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

Abstract

Regulatory activity at a locus depends on local sequence and on long-range interactions with sequence further away, but how long each kind of information stays shared between genomes over evolution is unknown. Here we use sequence-to-function models as a comparative assay, training FunFormer, a long-context binary classifier that reads 20 kb, together with an architecture- and parameter-matched control that reads only the central 1 kb, in six vertebrates spanning 435 million years, and measuring what each retains when applied to another genome. Within species, the long-context model outperformed three published 1 kb classifiers by 1.0 to 3.8 percentage points in every genome, a gain that came from context rather than architecture. Across species, its contribution decayed with divergence, was near zero at the amniote split and reversed between 320 and 435 Mya, while local sequence still transferred into fish. Zebrafish and Atlantic salmon, which share their divergence from tetrapods but differ by the salmonid genome duplication, behaved alike, so the reversal is not explained by the salmonid duplication. An in silico flank intervention showed that replacing the flank with GC-matched sequence from the target genome cost 8 percentage points among mammals. When a mammalian model was applied to fish, a flank drawn from the source genome on average performed as well as the target's own flank, so the model drew nothing from the fish flank that an arbitrary source-genome flank did not provide. The long-range information that these models learn therefore stops being shared between vertebrate genomes well before the local information does. The same boundary sets a practical rule for reusing models in species without their own data, where long context helps among close relatives but is a liability across deep splits, and in variant scoring most foreign models matched a native model in data-limited pig but none did in human.