Reproduce “DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models” on OpenResearch

About this paper

Yiming Qin, Kai Yi, Miruna Cretu, Sjors H.W. Scheres, Pietro Liò, Pascal Frossard

DBMol is a computational framework that designs high-affinity, target-specific small molecules by leveraging differentiable signals derived from advanced protein-ligand structure prediction models. This framework achieves competitive performance on held-out structural evaluation metrics, such as AlphaFold3-based pocket coverage, without requiring curated protein-ligand complex datasets or known reference ligands.

Key insights

• Modern protein-ligand structure prediction models can serve as effective differentiable signals to directly guide the generative process for *de novo* molecule design, rather than just acting as evaluators.
• The alternating optimization and learned projection strategy can effectively generate chemically valid small molecules while preserving structural signals derived from the differentiable structure prediction model.
• A balanced confidence objective within the loss function is crucial for ensuring that molecules optimized against a proxy model transfer effectively to independent structural assessment metrics.

Summary by alphaXiv.