Long-range regulatory sequence models reveal a phylogenetic boundary in vertebrate genome evolution

Long-context FunFormer models that read 20 kb retain long-range regulatory information across vertebrates only up to a phylogenetic boundary, near zero at the amniote split.

• Why it matters: It is unclear how long-range versus local sequence information remains shared between genomes over deep evolutionary time, limiting comparative interpretation and model reuse.
• What they did: FunFormer was trained as a long-context binary classifier and compared with an architecture- and parameter-matched 1 kb control across six vertebrates spanning 435 million years, then tested for transfer between genomes.
• The result: Long-range contribution decayed with divergence, reversed between 320 and 435 Mya, and flank interventions showed replacing GC-matched flanks cost 8 percentage points among mammals; model reuse was helpful for close relatives but a liability across deep splits.

AI summary of the abstract · may contain errors · not medical advice · read the paper.