We found similarities between un- and annotated domains (Frag1, Anthrax_toxA, and Gasdermin) with cluster-based domain decomposition. We found e.g. novel gasdermin domains in unannotated clusters similar to bacterial ones. Our study shows the potential for unknown functions. 7/n

We found clusters containing human and bacterial structures. Human histones (previously identified by Vikram Alva et al.) and innate immunity genes (BPI and AIM2) have cross-kingdom structurally similarity to bacteria, showing possible sharing of immunity-related proteins. 6/n

By mapping the clusters to the tree of life, we found that the majority of non-singleton clusters are likely to be ancient in origin, with 23%, 16.1%, and 13.5% conserved at the Cellular organism, Bacterial, and Eukarya levels, respectively. 5/n

We investigated our “dark clusters” with DeepFRI, predicted 5324 functions and identified 1770 pockets within 1707 well-predicted structures using AutoSite. We found evidence that these may contain a large number of membrane-bound proteins, inc. putative transporters. 4/n

Out of the 2.3M clusters we identified, over 700k clusters that do not have any annotation in Pfam/TigerFam or align to PDB using Foldseek. While these clusters tend overall to have fewer members than annotated ones, these still contain hundreds to thousands of structures. 3/n

We extended the MMseqs2/Linclust clustering to structures and implemented it in Foldseek. To cluster 214M structures, we first cluster at 50% sequence identity with MMseqs2 followed by Foldseek structural clustering. Resulting in 2.3M clusters with ≥2 structures. 2/n