Radiogenomic inference of transcription factor regulatory architecture associated with a high-risk CT phenotype in Non-Small Cell Lung Cancer.
TL;DR
A CT radiomics-defined poor-prognosis phenotype in NSCLC is associated with coherent transcriptomic programs and an interpretable TF regulatory architecture and supports regulatory-state-informed therapeutic hypothesis generation.
Abstract
Purpose
To determine whether a CT radiomics-defined poor-prognosis phenotype in NSCLC is associated with reproducible transcriptomic programs and distinct transcription factor (TF) activity patterns.
Methods
This retrospective radiogenomic study used public datasets from The Cancer Imaging Archive. A prognostic radiomics model was developed from pretreatment CT in 402 patients from the NSCLC-Radiomics collection and evaluated in an independent cohort of 110 patients from the NSCLC Radiogenomics collection. Seven tumor-centered intra- and peritumoral regions were analyzed. Features were extracted from original, wavelet-, and Laplacian-of-Gaussian-filtered CT images. After preprocessing and training-cohort standardization, a LASSO-penalized Cox model was developed. The locked model retained eight features, all from the GTV+10-mm region, and was applied to the internal and external validation cohorts without refitting. The training-cohort median RadScore was used unchanged for Kaplan-Meier visualization. Matched transcriptomic data were analyzed using preranked gene set enrichment analysis (GSEA), gene set variation analysis (GSVA), target-gene-based TF activity inference, TF co-activity networks, exploratory regulator prioritization, unsupervised TF-state clustering, and LINCS-based perturbational analysis.
Results
The radiomics model achieved Harrell concordance indices of 0.785 (95% confidence interval [CI], 0.767-0.793), 0.819 (95% CI, 0.788-0.833), and 0.785 (95% CI, 0.752-0.802) in the training, internal validation, and external validation cohorts, respectively. The high-risk phenotype showed exploratory enrichment of proliferative and stress-adaptive programs, led by HALLMARK_E2F_TARGETS (normalized enrichment score [NES] = 1.696, false discovery rate [FDR] = 0.045), with additional enrichment of MYC targets and inflammatory response. HALLMARK_HYPOXIA showed a positive directional trend but did not meet the prespecified exploratory threshold. GSVA provided complementary sample-level evidence for MYC targets, hypoxia, E2F targets, DNA repair, and cell-cycle activity. Eleven of 13 evaluated TFs met the prespecified FDR threshold, including YAP1, MYC, HIF1A, E2F3, STAT3, KLF4, E2F1, and TP53. Integrated analysis prioritized HIF1A, STAT3, TP53, YAP1, and E2F1/E2F3 as candidate regulatory nodes. Unsupervised clustering of all 110 radiogenomic patients, performed independently of radiomics risk labels, identified three exploratory TF-defined states: Hypoxia_Metabolic (n = 69), Cell_Cycle_E2F (n = 31), and Hippo_Mechanosensing (n = 10). LINCS-based analysis generated regulatory-state-associated compound hypotheses.
Conclusions
A CT radiomics-defined poor-prognosis phenotype in NSCLC is associated with coherent transcriptomic programs and an interpretable TF regulatory architecture. This framework links routine imaging phenotypes to latent tumor-control programs and supports regulatory-state-informed therapeutic hypothesis generation.