Vir2vec: A Genome-Wide Viral Embedding

Model description

Vir2vec is a viral genomic language model (gLM) designed to produce fixed-length, genome-level embeddings that can be fine-tuned across downstream tasks such as viral discrimination, host-range prediction, and variant typing. For more details and training scripts check GitHub

Intended use

Vir2vec embeddings are intended for tasks including (but not limited to):

• Virus vs non-virus genome/read discrimination
• DNA vs RNA virus classification
• Host-range prediction
• Intra-genus separation (e.g., HIV-1 vs HIV-2)
• Variant/subtype typing (e.g., SARS-CoV-2 lineages)
• Phenotypic signal detection (e.g., tissue tropism proxies)

Model sizes

All Vir2vec model scales produce fixed-length 4,096-dimensional genome-level embeddings via max-pooled vocabulary logits:

• 422M (main branch)
• 138M (revision="138M")
• 17M (revision="17M")

How to use

Load from Hugging Face

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("pabloarozarenad/Vir2vec", trust_remote_code=True) # Add revision=138M or revision=17M to change model size. 422M is default.
model = AutoModelForCausalLM.from_pretrained("pabloarozarenad/Vir2vec", trust_remote_code=True) # Add revision=138M or revision=17M to change model size. 422M is default.
model.eval()

Compute embeddings

dna = "ACGTAGCATCGCGATGACTGCATCACT"
inputs = tokenizer(dna, return_tensors="pt")

with torch.no_grad():
    # Forward pass to obtain vocabulary logits [batch, seq_len, 4096]
    outputs = model(**inputs)
    logits = outputs.logits

    # Max-pool logits over token positions -> 4,096-dimensional embedding
    embedding = torch.max(logits, dim=1).values[0]

print(embedding.shape)  # torch.Size([4096])

Access

Vir2vec can be loaded upon request, subject to providing an institutional email address, a brief description of the intended use, and the associated IRB protocol number. For more info contact pablo.arozarenadonel01@universitadipavia.it