Deep Neural Networks for Pose Validation, Affinity Prediction, and Input Attribution

Our inability to rapidly design small molecules that bind strongly to a macromolecule is a major bottleneck for understanding biological science. It typically still takes many years and large inter-disciplinary teams to create one such molecule. High-throughput facilities such as XChem at Diamond Light Source now enable the trivial identification of weakly binding small molecules (fragments). For this resource to be of benefit to biology we must now streamline the process of converting weak binders to strong binders. AI methods, being actively developed by BenevolentAI (BAI), have shown great promise in streamlining this process.

Technical summary: Currently decision making is subjective and time-consuming, even when an optimal decision is possible. Astonishingly, no freely-available computational methods exist that can provide automated solutions for this. We therefore propose a DPhil working actively with AI and Machine Learning Experts at BAI and the breakthrough XChem technology to develop such tools: Collating and generating large datasets of protein-fragment complexes Machine Learning (e.g. CNNs) for classifiers for determining if a ligand is present Artificial Intelligence to make use of classifiers to make automated decisions Project outcomes: This project will produce computational techniques and datasets that will enable improvements in data driven biology. Further, the work would greatly speed up the development of chemical probes used to study biological processes: such probes invariably have great impact as they enable fine-grained, systematic studies of phenotypic effects of specific protein targets. They would work in close-partnership with BAI, including at minimum 3 months spent on site.