Computational & AI Journal of Computer-Aided Molecular Design (Springer Nature)
Generative and foundation-model pipeline narrows millions of molecules to tested 3CLpro inhibitors
- Generative models, a biomedical foundation model and cheminformatics filters were chained to cut a huge, noisy pool of generated molecules to a short list.
- Cut-offs came from an in-house 3CLpro dataset, and a subset of survivors was ranked by free energy perturbation.
- All molecules tested experimentally kept target inhibition and one beat its parent lead; the foundation model also predicted activity well on a blind test.
A handful of assayed compounds all keeping activity is encouraging but weak evidence for the pipeline: without a control arm, such as random picks or simple property-filtered analogs of the same lead, we can't tell whether FEP or the foundation model added anything. I'd also want to see how far these sit from the parent lead, and whether the cut-offs from a proprietary dataset transfer to our target.
Paper: Journal of Computer-Aided Molecular Design (Springer Nature) ↗ In silico drug discovery framework leveraging biomedical foundation models and evaluating generative models—a use case on 3CLpro