CellPhenoX: An Explainable Machine Learning Method for Identifying Cell Phenotypes To Predict Clinical Outcomes from...

Single-cell technologies have transformed the understanding of disease heterogeneity, but linking cell-level phenotypic alterations to clinical outcomes becomes increasingly challenging as single-cell datasets continue to expand. This is further complicated by the lack of interpretability in existing methods and the difficulty of detecting interaction effects-nonlinear dependencies between factors like sex, age, and disease. To address this, a novel explainable machine learning method, CellPhenoX, is developed to identify cell-specific phenotypes and interaction effects linked to clinical outcomes. CellPhenoX integrates classification models, explainable artificial intelligence (AI) techniques, and a statistical framework to generate interpretable, cell-specific scores to uncover condition-associated cell populations. Extensive benchmarking and applications demonstrate the efficacy of CellPhenoX across diverse single-cell study designs, including the dedicated and disease-motivated simulations, binary disease-control comparisons, and severity-stratified patient cohorts. Notably, CellPhenoX identifies an activated monocyte phenotype in COVID-19, with expansion correlated with disease severity after adjusting for covariates and interactive effects. It also uncovers a fibroblast-specific state transition gradient predicting tissue inflammation in chronic diseases, and identifies therapy-induced T cell changes and biomarkers linked to the tumor microenvironment. By integrating interpretability into clinical classification, CellPhenoX offers a powerful framework for translating single-cell findings into clinical impact.