Foundation models · Viral genomics

HViLM

A foundation model for viral genomes that predicts pathogenicity, transmissibility and host tropism directly from sequence, released with the HVUE benchmark.

mSystems, accepted Built on DNABERT-2

Paper Models HVUE benchmark

3prediction tasks: pathogenicity, transmissibility, host tropism

92.4accuracy on held-out pathogenicity test set

91.3F1 for pathogenicity

83.1MCC for pathogenicity

Why it matters

Knowing whether a newly sequenced virus could cause disease, spread between people or infect human cells helps prioritize surveillance and research. HViLM learns these traits from the genome sequence itself.

How it works

Pre-train. HViLM-base continues pre-training from DNABERT-2 on viral genomes.
Fine-tune. Task-specific models are fine-tuned for pathogenicity, transmissibility and host tropism.
Benchmark. HVUE provides standardized train and test splits for all three tasks so viral models can be compared fairly.

Models

Model	Task
HViLM-base	Pre-trained viral genome model
HViLM-Patho	Pathogenic vs non-pathogenic
HViLM-R0	Transmissibility
HViLM-Tropism	Human-tropic vs non-human-tropic

Papers

Team

Pratik Dutta (lead), Jack Vaska, Pallavi Surana, Rekha Sathian, Max Chao, Matthew Obusan, Zhihan Zhou, Han Liu, Suman Ranjan Das, Ramana V. Davuluri

Cite

@article{dutta2026hvilm,
  title   = {HViLM: A Foundation Model for Viral Genomics Enables Prediction
             of Pathogenicity, Transmissibility, and Host Tropism},
  author  = {Dutta, Pratik and Vaska, Jack and Surana, Pallavi and Sathian, Rekha and
             Chao, Max and Obusan, Matthew and Zhou, Zhihan and Liu, Han and
             Das, Suman Ranjan and Davuluri, Ramana V.},
  journal = {mSystems},
  note    = {Accepted},
  year    = {2026}
}