#28 · ML / AI methods

Pretrained genomic language models are repurposed for expression and variant prediction

The dominant thread is reuse of large pretrained sequence models for prediction directly from DNA sequence. DNACHUNKER replaces fixed tokenization with a learnable adaptive segmentation module within a masked DNA language model pretrained on the human reference. Nucleotide GPT, a single-nucleotide decoder transformer, probes how pretraining and repetitive-element weighting shape learned genomic representations. Variformer, a deep sequence-based network fine-tuned on personal genomes and tissue expression, predicts individual gene expression and variant effects. What is new is the move from reference-based training toward individual-level inference, alongside treating tokenization as learnable rather than fixed. The available summaries do not establish how these models generalize beyond their reported settings, and Variformer's limitations are only partially described.