EquiBind

The foundational mathematical property enabling EquiBind to process 3D molecular structures correctly. A function is SE(3)-equivariant if rotating or translating the input protein-ligand system produces an identically transformed output pose. This eliminates the need for costly data augmentation and ensures the model inherently respects Euclidean geometry. EquiBind achieves this through tensor product layers operating on irreducible representations, ensuring predictions are physically consistent regardless of the input coordinate frame.

Instead of predicting atomic coordinates directly, EquiBind parameterizes the ligand pose using a set of learned 3D keypoints. These keypoints are matched to corresponding points on the protein surface via an optimal transport mechanism. This approach:

• Reduces the search space dimensionality
• Avoids the need for iterative sampling or refinement
• Enables direct-shot prediction in a single forward pass The final atomic pose is reconstructed from the matched keypoints using a deterministic geometric transformation.

The neural network building block that EquiBind uses to construct its SE(3)-equivariant layers. Tensor Field Networks (TFNs) learn to map point cloud features to higher-order tensor fields using learnable filters conditioned on relative positions. Key properties:

• Operate on continuous 3D coordinates, not discretized grids
• Preserve geometric relationships through Clebsch-Gordan tensor products
• Enable the model to reason about directional information (forces, orientations) TFNs provide the theoretical backbone for EquiBind's ability to reason about molecular geometry without breaking rotational symmetry.

EquiBind is designed for blind docking, meaning it predicts the binding pose without requiring a pre-specified binding pocket location. This contrasts with pocket-specific docking methods that require prior knowledge of the active site. EquiBind achieves blind docking by:

• Processing the entire protein surface simultaneously
• Using attention mechanisms to identify potential binding regions
• Learning to focus on geometrically and chemically complementary surface patches This capability is critical for real-world drug discovery where binding sites may be unknown or allosteric.

Traditional docking tools like AutoDock Vina or Glide rely on iterative stochastic sampling and scoring functions—generating thousands of candidate poses and ranking them. EquiBind fundamentally differs:

• Single forward pass: Predicts the pose directly, no sampling loops
• Speed: Orders of magnitude faster (seconds vs. minutes/hours per ligand)
• Learned scoring: The scoring function is implicitly learned end-to-end rather than hand-crafted
• Differentiable: Can be integrated into gradient-based optimization pipelines The trade-off is that EquiBind may sacrifice some atomic-level precision for dramatic speed gains, making it ideal for large-scale virtual screening.

EquiBind formulates the ligand-to-protein binding problem as an optimal transport problem. The model predicts a soft assignment matrix between ligand keypoints and protein surface points, minimizing a learned cost function. This approach:

• Handles variable numbers of points on both sides
• Produces a globally coherent matching rather than greedy local assignments
• Is computed efficiently using the Sinkhorn algorithm
• Enables end-to-end differentiability for training The optimal transport formulation elegantly solves the combinatorial matching problem at the heart of molecular docking.