ML Virtual Screening for Cancer Drug Discovery
Machine learning and molecular simulation to find natural breast cancer drug candidates in fungal chemical space.
99.98%
of the initial search space eliminated before lab validation
51,837
compounds screened down to 8 candidates
0.978
ROC-AUC, Extra Trees (Random Forest 0.898)
−8.96
kcal/mol, top hit NPA020664 on PDB 3EQM
A virtual screening pipeline I built during a funded research internship at King Mongkut's University of Technology Thonburi (KMUTT) in Bangkok, Thailand. It searches 51,837 fungal and mushroom compounds for new inhibitors of aromatase (CYP19A1), a key drug target in hormone receptor positive breast cancer, and narrows them down to 8 strong candidates.
Most breast cancers are hormone receptor positive, and aromatase inhibitors are a frontline treatment for them. Finding new inhibitors in the lab is slow and expensive: there are tens of thousands of natural compounds to consider, and only a handful can ever be tested experimentally.
I curated 5,308 compounds with known aromatase activity, trained Random Forest and Extra Trees classifiers on their molecular fingerprints, and used them to score the full MeFSAT fungal library. The highest-ranked compounds were docked into the aromatase structure (PDB 3EQM) with AutoDock Vina, and the best hits went through a full molecular dynamics production run in GROMACS.
The pipeline cut 51,837 compounds down to 8 final candidates, removing more than 99.9% of the search space before any lab work. Extra Trees reached a ROC-AUC of 0.978, and the top hit, NPA020664, binds at −8.96 kcal/mol and was carried through a full molecular dynamics production run. I am co-writing the journal paper with Dr. Thana Sutthibutpong, targeting a Q1 journal.
Curated 5,308 compounds with known aromatase activity to train the classifier
Extra Trees reached ROC-AUC 0.978; Random Forest 0.898
Top candidates validated with docking and full molecular dynamics
