JRSeek: Artificial Intelligence Meets Jelly Roll Fold Classification in Viruses

Area of Science:

• * Structural biology
• * Virology
• * Bioinformatics

Background:

• * The jelly roll (JR) fold is a prevalent structural motif in viral capsids and nucleocapsids across diverse viral families.
• * Predicting the JR fold from protein sequences is crucial for understanding viral structure-function relationships, especially for viruses with unresolved structures.

Purpose of the Study:

• * To develop and evaluate a sequence-based computational tool for predicting the presence of the JR fold in viral proteins.
• * To assess the accuracy and generalizability of large language model (LLM) embeddings combined with logistic regression (LR) for JR fold prediction.

Main Methods:

• * Trained logistic regression (LR) models using six different large language model (LLM) embeddings on a curated dataset of viral and non-viral protein sequences.
• * Utilized Principal Component Analysis (PCA) to visualize sequence embeddings and assess separability of different protein types.
• * Validated predictions using AlphaFold3 for unclassified viral sequences.

Main Results:

• * LR models achieved over 95% accuracy in distinguishing JR from non-JR sequences, independent of the specific LLM embeddings used.
• * PCA revealed inherent separability of some sequences, while LR models were essential for classifying ambiguous sequences.
• * The model demonstrated generalizability to double JR folds in some viral families and successfully predicted the JR fold in unclassified viruses, corroborated by AlphaFold3.

Conclusions:

• * Sequence-based LLM embeddings coupled with LR provide a highly accurate and efficient method for predicting the viral JR fold.
• * This approach is particularly valuable for analyzing viral sequences where experimental structural data is scarce.
• * Future work should focus on developing models specifically for more complex folds like double JR folds to enhance generalizability across all viral families.