This repository has a few changes to adapt ComBind to support RevBind scores applied to structural models.
To score a structural model with RevBind, follow the ComBind pose prediction for known binders instructions (below). This include data curation, data preparation (including docking), and featurization of docked poses.
For this follow-up manuscript, the code was minimally adapted so that running pose prediction ComBind will output the RevBind score in the top lines of out.csv.
combind pose-prediction features out.csv --max-poses 1
Setting --max-poses 1 uses only the top pose selected by the docking score (ComBind pose prediction is not run).
By default, ComBind considers up to 100 poses.
combind pose-prediction features out.csv
When running full ComBind pose selection, the top line of out.csv will include a RevBind score that integrates the ComBind score with the physics-based docking score, as described in the manuscript.
Running ComBind with pose selection (--max-poses 100, the default) generally provides the most reliable results for scoring models and identifying the best poses.
ComBind integrates data-driven modeling and physics-based docking for improved binding pose prediction and binding affinity prediction.
Given the chemical structures of several ligands that can bind a given target protein, ComBind solves for a set of poses, one per ligand, that are both highly scored by physics-based docking and display similar interactions with the target protein. ComBind quantifies this vague notion of "similar" by considering a diverse training set of protein complexes and computing the overlap between protein–ligand interactions formed by distinct ligands when they are in their correct poses, as compared to when they are in randomly selected poses. To predict binding affinities, poses are predicted for the known binders using ComBind, and then the candidate molecule is scored according to the ComBind score w.r.t. the selected poses.
First, see instructuctions for software installation at the bottom of this page.
Running ComBind can be broken into several components: data curation, data preparation (including docking), featurization of docked poses, and the ComBind scoring itself.
Note that if you already have docked poses for your molecules of interest, you can proceed to the featurization step. If you are knowledgable about your target protein, you may well be able to get better docking results by manually preparing the data than would be obtained using the automated procedure implemented here.
To produce poses for a particular protein, you'll need to provide a 3D structure of the target protein and chemical structures of ligands to dock.
These raw inputs need to be properly stored so that the rest of the pipeline can recognize them.
The structure(s) should be stored in a directory structures/raw.
Each structure should be split into two files NAME_prot.mae and NAME_lig.mae
containing only the protein and only the ligand, respectively.
If you'd prefer to prepare your structures yourself, save your
prepared files to structures/proteins and structures/ligands. Moreover,
you could even just begin with a Glide docking grid which you prepared yourself
by placing it in docking/grids.
Ligands should be specified in a csv file with a header line containing at least the entries "ID" and "SMILES", specifying the ligand name and the ligand chemical structure.