WaveSeekerNet: accurate prediction of influenza A virus subtypes and host source using attention-based deep learning

Area of Science:

• Virology
• Genomics
• Bioinformatics

Background:

• Influenza A virus (IAV) poses a global threat to animal health and pandemic preparedness.
• Accurate prediction of IAV subtypes and host source is critical for surveillance.
• Deep learning offers novel approaches for analyzing viral genomic sequences.

Purpose of the Study:

• To develop a novel deep learning model, WaveSeekerNet, for accurate and rapid prediction of IAV subtypes and host source.
• To leverage attention-based mechanisms and efficient token mixing (Fourier and Wavelet Transforms) for pattern recognition in viral sequences.

Main Methods:

• Developed WaveSeekerNet, a deep learning model utilizing attention mechanisms and Fourier/Wavelet Transforms.
• Evaluated model performance on diverse datasets, comparing it against existing methods like VADR and ESM-2.
• Assessed accuracy in predicting IAV subtypes and distinguishing between human, avian, and mammalian hosts.

Main Results:

• WaveSeekerNet demonstrated superior performance compared to traditional self-attention models.
• Achieved maximum scores (1.0) in subtype prediction, rivaling VADR.
• Outperformed ESM-2 in generalization and computational efficiency.
• Showcased high accuracy in host source prediction and flagging cross-species transmission events.

Conclusions:

• WaveSeekerNet offers superior performance, efficiency, and potential for real-time surveillance.
• The model's ability to flag cross-species transmission events is valuable for pandemic preparedness.
• Represents a significant advancement in deep learning for IAV classification, aiding epidemiological and public health efforts.