An interpretable AI framework for chemical retrosynthesis

Synthesis planning is a research priority that lies at the interface between chemistry, chemical engineering, and computer science. This process is predominantly carried out via retrosynthetic analysis where the desired chemical compound is iteratively broken down into smaller precursors until available reactants (e.g. industrial platform chemicals) are found. Empowered by the unprecedented development of artificial intelligence and new computational capabilities, machine learning based retrosynthesis has become the new norm for discovery of novel molecules, materials, and drugs. This is mainly accomplished via natural language processing by converting organic chemicals into SMILE strings or other molecular fingerprints. Despite its huge success, this approach is data-driven, thus it does not provide additional physical insight into the underlying reaction system. As the reactivity and chemical properties of organic compounds are greatly determined by their functional groups, a more promising alternative is to develop an automated framework which breaks down the whole molecules into functional groups to enable physical knowledge informed retrosynthesis. This can be achieved by combining cutting-edge techniques such as graph neural networks, graph/network theory, and symbolic regression. Furthermore, wider metrics such as compound toxicity and price can be estimated along with reactivity in the same manner so that multi-criteria decision-making can be addressed to rank new synthetic pathways at an early research stage. Success of this project will therefore deliver an interpretable AI based innovative retrosynthesis framework which accounts for environmental and economic sustainability.