Arc Institute preprint: an AI map of 376 million DNA regions between genes helped identify candidates for previously undescribed RNAs
Arc Institute preprint: an AI map of 376 million DNA regions between genes helped identify candidates for previously undescribed RNAs
On October 22, Arc Institute posted a preprint on bioRxiv describing COMPASS, a map of DNA regions between bacterial and archaeal genes. The system organized 376 million such regions into 5,14 million groups, which research agents then screened for candidates for previously undescribed RNAs.
These DNA regions often encode short RNAs: some regulate gene activity, others participate in chemical reactions, and others help proteins locate specific DNA regions. Distant relatives of these molecules can retain their biological roles even when their nucleotide sequences have diverged considerably. Conventional sequence comparisons often fail to detect these relationships.
The authors further trained Evo 2, a genomic language model trained on DNA sequences, on 750 million such regions. They named the resulting model Evo-PIN. It flagged positions where it confidently predicted the next DNA nucleotide; these positions were more likely to overlap with segments of known RNAs. COMPASS grouped similar sequences, then predicted an RNA structure and identified neighboring genes for each group. This gave researchers concrete features to use when selecting candidates.
The authors first tested the method on known RNA systems. Fewer than 1% of the groups already had a known annotation, but the method placed CRISPR, TIGR, and VIPR arrays into distinct groups along with their associated proteins. It also identified the known guide RNA of a bridge recombinase, an enzyme that uses this RNA to join specific DNA regions. For the IS621 system, the model predicted a combined length of 27 nucleotides for its guide regions; the experimentally measured length was 28. This tested the map against a characterized example.
Research agents used the map to search for new candidates. A researcher defined the evaluation criteria, a lead agent assigned the checks, and an independent agent repeated the key steps. During an autonomous search, the agents examined 14 722 groups and evaluated 11 849 hypotheses pairing a group with a motif. Independent checks retained 297 hypotheses for detailed review; 14 became preliminary candidates, and one met all the criteria.
Within this group, the authors identified a candidate RNA predicted to fold into a structure with several base-paired regions and located near the PhoH2 gene. In 148 of the 326 DNA regions in this group, PhoH2 is immediately adjacent to the RNA; in 228 cases, it lies within 5 thousand DNA nucleotides. In nine RNA sequencing datasets from the bacterium Pedobacter mucosus, the authors found evidence that this motif is transcribed into RNA. They now propose experiments to determine the RNA's function and its relationship to the PhoH2 protein.
“Experimental validation remains a bottleneck in any search based on genomic data,” the authors write in the preprint.
The open COMPASS database stores predicted RNA structures, neighboring genes, and other data for evaluating each group. The map gathers candidates, the agents assess the evidence, and laboratory experiments test the molecular hypothesis.