Binary Patterning as a Probe of Protein Language Model Representations

ESM2 embeddings capture the sequence grammar of binary patterning in four-helix bundle designs, revealing a jellyfish-shaped manifold across 128 heptad patterns.

• Why it matters: It tests whether protein language models encode the physical folding rules implied by amphipathic helix residue ordering, beyond generic sequence similarity.
• What they did: Synthetic libraries constrained by binary patterning were scored by ESMFold for helicity, compactness, and topology, and ESM2 embeddings were analyzed with principal component analysis across all 128 heptad binary patterns.
• The result: Representation geometry tracks hydrophobic composition as the dominant axis and encodes pattern phase in paired components, with manifold shape depending on amino-acid chemistry, repeat number, and linker sequence, and collapsing under arbitrary amino-acid category assignments.

AI summary of the abstract · may contain errors · not medical advice · read the paper.