Speaker
Description
Predicting patient-specific immune outcomes requires frameworks that are mechanistically grounded, scalable across disease contexts, and capable of producing interpretable clinical predictions. We present an integrated pipeline built around a large-scale logic-based mechanistic model of the human immune system.
The model encodes Boolean regulatory logic derived from 449 human experimental publications, representing 88 immune cell types across innate and adaptive compartments, 37 secretory factors, and 11 disease environments connected through 1,450 regulatory interactions. Validation against independent human data confirmed the model's capacity to reproduce pathogen-specific cytokine signatures and cell dynamics across nine pathogens, with agreement rates of 75–90%.
We leverage this validated model to generate synthetic datasets of in silico patient profiles by systematically varying immune state activity across the full spectrum from healthy to severely dysregulated. Machine learning models trained on these data identify patient-specific biomarkers predictive of pathogen clearance, hospitalization, and ICU admission across IAV, SARS-CoV-2, CMV, and Plasmodium falciparum. Critically, because each biomarker is a node in the underlying mechanistic model, its predictive weight can be traced directly to causal regulatory logic, addressing the interpretability gap that limits purely data-driven tools.
This pipeline provides a generalizable approach to mechanistically grounded biomarker discovery and a methodological foundation for personalized immune digital twins.
