Emma Smith King - Practical Machine Learning for Synthetic Chemistry - 6th MABC

6th Machine Learning and AI in Bio(Chemical) Engineering Conference (MABC)
07/July/2023 - Cambridge, UK
For more information: https://www.mabc-cambridge.ai/

Abstract:
Synthetic chemistry has many open challenges: how reaction yields change as reactants and conditions change, [1] how molecules interact with the human body, [2] or the full underlying mechanisms of some workhorse reactions. [3] Machine learning (ML) has seen enormous strides in modeling the world's "black boxes": from image processing and recognition that rival human ability, [4] consistently beating human playersin a variety of games, [5] to the amusing ruminations of the latest large language models. [6] Due to the low standardization of data, few large chemistry-focused datasets, and the mere fact that molecules are difficult systems to model, ML has historically struggled to make headway in the chemical sciences. [7] Recent developments in ML models and increased access to open-source chemistry datasets have opened the door to practical ML models, including DFT and molecular property predictions, activity predictions, and novel scaffold generation. Herein, we present two case studies utilizing recent and classic ML methods to further our predictive ability in and understanding of synthetic chemistry. First, we investigated ML applied to chemical transformations aimed at structural diversification of drug-like molecules, late stage functionalizations (LSFs). These types of reactions are a key component of drug discovery, capable of rapidly exploring the chemical space to yield pharmacokinetically ideal compounds. [8] However, predicting the regiochemical outcomes of LSFs is still an open challenge in the field. Notably, experimental data curation is labor-intensive and time consuming. We report the development of an approach that combines a message passing neural network and 13C NMR-based transfer learning to predict the atom-wise probabilities of functionalization. [9] We validated our model retrospectively and with a series of prospective experiments, showing that it accurately predicts the outcomes of Minisci-type and P450 transformations, outperforming state-of-the-art Fukui-based reactivity indices and other graph-based ML models (Figure 1A). [10] The second case study developed a dataset-ambivalent ML framework to analyze highthroughput experimentation (HTE) datasets. HTE hasthe potential to improve our understanding of organic chemistry by systematically interrogating reactivity across diverse chemical spaces. One notable bottleneck is the lack of facile analyzers which can interpret of these data's hidden chemical insights. [11] Herein we report the development of a High Throughput Experimentation Analyzer (HiTEA), a robust and statistically rigorous framework which is applicable to any HTE dataset regardless of size, scope, or target reaction outcome. [12] HiTEA is validated on previously proprietary medicinal chemistry data, elucidating hidden biases and relationships between reaction components (Figure 1B).