Small-Molecule AI Needs Protein-Scale Data: Researchers Debate the Missing Ligand Datasets
kaveh_Nasr · x · 2026-11-04
A technical AI-for-Science discussion on whether the AlphaFold-style "protein moment" can be replicated for small molecules. One argument: proteins succeeded thanks to genomics/metagenomics data, so the equivalent would be systematically identifying ligand presence/absence across every organism in the tree of life. A reply notes the search-space problem for exploring non-overlapping pockets scales as n·S/sl (proteins count × average surface area / ligand surface area), and argues better protein surface embeddings are needed. The core point: small-molecule drug discovery lacks a systematic data source comparable to protein sequences, which may be the main bottleneck.