CAPYBARA: A generalizable framework for predicting serological measurements across human cohorts

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Biological datasets are growing rapidly, necessitating methods to integrate past findings into new experiments.
• Serological studies, crucial for vaccine development, face challenges in reconciling data variations due to demographics or experimental design.
• Existing methods struggle to quantify the uncertainty of future serological measurements based on prior studies.

Purpose of the Study:

• To introduce CAPYBARA, a data-driven framework for quantifying and mapping serological relationships across different datasets.
• To assess the predictive power of CAPYBARA in inferring unknown serological measurements from a subset of data.
• To identify conserved serological patterns and cross-dataset trends for improved vaccine strain selection and study design.

Main Methods:

• Applied the CAPYBARA framework to 25 influenza hemagglutination inhibition (HAI) datasets spanning 1997-2023.
• Evaluated predictive accuracy by withholding and subsequently predicting HAI measurements, calculating mean absolute error.
• Utilized feature importance analysis within the interpretable CAPYBARA model to uncover cross-reactivity trends.

Main Results:

• CAPYBARA accurately predicted withheld HAI measurements with a 2.0-fold mean absolute error, comparable to experimental variability.
• Predictions maintained 2-3 fold accuracy across diverse conditions, including different age groups, vaccination vs. infection studies, and studies within a 10-year range.
• Identified global cross-reactivity trends through interpretable model analysis, highlighting conserved serological patterns.

Conclusions:

• CAPYBARA effectively quantifies and maps serological relations across diverse datasets, enabling prediction of antibody responses.
• The framework offers a valuable tool for vaccine strain selection and longitudinal studies by inferring broad responses from limited data.
• Understanding cross-dataset serological patterns enhances the utility of past research for future experimental design and prediction.