Machine learning for analysis of experimental scattering data in materials chemistry

Rebecca Manuela Neeser

Determining whether a molecule can be synthesized is crucial for many aspects of chemistry and drug discovery, allowing prioritization of experimental work and ranking molecules in de novo design tasks. Existing scoring approaches to assess synthetic feasibility struggle to extrapolate to out-of-distribution chemical spaces or fail to discriminate based on minor differences such as chirality that might be obvious to trained chemists. This talk will introduce the Focused Synthesizability score (FSscore), which learns to rank structures based on binary preferences using a graph attention network. The FSscore showcases how human expert and automated feedback can be utilized to optimize the assessment of synthetic feasibility for various applications. We will further discuss how sample efficiency of molecular generative models can be improved using a novel algorithm called Augmented Memory. Lastly, we will showcase a case study on how these two methods combined can lead to drastically improved synthesizability of generated molecules.