De novo molecular design can play an important role in the DMTA (Design, Make, Test, Analyze) cycle in drug discovery. Traditionally de novo design methods rely on a set of hard coded chemical reaction/transformation rules and available chemical reagents to enumerate molecules for searching and scoring. Thus the searchable space is limited to the set of molecules generated from the predefined set of reactions and building blocks. These types of rule based methods can only cover a very small part of the drug-like chemical space.<br/>Deep learning (DL) technology has recently gained remarkable success in various research areas of artificial intelligence. Inspired by the deep learning technology behind AlphaGo, we have developed a novel policy based deep reinforcement learning (DRL) method for automated molecular de novo design using generative sequence models from recurrent neural networks (RNN). The RNN based generative model doesn’t rely on any predefined transformation rules but learn from prior knowledge i.e. the underlying probability distribution generated from millions of molecules historically designed by chemists internally and from the public domain. We explored the capability of structure generation under a broad range of desirability functions including predicted target (and off-target) activity, drug-likeness scores etc. Our results show that the sequence based RNN is a viable method for generating valid chemical structures and combining with policy gradient based reinforcement learning methods, novel drug-like structures with desirable properties can be generated.

<i>De novo</i> molecular design using deep reinforcement learning methods

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.

De novo molecular design using deep reinforcement learning methods