Sequence Based Virus Host Prediction

Key Findings

• Created a comprehensive dataset of 58,046 virus genomes across 15 families.
• Neural network model accurately predicts human host compatibility based solely on sequence features.
• K-mer frequency patterns (5-mers) provided the strongest predictive signals.
• Evolutionary analysis revealed a gradient of human adaptation from highly-adapted viruses to those with minimal compatibility.
• Cross-family validation demonstrated model generalization to unseen virus families.
• Viruses with low adaptation signals often correlate with higher virulence in humans.

Dataset Composition

Our dataset comprises 58,046 complete virus genomes representing 15 distinct virus families, with an approximately equal distribution between human-associated (52.0%) and non-human-associated (48.0%) viruses.

Methods

Key Results & Visualizations

Model Performance Across k-mer Sizes

We evaluated the predictive power of k-mer frequency patterns as features for human host prediction. The Neural Network model consistently outperformed both Logistic Regression and Random Forest baselines across different k-mer sizes, with 5-mer features yielding the highest Matthews Correlation Coefficient (MCC = 0.820).

Family-Specific Performance Analysis

We conducted a detailed analysis of model performance across individual virus families, revealing substantial variation in how well sequence features predict human compatibility. This variability suggests differing levels of host-specific genomic signatures among virus groups, which may reflect their evolutionary histories and host adaptation strategies.

Embedding Analysis and Visualization

t-distributed stochastic neighbor embedding (t-SNE) is a widely used nonlinear dimensionality reduction technique that preserves local relationships between data points when projecting high-dimensional data into a lower-dimensional space. To visualize and analyze the relationships between virus sequences, we employed t-SNE to generate low-dimensional projections of both the raw k-mer frequency vectors and the learned embeddings from the neural network's final layer. This was done on both training and test datasets using scikit-learn.

The resulting visualizations were colored according to three different classification schemes: human/non-human host tropism, model prediction probabilities, and prediction uncertainty. The uncertainty u for a prediction probability p was calculated as:

where maximum uncertainty (1.0) occurs at p=0.5 and minimum uncertainty (0) at p=0 or p=1.

Pathogenicity and Adaptation Correlation

We observed a striking inverse relationship between predicted human adaptation and virulence. Viruses with high predicted human-host compatibility typically exhibit lower fatality rates in humans, while those with minimal adaptation signals (e.g., rabies virus, Marburg virus) tend to cause more severe or fatal disease when they do infect humans. This finding suggests potential evolutionary trade-offs between adaptation and virulence.

COVID-19 Adaptation Trend

Figure 4: Temporal analysis of SARS-CoV-2 genome sequences showing increasing human host adaptation probability from 2019 to 2024. Each point represents a viral genome isolate, with the trend line showing a consistent increase in predicted human adaptation.

Citation

Data Availability

• Data splits and vectorized dataset: GitHub
• Trained model weights: Hugging Face
• Parquet format with metadata: Hugging Face Dataset
• Demo space: Host Classifier

Metadata Explorer

Acknowledgements

We extend our gratitude to the data contributors and maintainers of public genomic databases. We also acknowledge the institutional support and feedback from colleagues during the development of this dataset.