Evaluating machine learning approaches for host prediction using H3 influenza genomic data.

Area of Science:

• Virology
• Bioinformatics
• Machine Learning

Background:

• Influenza A viruses (IAV) frequently cross species barriers, facilitating sustained transmission.
• Current machine learning (ML) host prediction for IAV often uses single genome segments or broad host categories.

Purpose of the Study:

• To develop and validate ML models for distinct host prediction of H3 IAV using sequence data from all eight genome segments.
• To identify IAV variants with increased between-species transmission potential.

Main Methods:

• Trained ML models (Random Forest, XGBoost) on k-mers and amino acid properties for each of the eight IAV genome segments.
• Validated models using a test dataset, analyzing predicted probabilities to investigate transmission patterns.
• Examined case studies including canine H3N8, swine H3N2, and duck H3 sequences.

Main Results:

• Models achieved high prediction accuracy (0.995-0.997) and kappa values (0.984-0.990) across all eight segments.
• Sequences with high predicted probabilities (>90%) often indicated between-species transmission events.
• Case study analyses confirmed model predictions align with known transmission patterns.

Conclusions:

• Developed models enable rapid, accurate host prediction for H3 IAV from any genome segment.
• Provides a framework for identifying IAV with high transmission potential.
• Suggests further refinement of training and validation processes is warranted.