Computational Methods for Integrative Cancer Data Analysis

Background: Gliomas are characterized by a high degree of molecular heterogeneity, which impairs the reproducibility of predictive biomarkers derived from bulk-based molecular profiling due to immune/stromal contamination of tumors and the high prevalence of the IDH mutation signature. Methods: In this study, we used MOFA+ to derive intrinsic molecular signatures from transcriptional, methylation, and genomic profiles of a cohort of 667 diffuse gliomas in the Cancer Genome Atlas database. Thereafter, factor scores were derived for two separate Chinese Glioma Genome Atlas batches (Batch 1, n = 325; Batch 2, n = 693) without any retraining on the model. The prognostic independence of identified molecular signatures was assessed using multivariable Cox regression adjusted for IDH mutation status and tumor purity; purity-residualized survival analyses; IDH-stratified Cox regression in each cohort; validation by concordance index against established molecular signatures; and survival extreme profiling. To characterize the biological significance of factor signatures, we projected gene set signatures corresponding to each factor signature onto a single-cell RNA-seq dataset of GBM (GSE131928). Results: MOFA+ identified 12 latent factors, of which a vascular–extracellular matrix (ECM) remodeling axis (Factor 1) explained the highest multi-omics variance (24.9%) and was the strongest independent prognostic factor. In multivariable Cox regression adjusting for IDH status and tumor purity, Factor 1 remained independently prognostic (HR = 1.67, 95% CI 1.27–2.20, p = 0.0002); in a fully-adjusted model additionally including age, WHO grade, MGMT methylation, and 1p/19q codeletion (plus radiotherapy and chemotherapy status in the CGGA cohorts), Factor 1 remained prognostic in both CGGA cohorts (CGGA1: HR = 1.50, p = 3.8 × 10⁻⁵; CGGA2: HR = 1.18, p = 0.003) but lost significance in TCGA (HR = 1.04, p = 0.83), consistent with the cohort-dependent magnitude reported in the IDH-stratified and meta-regression analyses below. Purity-residualized survival analysis showed negligible attenuation of the Factor 1 signal (raw HR = 3.57 vs. residualized HR = 3.72; concordance 96.5%). Within IDH-wildtype gliomas, Factor 1 was significant in both external validation cohorts (CGGA1: HR = 1.64, FDR = 4.6 × 10⁻⁶; CGGA2: HR = 1.20, FDR = 0.02), though the TCGA IDH-wildtype subgroup showed a trend that did not survive FDR correction (FDR = 0.060). All validation was performed without model retraining. Within IDH-mutant gliomas, Factor 1 was strongly prognostic in both CGGA cohorts but was not significant in TCGA (HR = 1.17, FDR = 0.33). These findings should therefore be interpreted as consistent in directionality across cohorts but not uniformly replicated at the FDR-adjusted significance threshold in the TCGA discovery dataset. Concordance index benchmarking on a matched subset (n = 503) showed Factor 1 achieved discrimination comparable to the Mesenchymal signature (C = 0.797 vs. 0.801; ΔC = −0.004) while outperforming four other established classifiers. Factor 1 consistently separated patients with extreme survival phenotypes (OS < 6 vs. >15 months) across all three cohorts (all log-rank p < 0.001). Projection onto a single-cell GBM atlas (GSE131928), supported by inferCNV-based malignant-cell classification, localized the Vascular–ECM program to malignant cells and the Immune–ECM axis to myeloid compartments. Conclusions: The Vascular–ECM axis is a consistent, prognostic program robust to purity adjustment for diffuse gliomas that remains relevant across IDH-defined subgroups in three independent datasets comprising 1685 patients. The Vascular–ECM axis is a reproducible, purity-robust prognostic program in diffuse glioma, with directionally consistent adverse effects across TCGA, CGGA Batch 1, and CGGA Batch 2 (pooled n = 1685). Given the strong co-loading of endothelial, ECM, and myeloid genes observed in the single-cell projection, Factor 1 is best interpreted as a vascular/ECM-associated tumor–microenvironment ecosystem program rather than a malignant-cell-autonomous signature. Its FDR-adjusted significance within IDH-stratified subgroups is cohort-dependent and robust in both CGGA cohorts but attenuated in the TCGA IDH-wildtype (FDR = 0.060) and TCGA IDH-mutant (FDR = 0.33) strata. The pooled signal should therefore be interpreted as evidence of a generalizable biological program rather than a uniformly replicated subgroup-specific biomarker. It is possible to calculate factor scores based on RNA sequencing alone using fixed loadings (Z = XWᵀ), which may have implications for future translational applications. All findings are correlative; a causal role for the Vascular–ECM program in glioma progression, invasion, or therapy resistance remains to be established through functional perturbation experiments. Full article

Background/Objectives: AP-2δ, encoded by TFAP2D, is one of the least characterized members of the AP-2 transcription factor family, although available evidence suggests biologically relevant roles in lung cancer that have not yet been thoroughly examined. The aim of the present study was to provide an integrated characterization of AP-2δ/TFAP2D in lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). Methods: LUAD and LUSC data were obtained from The Cancer Genome Atlas. The analysis comprised the expression profiling of AP-2δ target genes, survival-guided TFAP2D stratification, clinical profiling, differential expression and intersection analyses, methylation-derived chromatin compartment profiling, TFAP2D-associated cofactor rewiring, and genome-wide enrichment of AP-2δ targets. In parallel, pocket prioritization was performed using an AlphaFold model of AP-2δ with cross-tool consensus mapping. Results: TFAP2D stratification delineated biologically-distinct states in both histological subtypes (LUAD and LUSC). AP-2δ target genes showed subtype-specific expression patterns and functional organization. The consistent survival association was observed for progression-free interval rather than uniformly across all endpoints. Clinical profiling was more closely associated with molecular subtype composition than broad clinicopathological differences. Differential expression analyses identified both shared and histology-dependent programs associated with TFAP2D. In the chromatin-compartment analysis, LUSC showed a broader and more coherent footprint, whereas LUAD displayed more selective cofactor rewiring. Structure-based analysis prioritized a small set of reproducible candidate pockets concentrated within ordered regions of the TF_AP-2 domain. Conclusions: AP-2δ marks biologically meaningful but histologically non-uniform regulatory states in lung cancer. These findings provide an integrated framework for understanding TFAP2D-dependent regulation in LUAD and LUSC, highlighting AP-2δ as a candidate for future mechanistic and translational investigation. Full article