This repo contains code describing Fragment Ensemble Scoring (FRESCO), a method for hit-finding from a fragment screen. FRESCO uses unsupervised machine learning to model a distribution of pharmacophore combinations in 3D space, which is then used to conduct a virtual screen on the EnamineREAL library for discovery of moelcular hits. This method is unique in directly predicting hit compounds from a fragment screen without any assaying of binding activity required, and has demonstrated success computationally on a retrospective analysis of COVID Moonshot compounds as well as in prospectively discovering hit compounds for both SARS-CoV-2 Mpro and SARS-CoV-2 nsp3-Mac1.
The repo contains all of the code necessary for fitting a model on fragment screen data and scoring molecules, but not the EnamineREAL datadue to storage limitations.
flowchart TD
subgraph "FRESCO Workflow"
subgraph EnamineREAL
A[(VirtualFlow\nConformers)] --> B(Pharmacophore Descriptors)
end
subgraph Ensemble of Fragment-Protein Complexes
Mpro[(Mpro)] --> PharmD(Pharmacophore Distributions)
Mac-1[(Mac-1)] --> PharmD
PharmD -- scikit-learn + scipy interp1d--> KDE(KDEs of pharmacophore distances)
end
subgraph Virtual Screening
B --> E(Score Library)
KDE --> E
E --> F(Choose Top-50k)
F --> G(Filtering)
G -- "physchem (lead-like)"--> H(Clustering)
G -- PAINS --> H
G -- "at most one chiral center" --> H
G -- remove duplicate tautomers --> H
end
end
Functions for utilising FRESCO can be used as a python library by importing the fresco module via pip install .* in the root directory of the repo, which should only take several seconds.
The conda environment used for generating the results in the paper can be recreated by running conda env create -f environment.yml in the root directory of the repo.
This has been tested on macOS and on linux.
The demo notebook demo.ipynb contains a short example of how to train FRESCO on fragment-protein complexes, and use it to score a molecule.