VHSeek: biology foundation model-based virus representations revealing virus host associations
What happened
On November 20, 2026, researchers from the University of Southern California and Mohamed bin Zayed University of Artificial Intelligence introduced VHSeek, a new tool for predicting which hosts viruses may infect. VHSeek uses viral DNA and protein data to make its predictions, without needing genome sequences from potential host organisms. The tool was tested on large viral datasets and compared to existing methods. VHSeek was found to be more accurate and able to identify host associations for a greater number of viruses.
Key details
- VHSeek combines nucleotide data from AIDO.DNA and protein data from ESM2 to represent viruses.
- The tool improves genus-level maximum F1 score by 12%, Top-1 accuracy by 19%, and area under the precision–recall curve by 27% compared to iPHoP.
- On the IMG/VR v4.1 metagenomic collection, VHSeek increases effective host coverage from 36.5% to 64.5% using a confidence threshold.
- VHSeek can prioritize viruses and hosts with potential human health impacts, including multi-host associations.
- The research was partially funded by the National Science Foundation (NSF) and the National Institute of Allergy and Infectious Diseases.
Why it matters
Researchers studying virus transmission and those developing phage therapy now have a tool that can predict virus-host associations without needing host genome data, which is often unavailable. This expands the range of viruses that can be studied in large datasets and improves the accuracy of host predictions, supporting efforts in disease control and therapeutic development. Public health agencies and scientists can use VHSeek to better identify potential virus reservoirs and track viruses that may affect humans or other important species.
