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.

Use the following command to prepare the structural data using Schrodinger's prepwizard, align the structures to each other, and produce a docking grid.

combind structprep

In parallel, you can prepare the ligand data using the following command. By default, the ligands will be written to seperate files (one ligand per file). You can specify the --multiplex flag to write all of the ligands to the same file.

combind ligprep ligands.csv

Once the docking grid and ligand data have been prepared, you can run the docking. The arguments to the dock command are a list of ligand files to be docked. By default, the docking grid is the alphabetically first grid present in structures/grids; use the --grid option to specify a different grid.

combind dock ligands/*/*.maegz

combind featurize features docking/*/*_pv.maegz

combind pose-prediction features poses.csv

Optionally, you can extract the poses selected by ComBind to a single file. The resulting file will contain the protein structure followed by one pose (the one selected by ComBind) for each ligand.

combind extract-top-poses poses.csv docking/*/*_pv.maegz

To run virtual screening using ComBindVS, you must begin with a structure of the target protein, a set of helper ligands, and a library of compounds to screen.

The first two steps, which can be done in parallel, are to determine poses for the helper ligands using ComBind and to produce an initial set of docked poses for the library to be screened. Then, ComBindVS can be

Use ComBind to predict poses for the known binders and extract the selected poses to a single file, as described above. In the below, we'll assume that this file is named helpers_pv.maegz

The library to be screened can be docked the same way as described above, but here it is highly recommended that you use the --multiplex option during ligprep (to write all the compounds to one file) and the --screen option during docking, which will limit the number of poses per compound to 30 and not used enhanced pose sampling.

combind ligprep library.csv --multiplex
combind dock ligands/library/library.maegz --screen

To compute the ComBind scores for each pose, we need to compute the pairwise] features between each candidate pose to the helper ligand poses.

combind featurize --no-mcss --screen --max-poses 100000 features_screen docking/library-to-grid/library-to-grid_pv.maegz helpers_pv.maegz

With these features in hand, you can then compute the ComBind scores. The ComBind scores for each pose will be written to the indicated numpy file (here screen.npy).

combind screen screen.npy features_screen

It is often convenient to apply the scores to the original poseviewer file and use existing schrodinger utilities to sort the results.

combind apply-scores docking/library-to-grid/library-to-grid_pv.maegz screen.npy combind_scores_added_pv.maegz
$SCHRODINGER/utilities/glide_sort -best_by_title -use_prop_d r_i_combind_score -o combind_pv.maegz combind_scores_added_pv.maegz

See stats_data/pdbs_for_benchmark.csv for a list of PDBs used for benchmarking ComBind. The "query" column gives the PDB for the ligand being docked, the "grid" column gives the structure the query is docked to, and the "mcss<0.5" column indicates whether the query ligand shares a common substructure with the co-crystal ligand in the structure being docked to.

See stats_data/structures.tar.gz for the raw structural data used for benchmarking ComBind.

See stats_data/helper_best_affinity_diverse.csv and stats_data/helper_best_mcss.csv for a list of the "helper ligands" used when benchmarking ComBind. Each row lists a query ligand and one helper ligand; all the entries for each query ligand should be aggregrated. (Most query ligands have 20 associated helper ligands.)

Start by cloning this git repository.

ComBind requires access to Glide along with several other Schrodinger tools and the Schrodinger Python API.

First, make sure that you have a SCHRODINGER environmental variable set pointing to the root of the schrodinger software installation.

You can only access the Schrodinger Python API using their interpretter. Creating a virtual environment that makes their interpretter the default python interpretter is the simplest way to do this. To create the environment and upgrade the relevant packages run the following:

$SCHRODINGER/run schrodinger_virtualenv.py schrodinger.ve
source schrodinger.ve/bin/activate
pip install --upgrade numpy sklearn scipy pandas lmdb

To setup the environment before each use, run source schrodinger.vs/bin/activate to activate the environment and then run source setup.sh to set combind specific environmental variables.