{"id":"arxiv:2506.10271","kind":"preprint","title":"Evaluating DNA function understanding in genomic language models using evolutionarily implausible sequences","authors":["Shiyu Jiang","Xuyin Liu","Zitong Jerry Wang"],"author_count":3,"year":2025,"published":"2025-06-12","venue":null,"cited_by":null,"doi":null,"arxiv_id":"2506.10271","nct_id":null,"source_url":"https://arxiv.org/abs/2506.10271","archive_url":"https://space-kitty.com/papers/arxiv:2506.10271","updated":"2025-08-26","pmid":null,"pmcid":null,"openalex_id":null,"pdf_url":"https://arxiv.org/pdf/2506.10271","abstract":"Genomic language models (gLMs) hold promise for generating novel, functional DNA sequences for synthetic biology. However, realizing this potential requires models to go beyond evolutionary plausibility and understand how DNA sequence encodes gene expression and regulation. We introduce a benchmark called Nullsettes, which assesses how well models can predict in silico loss-of-function (LOF) mutations, in synthetic expression cassettes with little evolutionary precedent. Testing 12 state-of-the-art gLMs, we find that most fail to consistently detect these strong LOF mutations. All models show a sharp drop in predictive accuracy as the likelihood assigned to the original (nonmutant) sequence decreases, suggesting that gLMs rely heavily on pattern-matching to their evolutionary prior rather than on any mechanistic understanding of gene expression. Our findings highlight fundamental limitations in how gLMs generalize to engineered, non-natural sequences, and underscore the need for benchmarks and modeling strategies that prioritize functional understanding.","abstract_license":"CC0-1.0 (arXiv descriptive metadata)","license":null,"is_open_access":false,"retracted":false,"peer_reviewed":false,"categories":["q-bio.QM","cs.LG","q-bio.GN"],"details":null,"provenance":"arxiv","attribution":"Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.","topics":[{"slug":"synthetic-biology","name":"Synthetic Biology & Genome Engineering","domain_slug":"biotech"}],"cited_in_dossiers":[]}