Large Drug Discovery Model

Ilia Igashov, Arne Schneuing, Adrian W. Dobbelstein, Irina Morozova, Rebecca Manuela Neeser, Evgenia Elizarova, Philippe Schwaller, Michael M. Bronstein, Bruno Correia

Track: Machine learning: computational method and/or computational results

Nature Biotechnology: Yes

Keywords: drug discovery, generative modeling, molecules, design, docking, synthetic accessibility, fragment-based drug design, KRAS

Abstract: We introduce the Large Drug Discovery Model (LDDM), a generative framework for target-specific 3D molecule design, leveraging fragment-based masked modeling and large-scale training on the new synthetic dataset of protein-ligand complexes SynthDock. In addition to de novo drug design, LDDM is also able to solve more constrained drug discovery tasks, which allows users to interact with the model during the design process. We furthermore leverage this feature to introduce a new controlled sampling strategy for multi-objective optimization. We benchmarked LDDM across various tasks, including de novo generation, fragment-based drug design, and molecular docking. Finally, we experimentally validated LDDM-designed molecules that bind to the oncogenic target KRAS.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 56