README.md

---
library_name: transformers
base_model: duttaprat/HViLM-base
datasets:
  - duttaprat/HVUE-v2
pipeline_tag: text-classification
tags:
  - genomics
  - virology
  - dna
  - virus
  - host-tropism
  - hvue-v2
license: apache-2.0
---

# HViLM-Tropism

**HViLM-Tropism** is the official HViLM model for binary host tropism classification: human-tropic versus non-human-tropic viruses.

- **Fine-tuned from:** [duttaprat/HViLM-base](https://huggingface.co/duttaprat/HViLM-base)
- **Benchmark:** [duttaprat/HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2)
- **HVUE v2 configuration:** `Host_Tropism/standard_95_1000bp`
- **Checkpoint selection:** best validation F1 (`checkpoint-14000`)
- **Input:** virus nucleotide sequence
- **Output:** non-human-tropic vs. human-tropic

This repository contains a **standalone full fine-tuned checkpoint**, so users can load `duttaprat/HViLM-Tropism` directly without separately loading `HViLM-base`.

## Label Mapping

| ID | Label |
|---:|---|
| 0 | `NON_HUMAN_TROPIC` |
| 1 | `HUMAN_TROPIC` |

## Performance

Held-out HVUE v2 test set, standard 1000-nt configuration:

| Metric | Score |
|---|---:|
| Accuracy | 96.49 |
| F1 | 74.49 |
| MCC | 48.99 |
| Precision | 74.96 |
| Recall | 74.04 |

## Class Imbalance

The HVUE v2 Host Tropism benchmark is strongly imbalanced. Accuracy should therefore be interpreted together with F1 and MCC. HViLM retains substantially stronger minority-class discrimination than vanilla DNABERT-2 trained on the same imbalanced setting and avoids majority-class collapse without explicit class balancing.

## Training Details

- **Fine-tuning method:** LoRA
- **LoRA rank:** 8
- **LoRA alpha:** 16
- **Target modules:** query and value projections across all 12 transformer layers
- **Approximate trainable LoRA parameters:** ~0.3M
- **Learning rate:** 3e-5
- **Maximum input length:** 250 BPE tokens (approximately 1000 nt)
- **Early stopping:** patience 3, monitored using validation F1
- **Hardware:** NVIDIA A40 GPU

The released repository contains the full task-specific model weights rather than only the LoRA adapter.

## Usage

```python
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "duttaprat/HViLM-Tropism"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
    model_id,
    trust_remote_code=True,
)

sequence = "ATGCGTACGTTAGCCGATCGATTACGCGTACGTAGCTAGC"
inputs = tokenizer(
    sequence,
    return_tensors="pt",
    truncation=True,
    max_length=250,
)

with torch.no_grad():
    logits = model(**inputs).logits

prediction_id = logits.argmax(dim=-1).item()
print(model.config.id2label[prediction_id])
```

Possible outputs are `NON_HUMAN_TROPIC` and `HUMAN_TROPIC`.

## Intended Use

HViLM-Tropism is intended for research and benchmarking of virus sequence-based host association prediction. Host association is biologically context-dependent and may include multi-host, zoonotic, and reverse-zoonotic relationships; predictions should not be interpreted as definitive evidence of host range.

## Related Resources

- [HViLM-base](https://huggingface.co/duttaprat/HViLM-base)
- [HViLM-Patho](https://huggingface.co/duttaprat/HViLM-Patho)
- [HViLM-R0](https://huggingface.co/duttaprat/HViLM-R0)
- [HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2)
- [HViLM GitHub repository](https://github.com/duttaprat/HViLM)

## Citation

```bibtex
@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}
}
```