A CNN-Based Model of Cross-Immunity to Influenza A(H3N2) Virus: Testing Under "Real-World" Conditions.

Area of Science:

• Virology
• Computational Biology
• Immunology

Background:

• Influenza A(H3N2) poses a significant public health challenge due to rapid antigenic drift.
• Accurate prediction of cross-immunity is crucial for effective vaccine strain selection.
• Existing models often struggle with the complexity of real-world forecasting.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) model for predicting influenza A(H3N2) cross-immunity.
• To assess the model's performance in real-world forecasting scenarios.
• To evaluate the model's utility in aiding influenza vaccine strain selection.

Main Methods:

• Developed a CNN model utilizing hemagglutinin (HA/HA1) sequences encoded into 3D tensors.
• Derived antigenic distance from hemagglutination inhibition (HI) titers.
• Trained the model on WHO data (2011-2023) and tested on independent data (2022-2024).
• Evaluated performance using Accuracy, Sensitivity, Specificity, and Matthews Correlation Coefficient (MCC).

Main Results:

• The CNN model demonstrated high accuracy on a classic dataset (Accuracy = 0.9996, MCC = 0.9964).
• Testing on recent data yielded robust results (Accuracy: 0.73-0.81, MCC: 0.48-0.60), reflecting forecasting complexity.
• ROC analysis confirmed strong discriminative ability (AUC ≥ 0.805) and good calibration (Brier scores ≤ 0.192).
• A three-layer CNN architecture showed superior robustness on challenging datasets.

Conclusions:

• The developed CNN model is an effective tool for assessing influenza A(H3N2) antigenic distances.
• The model shows promise for integration into epidemiological models to support vaccine strain selection.
• Future improvements may involve modeling structural impacts of amino acid substitutions and polyclonal immune responses.