Multi-Task Multimodal Fusion with Tabular Foundation Models for Pertussis Booster Response Prediction

Pertussis booster vaccination produces immune responses that vary widely across individuals in both peak magnitude and long term durability, governed by partly distinct biological compartments. Most computational models target only one phase; jointly predicting both is non-trivial because the endpoints are biologically dissociated, samples are small, and modalities have structured missingness. We propose a multi-task contrastive fusion architecture combining frozen TabPFN-v2 per-modality encoders, a dual-label supervised contrastive loss, modality dropout calibrated to empirical missingness, and missingness-masked attention fusion. On the CMI-PB pertussis booster subset (four modalities: antibody titers, cytokines, cell frequencies, gene expression; 44.9\% of Task 1 subjects and 39.6\% of Task 2 subjects missing at least one modality), we jointly predict peak response ($\log_2$(day 14 / day 0) IgG anti-pertussis-toxin fold change, n = 158) and durability ($\log_2$(day 120 / day 30) retention, n = 96). The model achieves test AUROC 0.797 (95\% CI [0.621, 0.948]) for peak and 0.755 ([0.519, 0.945]) for durability, both significant under joint label permutation (p = 0.002, p = 0.045). Across logistic regression, XGBoost, and MLP baselines on raw features and TabPFN embeddings, the proposed model is the only method whose 95\% CIs lie above chance on both tasks simultaneously.