Application of Machine Learning Approaches to Identify New Anticonvulsant Compounds Active in the 6 Hz Seizure Model

Abstract

Epilepsy is the second most common chronic brain disorder, affecting 65 million people worldwide. According to the NIH’s Epilepsy Therapy Screening Program, evaluation of potential new antiepileptic drug candidates begins with assessment of their protective effects in two acute seizure models in mice, the Maximal Electroshock Seizure test and the 6 Hz test. The latter elicits partial seizures through an electrical stimulus of 44 mA, at which many clinically established anti-seizure drugs do not suppress seizures. The inclusion of this “high-hurdle” acute seizure assay at the initial stage of the drug identification phase is intended to increase the probability that agents with improved efficacy will be detected. In this work, we have used machine learning approximations to develop in silico models capable of identifying novel anticonvulsant drugs with protective effects in the 6 Hz seizure model. Linear classifiers based on Dragon conformation-independent descriptors were generated through an in-house routine in R environment and validated through standard validation procedures. They were later combined through different ensemble learning schemes. The best ensemble comprised the 29 best-performing models combined using the MIN operator. With the objective of finding new drug repurposing opportunities (i.e. identifying second or further therapeutic indications, in our case anticonvulsant activity, in existing drugs), such model ensemble was applied in a virtual screening campaign of DrugBank and Sweetlead databases. 28 approved drugs were identified as potential protective agents in the 6 Hz model. The present study constitutes an example of the use of machine learning approximations to systematically guide drug repurposing projects.