Retrieval and competition: how a protein foundation model starts a protein

Researchers traced how ESM2-8M, a protein language model, predicts that proteins begin with methionine—a near-universal biological rule. The analysis reveals the model doesn't recognize methionine through direct evidence detection, but rather retrieves it via a distributed computational circuit anchored at the sequence start token. Critically, the model fails on sequences where biology diverges from the statistical default, suggesting that model confidence may not reflect genuine biological understanding.

This mechanistic analysis of ESM2-8M exposes a fundamental vulnerability in protein language models that increasingly guide experimental and clinical decisions. Rather than detecting methionine through learned biological recognition, the model relies on a positional-prior retrieval circuit—essentially pattern-matching to statistical defaults. The researchers employed sophisticated interpretability techniques, including norm-direction decomposition of attention scores within rotary frequency bands, to trace how positional information flows through the network's computational pathways.

The work builds on growing concerns about interpretability in deep learning systems, particularly in high-stakes domains like drug discovery and molecular biology. Previous research has highlighted the gap between model confidence and true understanding, but this paper provides granular circuit-level evidence for how even simple, well-established biological rules can be learned through mechanisms divorced from actual feature detection.

For the AI and biotech industries, these findings carry substantial implications. Protein language models are already deployed in research pipelines and clinical workflows, yet this analysis demonstrates that confident predictions may mask brittle, distribution-dependent computations. Organizations relying on these models for decision-making should implement mechanistic verification protocols rather than treating model outputs as ground truth. The distributed nature of the methionine-prediction circuit—spanning multiple layers, frequency bands, and query compositions—suggests that more complex biological predictions will be even harder to verify through traditional performance metrics alone.

Looking forward, this work signals that interpretability research must become a standard component of model validation, particularly before deployment in domains where prediction accuracy has real biological consequences. The research community may need to develop new verification frameworks that go beyond benchmark performance to assess whether models have learned genuine biological principles or merely statistical associations.

• →ESM2-8M predicts methionine starts through positional retrieval circuits rather than direct biological feature detection.
• →The model fails on sequences where the true N-terminus is not methionine, revealing brittleness in statistical-default learning.
• →Mechanistic interpretability analysis shows the prediction emerges from distributed computations across multiple layers and attention frequency bands.
• →Model confidence does not correlate with underlying biological evidence, posing risks for clinical and experimental applications.
• →Verification of protein language model predictions will require circuit-level analysis, not just performance benchmarking.