CovCysPredictor: Predicting Selective Covalently Modifiable Cysteines Using Protein Structure and Interpretable Machine Learning

Targeted covalent inhibition is a powerful therapeutic
modality
in the drug discoverer’s toolbox. Recent advances in covalent
drug discovery, in particular, targeting cysteines, have led to significant
breakthroughs for traditionally challenging targets such as mutant
KRAS, which is implicated in diverse human cancers. However, identifying
cysteines for targeted covalent inhibition is a difficult task, as
experimental and in silico tools have shown limited accuracy. Using
the recently released CovPDB and CovBinderInPDB databases, we have
trained and tested interpretable machine learning (ML) models to identify
cysteines that are liable to be covalently modified (i.e., “ligandable”
cysteines). We explored myriad physicochemical features (pKa, solvent exposure, residue electrostatics,
etc.) and protein–ligand pocket descriptors in our ML models.
Our final logistic regression model achieved a median F1 score of 0.73 on held-out test sets. When tested on a small sample
of holo proteins, our model also showed reasonable
performance, accurately predicting the most ligandable cysteine in
most cases. Taken together, these results indicate that we can accurately
predict potential ligandable cysteines for targeted covalent drug
discovery, privileging cysteines that are more likely to be selective
rather than purely reactive. We release this tool to the scientific
community as CovCysPredictor.