Development and cross-cohort evaluation of a DLX4-associated epithelial-mesenchymal transition pathomics signature for overall survival in clear cell renal cell carcinoma

Background: Clear cell renal cell carcinoma (ccRCC) is biologically heterogeneous, and epithelial-mesenchymal plasticity contributes to invasion and treatment resistance. We aimed to develop a DLX4-anchored pathomics model for overall survival (OS) and to relate routine hematoxylin and eosin morphology to malignant epithelial and spatial states.

Methods: This retrospective secondary analysis used public TCGA-KIRC and CPTAC ccRCC data accessed and frozen in June 2026. Patients with ccRCC, positive OS time, known vital status, and matched whole-slide-image pathomics features were eligible. TCGA was used for model fitting (506 patients; 167 deaths). CPTAC (94 patients; 20 deaths) contributed to cross-cohort feature confirmation and algorithm ranking and therefore was not an untouched external-validation cohort. EMT genes were screened using dbEMT and LASSO-Cox. Direction-consistent DLX4-associated features were filtered in TCGA and evaluated in a 101-style survival machine-learning library. The final eight-component partial least squares Cox (plsRcox) model used 705 standardized predictors. Bulk, single-cell, copy-number, trajectory, and spatial analyses were used for biological interpretation.

Results: LASSO-Cox retained 19 EMT genes, and DLX4 was prioritized as an adverse-prognosis anchor. Of 1,192 direction-consistent DLX4-pathomics features, 705 passed TCGA univariable Cox filtering. Of 117 prespecified parameterized jobs, 63 generated finite paired TCGA-CPTAC risk scores and were evaluable. The selected plsRcox model yielded an apparent TCGA C-index of 0.877 (95% confidence interval [CI], 0.853-0.900) and a CPTAC cross-cohort C-index of 0.775 (95% CI, 0.617-0.915). The fixed TCGA median cutoff separated CPTAC low-risk (44 patients; 4 deaths) and high-risk (50 patients; 16 deaths) groups (log-rank P=0.0078). Adding the pathomics score to age, stage, and grade increased the C-index in TCGA by 0.121 (95% CI, 0.089-0.158), but this increment was not reproduced in CPTAC (-0.019; 95% CI, -0.184 to 0.118).

Conclusions: The DLX4-associated pathomics score was associated with OS and with malignant epithelial and spatial remodeling in ccRCC. Because CPTAC influenced feature confirmation and algorithm ranking, its performance is a cross-cohort evaluation rather than unbiased external validation. The model and its incremental clinical value require locked, independent multicenter evaluation before clinical use.