Oxford Finds ESM2 Fakes Biology Knowledge Through Statistical Pattern Matching
A new mechanistic study reveals ESM2 predicts methionine at protein starts not by recognition, but by retrieving a statistical default — and fails when biology diverges from the average.
Takeaways
- ESM2-8M predicts methionine at protein N-termini not by detecting it, but by retrieving a statistical default from the beginning-of-sequence (BOS) token.
- A distributed, multi-layer circuit spanning all 6 layers implements this prediction; a single attention head (L6H8) writes the final methionine signal.
- The model fails on proteins whose true N-terminus is not methionine (e.g., after post-translational cleavage), predicting methionine anyway with high confidence.
- Researchers introduce a new norm-direction decomposition of RoPE frequency bands to trace how positional information flows through attention layers.
- The paper argues that high model confidence does not imply biological reasoning, and circuit-level verification is necessary for high-stakes predictions.
- Full paper available at arxiv.org/abs/2605.16331; no code or weights released.
Protein language models like ESM2 are increasingly trusted to guide drug discovery, protein engineering, and clinical decisions. But a new paper from Oxford asks a deceptively simple question: when ESM2 confidently predicts that a protein starts with methionine, does it actually know that, or is it just pattern-matching to a statistical average? The answer, it turns out, is the latter , and the implications are significant.
The simplest rule in biology
Every protein-coding gene begins with an AUG start codon, which encodes methionine. This is one of the most universal rules in molecular biology. Proteins typically begin with methionine, reflecting translation initiation at the AUG start codon. When the first residue is masked, protein language models overwhelmingly predict methionine, even for sequences whose mature N-terminus is something else. The initial methionine is often cleaved off after translation, meaning the "true" first amino acid of many mature proteins is not methionine at all.
This makes the prediction a perfect test case. Of all PLM predictions, this is the one a reader would most confidently expect the model to implement through recognition: detect the position-0 identity, write methionine because it belongs there. The rule is trivial; one might expect an equally simple implementation. The researchers find the opposite.
A distributed circuit, not a local detector
In ESM2-8M, a six-layer transformer with 8 million parameters, the rule is implemented by a distributed, multi-layer circuit , a set of interacting components that spans the entire depth of the model. The researchers traced the full computational pathway, and what they found was surprising.
The model does not detect methionine at the masked position. Instead, it retrieves a methionine-favouring signal from a reference representation at the beginning-of-sequence token via a position-specific query assembled across layers, with the final output emerging through competition with context-dependent circuits.
Think of it this way: the model has a special "start of sequence" token (BOS) that acts like a memory slot. Early layers build up a stable methionine-biased representation at that BOS token. Later layers then query that slot from the masked position using positional information encoded in the attention mechanism , and retrieve the methionine signal. The model never actually inspects the amino acid at position zero; it just looks up what position zero is supposed to be.