BCC2020

→ Abstract


The presenter(s) will be available for live Q&A in this session (BCC West).

Simon Bray 1, Tim Dudgeon 2, Björn Grüning 3

1 Bioinformatics Group, University of Freiburg. Email: sbray@informatik.uni-
freiburg.de
2 Informatics Matters, Oxford.
3 Bioinformatics Group, University of Freiburg.

Project Website: https://cheminformatics.usegalaxy.eu (webserver);
https://covid19.galaxyproject.org/cheminformatics (project documentation)
Source Code: https://github.com/bgruening/galaxytools/tree/master/chemicaltoolbox;
https://github.com/galaxycomputationalchemistry/galaxy-tools-compchem
License: Apache 2.0 (tools); MIT (documentation)

In silico analysis plays a vital role in drug discovery. Computational work allows
simulation and study of potential drug candidates on a scale that is impossible
experimentally. There are many different steps in virtual screening, each with a
different use-case, accuracy and computational cost. Thus, tools are often linked
into workflows, starting with low-accuracy, low-cost methods and filtering outputs
before applying high-accuracy, high-cost methods. The problem of how to link
multiple tools into a single workflow lends itself ideally to workflow management
systems such as Galaxy or Nextflow. There are already a wide range (over 100) of
cheminformatics tools available in Galaxy under the aegis of the ChemicalToolbox
(Bray et al., 2020, submitted, available at https://cheminformatics.usegalaxy.eu).
These draw on a range of open-source cheminformatics (OpenBabel, RDKit),
molecular dynamics (GROMACS) and molecular docking (rDock, AutoDock Vina)
libraries.

Here, we present a case study in which the ChemicalToolbox was used to assemble
and run workflows for virtual screening of the SARS-CoV-2 main protease (Mpro).
Mpro is one of the main druggable targets for the SARS-CoV-2 virus, and in March
2020 the Diamond Light Source released the results of a crystallographic fragment
screen, providing crystal structures of Mpro in complex with over 50 small organic
molecules. These were used to calculate a list of ~40k candidate molecules and
Galaxy workflows were constructed for charge enumeration, 3D conformer
generation, docking into the Mpro active site with rDock, pose scoring with the
TransFS deep learning approach, and validation against the experimental
structures using the SuCOS structural similarity measure. Running these
workflows led to selection of a shortlist of 500 most promising molecules, to be
purchased for further in vitro study. During the course of the study, over 25
CPU/GPU years, provided by a Europe-wide compute infrastructure network, were
used to generate and score ~80 million docking poses.

In this presentation, the ChemicalToolbox will be briefly introduced, followed by
discussion of the workflow and results.