Scaffold-Constrained Molecular Generation

Abstract

One of the major applications of generative models for drug discovery targets the lead-optimization phase. During the optimization of a lead series, it is common to have scaffold constraints imposed on the structure of the molecules designed. Without enforcing such constraints, the probability of generating molecules with the required scaffold is extremely low and hinders the practicality of generative models for de novo drug design. To tackle this issue, we introduce a new algorithm, named SAMOA (Scaffold Constrained Molecular Generation), to perform scaffold-constrained in silico molecular design. We build on the well-known SMILES-based Recurrent Neural Network (RNN) generative model, with a modified sampling procedure to achieve scaffold-constrained generation. We directly benefit from the associated reinforcement learning methods, allowing to design molecules optimized for different properties while exploring only the relevant chemical space. We showcase the method's ability to perform scaffold-constrained generation on various tasks: designing novel molecules around scaffolds extracted from SureChEMBL chemical series, generating novel active molecules on the Dopamine Receptor D2 (DRD2) target, and finally, designing predicted actives on the MMP-12 series, an industrial lead-optimization project.

LibINVENT: Reaction-based Generative Scaffold Decoration for in Silico Library Design

“…There are many ways of slicing a molecule to obtain scaffold-decoration pairs for training a decorator model 22 . Recently, exhaustive slicing of single-bonds according to RECAP 23 rules has been explored 17,18 . While this approach appears natural at a glance, it is not always effective for a wet lab chemist attempting to synthesise the proposed compounds 24 .…”

“…Because the aim of our experiments is to showcase some of the capabilities of the Lib-INVENT generative model against publicly known and commonly discussed DRD2 target 17,18 , we remove all compounds sharing a scaffold with compounds found in the dataset obtained by slicing the DRD2 scaffolds according to the set of reactions previously used to slice ChEMBL 35 . This way, the training and validation sets are kept independent and the subsequent validation on the DRD2 target remains unbiased.…”

“…These experiments clearly show that the use cases of the decorator model include working with scaffolds of varying numbers of attachment points. In contrast with prior work on scaffold decoration, which trained the models on data obtained using RECAP rules for slicing 17,18 , Lib-INVENT has been trained on compounds sliced according to chemical reactions. This crucially affects the speed and reliability with which the model adjusts to reaction filter restrictions and is a significant step towards the automation of the symbiosis between in silico and in vitro library generation.…”

“…The algorithm introduced here, called Lib-INVENT, takes the work further and closer towards utilization of chemistry automation platforms by building focused, easy synthesisable libraries. Related models have appeared in the literature over the recent years, focusing both on scaffold decoration itself or on the usage of reinforcement learning to guide the decorative process 14,17,18 . The major enhancement Lib-INVENT brings to these methods lies in the volume and diversity of its output within a focused chemical space.…”

Lib-INVENT: Reaction Based Generative Scaffold Decoration for in silico Library Design

“…The algorithm introduced here, called LibINVENT, takes the work further and closer towards utilization of chemistry automation platforms by building focused, easy synthesizable libraries. Related models have appeared in the literature over the recent years, focusing both on scaffold decoration itself or on the usage of reinforcement learning to guide the decorative process 14,16,17 . The major enhancement LibINVENT brings to these methods lies in the volume and diversity of its output within a focused chemical space.…”