A novel Vector-Symbolic Architecture for graph encoding and its application to viral pangenome-based species...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Viral species classification is vital for epidemiology and diagnostics.
• Traditional sequence similarity methods struggle with rapidly evolving viruses.
• Pangenomes offer comprehensive genomic diversity but require advanced computational analysis.

Purpose of the Study:

• To investigate Hyperdimensional Computing (HDC) for encoding and classifying viral pangenomes.
• To develop a novel method for representing graph-structured pangenomes using high-dimensional vectors.
• To evaluate different HDC model architectures for viral species classification.

Main Methods:

• Viral pangenomes were encoded as weighted de Bruijn graphs using k-mers.
• Species taxonomic labels were represented as specific high-dimensional vectors (species hypervectors).
• Three classification strategies were tested: flat species-level, flat genus-level, and hierarchical models.

Main Results:

• The flat species-level model achieved the highest accuracy at 87.08%.
• Genus-level and hierarchical models showed significantly lower accuracies (60.51% and 33.57%).
• Model performance was sensitive to architecture and routing strategies, indicating limitations in multi-step models.

Conclusions:

• HDC offers a promising new approach for viral classification using pangenomic data.
• The study highlights the importance of model architecture and routing strategies in HDC applications.
• This method provides insights into genomic taxonomy challenges and potential solutions.