Answers are grounded in this paper and cite their sources.
Deciphering the mapping between biomolecular sequences and complex biological functions is a central challenge in genomics. While existing generative nucleotide models capture sequence distributions, they lack explicit access to gene-associated biological semantics, limiting semantics-guided design. Here, we present GenNA, a generative nucleotide foundation model guided by natural language annotations. Pretrained on a multimodal nucleotide-text corpus spanning 2221 eukaryotic species (∼416 billion characters), GenNA internalizes relationships between sequence syntax and functional semantics within a unified autoregressive framework. Without explicit rule-based supervision, GenNA yields distinguishable perplexity responses to sequence-annotation mismatches, mutational variants, and species-label perturbations. Furthermore, across diverse generation tasks, GenNA produces statistically plausible nucleotide sequences consistent with target semantics, structural constraints, and species contexts. Overall, GenNA provides a versatile framework for controllable nucleotide generation, establishing a practical route to integrate functional descriptions with sequence design.
No takes yet. Share an insight, caveat, or question.
Shen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: