GenART: Reading the genome's language with adaptive "words"

Tokenization is both a prerequisite and a central challenge for genomic language models, due to the inherent difficulty of delineating meaningful segments within continuous DNA sequences. We present GenART, a genomic language model framework that dynamically segments DNA into variable-length words directly from raw sequences without manual annotation. Its key adaptive tokenization module infers biologically meaningful boundaries from multi-scale contextual signals during pretraining. GenART demonstrates competitive performance across diverse tasks, outperforming leading approaches. Word boundaries derived from unsupervised tokenization show strong correspondence with base-resolution regulatory signals, such as DNA methylation, and with multi-nucleotide functional elements annotated in GENCODE v.49. These findings highlight adaptive tokenization as a powerful strategy for building interpretable and biologically responsive genomic language models.