HVUE: Human Virome Understanding Evaluation Benchmark

Current release: v2.0

Overview

HVUE is a rigorously constructed benchmark for evaluating DNA language models on three epidemiologically relevant viral prediction tasks:

1. Pathogenicity Classification — Distinguishing disease-causing from benign viral strains
2. Transmissibility Prediction: R₀-informed binary classification of documented sustained human transmission capacity
3. Host Tropism Prediction — Identifying human-infecting viruses from non-human-tropic viruses

Six-Stage Pipeline

HVUE eliminates train-test overlap through a six-stage pipeline:

1. Consolidation: Source genomes collected and deduplicated at genome level
2. MMseqs2 Clustering: Genomes clustered at 95% identity (standard) or 70% identity (hard)
3. Cluster-Aware Splitting: Train/val/test splits (70/15/15) at the cluster level — all sequences in a cluster go to the same split
4. Chunking: After splitting, genomes segmented into fixed-length windows (500/1000/2000 bp)
5. Deduplication: SHA-256 hash-based exact-match removal within splits
6. Leakage Audit: Automated verification of zero cross-split overlap

Splitting at the cluster level before chunking is what prevents homologous fragments of the same genome from appearing on both sides of a split.

Configurations

Pathogenicity

Configuration	Description
standard_capped_500bp	Standard leakage-controlled benchmark, 500-bp fragments
standard_capped_1000bp	Standard leakage-controlled benchmark, 1000-bp fragments
standard_capped_2000bp	Standard leakage-controlled benchmark, 2000-bp fragments
hard_capped_1000bp	More stringent hard-split benchmark, 1000-bp fragments
standard_temporal_1000bp	Temporal benchmark, 1000-bp fragments

Transmissibility

Configuration	Description
standard_capped_500bp	Standard leakage-controlled benchmark, 500-bp fragments
standard_capped_1000bp	Standard leakage-controlled benchmark, 1000-bp fragments
standard_capped_2000bp	Standard leakage-controlled benchmark, 2000-bp fragments
hard_capped_1000bp	More stringent hard-split benchmark, 1000-bp fragments
standard_temporal_1000bp	Temporal benchmark, 1000-bp fragments

Host Tropism

Configuration	Description
standard_95_500bp	95% cluster-aware benchmark, 500-bp fragments
standard_95_1000bp	95% cluster-aware benchmark, 1000-bp fragments
standard_95_2000bp	95% cluster-aware benchmark, 2000-bp fragments
hard_70_1000bp	70% cluster-aware hard benchmark, 1000-bp fragments

Temporal splitting unavailable for Host Tropism (source data lacks collection-date metadata).

Directory Structure

HVUE/
├── Pathogenicity/
│   ├── standard_capped_500bp/
│   ├── standard_capped_1000bp/
│   │   ├── train.csv
│   │   ├── dev.csv
│   │   └── test.csv
│   ├── standard_capped_2000bp/
│   ├── hard_capped_1000bp/
│   └── standard_temporal_1000bp/
├── Transmissibility/
│   └── ...
├── Host_Tropism/
│   ├── standard_95_500bp/
│   ├── standard_95_1000bp/
│   ├── standard_95_2000bp/
│   └── hard_70_1000bp/
└── README.md

CSV Format

Each CSV contains two columns:

• sequence: DNA sequence (A/T/G/C)
• label: Binary label (0 or 1)

Usage

import pandas as pd
from datasets import load_dataset

# Load a specific config
ds = load_dataset("duttaprat/HVUE",
                  data_dir="Pathogenicity/standard_capped_1000bp",
                  data_files={"train": "train.csv", "validation": "dev.csv", "test": "test.csv"})

# Or load directly
train = pd.read_csv("Pathogenicity/standard_capped_1000bp/train.csv")

Leakage Audit

Zero exact-match or near-match overlap confirmed across all configurations. Audit scripts available at: https://github.com/duttaprat/HViLM

Citation

@article{dutta2026hvilm,
  title={HViLM: A foundation model for viral genomics enables multi-task prediction of pathogenicity, transmissibility, and host tropism},
  author={Dutta, Pratik and Vaska, Jack and Surana, Pallavi and Sathian, Rekha and Chao, Max and Zhou, Zhihan and Liu, Han and Davuluri, Ramana V},
  journal={bioRxiv},
  pages={2026--03},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}

License

MIT License

Related Resources

• HViLM model weights: duttaprat/HViLM-base
• Code: github.com/duttaprat/HViLM