Advanced Machine Learning (ML) Across Multiple Vaccines to Predict Durable Immunogenicity using the Immune Signatures Data Resource (ISDR) 4279

Abstract Description Understanding vaccine durability is key to designing vaccines with long-term efficacy. Using the Immune Signatures Data Resource, a compendium of transcriptomic and immune responses from 820 healthy adults (18–50 years) across 13 vaccines, we analyzed responses to live (e.g. yellow fever, smallpox), recombinant viral vectors (e.g. Ebola), inactivated (e.g. influenza), and glycoconjugate (e.g. pneumococcal) vaccines. Data preparation included normalizing input features across time points. Post-vaccination timepoints accounted for 15% of variance, indicating shared immune kinetics, while age and sex contributed minimally. Advanced machine learning (ML) techniques predicted vaccine immunogenicity and durability. For high-dimensional, low-sample-size multi-omics datasets, we used HSIC Lasso to identify features predictive of antibody responses. These features were then used as input for ML algorithms, such as RF, Lasso, Ridge, DT, ET, XGB, and CatBoost. We compared single-target models versus stacked, multi-output, and chained multi-output wrappers, with CatBoost wrappers yielding the most accurate predictions across 13 vaccines. Moreover, we trained a neural network using a multi-layer architecture with two hidden layers. Key findings included conserved interferon signaling and plasmablast-related genes as strong predictors of antibody durability. This study identified key features critical for developing vaccines with durable responses, advancing systems vaccinology. Funding Sources NIH/ NIAID U19AI168643 ; 3U01AI67892-03S2 Topic Categories Computational and Systems Immunology (COMP)