Investigating recurrent neural networks for feature-less computational drug design

Abstract

This paper investigates Recurrent Neural Networks (RNNs) in the context of virtual High-Throughput Screening (vHTS). In the proposed approach, RNNs, particularly Bidrectional Dynamic Cortex Memories (BDCMs), are trained to derive the chemical activity of molecules directly from human readable strings (SMILES), uniquely describing entire molecular structures. Thereby, the so far obligatory procedure of computing task-specific fingerprint features is omitted completely. Moreover, it is shown that RNNs in principle are capable to incorporate contextual information even over entire sequences. They can not only gain information from this raw string representation, they are also able to produce comparably reliable predictions, i.e. yielding similar and partially even better AUC rates, as previously proposed state-of-the-art methods. Their performance is confirmed on different publicly available data sets. The research reveals a great potential of RNN-based methods in vHTS applications and opens novel perspectives in computational drug design.