Structural diversification of lead molecules is a key component of drug discovery to explore close-in chemical space. Late stage functionalizations (LSFs) are versatile methodologies capable of installing functional handles on richly decorated intermediates to deliver numerous diverse products in a single reaction. Predicting the regioselectivity of LSF is still an open challenge in the field. Numerous efforts from chemoinformatics and machine learning (ML) groups have made significant strides in this area. However, it is arduous to isolate and characterize the multitude of LSF products generated, limiting available data and hindering pure ML approaches. We report the development of an approach that combines message-passing neural network and an 13C NMR-based transfer learning to predict the atom-wise probabilities of functionalization. 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.

Meaningful exploration of the chemical space of druglike molecules in drug design is a highly challenging task due to a combinatorial explosion of possible modifications of molecules. In this work, we address this problem with transformer models, a type of machine learning (ML) model originally developed for machine translation. By training transformer models on pairs of similar bioactive molecules from the public ChEMBL data set, we enable them to learn medicinal-chemistry-meaningful, context-dependent transformations of molecules, including those absent from the training set. By retrospective analysis on the performance of transformer models on ChEMBL subsets of ligands binding to COX2, DRD2, or HERG protein targets, we demonstrate that the models can generate structures identical or highly similar to most active ligands, despite the models having not seen any ligands active against the corresponding protein target during training. Our work demonstrates that human experts working on hit expansion in drug design can easily and quickly employ transformer models, originally developed to translate texts from one natural language to another, to "translate" from known molecules active against a given protein target to novel molecules active against the same target.

Structural diversification of lead molecules is a key component of drug discovery to explore close-in chemical space. Late stage functionalizations (LSFs) are versatile methodologies capable of installing functional handles on richly decorated intermediates to deliver numerous diverse products in a single reaction. Predicting the regioselectivity of LSF is still an open challenge in the field. Numerous efforts from chemoinformatics and machine learning (ML) groups have made significant strides in this area. However, it is arduous to isolate and characterize the multitude of LSF products generated, limiting available data and hindering pure ML approaches. We report the development of an approach that combines message-passing neural network and an 13C NMR-based transfer learning to predict the atom-wise probabilities of functionalization. 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.

Meaningful exploration of the chemical space of druglike molecules in drug design is a highly challenging task due to a combinatorial explosion of possible modifications of molecules. In this work, we address this problem with transformer models, a type of machine learning (ML) model, with recent demonstrated success in applications to machine translation and other tasks. By training transformer models on pairs of similar bioactive molecules from the public ChEMBL dataset, we enable them to learn medicinal-chemistry-meaningful, context-dependent transformations of molecules, including those absent from the training set. Most generated molecules are highly plausible and follow similar distributions of simple properties (molecular weight, polarity, hydrogen bond donor and acceptor numbers) as the training dataset. By retrospective analysis of the performance of transformer models on ChEMBL subsets of ligands binding to COX2, DRD2, or HERG protein targets, we demonstrate that the models can generate structures identical or highly similar to highly active ligands, despite the models having not seen any ligands active against the corresponding protein target during training. Thus, our work demonstrates that transformer models, originally developed to translate texts from one natural language to another, can be easily and quickly extended to “translations” from known molecules active against a given protein target to novel molecules active against the same target, and thereby contribute to hit expansion in drug design.