The abstract is read at the publisher; the summary is JClub's.
On coding sequences of 3,244 human genes with a homology-aware split, the frozen encoder NT-v2 recovers coarse protein-family signal only at about nucleotide-composition level (macro-F1 0.693 vs 0.639).
- Why it matters: Frozen DNA embeddings are used as fixed features, but it is unclear what linearly accessible information they already contain versus simple sequence-composition baselines.
- What they did: MINA benchmarks two linear probes on 3,244 human genes: 5-way coarse protein-family classification and regression to each gene’s GenePT text-summary embedding, comparing frozen encoders to nucleotide and amino-acid composition.
- The result: Encoder performance is highly sensitive to evaluation design: pooling and transcription-start-site windows can swing macro-F1 (e.g., NT-v2 0.504→0.698; 0.694→0.300), while random splits inflate results (0.693→0.801) via homologous leakage.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.