ViTrace detects viral signatures in tumor transcriptomes using a hybrid language model

Area of Science:

• Oncology
• Virology
• Bioinformatics

Background:

• Identifying viruses in tumor transcriptome is crucial for understanding oncogenesis and tumor progression.
• Current virus identification tools using sequence alignment struggle with viral mutations, divergence, and incomplete genomes.

Purpose of the Study:

• Develop ViTrace, a novel hybrid language representation learning model for detecting viral sequences in human transcriptomic data.
• Evaluate ViTrace's performance on diverse, previously unseen viral strains and its application in cancer research.

Main Methods:

• ViTrace integrates DNA contexts, position relationships, and amino acid coding information.
• The model was trained on limited viral species but tested on a broad range of absent-in-train viruses.
• Applied to single-cell RNA-seq data from esophageal and oropharyngeal squamous cell carcinomas.

Main Results:

• ViTrace achieved 86.39% recall on 1179 absent-in-train virus strains across 935 species and 167 genera.
• The model revealed tumor-, cell-, and patient-specific viral colonization patterns in cancer data.
• Uncovered both known and previously unreported viruses in the analyzed cancer types.

Conclusions:

• ViTrace offers a scalable framework for virus detection in transcriptomic data.
• The tool facilitates precision oncology by identifying tumor-associated viruses.
• ViTrace aids in the discovery of novel, uncharacterized tumor-associated viruses.