Patient-Level Transcriptomic Backtracking of Immune-Pathway Deficiencies in Immune Checkpoint Inhibitor Non-Responders: A Multicohort Secondary Analysis

Abstract Objectives To determine whether pretreatment tumor transcriptomes can identify patient-level immune deficits and organize mechanistically distinct forms of immune checkpoint inhibitor non-response. Design Secondary analysis using a fixed 16-module, 195-gene panel in four initial cohorts, external application of the unchanged panel and four-state rules in IMvigor210, and a five-cohort meta-analysis of raw IFN/STING discrimination. Setting Publicly archived pretreatment tumor-transcriptomic cohorts from immune checkpoint inhibitor-treated patients with multiple solid tumors. Participants Five cohorts comprising 474 tumors (120 responders and 354 non-responders); four initial cohorts included 176 tumors, and IMvigor210 included 298. Main outcome measures Cohort-level response AUCs, patient-level CD45-adjusted module deficits, four-state distributions, and concordance with independent perturbation-derived footprints. Results In the 85-tumor Gene Expression Omnibus accession GSE176307 urothelial cohort, the raw IFN/STING AUC was 0.726 (q=0.064); after CD45 adjustment, IFN/STING (AUC 0.764; q=0.014) and antigen-presentation machinery (AUC 0.720; q=0.042) were associated with response. The four-state hierarchy classified 124 initial-cohort non-responders as immune-desert (26.6%), inflamed-excluded (12.9%), ignition-deficit (21.0%), or substrate-adequate (39.5%); substrate-adequate tumors retained additional pathway heterogeneity. Applying the fixed rules to 230 IMvigor210 non-responders produced a similar overall distribution, and PD-L1 immune-cell staining varied by state. IMvigor210 showed nominal IFN/STING discrimination (AUC 0.579; p=0.048; q=0.147; permutation p=0.024). Across five cohorts, the pooled random-effects IFN/STING AUC was 0.606 (95% CI 0.483 to 0.716; modified Hartung–Knapp). Independent innate-ignition and anti-VEGFA endothelial footprints correlated with their corresponding modules. Conclusions Patient-level module deficits considered with leukocyte infiltration distinguish biologically interpretable non-responder states and nominate candidate failure mechanisms. The resulting deficit–function framework supports prospective testing of function-matched adjunctive strategies.