A Medicinal Chemistry Perspective on Generative AI Synthesis Predictions
Thane Jones, Sean Ekins · 2024
Over the past few decades, various computational approaches have been increasingly used in drug discovery to identify hits and perform lead optimization. 1 , 2 More recently, machine learning models have also been used to propose molecules that have then been tested in vitro . 3–5 In high throughput screening, there is a heavy reliance on vendor-available compounds, while virtual screening may be too limiting as companies try to discover new chemistries for existing or novel targets. Hence, there is an increasing interest in the development and application of generative models 6 to design molecules de novo , 7–13 which possess desirable physicochemical and ADME/Tox properties. 14–17 While the generation of novel molecule designs is important, their synthesizability is also critical if they are to be useful. Ideally, machine learning models could suggest synthetic pathways in order to reduce them to practice. 10 The methods for predicting the synthetic feasibility of compound syntheses have been developing for years, 18–22 although implementation and applications have been quite limited until the advent of generative drug discovery which seems to have reawakened interest in the field. 23 The earliest efforts in synthesis planning, reaction prediction and synthetic feasibility assessment had used rule-based approaches, namely LHASA, CAMEO and CAESA 24 which required a chemist to obtain optimal results. 24