MINA: linear probes reveal coding-sequence family signal in frozen DNA encoders up to protein composition

Abstract

Frozen DNA encoders are widely used as feature extractors, but fine-tuning hides what their fixed embeddings already make linearly accessible. We introduce MINA (Model Interrogation of Nucleotide Architectures), a lightweight benchmark testing whether frozen DNA embeddings can recover (i) a 5-way protein-family label and (ii) the 1,536-d GenePT natural-language embedding for each gene, comparing coding-sequence (CDS) and transcription-start-site (TSS) contexts under a homology-aware split, against translated-composition and protein-language-model baselines. Across 3,244 human genes from five families, the best DNA encoder recovers CDS family labels above 4-mer composition (NT-v2 macro-F1 0.727 vs. 0.633) but only reaches translated amino-acid composition (AA 2-mer 0.735), far below ESM-2 on translated proteins (0.960). GenePT recovery is weak (best DNA R2 = 0.077), and TSS family recovery collapses under homology control (NT-v2 0.447 [->] 0.313). Pooling and encoder choice matter as much, so encoders are not interchangeable: frozen DNA encoders expose coding-sequence protein-compositional signal from raw nucleotides, not robust gene-function signal from arbitrary genomic context.