bioRxivJClub
A multispecies foundation model for genome function prediction, gene annotation and sequence design
bioRxiv · · Preprint · Open access
Boshar, Evans + 22 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
Nucleotide Transformer v3 (NTv3) is a long-context cross-kingdom foundation model that matches or exceeds published functional-track models across 19 species, using sequences up to 1Mb.
- Why it matters: Genomic models have not combined functional prediction, gene annotation, and controllable sequence design across many species, limiting comparative genomics and synthetic biology.
- What they did: NTv3 is pre-trained on more than 128,000 genomes and post-trained on 16,262 functional tracks across 15 animal and plant genomes, then evaluated via matched human assays and cross-species comparisons.
- The result: NTv3 improves track prediction with genome pre-training and expanded functional supervision, outperforms ab initio gene predictors with fewer spurious models and better long-gene recovery, and designs regulatory DNA with specific activity levels validated in plant and fly enhancers.
The findingWhy it mattersWhat they didThe result
- Open access
- 21 cites