vir2vec: A Viral Genome-Wide Viral Embedding.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic language models (gLMs) show promise for DNA analysis but lack viral-specific architectures and benchmarks.
• Existing models often focus on human DNA or limited viral datasets.
• A comprehensive evaluation framework for viral genome representation learning is needed.

Purpose of the Study:

• Introduce vir2vec, a large-scale genomic language model pretrained on a diverse pan-viral corpus.
• Present vGUE, a unified benchmark for assessing viral genome understanding and representation learning.
• Evaluate vir2vec's performance on various viral genomic prediction tasks.

Main Methods:

• Continual pretraining of Mistral-DNA on 565,747 complete viral genomes from 295 species to create vir2vec (422M parameters).
• Development of vGUE, a benchmark using vir2vec embeddings fed into classifiers (logistic regression, SVM, random forests) under nested cross-validation.
• Assessment of prediction tasks including organism discrimination, evolutionary fingerprinting, species separation, variant typing, and phenotypic context detection.

Main Results:

• vir2vec achieved superior balanced accuracy on seven out of eight diverse viral classification tasks.
• The model consistently outperformed human-DNA-trained and existing viral-specific genomic foundation models.
• Embeddings from vir2vec effectively captured biologically relevant viral variation across multiple tasks.

Conclusions:

• vir2vec and vGUE establish a robust foundation for viral genomic modeling, surveillance, and discovery.
• The developed tools offer improved capabilities for understanding viral diversity and evolution.
• Responsible deployment of vir2vec necessitates ethical considerations and governance oversight due to its potential dual-use implications.