Molecular docking helps us understand how small-molecules interact with proteins. This is especially useful in early drug development stages such as target identification and compound screening. Quick and accurate docking software allows researchers to focus their attention on a smaller set of lead molecules for further testing. Traditionally, docking software has employed first principles from physics and chemistry. Recently, deep learning has become all the rage for molecular docking, maybe motivated by the successful application of deep learning to molecular folding.

Method

EquiBind is a deep learning unconstrained docking method which models a fixed receptor and a ligand with selected rotatable bonds. It predicts the binding pocket and the ligand’s conformation within the pocket in one go. Under the hood, EquiBind employs two great ideas from a recent ICLR 2022 Paper: a SE3-invariant graph neural network based architecture and the idea to generate fixed sets of matching key points to define a rotation and translation between receptor and ligand. In addition, the authors innovate a fast method to project a deformed ligand onto the space spanned by the rotatable bonds of a pre-generated ligand conformation.

Performance

The method performs well and fast on the chosen test set (363 randomly selected PDBBind complexes from 2019 or later). The unconstrained method, that is without projecting the deformed ligand onto a more realistic space, leads to very good performance but unrealistic results. Fine-tuning the results with other docking routines such as Smina leads to better performance, but the fast processing advantage is lost.

Table 1 from EquiBind. The green method is the fast unconstrained model producing unrealistic ligand conformations. The blue method also takes 0.1 seconds but a semi-flexible ligand is projected onto the predicted ligand leading to realistic ligands . The orange method uses Smina for fine-tuning but takes 51 seconds – the same as Smina alone.

Playground

As a quick sensitivity check, I used RDKit to generate 5 different ligand conformations for the test set cases and used EquiBind to predict a docking pose for each conformation. For some ligands, the predicted docking pose is quite stable, but for other complexes, the predicted pose varies a lot depending on the initial conformation used.

Take away

EquiBind is a fantastic deep learning docking method – it is fast and performs well. Like many other docking routines, it is not robust in terms of ligand preparation. For example, the ligand conformation seems to affect final results a lot for some compounds in the test set. Nonetheless, we will probably see more deep learning docking methods which build on this architecture.