Virus-Host-Genomes Dataset

Dataset Summary

Virus-Host-Genomes is a comprehensive collection of viral genomic sequences paired with host information, containing 58,046 viral sequences. The dataset includes metadata such as viral taxonomy (family, genus), host information, geographic data, isolation sources, and various annotations including zoonotic potential indicators. This dataset was put together to support investigations into genetic determinants of host specificity, zoonotic potential, and genome-based classification models.

This repository hosts the original dataset associated with the paper. An actively maintained version of the dataset is available at hiyata/Virus-Host-Genomes-updates-v2 and is updated monthly. The latest maintained version is always available on Hugging Face.

Citation Information

If you use this dataset, please cite:

@article{carbajo2026sequence,
  author    = {Carbajo, Alan L and Vensko, Taylor A and Pellett, Philip E},
  title     = {Sequence Based Virus Host Prediction: A Curated Dataset and Generalizable Framework for Training Artificial Intelligence to Identify Viruses of Humans},
  journal   = {Virus Evolution},
  year      = {2026},
  pages     = {veag009},
  publisher = {Oxford University Press},
  doi       = {10.1093/ve/veag009},
  url       = {https://doi.org/10.1093/ve/veag009}
}

Supported Tasks

• Host Prediction: Using viral sequences to predict potential hosts
• Zoonotic Potential Assessment: Identifying viruses with potential to cross between species
• Taxonomic Classification: Classifying viruses based on genomic sequences
• Sequence Analysis: Extracting sequence features like k-mer frequencies for analyses or preprocessing

Dataset Structure

Data Instances

A typical data instance contains a virus genome sequence with taxonomic classification and host information:

{
  'sequence': 'CCATTCCGGG...', # Viral genomic sequence
  'virus_name': 'Human betaherpesvirus 5', # Common virus name
  'host': 'human',            # Primary host (human or non-human)
  'zoonotic': False,          # Whether virus is known to be zoonotic
  # See Data Fields below for the full schema
}

Data Fields

The dataset contains the following fields:

Field Name	Type	Description	Example
sequence	string	Genomic sequence of the virus	"CCATTCCGGG..."
family	string	Taxonomic family of the virus	"Orthoherpesviridae"
accession	string	Database accession number	"AY446894.2"
host	string	Primary host (human or non-human)	"human"
genus	string	Taxonomic genus of the virus	"Cytomegalovirus"
isolation_date	string	Date when virus was isolated	"1999"
strain_name	string	Strain or isolate identifier	"Merlin"
location	string	Geographic location of isolation	"United Kingdom: Cardiff"
virus_name	string	Common name of the virus	"Human betaherpesvirus 5"
isolation_source	string	Source material of isolation	"urine from a congenitally infected child"
lab_culture	bool	Whether isolated from lab culture	true/false
wastewater_sewage	bool	Whether isolated from wastewater	true/false
standardized_host	string	Standardized host taxonomy	"Homo sapiens"
host_category	string	Category of host organism	"Mammal"
standardized_location	string	Standardized location	"United Kingdom"
zoonotic	bool	Known to cross species barriers	true/false
processing_method	string	How sequence was processed	"NGS"
gemini_annotated	bool	Annotated with Gemini AI	true/false
is_segmented	bool	Whether virus has segmented genome	true/false
segment_label	string	Label for genome segment	"NA"

Data Splits

The dataset contains train and test splits:

Split Name	Number of Instances
train	51,935
test	6,111

Dataset Creation

Source Data

This dataset compiles virus sequences from multiple public repositories, including:

• NCBI Virus
• GenBank

Data Processing

The dataset has undergone several processing steps:

• Sequence standardization (using only unambigious IUPAC nucleotide characters)
• Host information standardization
• Geographic location normalization
• Additional annotations including zoonotic potential labeling
• Quality filtering to remove low-quality or incomplete sequences

Host labels were generated through a tier-based approach:

1. Approximately 10,000 sequences were manually labeled by experts
2. First-tier automated labeling used direct string matching against known host names
3. Second-tier labeling employed pattern recognition from a species dictionary
4. For sequences that couldn't be classified by either tier, Google Gemini was used to analyze available metadata and assign host labels

Some sequences were annotated using the Gemini AI system to provide additional metadata where information was incomplete.

Considerations for Using the Data

Limitations and Biases

• Sampling Bias: The dataset may overrepresent viruses of clinical importance and underrepresent environmental viruses.
• Temporal Distribution: More recent viruses (especially those causing outbreaks) may be overrepresented.
• Geographic Bias: Samples from regions with stronger research infrastructure may be overrepresented.
• Host Bias: Human viruses and viruses from domestic/agricultural animals may be overrepresented.
• Annotation Quality: Some metadata fields are incomplete or may contain uncertainties.

