Planning And Interpretation Of Late Stage Functionalisation Reactions

With surging costs of drug discovery, we innovate to improve efficiency for every step of the process. Synthetic chemistry delivers drug candidates in a fast and efficient manner. Late stage functionalisation (LSF) is poised to accelerate the process by making chemical synthesis modular so most of the molecules are assembled via standard protocols. Predicting the major reaction site is not straightforward due to multitude of competing factors. Once a reaction is executed, NMR spectroscopy is typically required to establish its outcome. First chapter covers a data-driven mechanism-agnostic approach for predicting selectivity of C–H borylation using a multi-task language model. We show it is possible to predict the borylation product with accuracy comparable to that of a synthetic chemistry expert. We compare performance of our language model with a graph neural network, semiempirical quantum calculations, and a simple baseline with rules-based featurisation. Second chapter describes a DFT-free reformulation of DP5 probability. We develop and use an end-to-end neural network to predict both chemical shifts and associated uncertainties. The new approach delivers rapid confirmation of a structure candidate, bypassing expensive DFT calculations. We demonstrate its effectiveness in large-scale combinatorial studies. The revised model is deployed for the structure revision of 24 natural products, where 23 of 24 correct structures had a higher DP5 score. We also study its application to relative stereo- chemistry determination on 42 complex examples, achieving an unprecedented performance, with 2.9 × 10−8 chance of replicating it by a random guess. Third chapter presents the attempts to develop an automated method for refining structural proposals of complex organic molecules using 13C NMR data. We explore the application of molecular optimisation methods to this problem. We find that specifying the objective function is a major challenge, as the NMR error alone is insufficient to guide the optimisation. Overall, the methods developed in this work solve multiple problems in planning and analysing pharmaceutically important reactions.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Kotlyarov, Ruslan

Advisor dc:contributor.advisor

• Goodman, jonathan

• late-stage functionalisation
• machine learning
• selectivity prediction
• Reaction site prediction
• Automated structure verification
• Chemical language models
• Graph neural networks

Licence

• https://www.repository.cam.ac.uk/bitstreams/3d4d9400-4f3c-4608-b5f6-7e9ac20f9493/download
• https://creativecommons.org/licenses/by-sa/4.0/

Language dc:language

eng

Author Identifier

0000-0002-2519-2358

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/395780