Computer Aided Synthesis Prediction to Enable Augmented Chemical Discovery and Chemical Space Exploration
The drug-like chemical space is estimated to be 10 to the power of 60 molecules, and the largest generated database (GDB) obtained by the Reymond group is 165 billion molecules with up to 17 heavy atoms. Furthermore, deep learning techniques to explore regions of chemical space are becoming more popular. However, the key to realizing the generated structures experimentally lies in chemical synthesis. The application of which was previously limited to manual planning or slow computer assisted synthesis planning (CASP) models. Despite the 60-year history of CASP few synthesis planning tools have been open-sourced to the community. In this thesis I co-led the development of and investigated one of the only fully open-source synthesis planning tools called AiZynthFinder, trained on both public and proprietary datasets consisting of up to 17.5 million reactions. This enables synthesis guided exploration of the chemical space in a high throughput manner, to bridge the gap between compound generation and experimental realisation.
I firstly investigate both public and proprietary reaction data, and their influence on route finding capability. Furthermore, I develop metrics for assessment of retrosynthetic prediction, single-step retrosynthesis models, and automated template extraction workflows. This is supplemented by a comparison of the underlying datasets and their corresponding models.
Given the prevalence of ring systems in the GDB and wider medicinal chemistry domain, I developed â€˜Ring Breakerâ€™ – a data-driven approach to enable the prediction of ring-forming reactions. I demonstrate its utility on frequently found and unprecedented ring systems, in agreement with literature syntheses. Additionally, I highlight its potential for incorporation into CASP tools, and outline methodological improvements that result in the improvement of route-finding capability.
To tackle the challenge of model throughput, I report a machine learning (ML) based classifier called the retrosynthetic accessibility score (RAscore), to assess the likelihood of finding a synthetic route using AiZynthFinder. The RAscore computes at least 4,500 times faster than AiZynthFinder. Thus, opens the possibility of pre-screening millions of virtual molecules from enumerated databases or generative models for synthesis informed compound prioritization.
Finally, I combine chemical library visualization with synthetic route prediction to facilitate experimental engagement with synthetic chemists. I enable the navigation of chemical property space by using interactive visualization to deliver associated synthetic data as endpoints. This aids in the prioritization of compounds. The ability to view synthetic route information alongside structural descriptors facilitates a feedback mechanism for the improvement of CASP tools and enables rapid hypothesis testing. I demonstrate the workflow as applied to the GDB databases to augment compound prioritization and synthetic route design.