EquiBind

EquiBind operates on 3D molecular coordinates while respecting the symmetries of Euclidean space. The model's predictions are equivariant to rotation and translation—if you rotate the input protein, the predicted ligand pose rotates identically. This is achieved through tensor field networks and irreducible representations that process geometric features without requiring data augmentation across arbitrary orientations. The architecture ensures that the physical laws governing molecular interactions are baked into the model's inductive biases.

Unlike traditional docking tools that require users to manually define a bounding box around a known binding site, EquiBind performs blind docking—it identifies the binding pocket and predicts the pose simultaneously. The model learns to detect potential binding regions across the entire protein surface by processing all alpha-carbon coordinates and their spatial relationships. This capability is critical for drug repurposing and target fishing applications where the binding site may be unknown or allosteric.

EquiBind represents the ligand as a set of ordered keypoints rather than an atomic graph. These keypoints correspond to rigid structural motifs within the molecule, and the model predicts their 3D coordinates directly. The approach:

• Avoids iterative conformational sampling of rotatable bonds
• Handles ligand flexibility through learned geometric priors
• Enables differentiable end-to-end training without reinforcement learning
• Reduces the docking problem to a point cloud registration task between protein surface points and ligand keypoints

EquiBind predicts a complete binding pose in one forward pass through the neural network, eliminating the need for Monte Carlo sampling, genetic algorithms, or scoring function evaluations that dominate traditional docking. This architectural choice yields inference times of less than one second per complex on a single GPU, compared to minutes for conventional methods. The speed enables virtual screening of billion-compound libraries and real-time interactive docking applications that were previously computationally prohibitive.

The model employs a specialized attention mechanism that operates on pairwise distances between protein residues and ligand keypoints. This distance-aware cross-attention allows the network to learn which protein surface features are relevant for binding each ligand substructure. By operating in distance space rather than Cartesian coordinates, the attention scores become rotationally invariant while preserving the geometric information necessary for pose prediction. The mechanism naturally captures hydrogen bond geometries, hydrophobic contacts, and pi-stacking interactions.

A defining characteristic of EquiBind is that it is trained directly on experimental holo structures from the PDBbind database without requiring pre-generated docked poses as labels. The loss function compares predicted keypoint coordinates to the crystallographic ligand coordinates using a soft bipartite matching algorithm that handles keypoint permutation invariance. This eliminates the circular dependency of training on the outputs of legacy docking software and allows the model to learn binding geometries that may differ from classical scoring function optima.

A geometric deep learning property ensuring that when a 3D input (protein-ligand system) is rotated or translated, the model's output transforms identically. EquiBind enforces this via tensor field networks and irreducible representations, eliminating the need for data augmentation and guaranteeing physically consistent predictions regardless of the coordinate frame. This is distinct from invariance, where the output remains unchanged under transformations.

EquiBind identifies the binding site by predicting a set of ordered keypoints on the protein surface that geometrically correspond to ligand atoms. These keypoints are generated via a cross-attention mechanism between protein residue embeddings and ligand atom embeddings, creating a structural template that guides the final pose. This bypasses traditional pocket-finding algorithms like Fpocket or DeepSite.

Blind docking predicts the ligand pose without prior knowledge of the binding site, searching the entire protein surface. Pocket-constrained docking restricts the search to a predefined region. EquiBind performs blind docking natively by simultaneously identifying the pocket and the pose, making it suitable for target fishing and polypharmacology studies where the binding site is unknown.

EquiBind handles ligand flexibility by predicting torsional angles for rotatable bonds in a single forward pass. The model outputs a torsion score for each bond, representing the preferred dihedral angle. This contrasts with traditional docking tools like AutoDock Vina that iteratively sample torsional space, and with DiffDock which uses a diffusion process over torsion angles.

Traditional docking pipelines use a scoring function (e.g., Vina, GlideScore) to rank thousands of generated poses. EquiBind eliminates this by directly regressing the 3D coordinates of the docked ligand from the protein-ligand pair. The model learns an implicit energy landscape through its equivariant message-passing layers, outputting a single high-confidence pose without iterative sampling or rescoring.

EquiBind's performance is evaluated using Ligand RMSD (Root Mean Square Deviation of atomic positions after symmetry-aware alignment) and Centroid Distance (Euclidean distance between predicted and ground-truth ligand centers). The model reports median RMSD across test sets like PDBbind, with symmetry-corrected RMSD accounting for chemically equivalent atom permutations in symmetric functional groups.