JRSeek

The jelly roll is the most common structural motif in viral capsid and nucleocapsid proteins — eight antiparallel β-strands curled into an incomplete barrel, found across families that share almost no recognizable sequence similarity. That last point is the problem. Southern bean mosaic virus and tomato bushy stunt virus were among the first icosahedral viruses solved and both build capsids from jelly roll folds, yet their coat protein sequences look unrelated. Alignment-based annotation therefore fails on exactly the cases where a structural label would be most useful: newly sequenced, unclassified viruses from metagenomic surveys.

JRSeek, from the Li and Xiao labs at The University of Texas at El Paso, answers the question from sequence alone. It trains a logistic regression classifier on frozen embeddings from a pretrained protein language model, and reports over 95% accuracy separating jelly roll from non-jelly-roll viral proteins — including from immunoglobulin-like β-sandwich proteins, structurally close enough to serve as the hardest available negatives.

Structure predictors can of course reveal a jelly roll, but confirming one requires inspecting each prediction, which does not scale to a virome. JRSeek is the cheap upstream filter: run it over thousands of candidate sequences, then fold only the hits. The paper does exactly that, screening NCBI's unclassified viruses and corroborating two predictions with AlphaFold 3, which placed the predicted jelly roll regions at pLDDT above 90.

#Key Features

• Sequence-only detection: No alignment, homology search, or structure prediction is required, so the method works on viral families with no close characterized relatives.
• Discriminates the hardest negatives: The training set deliberately includes immunoglobulin-like β-sandwich proteins from SCOP2, whose fold closely resembles the jelly roll, alongside α-helical and other viral non-jelly-roll motifs.
• Embedding-agnostic performance: Six different pretrained embeddings were tested and all yielded comparable accuracy, indicating that data quality and diversity, not model choice, are the binding constraint.
• Cheap to run: The classifier is one linear layer plus a sigmoid; the only real cost is a single forward pass through the embedding model.
• Screens uncharacterized viruses: Applied to NCBI sequences with no subordinate taxonomic classification, it nominates jelly roll capsids for structural follow-up.

#Technical Details

Six pretrained transformers from ProtTrans supply the features: ProtBert-UniRef100, ProtBert-BFD100, ProtAlbert-UniRef100, ProtT5-XL-UniRef50, ProtT5-XL-BFD100, and ProtXLNet-UniRef100, abbreviated in the paper as Bert-U100, Bert-BFD, Albert-U100, T5-U50, T5-BFD, and XLNet-U100. Per-residue embeddings are mean-pooled to a sequence vector, standardized, and fed to a logistic regression head — one linear layer with a sigmoid, trained with binary cross-entropy and Adam at a learning rate of 0.0001 for 1,000 full-batch epochs. Viral sequences came from NCBI Protein, filtered by FASTA header keywords and capped at 1,000 amino acids; the balanced dataset pairs 2,225 immunoglobulin-like β-sandwich sequences from SCOP2 with an equal number of viral non-jelly-roll sequences and a matching set of single-jelly-roll sequences, split 70/30 with stratification.

Test-set accuracy ranged from 0.95 to 0.99 with ROC AUC of 0.99 to 1.0 across all six embeddings. Bert-U100 was selected for downstream analysis on speed and false positive rate rather than top accuracy: its table values are 0.97 accuracy with a 0.02 false positive rate and 9 minutes to embed the dataset, against 0.99 accuracy for T5-U50 at 29 minutes and 0.98 for Albert-U100 at 48 minutes. On a held-out generalizability set of double jelly roll and unseen β-strand-dominated sequences, Bert-U100 reached 0.92 accuracy with a 0.06 false positive rate but a 0.52 false negative rate; per-family recall ranged from 100% for Asfarviridae, Marseilleviridae, and Turriviridae down to 0% for Ascoviridae and Corticoviridae.

#Applications

The intended use is triage. Metagenomic virome surveys generate large numbers of viral sequences with no structural annotation — one freshwater mussel study cited in the paper contributed 451 viruses, roughly a fifth unclassified — and JRSeek assigns a probable capsid fold to each in one pass. That supports evolutionary work on capsid lineages, prioritization of targets for cryo-EM or structure prediction, and flagging of uncharacterized viruses whose capsid architecture resembles known human pathogens.

#Impact

JRSeek is a small, focused demonstration with a clear result: jelly roll folds carry a sequence signature detectable by a linear probe on general-purpose protein language model embeddings, even where pairwise identity is near zero. That accuracy is essentially independent of which embedding is used is a useful negative result for embedding-plus-classifier tooling generally. Trained weights, the scaler, and prediction scripts are on GitHub under GPL v3. Two limits are real: the model was trained on single jelly rolls and misses roughly half of double jelly rolls, with recall varying sharply by viral family, and the work remains a preprint. The authors also note the classifier may be partly keying on viral versus non-viral sequence character — aligned with its purpose here, but a bound on reading it as a general fold detector.