
DNA replication origins in mammals are hard to characterise at the sequence level. Unlike lower eukaryotes and lower life forms, no consensus sequence motif has been identified that reliably marks where replication initiates across mammalian genomes, leaving the molecular basis of origin specification poorly understood. To address this, we leveraged a comprehensive mammalian replication origin database developed in our lab, integrating origin calls across multiple experimental datasets and cell types annotated by the degree of experimental support, and combined it with HyenaDNA, a large language model pre-trained on genomic sequences, to learn sequence representations of replication origins and distinguish them from non-origin regions.
The model achieves strong classification performance, suggesting that discriminative sequence signals exist. Using a systematic feature elimination strategy, we find that two well-known genomic properties, GC content and replication timing, together account for approximately 95% of origin calls. Strikingly, after explicitly controlling for these confounders through a two-dimensional matching strategy and recalibrating the model, classification performance is retained at 75% AUC, indicating that genuine sequence-level information remains. We are currently dissecting the sequence features driving this residual signal, with the goal of identifying the first experimentally grounded sequence determinants of mammalian replication origin specification.
These results reveal that much of what deep learning models interpret as "origin sequence features" reflects these pervasive genomic biases rather than novel origin-specifying motifs, highlighting the importance of confounder-aware evaluation in genomic deep learning and providing a refined framework for isolating genuine sequence determinants of replication origin specification in mammals.