Two AI models, one for DNA and one for protein, build the same picture of protein biology

One model has only ever read DNA, the other only proteins. Train a translator between them and they turn out to describe the same biology — but only their chemistry can be pushed.

A gene is a recipe and a protein is the dish. Two of today's large biological models each learned one half of that relationship and never saw the other: Evo 2 reads raw genome sequence, ESM-C reads amino-acid chains. This work tested whether they nevertheless built the same internal picture of protein biology, and whether that picture can be used to control them.

Start with the sharpest version of the question. Pick one residue in one gene and change the DNA underneath it, in one of two ways. A silent edit swaps a DNA letter but leaves the amino acid identical — the recipe is rewritten, the dish is unchanged. A meaningful edit changes the amino acid too. Then watch where each model's internal representation of that residue moves.

drag either panel to turn it

the DNA model

the protein model

The asymmetry is the point. For a silent edit the protein model does not move at all, because its input genuinely has not changed — that zero is structural, not a discovery. The DNA model does move, since its input changed; but 85.60% of that motion runs sideways to the direction that carries protein meaning. It notices the edit and files it somewhere that does not disturb the protein interpretation.

The two models agree about protein biology in a way that goes beyond coincidence: their shared features fire on the same annotations, and the shared space behaves like a coordinate system for protein meaning. But that agreement is recognition, not control. Of everything tested, only continuous chemistry — how oily or how charged a residue is — moved both models when it was pushed. The rest are read-outs.

Why agreement between these two models matters

Any single model's internal structure might be an artifact of how it was built: its architecture, its training data, its tokenizer. When two models trained on opposite sides of the same biological relationship converge on the same organization, that organization becomes a candidate for a property of the biology rather than a quirk of one network. It also has a practical edge: a structure that survives translation between a genome model and a protein model is a structure you can look for in the next model, or use to compare them.

A translator that only counts genuine bridging

Both models were read out through a shared dictionary of features — think of it as a vocabulary of recurring patterns, learned so that the same word can describe activity in either model. The design choice that mattered was the training objective. A dictionary trained only to compress each model's activity faithfully learns two vocabularies that happen to sit side by side. So the objective here paid off only when one model's feature activity genuinely predicted the other's.

That distinction is not subtle. Scored on 431,802 held-out residues from genes that never appeared in training, the cross-prediction dictionary predicts protein-model activity from DNA-model activity at 0.3177, where a reconstruction-only dictionary manages 0.0145 — essentially nothing.

0.0 0.1 0.2 0.3 reconstruction-only dictionary 0.0145 linear ridge floor 0.1015 dense-network ceiling 0.2137 cross-prediction dictionary 0.3177
Predicting one model's features from the other's, in R² per dimension, on held-out genes. Two reference points keep the headline number honest: a plain linear ridge regression between the two models' raw activations reaches 0.1015, and a large dense network trained for the same job reaches 0.2137. The sparse feature bridge at 0.3177 clears both. The reverse direction, protein to DNA, is harder: 0.186 against a floor of 0.1082.

The harder reverse direction makes sense. DNA carries information the protein simply does not have — which of several synonymous codons was used, where the exon boundaries fall — so predicting genome-model activity from protein-model activity has a lower ceiling built into it.

Shared features mean the same thing in both models

A feature counting as "shared" by the objective above could still be shared for uninteresting reasons. The check is semantic: take a feature, find what it fires on inside the DNA model, find what it fires on inside the protein model, and ask whether those are the same annotation — a transmembrane helix, a disulfide-rich domain, a signal peptide, a beta strand.

0% 25% 50% 75% 100% shared features 95.80% protein-model-only 57.80% DNA-model-only 33.30%
Out of a sample of 300 features, the 120 shared ones almost always name the same annotation in both models. Features private to one model agree far less often, which is what makes the shared partition worth trusting rather than a labeling artifact.

The shared vocabulary also names biology the reference labels had missed. Of 100 shared features with no matching annotation in the reference set, 42 turned out to correspond to a real, externally documented protein concept — collagen's Gly-X-Y triple-helix repeat, the glycine hinge between ankyrin repeats, the coenzyme-binding fold of a dehydrogenase family — when checked against outside databases such as Pfam and InterPro. 30 of those fire across several unrelated genes, so they describe a concept rather than memorizing one protein. The same pipeline run on shuffled labels validates nothing at all.

Recognition, not control

Reading a representation and steering it are different capabilities, and the difference decides whether a shared feature is a scientific instrument or just a nice description. The test pushes directly on a shared feature inside one model to see whether the behavior it supposedly encodes changes, then does the same in the other model. A shared feature that is really a shared control should move both.

Six shared amino-acid features were pushed this way, each with a matched random direction as the control. None of them raised its own amino acid in both models. The two that clearly moved the protein model, features for glycine and proline, moved the DNA model in the opposite direction.

-0.1 +0.0 +0.1 +0.2 +0.3 +0.4 glycine feature, protein model 0.3539 glycine feature, DNA model -0.09306 proline feature, protein model 0.1176 proline feature, DNA model -0.04028 GC control, DNA model 0.1074
Change in how often the target appears in what the model produces, relative to a matched random direction of the same size; positive means the intended change happened. The glycine and proline features raise their amino acid in the protein model and lower it in the DNA model. The GC-content feature at the bottom is a positive control: the DNA model can be steered, just not by these features.

