Capsid-specialized protein language models reveal higher-order viral architecture from sequence
- 1Tongji university
Abstract
Viral capsid proteins preserve information on higher-order shell architecture and deep evolutionary history, yet current capsid annotation relies predominantly on homology-based methods that have reduced sensitivity across highly divergent environmental sequences. Here we develop ESMCapsid, a capsid-specialized protein language model for remote capsid detection and architecture-aware representation learning. Screening 343 million representative metagenomic protein clusters revealed a large homology-dark capsid repertoire, with approximately 62% of candidates lacking matches to existing reference databases. Sparse autoencoder decomposition identified recurrent semantic motifs linking homology-dark proteins to known structural lineages, suggesting that interpretable higher-order architectural information can be recovered directly from capsid sequences at metagenomic scale. Mapping conserved motif cores onto resolved viral shells showed spatial clustering and restricted radial positions, indicating that ESMCapsid captures geometric constraints beyond sequence similarity alone. Together, our findings establish a sequence-based route to organize homology-dark viral diversity through conserved architectural principles, extending viral discovery beyond sequence homology.
Related
- pLM-bVAE: protein language model-enhanced binary representation improves quantum annealing-based protein sequence design
- Steering sequence generation in protein language models through Iterative Lookback Monte Carlo Sampling
- Language-guided domain adaptation improves protein sequence rep-resentations for antimicrobial resistance analysis
- Pruning the Search, Not the Signal: Adaptive-Banding Needleman-Wunsch Sequence Alignment via Protein Language Model Confidence
- Sequence-only prediction of VHH nanobody epitopes and paratopes on HER2 with a frozen protein language model ( ESM2) and a structure-derived residue prior
BibTeX
@misc{liu2026capsidspecialized,
title = {{Capsid-specialized protein language models reveal higher-order viral architecture from sequence}},
author = {S. Liu and S. Xia and H. Wang},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.06.749605},
url = {https://www.biorxiv.org/content/10.64898/2026.09.06.749605v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Capsid-specialized protein language models reveal higher-order viral architecture from sequence
AU - S. Liu
AU - S. Xia
AU - H. Wang
PY - 2026
DA - 2026/09/06
PB - bioRxiv
DO - 10.64898/2026.09.06.749605
UR - https://www.biorxiv.org/content/10.64898/2026.09.06.749605v1
AB - Viral capsid proteins preserve information on higher-order shell architecture and deep evolutionary history, yet current capsid annotation relies predominantly on homology-based methods that have reduced sensitivity across highly divergent environmental sequences. Here we develop ESMCapsid, a capsid-specialized protein language model for remote capsid detection and architecture-aware representation learning. Screening 343 million representative metagenomic protein clusters revealed a large homology-dark capsid repertoire, with approximately 62% of candidates lacking matches to existing reference databases. Sparse autoencoder decomposition identified recurrent semantic motifs linking homology-dark proteins to known structural lineages, suggesting that interpretable higher-order architectural information can be recovered directly from capsid sequences at metagenomic scale. Mapping conserved motif cores onto resolved viral shells showed spatial clustering and restricted radial positions, indicating that ESMCapsid captures geometric constraints beyond sequence similarity alone. Together, our findings establish a sequence-based route to organize homology-dark viral diversity through conserved architectural principles, extending viral discovery beyond sequence homology.
ER -