HViLM: A Foundation Model for Viral Genomics Enables Multi-Task Prediction of Pathogenicity, Transmissibility, and...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Emerging viral pathogens pose significant global health risks.
• Current computational tools for viral risk assessment are often virus-specific and require extensive retraining.
• Rapid characterization of novel viruses is crucial for effective public health responses.

Purpose of the Study:

• To develop a versatile computational model for rapid, multi-dimensional analysis of viral genomes.
• To establish a benchmark for evaluating viral characterization models.
• To improve prediction of pathogenicity, host tropism, and transmissibility for emerging viruses.

Main Methods:

• Continued pre-training of DNABERT-2 on 5 million viral sequences to create the Human Virome Language Model (HViLM).
• Development of the Human Virome Understanding Evaluation (HVUE) benchmark with seven curated datasets.
• Parameter-efficient fine-tuning using LoRA for pathogenicity, host tropism, and transmissibility prediction.

Main Results:

• HViLM achieved state-of-the-art performance on the HVUE benchmark with average accuracies of 95.32% (pathogenicity), 96.25% (host tropism), and 97.36% (transmissibility).
• The model demonstrated robust cross-family generalization, outperforming existing baselines.
• Interpretability analysis revealed HViLM identifies biologically relevant determinants, including immune evasion strategies via molecular mimicry.

Conclusions:

• HViLM represents a significant advancement in panviral genomic analysis, offering a generalized approach to characterizing emerging viruses.
• The developed HVUE benchmark provides a standardized evaluation framework for viral prediction models.
• HViLM's findings offer insights into viral evolution and host-pathogen interactions, aiding in pandemic preparedness.