ViSTA: variant-integrated sequence transformer architecture learns DNA mutational contexts for breast cancer subtyping

While DNA language models have shown promise in regulatory and functional prediction tasks, patient-specific mutation profiles remain underutilized due to the difficulty of modeling complex and vast genomic data. We introduce ViSTA, a BERT-based DNA language model pretrained and fine-tuned on variant-centered nucleotide segments from tumor exome sequences to capture mutation-aware sequence patterns linked to clinical phenotypes. ViSTA accurately predicts breast cancer subtypes using only exome variant data and revealed biologically relevant embeddings, subtype-specific hotspots, and oncogenic mutational signatures. Our study demonstrates the potential of variant-aware DNA language models for interpretable and solely sequence-based cancer subtyping.