Abstract
Understanding the host-specificity of different families of viruses sheds light on the origin of zoonotic pathogens like SARS-CoV-2 and rabies. Identifying these reservoirs enables epidemiologists and medical professionals to curb existing epidemics and prevent future ones. We propose Virus2Vec, a feature-vector representation for viral sequences that enables machine learning models to identify viral hosts without computationally expensive sequence alignment. Virus2Vec leverages the power of both minimizers and Position Weight Matrices (PWM) to generate compact, informative feature vectors. Evaluated on real-world Coronaviridae spike proteins and rabies virus sequences, Virus2Vec demonstrates superior predictive accuracy and significant speed improvements over baseline and state-of-the-art methods.
Introduction
The identification of animal reservoirs for zoonotic viruses—such as SARS-CoV-2 or Rabies—is a critical component of pandemic preparedness and response. Traditional bioinformatics pipelines rely heavily on sequence alignment, a process that arranges sequences to identify regions of similarity. However, alignment is an NP-hard problem that is computationally expensive and difficult to scale to the millions of sequences generated during global health crises.
Existing machine learning approaches often utilize $k$-mers (short substrings of length $k$), but these can result in massive, sparse feature vectors that are memory-intensive and noisy. Virus2Vec addresses these limitations by creating compact, numerical representations of viral sequences that bypass the alignment step while retaining the biological "signal" necessary for accurate host prediction.
Methods: The Virus2Vec Pipeline
Virus2Vec is an alignment-free feature engineering pipeline that transforms raw protein or nucleotide sequences into fixed-length vectors. The process is summarized in the following flowchart:
The pipeline consists of several key stages:
1. Minimizer Computation
To reduce data dimensionality, Virus2Vec identifies "minimizers"—the lexicographically smallest $m$-mer within a sliding window of $k$-mers. This acts as a lossy compression technique, preserving a representative fraction of the sequence while drastically reducing the number of features the model must process. This allows the model to focus on the most "stable" parts of the viral signature.
2. Position Weight Matrix (PWM) Scoring
To provide biological context, the model calculates a Position Weight Matrix (PWM). This matrix represents the log-likelihood of a specific amino acid or nucleotide appearing at a certain position compared to a random background distribution. Unlike simple one-hot encoding, PWM allows the model to understand that some mutations are more biologically significant than others. By summing the PWM scores for each minimizer, Virus2Vec generates a compact, dense feature vector.
Experimental Evaluation
The researchers evaluated Virus2Vec on two primary real-world datasets:
- Coronavirus Host Data: 5,558 Spike protein sequences covering 22 different host classes (e.g., humans, bats, camels, cats).
- Rabies Virus (RABV) Data: 20,051 nucleotide genome sequences representing 12 host classes (e.g., dogs, foxes, bats).
Key Results
Virus2Vec consistently outperformed baselines such as OHE, Spike2Vec, and PWM2Vec across multiple metrics:
- Accuracy: On unaligned Coronavirus data, Virus2Vec achieved an F1-score of 0.88 using a Logistic Regression classifier.
- Speed: The efficiency gains were dramatic. For the Coronavirus dataset, Virus2Vec generated feature vectors in just 90.65 seconds, compared to 1,506 seconds for PWM2Vec and 379 seconds for the Approximate Kernel method.
- Robustness: The model maintained high performance on both amino acid (protein) and nucleotide (DNA/RNA) sequences, demonstrating versatility across different biological data types.
Conclusion and Future Work
Virus2Vec provides a scalable, efficient, and accurate framework for viral host identification. By bypassing the alignment bottleneck and utilizing compressed biological signatures, it enables rapid analysis of genomic data during outbreaks. Future research will focus on testing the framework's performance on "unassembled" short-read data directly from sequencing machines and scaling the system to handle millions of sequences for real-time global monitoring.
Summary updated