scVIP: personalized modeling of single-cell transcriptomes for developmental and disease phenotypes

Single-cell transcriptomics resolves cellular heterogeneity within individuals, but connecting molecular states to individual-level phenotypes requires frameworks that explicitly bridge these scales. We present scVIP, a generative model that links gene expression, cell-type composition, and phenotypic measurements within a single probabilistic model, which enables accurate phenotype prediction and interpretable trajectory inference. A cell-type-aware multi-instance learning architecture learns donor embeddings that capture progression while localizing phenotype-associated signals to specific cell populations. Applied across four settings, scVIP accurately predicts cortical developmental age (Pearson r = 0.95), characterizes Huntington's disease progression (concordance correlation coefficient = 0.90), integrates two Alzheimer's disease cohorts recovering disease-relevant microglial and astrocytic programs, and distinguishes healthy from ACPA-positive individuals and non-progressors from early RA individuals, identifying inflammatory T cell programs associated with disease. scVIP enables principled analysis of how cellular states collectively shape organism-level phenotypes across development and disease.