scTIME: single-cell-informed interpretable modeling of immunotherapy-associated immune states in non-small cell lung cancer

Immune checkpoint inhibitor response in non-small cell lung cancer (NSCLC) reflects coordinated tumor and immune states that are not fully captured by single biomarkers. We developed scTIME, an interpretable model that translates single-cell immune programs into bulk transcriptomic scores, and integrated public single-cell, T-cell receptor, transcriptomic, survival, and proteogenomic cohorts. Response modeling used two cohorts with different endpoints (40 patients) and compared six algorithms. ElasticNet was selected as a sparse reporting model, achieving a pooled five-fold out-of-fold ROC-AUC of 0.810; nested model selection yielded 0.772. Matched held-out stratification identified response-enriched scTIME groups in a direct descriptive comparison with CD274, while the fitted score emphasized cytotoxic CD8, natural killer, SPP1 macrophage, interferon-response, and antigen-presentation programs. Single-cell analyses and external projections linked these programs to CXCL10-CXCR3 recruitment, exclusion-associated states, and T-cell receptor ecology across molecular contexts. Exploratory survival and cross-omics analyses extended the framework beyond the response cohorts but were interpreted as projection rather than independent validation. Together, the results show how cell-resolved immune biology can support transparent, multi-omic biomarker development without reducing response to a single pathway. scTIME remains exploratory because model choice was post hoc , cohorts were small and clinically heterogeneous, and endpoint-matched external validation was unavailable.