Identifying novel CDK9 inhibitors by traditional virtual screening and AI-based approaches
Kyryl Popov, Natasha Kamerlin, Johanna Åberg, Emelie Andersson, Mahtab Azimi, Mattias Jönsson, Areej Khabut, Filip Paulsen, Vuslat Sayin, Johan Evenäs, Johan Gudmundsson, Michael Green, S. Jalil Mahdizadeh, Leif A. Eriksson, Tomas Friman, Christina Velasco, Oskar Alsing
Poster presented at Drug Discovery Chemistry (DDC) Europe 2024 in collaboration with RG Discovery, Anyo Labs, and Pelago Bioscience.
Summary
CDK9 is a validated oncology target, and we compared three approaches for finding novel CDK9 inhibitors in a 24,000-compound Enamine hinge-binder library:
- Classic virtual screening (docking + pose / hit selection): 29.4% hit rate at >30% inhibition @ 10 µM on confirmed VS hits.
- AIChemy (Desupervised), an equivariant deep probabilistic graph neural network that quantifies uncertainty on every prediction: 15.6% hit rate on the same library subset.
- Anyo Labs hybrid AI model: 8.9% hit rate.
Nine of ten confirmed VS hits showed clear dose-response curves, with IC₅₀ values ranging from 0.48 µM (RG-3290) down to 22.5 µM (RG-3294). Fragment-sized hits were prioritised for hit expansion; SAR-by-catalogue and a 25-compound focused library were synthesized at RG Discovery, yielding RG-1003686 with CDK9 IC₅₀ = 757 nM, LE 0.48, LLE 4.6, and improved physchem properties (solubility >95 µM, microsomal CL <10 µL/min/mg) over the RG-3284 starting point. CETSA® Explore selectivity profiling at Pelago Bioscience confirmed CDK9 engagement and revealed on/off-target kinase engagement for both series.
The work demonstrates that AIChemy’s probabilistic GNN-based virtual screening can complement classical docking pipelines and surface chemically novel hit series in a production hit-finding campaign.