Machine Learning applied to drug discovery. A QSAR model to identify novel TRPV1 channel antagonists.

LLANOS MANUEL; GAVERNET LUCIANA

Buenos Aires

Workshop; The Machine Learning Summer School 2018; 2018

We have developed two Random Forest based QSAR classification models capable of identify TRPV-1 antagonists. For modeling purposes, a dataset consisting of 583 actives and 208 inactives compounds was rationally partitioned by means of two consecutive clustering instances. It allow us to exploit the molecular diversity available in the dataset, deriving a balanced Training Set (156 antagonists + 156 inactives), a Test Set of 104 molecules (52 antagonists and 52 inactives) and a simulated database of 30.325 compounds, with a low proportion of actives spreaded within DUD-E generated inactive decoys (1.1 % actives). The algorithm hyperparameters were fine-tuned performing K-fold cross validation on the Training Set, through a brute force approach. Two differents models were derived, based either on Molecular Descriptors or Fingerprints. The models performs very good on all the validations , with AUC-ROC on the Test Set of 0.983 and 0.992 for the fingerprints and descriptors model, respectively. Enrichment factors of EF1% = 72.27, EF10% = 8.85, and BEDROC = 0.843 on the simulated database also supports the model?s usefulness for virtual screening purposes. Therefore, both models were jointly applied to screen the DrugBank 5.0.10 Database. As a result, 413 molecules were predicted as antagonist by the models. These results were confirmed by a docking model developed by us, and then some interesting candidates were acquired. We are currently testing these compounds as TRPV-1 antagonists by in-vitro assays, in order to experimentally validate the models.