AFESM: a metagenomic guide through the protein structure universe! We clustered 821M structures (AFDB&ESMatlas) into 5.12M groups; revealing biome-specific groups, only 1 new fold even after AlphaFold2 re-prediction & many novel domain combos. 🧵 🌐 afesm.foldseek.com 📄 www.biorxiv.org/content/10.1... - ThreadSky

martinsteinegger.bsky.social • 4 days ago

AFESM: a metagenomic guide through the protein structure universe! We clustered 821M structures (AFDB&ESMatlas) into 5.12M groups; revealing biome-specific groups, only 1 new fold even after AlphaFold2 re-prediction & many novel domain combos. 🧵
🌐 https://afesm.foldseek.com
📄 https://www.biorxiv.org/content/10.1101/2025.04.23.650224v1

Comments

martinsteinegger.bsky.social•4 days ago

We annotated all ESMatlas entries with MMseqs2 taxonomy (93% coverage) and computed lowest common ancestors (LCA). Most LCAs at superkingdom level, indicating structures shared across domains. Avg. clusters/genus: Bacteria 1,557; Archaea 723; Viruses 17; Eukaryotes 2 (sampling bias). 2/n

martinsteinegger.bsky.social•4 days ago

ESMatlas uses MGnify environmental labels. Leveraging this, we computed the lowest common biomes per structural cluster, revealing protein adaptations unique to specific environments, especially extreme ones like hyperthermal, hypersaline, and glaciers. 3/n

martinsteinegger.bsky.social•4 days ago

We annotated domains in ESM-only clusters using the TED workflow and found 0 new folds. Re-modelling ~25% of proteins with AlphaFold2 revealed 1 novel fold; unlike AFDB’s >7k novel folds surprisingly, hinting at a saturating fold space or predictor limits.
📄 https://doi.org/10.1126/science.adq4946 4/n

martinsteinegger.bsky.social•4 days ago

We identified 11,941 novel multi-domain combinations. We found membrane-associated domains (e.g., TonB dependent receptor, highlighting domain recombination rather than new folds as a driver of structural innovation. 5/n

martinsteinegger.bsky.social•4 days ago

However, these novel domain combinations explain only a small fraction (0.3%) of the ESM-only clusters. The remainder are mostly low-quality predictions (53%), fragments (16%), known domains with potential unknown extensions (19%), or without identifiable domains (9.3%). 6/n

martinsteinegger.bsky.social•4 days ago

Explore AFESM with our website! You can search your favorite proteins from ESMatlas or AFDB using their identifiers. It's still a work in progress, with many exciting features on the way! 7/n
🌐 https://afesm.foldseek.com

ewanbirney.bsky.social•3 days ago

Wow. No - or 1 perhaps - new folds on the addition of 600 million metagenomic proteins! We have spanned the AI-predictable natural universe (of course, there might be some novel folds that are not AI predictable. Feels like this might be a quest for unicorns of course)

saschaindra.bsky.social•3 days ago

Thank you for sharing

chasingmicrobes.bsky.social•2 days ago

📌

Comments

Posting Rules

Reply