Virtual, make-on-demand chemical libraries are essential for early stage drug discovery. In recent years, their size has steadily grown from millions to billions to trillions of compounds. The sheer size of such libraries poses new challenges for hit discovery and potency optimization through analogue search using enumerative strategies. While a library of trillions of compounds may be non-enumerable, their construction from a reaction-synthon rule based hierarchical structure suggests efficient alternative search strategies. Here, we propose synthon-based generative models to create and search enriched latent representations of rule-based libraries beyond enumeration.<br/>We demonstrate that such models can learn the rules of molecular construction and provide substantial coverage of the chemical space spanned by the library. We use a two-step reinforcement learning approach, based upon removal and recovery of synthons. We first pre-train the model to learn the library, and encourage validity and representability of molecules in the library. We then fine-tune the model to generate molecules with specific properties via a multi-objective optimization problem. We demonstrate our final model is capable of sampling molecules from across the library that also meet our desired property optimization threshold. With this approach, we pave a path towards unlocking ultra large non-enumerable libraries for virtual screening and de novo drug design.

Generative models to unlock exascale compound libraries

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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