Binary Patterning as a Probe of Protein Language Model Representations

Do protein language models (pLMs) encode the physical rules that govern protein folding? The four-helix bundle provides a classic test case because binary patterning, the ordered placement of polar and nonpolar residues along an amphipathic α-helix, is sufficient to generate folded proteins with high probability. Here, we ask whether ESM2 captures the sequence grammar underlying this design strategy. Libraries of synthetic sequences constrained only by binary patterning were predicted by ESMFold to form four-helix bundles. Predicted helicity, compactness, and topological diversity changed with sequence length in the same direction as experimentally characterized libraries, although the model did not distinguish successful from unsuccessful individual designs. Principal component analysis of ESM2 embeddings revealed representations of hydrophobic patterning, helix-loop boundaries, and heptad register. Extending the analysis to all 128 heptad binary patterns produced a jellyfish-shaped manifold in which hydrophobic composition formed the dominant axis, while paired components encoded pattern phase. Manifold geometry depended on amino-acid chemistry, repeat number, and linker sequence, and collapsed when amino acids were assigned to arbitrary categories. ESM2 therefore combines general repeat recognition with protein-specific physicochemical information. More broadly, synthetic sequence libraries provide experimentally grounded probes of protein language model representations and their relationship to protein biophysics.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-13
DOI
https://doi.org/10.64898/2026.11.11.765943
Primary Topic
Protein Structure and Dynamics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Binary Patterning as a Probe of Protein Language Model Representations

Vikas Nanda, Dylan Klein, Michael H. Hecht
1 links
bioRxiv (Cold Spring Harbor Laboratory)
Protein Structure and Dynamics
preprint

Binary Patterning as a Probe of Protein Language Model Representations

Vikas Nanda, Dylan Klein, Michael H. Hecht
preprint en

Abstract

Do protein language models (pLMs) encode the physical rules that govern protein folding? The four-helix bundle provides a classic test case because binary patterning, the ordered placement of polar and nonpolar residues along an amphipathic α-helix, is sufficient to generate folded proteins with high probability. Here, we ask whether ESM2 captures the sequence grammar underlying this design strategy. Libraries of synthetic sequences constrained only by binary patterning were predicted by ESMFold to form four-helix bundles. Predicted helicity, compactness, and topological diversity changed with sequence length in the same direction as experimentally characterized libraries, although the model did not distinguish successful from unsuccessful individual designs. Principal component analysis of ESM2 embeddings revealed representations of hydrophobic patterning, helix-loop boundaries, and heptad register. Extending the analysis to all 128 heptad binary patterns produced a jellyfish-shaped manifold in which hydrophobic composition formed the dominant axis, while paired components encoded pattern phase. Manifold geometry depended on amino-acid chemistry, repeat number, and linker sequence, and collapsed when amino acids were assigned to arbitrary categories. ESM2 therefore combines general repeat recognition with protein-specific physicochemical information. More broadly, synthetic sequence libraries provide experimentally grounded probes of protein language model representations and their relationship to protein biophysics.

bioRxiv (Cold Spring Harbor Laboratory)
Rutgers, The State University of New Jersey (US), Princeton University (US)
Protein Structure and Dynamics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon