Challenge #2

Hit Identification
Method type (check all that applies)
Deep learning
High-throughput docking
Machine learning
Description of your approach (min 200 and max 800 words)

Our goal with this competition is to evaluate our released tools. Namely, gnina (https://github.com/gnina) and pharmit (http://pharmit.csb.pitt.edu).

We will perform hierarchical, "blind" high-throughput docking on the Enamine Real Database. We will iteratively dock batches of ligands from Enamine Real to a single target structure using gnina and select subsequent batches of ligands based on the resulting docking scores. New batches of Enamine Real compounds will either be selected by similarity to the best-scoring docked ligands or by a simple ML model trained to predict docking scores from ligand fingerprints. The single target structure used for docking will be selected from among the several available crystal structures based on druggability scores.

Pharmit will be used to perform pharmacophore screens of purchasable compounds. Pharmacophores will be constructed in a mostly manual process through inspection of known binders and docked fragments. We will screen all the commercial libraries in Pharmit (ZINC, MolPort, MCULE, Chemspace, LabNetwork).

Hit compounds will be ranked using the default deep learning scoring function of gnina. Multiple conformations of the target structure will be used for final pose optimization and ranking. We will merge the ranked lists from high-throughput docking and pharmacophore screening, but ensure that compounds from both are included in our submission.

What makes your approach stand out from the community? (<100 words)

Pharmit (193 citations) and gnina (44 citations) are established, open-source methods that are currently used by researchers. As such, their performance in this competition will serve as an important baseline to which novel methods can be compared.  

Method Name
gnina+pharmit
Virtual screening of merged selections
Method type (check all that applies)
Deep learning
High-throughput docking
Machine learning
Hybrid of the above
High-throughput docking that uses a deep-learning based scoring function.
Description of your approach (min 200 and max 800 words)

We will dock all of the compounds into the receptor using gnina. We will dock each compound into all of the receptor conformations available from crystal structures. We may also dock the molecules into an ensemble of receptor structures from MD simulations.

Compounds will be ranked by their docking scores.

Method Name
gnina
Free software packages used
Hit Optimization Methods
Method type (check all that applies)
Deep learning
High-throughput docking
Machine learning
Description of your approach (min 200 and max 800 words)

The exact method of hit optimization will depend on the nature of the hits.  Assuming they present a consistent and compelling structural hypothesis for ligand binding, the most likely approach will be to perform template docking using all compounds with similar scaffolds from the relevant made-to-order library.  Template docked compounds will be optimized and ranked using gnina as described for hit identification.  For this step we may utilize neural network scoring functions designed for ΔΔG prediction that currently aren't part of the gnina release (although they might when the work is performed).

If hits do not suggest a compelling structural hypothesis (e.g. lack of consistency in binding modes/interactions), we will endeavor to construct such a hypothesis through enhanced sampling techniques that include receptor flexibility (both docking and simulation based) and generally banging our heads against the problem until we have multiple reasonable structural hypotheses that are mostly consistent with the data.  Template docking will then be performed to suggest compounds compatible with each hypothesis.

During selection of final compounds will try to maximize the informativeness of the resulting SAR. For example, we will also endeavor to select "activity cliff" compounds where a small chemical change is expected to remove activity as this would help validate/reject  the structural hypothesis of ligand binding.  We will also select a diversity of modifications (e.g. not just target one R group).

What makes your approach stand out from the community? (<100 words)

Our approach uses existing, open-source tools; their performance in this competition will serve as an important baseline to which novel methods can be compared.  

Method Name
template+gnina
Free software packages used

gnina