Chemistry behaves differently, and the transfer is what makes the case. A hydrophobicity direction fitted inside one model, then carried over and applied unchanged inside the other, still steers it.

0 1 2 3 4 5 hydrophobicity, protein model 4.235 hydrophobicity, DNA model 2.009
Shift in the hydrophobicity of what each model produces when the transferred direction is pushed, again relative to a matched random direction, with both effects comfortably clear of that control (the DNA-side effect's lower bound is 1.35). This is a shift on a chemical property scale, not a change in a probability, so its size is not comparable with the chart above — what matters is that both bars are positive.

That carried-over direction is cosine 0.9945 and 0.9946 to the direction each model would have produced natively. The two models are not merely correlated about oiliness; they use very nearly the same vector for it.

One caveat keeps that from being eight separate control knobs. Eight biochemical properties were mapped this way, and they are not independent axes: pushing any one of them measurably shifts 54 of the other 56 property readings. Increasing hydrophobicity alone shifts polarity by -1.039, sheet propensity by 1.49 and flexibility by -0.9835. What transfers is one correlated chemical subspace, not a mixing desk.

The shared features help with disease mutations, and stay behind the specialist

The natural application is clinical: deciding whether a mutation is harmful. Using only the shared features as input, a simple classifier was asked to separate 43,258 human variants labeled pathogenic or benign in ClinVar, spanning 6,434 genes.

0.5 0.6 0.7 0.8 0.9 1.0 coin flip features in neither bank 0.614 conservation only 0.8541 conservation + composition 0.8771 shared features 0.9019 DNA-model-private features 0.9062 protein-model-private features 0.936 whole feature bank 0.9388 AlphaMissense 0.9578
Separating pathogenic from benign variants, as AUROC (1.0 is perfect, 0.5 is a coin flip). The shared features clear a deliberately strong shortcut baseline built from evolutionary conservation plus sequence composition, so they are not merely restating conservation. They also stay below the whole feature bank, below the protein model's own private features, and below AlphaMissense, a predictor built specifically for this task.

The shared features are genuinely useful here, and not the best tool available. Two things temper that further. This measurement is taken near the protein model's output, where features are closest to its own pathogenicity prediction, and the shared bank there is the smaller partition — 2,709 features shared against 2,023 private to the protein model, 1,764 private to the DNA model and 9,888 assigned to neither. And much of what the shared disease features read is evolutionary conservation: their scores track a standard conservation score at correlation 0.735, and when the top shared features were described one by one, 6 of eighteen were labeled as reading conservation rather than a specific mechanism.

The shape of the shared space

Return to the explorer at the top, and ask what the movement means rather than how big it is. Three properties, each measured against shuffled controls, say the shared space behaves like a coordinate system for protein meaning.

First, size tracks consequence: a meaningful edit displaces the representation 2.758 times as far as a silent one (11.52 against 4.175 in median). Second, the two models move the same way.

0.0 0.1 0.2 0.3 0.4 random directions 0.0002539 shuffled edit pairs 0.0008608 shuffled positions 0.001963 the real pairing 0.3216
Agreement between the direction the DNA model moves and the direction the protein model moves, as a median cosine over 9,923 meaningful edits, against three separate shuffled controls: pairing each edit with another edit's movement, shuffling residue positions, and random directions. Agreement is positive for 98.26% of individual edits.

Third, and most tellingly, the movement points somewhere specific. Each of the twenty amino acids occupies its own region of the shared space. When a residue is mutated, the direction it travels lands nearest the region of the amino acid it became.

0% 25% 50% 75% 100% chance 5.70% observed 82.88%
How often the destination amino acid's region is the nearest of all twenty after a meaningful edit. The chance rate shown is measured by randomizing where the movement points rather than assumed to be one in twenty, because the twenty regions differ in size and spacing. For reference, the movement aligns with the straight line from the old amino acid's region to the new one at median cosine 0.5567, where shuffling gives 0.2744 rather than zero, because amino-acid regions are not randomly arranged to begin with. Silent edits, which have no destination amino acid, land on a given region 14.69% of the time.

That is a strong statement about the geometry, and it is also the most conservative of the findings: the amino-acid organization it describes follows largely from structure established earlier in this work. It confirms the picture rather than revealing a new mechanism.

What this shows, and what it does not

A DNA model and a protein model, trained on opposite halves of the relationship between genes and proteins, can be shown to hold the same internal picture of protein biology — legible in the same annotations, organized as a usable coordinate system, and rich enough to name real biology the reference labels had missed. That agreement does not come with a control panel. Where it was tested directly, only chemistry moved both models.

The honest limits:

0.0 0.1 0.2 0.3 DNA model, before 0.3222 DNA model, after 0.231 protein model, before 0.2144 protein model, after 0.04894
Alignment between features and codon-level concepts, before and after conditioning out which amino acid is present. The protein model keeps almost nothing; the DNA model keeps most of what it had. Measured on a later, separately trained dictionary, so read it as a direction, not as a correction to the numbers above.

The next question this raises is a comparison, not a refinement. If a shared chemical subspace is what genuinely transfers between a genome model and a protein model, then it is the thing to look for first in the next pair — and the honest test of the shared picture is whether a third model, trained differently again, lands in the same place.