ESMFold author proposes tree-of-life ligand survey to replicate AlphaFold success for small molecules

sokrypton (author of ESMFold), in a discussion on AI for Science, argued that replicating AlphaFold's protein-structure-prediction success for small molecules/ligands would require data at the same scale that genomics/metagenomics once provided for proteins—mapping the presence and absence of ligands (small molecules) in every organism across the entire tree of life, effectively building an "evolutionary record" for small molecules.

Confirmed

• sokrypton believes the key to AlphaFold's success was exploiting covariation statistics from unlabeled protein sequences across thousands of organisms; he speculates that if supplementary data on these organisms could be collected, a similar approach could transfer to small molecules
• He summarized his position clearly: the goal of such methods is not to "solve" the problem (just as AlphaFold did not fully solve structure prediction), but to leverage the greatest experiment on Earth—billions of years of evolution—to drastically reduce the search space
• proteinrosh joined the discussion, framing the topic as "how small-molecule drug design can replicate an AlphaFold-style breakthrough"; kavehNasr relayed and took part in the debate over the small-molecule data bottleneck

Why it matters

• The protein structure prediction breakthrough is believed to stem from co-evolutionary signals in genomic/metagenomic data, while the small-molecule field currently lacks a data foundation of comparable scale—something sokrypton sees as the core bottleneck
• The proposal points to a route distinct from simply scaling up models: trading data engineering (a tree-of-life-scale ligand census) for search-space compression, which if feasible could reshape the paradigm of AI-driven small-molecule drug discovery

2026-11-04 ~ 2026-11-04 · 8 related posts

• [source] sokrypton: small-molecule AlphaFold needs ligand data across the tree of life — sokrypton · 2026-11-04
• Small-Molecule AI Needs Protein-Scale Data: Researchers Debate the Missing Ligand Datasets — kaveh_Nasr · 2026-11-04
• ESMFold author: AlphaFold-style covariation could unlock small-molecule prediction — sokrypton · 2026-11-04
• [source] sokrypton: the goal is shrinking the search space using billions of years of evolution — sokrypton · 2026-11-04
• Covariation made AlphaFold and LLMs possible — small molecules need equivalent data — proteinrosh · 2026-11-04
• sokrypton: we need cheap ways to capture biodiversity data before it goes extinct — sokrypton · 2026-11-04
• [source] sokrypton: covariation in data underlies both LLMs and AlphaFold-style breakthroughs — sokrypton · 2026-11-04
• ESM researcher: covariation in data made LLMs and AlphaFold possible — other modalities need it too — sokrypton · 2026-11-04