A fitted DNA ‘plug’ lets frozen language models answer gene questions

A language model may already know a gene’s function yet lack a way to connect that knowledge to an unnamed DNA sequence. In the preprint A Training-Free Socket for DNA in Frozen Language Models, Arash Fatehi tests a relatively simple bridge: map a DNA encoder’s output into the language model’s own internal representation of gene information.

The reported gains are substantial on a multiple-choice benchmark. But the paper draws an important boundary around them: the bridge appears to communicate which gene, or kind of gene, a sequence represents. It does not demonstrate that the language model acquires biological knowledge absent from its existing text-based knowledge.

What ‘training-free’ means here

Both the DNA encoder, Evo 2 7B, and the receiving language model remain frozen: their parameters do not change. The connector is nevertheless fitted using paired examples. ‘Training-free’ therefore does not mean an absence of data fitting or computation.

For 8,319 training genes, the method pairs DNA features with internal language-model representations of the gene’s official symbol, protein name and description. A ridge regression—a regularized linear mapping—connects the two. Its penalty is selected using training genes only, and its output is rescaled to match the typical size of the target representations’ deviations from their mean.

When answering a question, the language model first processes a prompt containing placeholder positions for DNA. After a selected early layer, the method replaces those positions’ internal states with the mapped DNA representation. Processing then continues normally. No gradient-based optimization passes through the host model.

Four hosts, one held-out evaluation

The main evaluation used GeneHop, a benchmark with five answer options per question. Its four question families concern pathways, disease associations, processes that must be blocked to disable a gene product, and processes shared by two gene products. The gene is identified to the model through DNA rather than its name.

Training, development and test splits keep whole groups of related genes together under the benchmark’s specified homology threshold. Wrong options are selected without reference to the correct answer’s properties, reducing a text-only shortcut found in an earlier version of the benchmark.

In one held-out run covering 7,208 test items, the paper reports these accuracy gains for real DNA over shuffled DNA:

• Qwen3.5-4B: 22.2 percentage points.
• Qwen3.6-27B: 26.1 points.
• Mistral-7B: 16.7 points.
• OLMo-2-7B: 18.4 points.

With real DNA, absolute accuracy ranged from 36.8% to 46.0%; shuffled-DNA accuracy stayed near the 20% chance level. All four primary effects survived Holm correction for comparisons across hosts. Their 95% bootstrap intervals, calculated by resampling homology groups, also remained above the paper’s four-point threshold for a meaningful effect.

Mostly an identity channel

Several controls help explain the improvement. Shuffled DNA preserves every five-base sequence count while disrupting longer-range ordering. Other controls use mismatched gene–text pairings, another gene’s mapped representation, or a random mapping. On the test set, these null controls all stayed within two percentage points of the zero-write baseline.

A simpler nearest-neighbour method recovered 72–84% of the socket’s reported gain: it located the closest training gene in DNA-feature space and inserted that gene’s text representation. The fitted plug still exceeded this baseline by 3.0–5.2 points across hosts. Those secondary comparisons had positive paired confidence intervals, although they were not corrected across hosts.

The evidence therefore supports more than copying the nearest training example, but not general sequence interpretation. Part of the benefit also tracks words shared between gene descriptions and correct answers. Disease questions showed particularly weak gains.

Portability, with limits

A second mapping, fitted only between hosts’ gene-text representations, transferred plugs across all twelve ordered host pairs. Reported test gains stayed within 0.9 points of each receiving host’s own plug. This demonstrates portability among the tested models, not a universal interface.

Choosing the insertion layer remains important. A readability-based rule selected using three hosts predicted a usable layer on OLMo-2-7B before DNA-plug scoring there. However, that host had a broad working plateau, and its inclusion in the test plan was approved after successful development results were known.

The study is a preprint, with one DNA encoder layer, one benchmark and one held-out test run. Comparisons with trained connectors are descriptive: those checkpoints used an earlier option builder, and a disclosed scheduling error reduced their intended question-training exposure. An exploratory protein extension used development data only.

The result is a useful baseline for connecting biological encoders to language models. It is not evidence of variant-effect prediction, previously unknown sequence-level understanding or clinical readiness.