#33 · Single-cell

Personalized and reference-based models for high-dimensional biomedical data

The dominant thread is learning structured models that separate shared population-level structure from individual or disease-specific deviation. A parameter decomposition into shared and patient-specific components is applied to calibrating dynamical cancer models under data scarcity. A subspace one-class classifier is trained on healthy plasma proteomic profiles to flag disease anomalies, while GUT-FORMer, a transformer autoencoder, jointly models taxonomic and functional metagenomic profiles for disease classification and phenotype discovery. MOZAIC applies AI to donor-recipient microbiome matching across 515 pre/post-FMT metagenomes from 30 cohorts. What is new is the explicit shared-plus-individual decomposition for scarce data and the joint modeling of taxonomic and functional layers.