CT-based immune radiomic signature for prognosis and prediction of immunotherapy and anticancer drug response in NSCLC

Background Non-small cell lung cancer (NSCLC) accounts for more than 85% of lung cancers and remains the leading cause of cancer-related mortality worldwide, with dismal prognosis and pronounced inter-patient heterogeneity in responses to immune checkpoint inhibitors, chemotherapy and targeted therapies. Current approaches to characterising the tumour microenvironment (TME) immune landscape rely on invasive tissue biopsies, which are constrained by sampling bias and an inability to capture spatial and temporal heterogeneity. CT-based radiomics, which derives high-dimensional quantitative features from routine imaging, offers a non-invasive alternative; however, most existing radiomic signatures lack robust biological grounding and do not provide integrated prediction of multiple treatment responses. Objective To develop and validate a non-invasive, immune-informed CT-based radiomic signature (CT-RadScore) that reflects TME immune phenotypes and to evaluate its utility for prognostic stratification and prediction of responses to immunotherapy and anticancer drugs in NSCLC. Methods We analysed multiple NSCLC cohorts, including TCIA NSCLC Radiogenomics, TCGA LUAD+LUSC and an in-house cohort from Nanyang Central Hospital. Consensus clustering of immune profiles was used to derive immune subtypes, and weighted gene co-expression network analysis (WGCNA) linked immune infiltration patterns to gene modules. CT radiomic features were extracted using PyRadiomics, and machine-learning survival models (random survival forest [RSF], LASSO–Cox, Elastic Net–Cox) were compared to construct CT-RadScore. Functional validation included Tumour Immune Dysfunction and Exclusion (TIDE) analysis, Tracking Tumour Immunophenotype (TIP) profiling, in silico drug sensitivity prediction, single-cell RNA sequencing (scRNA-seq) and GPX2 knockdown experiments. Results Consensus clustering identified two reproducible immune subtypes (immune-hot and immune-cold), which were validated across seven independent immune deconvolution algorithms. WGCNA revealed an immune-related gene module enriched for pathways governing innate and adaptive immunity. An RSF-integrated Cox model (COX+RSF) yielded CT-RadScore, composed of 12 non-redundant radiomic features, which demonstrated robust and externally validated prognostic performance (C-index: 0.791 in TCIA, 0.729 in TCGA and 0.844 in the in-house cohort). High CT-RadScore was associated with significantly worse overall survival (all log-rank P<0.0017) and remained an independent prognostic factor after adjustment for clinicopathological covariates. Low CT-RadScore predicted improved immunotherapy response (response rates 54.7–75% versus 30–34.4% in high-score groups; in-house AUC = 0.817) and was associated with lower TIDE scores. CT-RadScore further correlated with predicted sensitivity to key agents (for example, paclitaxel, gefitinib and carboplatin) and was mechanistically linked to GPX2-driven DNA replication licensing programmes and impaired antitumour immune cycling. Conclusion CT-RadScore is a non-invasive, biologically anchored radiomic signature that captures TME immune phenotypes and tumour proliferative programmes. It provides independent prognostic information and robust prediction of immunotherapy and anticancer drug responses, with potential to refine risk stratification and guide personalised treatment selection in NSCLC. Prospective multicentre validation and methodological standardisation will be essential for clinical translation.