The immunotherapy response problem: Checkpoint immunotherapy has transformed advanced NSCLC treatment, but only 20-30% of unselected patients respond. Current biomarkers like PD-L1 and TMB are imperfect predictors. Better predictors that are easily available before treatment would help avoid treating non-responders with toxic, expensive drugs.
The immune-radiomics approach: This study took an innovative approach: first identify immune-related genes (IRGs) that predict immunotherapy response using transcriptomic data, then find CT radiomics features that correlate with those genes - creating an imaging-based surrogate for the tumor immune microenvironment.
Study cohort: 210 advanced NSCLC patients (aNSCLC) were enrolled. The combined model was developed using the training cohort and validated in an independent validation cohort.
Key result: The combined clinical-radiomics model achieved AUC of 0.725 in the training cohort and 0.706 in the validation cohort for predicting immunotherapy response - modest but clinically meaningful improvement over clinical variables alone.
What is WGCNA: Weighted Gene Co-expression Network Analysis (WGCNA) identifies groups of genes that are co-expressed - meaning they turn on and off together. These co-expression modules often reflect shared biological functions, such as immune signaling, inflammation, or cytotoxicity.
Identifying immunotherapy-relevant modules: The researchers applied WGCNA to NSCLC gene expression datasets from TCIA and GEO databases. Co-expression modules most correlated with known immunotherapy response signatures were identified as the most biologically relevant.
84 immune-related genes identified: Within the most relevant co-expression modules, 84 IRGs showed strong correlation with immunotherapy response - including genes involved in T cell activation, interferon signaling, and checkpoint pathways. These genes represent the molecular underpinning of immunotherapy sensitivity.
From genes to imaging: The key innovation was correlating each of these 84 IRGs with CT radiomics features - looking for imaging patterns that reflect high vs. low expression of immune activation genes. This bridges molecular biology and radiology.
CT imaging dataset: 94 radiomics features were extracted from CT scans in the TCIA Lung_3 dataset - a public repository of annotated lung cancer CT images. Features covered shape, first-order statistics, and texture matrices (GLCM, GLRLM, GLSZM).
LASSO feature selection: With 94 features and the risk of overfitting, LASSO (Least Absolute Shrinkage and Selection Operator) regression was used to select the most informative features while penalizing model complexity. This resulted in 7 radiomics features forming the final signature.
Radiomics score computation: Each patient's seven selected radiomics features were weighted by their LASSO coefficients and summed to produce a single 'Rad-Score.' Patients were stratified into high vs. low Rad-Score groups with distinct immunotherapy response rates.
Combined model construction: The Rad-Score was combined with clinical variables (stage, PD-L1 expression, ECOG performance status, treatment type) in a multivariable logistic regression model to create the final combined predictor.
Clinical-radiomics combined model performance: The combined model achieved AUC of 0.725 in the training cohort and 0.706 in the validation cohort. While not dramatically high, this represents meaningful discrimination beyond what clinical variables (PD-L1, stage) achieve alone.
Nomogram design: The model was visualized as a nomogram - a graphical tool that allows clinicians to input patient parameters and read off the predicted probability of immunotherapy response. Each variable's contribution is shown as a scale with associated points.
Decision curve analysis: Decision curve analysis showed that the combined clinical-radiomics model provided net clinical benefit across a range of probability thresholds, meaning it would lead to better treatment decisions than either treating all patients or treating based on clinical variables alone.
High Rad-Score patients: Patients with high Rad-Scores showed significantly lower immunotherapy response rates and shorter progression-free survival - consistent with the hypothesis that their imaging reflects an immunosuppressive tumor microenvironment.
Texture features and immune exclusion: Several of the 7 selected radiomics features measured intratumoral heterogeneity in CT density - reflecting the patchy, heterogeneous nature of immune cell infiltration. Tumors with high textural heterogeneity may harbor immune 'deserts' alongside infiltrated regions.
Correlation with immune gene modules: The selected radiomics features correlated significantly with WGCNA immune co-expression modules - particularly those involving interferon-gamma signaling and cytotoxic T lymphocyte (CTL) activity. High Rad-Score correlated with low CTL activity gene modules.
Tumor border characteristics: Features capturing tumor border irregularity correlated with genes involved in epithelial-mesenchymal transition (EMT) - a process associated with immune evasion and reduced immunotherapy sensitivity.
Density and vascularity: CT density features correlated with genes in angiogenesis and hypoxia pathways. High tumor vascularity and hypoxia create physical barriers to T cell infiltration, which may explain why density features predict immunotherapy resistance.
Modest AUC performance: An AUC of approximately 0.71 is clinically useful but leaves significant room for improvement. Future models could incorporate additional biomarker types (cfDNA, protein markers, metabolomics) to achieve higher predictive accuracy.
Retrospective study limitations: Using retrospective imaging data means CT acquisition parameters varied, which can affect radiomics feature reproducibility. Prospective studies with standardized CT protocols would produce more reliable radiomics features.
Immunotherapy type heterogeneity: The cohort included patients receiving different immunotherapy agents (PD-1 and PD-L1 inhibitors) with and without chemotherapy. Future studies should stratify by specific treatment regimen to assess whether radiomics prediction differs by treatment type.
Dynamic radiomics: Single pre-treatment CT scans capture only a static snapshot. Sequential CT scans during treatment - with early radiomics changes analyzed for response patterns - could achieve much higher predictive accuracy by incorporating tumor response dynamics.