saved
Systematic mapping of the prokaryotic noncoding sequence space with a genomic language model
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
An Arc Institute bioRxiv preprint led by Patrick Hsu uses Evo-PIN, a tuned Evo 2 DNA language model, to sort 376 million noncoding regions into clusters. Evo-PIN was trained further on 750 million intergenic sequences, the stretches between genes, from 1.3 terabase pairs of prokaryotic DNA. Its pipeline, COMPASS, built 5.14 million clusters, fewer than 1% with a known noncoding RNA. It nominated 20,149 clusters of candidate structured RNAs with a likely protein partner, and AI agents mined it for divergent sensor and catalytic RNAs. The candidates are computational predictions, and the preprint is not peer reviewed.
ideas
- Fine-tuning on intergenic DNA improved clustering of known RNAs. A cluster quality score rose from 0.43 for mean-pooled Evo 2 embeddings to 0.60 for Evo-PIN, and to 0.65 when only low-entropy positions, where the model is most certain of the next base, were pooled.
- Clusters group sequences that alignment cannot. Median pairwise identity within a cluster was 28.5%, below the 50-60% range where alignment tools become unreliable for RNA, while known annotations were 0.92 consistent across 24,837 annotated clusters.
- Periodic model uncertainty recovers repeat arrays. Of 1,672 clusters with a consistent token-entropy period, 96.1% fell at 60-79, 18 or 5 nucleotides, matching the CRISPR, TIGR and VIPR arrays of RNA-guided systems, and the flanking genes matched each system's known effector protein.
- Bridge RNAs with longer guides turned up. Across 32,190 loci of bridge RNAs, the guides for DNA-joining bridge recombinases, predicted programmable nucleotides ranged from 12 to 40, and 6,113 exceeded the 28 measured for the characterized IS621 bridge RNA.
- Agent teams screened thousands of hypotheses. Agents flagged a divergent fluoride riboswitch and a hammerhead ribozyme that fall below Rfam family thresholds, and a rubric-guided run over 11,849 hypotheses left one cluster, tied to the PhoH2 nuclease-helicase protein, meeting every criterion.
quotes
“fewer than 1% of the clusters contained known ncRNA annotations”
“The median pairwise sequence identity between sequences within the same cluster was 28.5%”
“the ncRNA candidates reported here are computational predictions, and experimental testing remains the rate-limiting step in any genome mining campaign”