equivariant layers.This geometric inductive bias targets consistent predictions under changes in the molecules’ initial spatial positioning and orientation.
  • Binding-site and pose prediction: Predicted coordinate outputs identify the rigid SE(3) transformation, binding site, and deformed ligand atomic point cloud.Ligand and receptor keypoints are trained to match ground-truth binding pocket points using an optimal transport loss; perfect keypoint alignment would recover the docking transformation.
  • Ligand flexibility: Ligand flexibility is modeled by changing rotatable-bond torsion angles while keeping local atomic structures, including bond lengths and adjacent bond angles, fixed.The method follows the assumption that rotatable bonds are flexible whereas local atomic structures and small rings are mostly rigid.
  • Distance geometric constraints: LAS distance-geometry projection applies a fixed number of gradient-descent layers to minimize a distance-constraint objective after coordinate transformations.The soft projection is designed to enforce local atomic-structure constraints, but its gradient-descent form is not guaranteed to enforce them exactly.
  • Fast point-cloud ligand fitting: A closed-form von Mises fitting procedure computes torsion angles for chemically plausible conformers, recovering 44 rotatable-bond angles in 0.04 seconds versus 3143 seconds for differential evolution.The procedure accounts for coupling among all dihedral angles associated with the same rotatable bond; unconstrained point-cloud projection can otherwise produce implausible conformers.

4. Experiments

Experiments evaluate EquiBind on time-split PDBBind complexes using blind self-docking and re-docking, comparing direct predictions and fine-tuned variants with classical and recent baselines. EquiBind offers substantial speed advantages while maintaining strong binding-location and pose quality, though receptor side-chain representation limits precise docking.

  • Dataset and evaluation: 19,443 PDBBind v2020 complexes include 3,890 unique receptors and 15,193 unique ligands; the final test set contains 363 complexes from 125 proteins discovered in 2019 or later.The split excludes test-set ligands from training and uses older complexes for training and validation.
  • Dataset and evaluation: Flexible blind self-docking gives every method a random RDKit conformer and evaluates binding site, orientation, and conformation using L-RMSD, centroid distance, and Kabsch RMSD.Centroid distance measures binding-pocket localization, while RMSD metrics assess structural agreement after hydrogens are removed.
  • Flexible blind self-docking: Vanilla EQUIBIND identifies approximate binding locations well and outperforms baselines on metrics other than the 25th RMSD percentile and the fraction of predictions below 2 Å RMSD.The fine-tuning variants EQUIBIND + Q and EQUIBIND + Q2 retain significant inference speed-ups, while fine-tuned models generally improve quality.
  • Flexible blind self-docking: EQUIBIND is three orders of magnitude faster than the fastest baseline, improves predictions in the > 4 Å RMSD regime, and matches baselines in the low centroid-distance regime without fine-tuning.SMINA fine-tuning improves performance in the low-RMSD regime, and the paper attributes this benefit to the difficulty of predicting torsions, binding location, and orientation jointly.
  • Rigid re-docking: EQUIBIND-R provides much faster rigid re-docking while outperforming baselines on metrics other than the 25th percentiles and the fraction of predictions below 2 Å error.Docking 10 times as many conformers could potentially preserve a 10 times speed-up over the fastest baseline.
  • Qualitative analysis: EQUIBIND usually remains near the true ligand, but can miss exact torsion angles; GNINA is better in some examples, whereas EQUIBIND still finds the correct approximate location.The method’s fine-tuning and conformer-fitting extensions address pose refinement while retaining the direct prediction workflow.
  • Limitations: A limitation is that EQUIBIND implicitly represents receptor side-chain atom positions, and explicitly representing them might improve precise docking.Experiments with surface atoms and receptor atom subgraphs produced only small or no improvements.

5. Conclusion

EQUIBIND is a deep neural model that uses SE(3)-equivariant graph neural networks to predict bound protein–ligand conformations in a single shot. It shows strong empirical performance against state-of-the-art baselines and can be combined with fine-tuning methods.

  • EQUIBIND uses SE(3)-equivariant graph neural networks to predict bound protein–ligand conformations in a single shot.
  • EQUIBIND shows strong empirical performance against state-of-the-art baselines.
  • Combining EQUIBIND with existing fine-tuning methods demonstrates its potential in a hybrid workflow.

A. Additional Results

Additional results examine EQUIBIND’s accuracy, robustness, runtime, receptor representations, and qualitative behavior across docking settings. Fine-tuning improves low-L-RMSD performance, while vanilla EQUIBIND can outperform baselines above a 3.8 Å cutoff.

  • Additional analyses: The additional figures assess conformer sensitivity, molecule-size correlations, symmetric examples, and qualitative predictions against ground-truth conformers.
  • Receptor representations: EQUIBIND-SA adds receptor surface atoms, whereas EQUIBIND-A refines EQUIBIND-R predictions using an all-atom receptor subgraph within 10 Å.Both variants use more explicit receptor atom information than standard EQUIBIND.
  • Flexible blind self-docking: Fine-tuning enables EQUIBIND to match or outperform baselines across all low-L-RMSD ranges, while vanilla EQUIBIND outperforms them above 3.8 Å.This trades prediction speed for additional accuracy through classical physics-based fine-tuning.
  • Flexible blind self-docking: EQUIBIND’s flexible self-docking evaluation gives random ligand conformers and requires prediction of the binding site, orientation, and conformation.
  • New receptors: New-receptor evaluation removes test complexes whose receptors appeared in complexes older than 2019, leaving a stricter test subset.The corresponding table reports results for 144 remaining complexes.

B. Dataset

The dataset uses PDBBind v2020 complexes processed for time-based evaluation, with preprocessing steps addressing invalid structures and symmetric receptor pathologies. Receptor and ligand frequency distributions are also examined.

  • Preprocessing: RDKit filtering reduces the dataset from 19 443 to 19 119 processable complexes before OpenBabel processing and receptor hydrogen correction.
  • Data pathologies: Symmetric receptors can create multiple equally valid ligand positions even though the ground-truth ligand occupies only one location.The pipeline addresses most such cases by retaining receptor connected components near the ligand.
  • Dataset statistics: Histograms report how often each unique ligand and receptor appears in the PDBBind dataset.

C. Implementation Details

Implementation details specify geometric molecular features, runtime handling for docking baselines, hyperparameter schedules, and the validation-based model-selection procedure. The tested search space determines the final EQUIBIND configuration.

  • Featurization: Receptor α-carbon graphs use residue types, Gaussian-basis interatomic distances, and local-frame orientation encodings.
  • Featurization: Ligand atoms are represented by chemical, charge, valence, hydrogen, hybridization, aromaticity, and ring-membership features.
  • Runtime: GLIDE processes each complex on a single thread, while parallel execution across complexes requires separate processes and software licenses.
  • Training: EQUIBIND uses learning rate 10^-4 and EQUIBIND-R uses 3 × 10^-4, with reductions after 60 epochs without validation improvement.Models with the best validation percentage of predictions below 2 Å are tested on the time-based test set.
  • Hyperparameter search: The final standard EQUIBIND parameters are selected from the reported hyperparameter search space using bolded entries.
Loading 2202.05146v4…