A genome language model for mapping DNA replication origins

Origin firing is a central process during DNA replication, but specific sequences defining replication origin usage have not been defined in human cells. Here, we show that a genome language model can accurately predict which sequences can act as an origin of replication, thereby enabling the fast and cost-effective creation of genome-wide replication origin maps. We fine-tuned a genome language model on the primary sequence of mapped human origins to establish ORILINX (ORIgin of replication Language-model Inference via Nucleotide conteXt) and found that it learns a rich representation of sequence features linked to replication initiation, extending beyond known predictive features such as GC-content and G-quadruplex motifs. When applied genome-wide, the models sequence-derived origin calling closely mirrors origin efficiency inferred from replication timing, suggesting that intrinsic sequence context encodes information relevant to initiation frequency. Furthermore, we performed Short Nascent Strand sequencing (SNS-seq) and Repli-seq to demonstrate that ORILINX can generalise to other mammalian genomes, such as those of mice and sheep, as well as other vertebrates such as chickens. Finally, we packaged ORILINX into a simple, easy-to-use tool which is available at https://github.com/Pfuderer/ORILINX.git.