DeepFlu

DeepFlu predicts susceptibility to symptomatic influenza A infection from pre-exposure human gene expression profiles.

Key Features:

• Deep Learning Approach: Uses a deep neural network architecture that outperformed convolutional neural networks, random forests, and support vector machines in comparative tests.
• Gene Expression Data Utilization: Analyzes human pre-exposure gene expression profiles to identify patterns associated with susceptibility to symptomatic IAV infection (H1N1 and H3N2).
• Performance Metrics: In leave-one-person-out cross-validation, achieved for H1N1 an accuracy of 70.0%, AUROC 0.787, AUPR 0.758, and for H3N2 an accuracy of 73.8%, AUROC 0.847, AUPR 0.901, with comparable external validation and comparisons to the biomarker KLRD1.
• Model Training Insights: Training exclusively on pre-exposure data yielded better predictive performance than mixed time-span data, and combining H1N1 and H3N2 datasets did not improve accuracy, indicating subtype-specific models are important.

Scientific Applications:

• Prospective risk assessment: Forecasts individual risk of symptomatic influenza A infection prior to exposure to inform public health surveillance.
• Targeted interventions: Identifies individuals for tailored interventions to reduce transmission risk.
• Healthcare planning: Supports allocation of healthcare resources by predicting likely symptomatic cases before exposure.
• Research on host immunity: Facilitates investigation of genetic determinants and host immune signatures associated with influenza A susceptibility.

Methodology:

Trains a deep neural network on pre-exposure human gene expression data from individuals exposed to IAV subtypes H1N1 and H3N2, evaluated with leave-one-person-out cross-validation, compared against convolutional neural networks, random forests, and support vector machines, and externally validated.