Predictive chemistry: machine learning for reaction deployment, reaction development, and reaction discovery

Abstract

The field of predictive chemistry relates to the development of models able to describe how molecules interact and react. It encompasses the long-standing task of computer-aided retrosynthesis, but is far more reaching and ambitious in its goals. In this review, we summarize several areas where predictive chemistry models hold the potential to accelerate the deployment, development, and discovery of organic reactions and advance synthetic chemistry.

Fig. 1. Overview of the three main categories of predictive chemistry tasks discussed throughout this review: reaction deployment, development, and discovery. It is useful to consider the extent to which each task represents an extrapolation from known reactivity to new reactivity.

Fig. 2. Overview of five key reaction deployment tasks. Reaction outcome prediction aims to predict the major product given the reactants. One-step retrosynthesis is the reverse task of proposing reaction precursors for new targets. The one-step models are called at each step of multi-step planning, which aims to propose synthesis routes that end in commercially/experimentally accessible building blocks. Atom mapping aligns the atoms on both sides of a reaction, and reaction classification maps reactions into distinct (human-interpretable) classes, both of which are complementary to the core synthesis planning workflow.

Fig. 3. A sample iteration of multi-step planning, which takes a partially-expanded synthetic tree and chooses one chemical node to expand further.

Fig. 4. Overview of key reaction development tasks. Condition recommendation and optimization models can be built based on existing literature and electronic lab notebook data. Substrate scope assessment models have so far mainly been designed based on high-throughput experimentation results, where combinations of two or more reactant types are tested exhaustively. Catalyst/ligand design has been approached either through exhaustive screening campaigns, where ligand combinations are exhaustively enumerated from a library, or through generative modelling in recent years.

Fig. 5. Systematic design of a substrate scope promoting diversity in descriptor space. Reproduced with permission from Kariofillis et al. Copyright 2022 American Chemical Society.

Fig. 6. Overview of key reaction discovery tasks. Mechanism elucidation involves the explicit mapping of elementary reaction steps, and intermediates formed along the way, to achieve atomistic understanding of the chemical process under study. New method development involves the proposal of unprecedented reactivity by machine learning models that transcends trivial modifications of known templates.

Fig. 7. Computed mechanisms for the previously unknown chain-transfer to monomer pathway, competing with the regular chain-growth catalytic cycle, identified through reaction discovery computations. Reproduced with permission from Smith et al. Copyright 2016 American Chemical Society.

Zhengkai Tu

Thijs Stuyver

Connor W. Coley