Usage Examples

Data Preparation and K-mer Vectorization

import numpy as np
from datasets import load_dataset
from itertools import product
from sklearn.preprocessing import StandardScaler, LabelEncoder
import joblib
from tqdm import tqdm

# Load dataset
virus_dataset = load_dataset("hiyata/Virus-Host-Genomes")
train_dataset = virus_dataset['train']
test_dataset = virus_dataset['test']


# Generate k-mer dictionary once
def generate_kmer_dict(k):
    return {''.join(kmer): i for i, kmer in enumerate(product('ACGT', repeat=k))}

# Calculate k-mer frequency
def calculate_kmer_freq(seq, k, kmer_dict):
    freq = np.zeros(4**k)
    total_kmers = len(seq) - k + 1
    for i in range(total_kmers):
        kmer = seq[i:i+k]
        if 'N' not in kmer and all(base in 'ACGT' for base in kmer):
            freq[kmer_dict[kmer]] += 1
    return freq / total_kmers if total_kmers > 0 else freq

# Vectorize dataset
def vectorize_dataset(dataset, k=4):
    kmer_dict = generate_kmer_dict(k)
    num_samples = len(dataset['sequence'])
    X = np.zeros((num_samples, 4**k))
    y = np.array(['human' if host.lower() == 'human' else 'non-human' for host in dataset['host']])

    for idx, seq in enumerate(tqdm(dataset['sequence'], desc="Vectorizing sequences")):
        X[idx] = calculate_kmer_freq(seq.upper(), k, kmer_dict)

    return X, y


X_train, y_train = vectorize_dataset(train_dataset)
X_test, y_test = vectorize_dataset(test_dataset)

# Standard Scaler
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)

# Save scaler
joblib.dump(scaler, 'standard_scaler.joblib')

# Label encoding
le = LabelEncoder()
y_train_enc = le.fit_transform(y_train)
y_test_enc = le.transform(y_test)

print("Vectorization complete.")

Neural Network Training for Host Classification

import torch
from torch import nn, optim
from torch.utils.data import DataLoader, TensorDataset

# Define your neural network
class VirusClassifier(nn.Module):
    def __init__(self, input_shape: int):
        super(VirusClassifier, self).__init__()
        self.network = nn.Sequential(
            nn.Linear(input_shape, 64),
            nn.GELU(),
            nn.BatchNorm1d(64),
            nn.Dropout(0.3),

            nn.Linear(64, 32),
            nn.GELU(),
            nn.BatchNorm1d(32),
            nn.Dropout(0.3),

            nn.Linear(32, 32),
            nn.GELU(),

            nn.Linear(32, 2)
        )

    def forward(self, x):
        return self.network(x)

# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# DataLoader setup
train_loader = DataLoader(TensorDataset(
    torch.tensor(X_train, dtype=torch.float32),
    torch.tensor(y_train_enc, dtype=torch.long)
), batch_size=64, shuffle=True)

test_loader = DataLoader(TensorDataset(
    torch.tensor(X_test, dtype=torch.float32),
    torch.tensor(y_test_enc, dtype=torch.long)
), batch_size=64, shuffle=False)

# Initialize the model
model = VirusClassifier(input_shape=X_train.shape[1]).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Training loop
epochs = 15
for epoch in range(epochs):
    model.train()
    epoch_loss = 0
    for X_batch, y_batch in train_loader:
        X_batch, y_batch = X_batch.to(device), y_batch.to(device)
        optimizer.zero_grad()
        outputs = model(X_batch)
        loss = criterion(outputs, y_batch)
        loss.backward()
        optimizer.step()
        epoch_loss += loss.item()
    avg_loss = epoch_loss / len(train_loader)
    print(f"Epoch [{epoch+1}/{epochs}], Loss: {avg_loss:.4f}")

# Save the trained model
torch.save(model.state_dict(), 'virus_classifier_model.pth')

Model Evaluation with Matthews Correlation Coefficient

from sklearn.metrics import classification_report, matthews_corrcoef

model.eval()
y_preds = []
y_true = []

with torch.no_grad():
    for X_batch, y_batch in test_loader:
        X_batch = X_batch.to(device)
        outputs = model(X_batch)
        preds = torch.argmax(outputs, dim=1).cpu().numpy()
        y_preds.extend(preds)
        y_true.extend(y_batch.numpy())

# Classification Report
report = classification_report(y_true, y_preds, target_names=['human', 'non-human'])
print("Classification Report:\n", report)

# MCC Score
mcc = matthews_corrcoef(y_true, y_preds)
print(f"Matthews Correlation Coefficient (MCC): {mcc:.4f}")