In plain language: Generative AI can propose huge numbers of new molecules, but many of them never bind the target or cannot be made. Our team combined a generative model with physics-based simulations in two nested learning loops, so each round keeps only designs predicted to bind well and be synthesizable. For CDK2, 8 of the 9 molecules synthesized were active, one with nanomolar potency, and the workflow also produced new scaffolds for KRAS.
Why it matters to clients: Fewer, better compounds to make: novel chemotypes with a high hit rate, which shortens the design-make-test cycle.
Published in Communications Chemistry, this work by Nostrum Biodiscovery, the Barcelona Supercomputing Center and collaborators at Superluminal Medicines and RA Capital describes the generative workflow behind ALGen, part of our AI platform. The model explores chemical space beyond known scaffolds for each target, while the physics-based loop keeps designs grounded in predicted target engagement, drug-likeness and synthetic accessibility.
Read the full article in Communications Chemistry: https://doi.org/10.1038/s42004-025-01635-7
