Integrating chemical reasoning with generative design of small molecules

Drug discovery is a challenging process that can take 10-years and over $1 billion to bring a new drug to market. Many stages are involved in this pipeline including target discovery, hit identification, hit-to-lead, lead optimisation, and clinical trials; this project focuses on the "hit-to-lead" and "lead optimisation" stages. During these stages, the aim is to design a molecule optimised for multiple properties such as potency against a target, metabolism, and toxicity.

Molecular optimisation is typically achieved through the "design, make, test, analyse (DMTA) cycle" where chemists iteratively propose molecules, test their properties, and suggest follow-up compounds based on the outcome of the previous experiments.

Artificial intelligence and machine learning have recently gained significant attention in a broad range of application areas and there is widespread hope that AI tools could accelerate the drug discovery process. Within small molecule drug discovery, generative AI tools have been proposed that can rapidly generate many drug-like structures of interest.

This includes methods that have been trained on natural language representation of molecules, molecular graphs, and 3D molecular representations in either a supervised manor or as reinforcement learning agents that learn to generate desirable molecules through iterative generation and scoring of molecules.

While AI tools have gained much attention, the impact on drug discovery thus far has been relatively limited. A key barrier to impact and adoption is that many current methodologies do not easily fit into the existing DMTA cycle: models typically are not readily steerable by the user or cannot incorporate existing knowledge from previous experiments.

Furthermore, chemists struggle to interpret the results of the models and contextualise the proposed designs with previous data. To improve the uptake and impact of such methodologies in drug discovery, methods that bridge the usability and communication gaps between generative models and chemists are vital. In particular, a desirable model would be one that can justify its own suggestions and provide hypotheses for each compound it is proposing the chemist should make next.

This work will focus on developing tools and methodologies to transform the impact that advances in AI can have on the drug discovery process. This project will focus on bridging the gap between generative artificial intelligence tools and their users, medicinal chemists. We aim to develop a framework capable of performing chemical reasoning and justifying to chemists why certain molecules should be synthesised.

Such a model should be capable of forming an argument consistent with medicinal chemistry expectations, and the chemist should be able to interact with the model to iterate ideas and arrive at productive hypotheses. We aim to make this method general and modular, allowing drug discovery teams to easily incorporate this tool into the DMTA cycle.

This project will explore knowledge representation strategies for chemistry, artificial intelligence architecture such as large language models, and mathematical constructs such as trees and graphs. The project falls within the following EPSRC research areas "biological informatics", "computational and theoretical chemistry" and "Artificial intelligence technologies".