Coevolutionary mining of prokaryotic non-coding elements with a genome language model

Key results

1. 01

   ShownLightweight interaction heads on Minerva-MLM attention beat scalar categorical Jacobian methods in AUPRC on all three benchmarks and run about five orders of magnitude faster, which makes genome-wide scanning practical.

   Interaction heads exceed all other methods on more than 84% of tested Rfam families. They take 79 ms per sequence, versus over 1.5 hours for the accelerated mask-based Jacobian and 6 hours for the original. They beat RiNALMo on isolated RNA; ESM-2 beats both genome models on protein contacts. · Minerva accurately predicts pairwise…; Evaluation metrics and model comparisons; Computational efficiency evaluation

2. 02

   ShownFamily-specific fine-tuning sharpens RNA structure predictions at a locus (recovering the DRT2 pseudoknot). Codon periodicity appears in the base model and helps repair mis-tokenized ORFs.

   Fine-tuning on 13,034 DRT2 loci recovered the pseudoknot, and the signal appeared with as few as 200 loci. For several mis-tokenized loci, extending the protein to the nearest in-frame start gave full-length ORFs matching RefSeq proteins with 100% identity by BLASTp. · Fine-tuning improves interaction predictions and…; Family-specific fine-tuning of Minerva-MLM; Fine-tuning data-efficiency analysis; Mistokenized ORF analysis

3. 03

   ShownA scan of 150 bacterial genomes predicts tens of thousands of multi-hairpin loci. 69.5% of the 2+ hairpin loci are outside existing annotations (41.1% for the 4+ hairpin loci). In five genomes tested against HMM-generated null genomes, these loci are strongly enriched.

   57,787 loci have at least two neighboring hairpins, 40,175 (69.5%) of them unannotated. 9,044 loci have at least four hairpins, 41.1% unannotated. In five randomly chosen genomes versus HMM-generated nulls, the 2+ hairpin loci are enriched 14-fold (FDR 0.07) and the 4+ hairpin loci 132-fold (FDR 0.008). The scan took 100 minutes on one H100 GPU. · Genome-scale coevolutionary mining reveals…; Genome selection, annotation, and scanning; Multi-hairpin enrichment analyses

4. 04

   ShownTwoAYGGAY RNAs in Pseudomonadota carry a larger secondary structure than the Rfam model describes. They occur in highly variable copy number, often as divergent doublets, with flanking short repeats.

   The Minerva-extended covariance model found 16,086 matches in 1,148 of 1,324 Pseudomonas strains. Divergent orientation accounted for 68.4% of doublets (2,164 of 3,166), a 2.7-fold enrichment over the 25% expected under independent strand assignment (χ² p < 10^-300). Doublet formation was enriched against an empirical null (p < 10^-4). Mean upstream 5' repeat motif count was 6.18, versus 2.84 in random regions. · Structure extension and repeat association in…; TwoAYGGAY genomic architecture analyses; TwoAYGGAY repeat analyses

5. 05

   ShownUG27 reverse transcriptase systems in prophages encode arrays of sequence-diverse but structurally conserved ncRNA units. In heterologous E. coli expression, the central hairpins of these units are reverse transcribed into short cDNAs, and the accessory proteins were required in the two systems tested.

   A covariance model found 82,314 units across 14,118 loci, with lengths concentrated near 149 and 167 nt. All five tested systems gave 50-100-nt ssDNA products mapping to the central hairpins. Mutating the RT active site ablated products. Deleting either accessory ORF, tested in two systems (Blautia hansenii and Bacteroides sp. 2-1-16), abolished cDNA production. · Minerva reveals a variable ncRNA array…; UG27 systems reverse transcribe ncRNA hairpins…; UG27 ncRNA covariance model construction and…