6-Parameter Model Maps Molecular Co-Folding to Cell Phenotypes

dom_beaini · x · 2026-08-26

The author built a minimal "toy" Virtual Cell model with only 6 trainable parameters. It combines 34 million protein-ligand and protein-protein co-foldings with gene expression and phenomics maps from Recursion. Despite its simplicity, the biologically grounded model achieved an average AUROC of 71% across 11 blinded cell lines. This prototype demonstrates how predicted molecular interactions can be mapped to cellular phenotypes.

Related event: AI virtual cell models narrow gap between drug affinity and biological activity(2 posts)→

Original post →

More from Research

Research channel →