From nucleotides to semantics: genomic representation learning via joint-embedding predictive architecture

Decoding the regulatory syntax encoded in genomic sequences is a central objective in computational biology. Most existing genomic foundation models treat DNA as a language and adopt pretraining objectives from natural language processing. DNA sequences, however, lack explicit semantic boundaries and contain substantial evolutionary noise. Nucleotide-level reconstruction in a low-dimensional input space can therefore increase computational overhead and may yield representations with limited discriminative capacity. Downstream tasks often depend on expensive finetuning, which restricts practical use in many biology laboratories. Here we present GenoJEPA, a genomic representation learning framework based on joint-embedding predictive architecture. GenoJEPA combines continuous patching with semantic alignment, shifting the optimization from local base reconstruction to semantic alignment in latent space. Across 55 downstream tasks, GenoJEPA shows strong representational capacity and robust generalization while reducing parameter count and computational cost. The resulting semantic vectors from frozen GenoJEPA support lightweight GPU-free classifiers to achieve competitive accuracy. These results suggest a practical route towards efficient training and broad application of larger-scale genomic foundation models.