Papillary renal cell carcinoma (pRCC) is the second most common subtype of kidney cancer after clear cell RCC, accounting for approximately 15 percent of all renal malignancies. Despite its significant prevalence, robust and validated tools for predicting recurrence risk after surgical resection are lacking, leaving oncologists without reliable instruments for post-surgical surveillance intensity or adjuvant therapy selection decisions.
Existing prognostic approaches for pRCC rely primarily on clinicopathological variables such as tumor stage and histological grade. These variables capture important but incomplete information about the underlying molecular biology of individual tumors, resulting in heterogeneous outcomes within the same clinical stage category that cannot be resolved by staging systems alone.
Advances in molecular profiling, digital pathology, and machine learning have created opportunities to develop multi-dimensional prognostic tools that integrate complementary data types. A system combining molecular biomarkers derived from RNA expression, deep learning applied to histological images, and standard clinical variables could achieve substantially better prognostic discrimination than any single approach.
A total of 793 pRCC patients were distributed across a training cohort, an independent validation cohort, and a TCGA (The Cancer Genome Atlas) validation cohort for model development and assessment. Three independent prognostic classifiers were developed using distinct data modalities: a four-lncRNA molecular classifier, a deep learning whole-slide image classifier, and a clinicopathological classifier based on histological grade and pathological stage.
The molecular classifier identified four long non-coding RNAs (lncRNAs) with strong independent prognostic associations through univariate and LASSO Cox regression: CYTOR, LUCAT1, AC099850.3, and lnc-TRDMT1-5. Each patient received a lncRNA risk score calculated from the weighted sum of expression levels of these four lncRNAs according to their LASSO model coefficients.
The digital pathology classifier was built using MobileNetV3, a lightweight convolutional neural network optimized for efficient image classification. Whole-slide hematoxylin and eosin images from pRCC resection specimens were divided into tiles, and the neural network was trained to predict recurrence risk from histological texture patterns visible in tumor tissue, capturing morphological information not represented in discrete pathological grade categories.
The three independent classifiers were combined into a unified multi-classifier scoring system by assigning each patient a composite score integrating risk scores from all three modalities. Patients were stratified into risk tiers based on their combined scores, enabling finer prognostic discrimination than possible with any individual classifier. The C-index, a measure of discriminatory ability analogous to AUC for survival data, was the primary performance metric.
The multi-classifier system achieved C-index values ranging from 0.831 to 0.858 across the training and validation cohorts. Individual single classifiers achieved substantially lower C-indices ranging from 0.642 to 0.777, demonstrating that integration of complementary data types provides a synergistic improvement in predictive accuracy that is clinically meaningful.
Kaplan-Meier survival analysis confirmed that the multi-classifier system stratified pRCC patients into groups with significantly different recurrence-free survival. The separation between risk groups was substantially wider and more statistically robust for the multi-classifier compared to each individual component classifier, supporting the value of combining molecular, imaging, and clinical data types in a unified prognostic framework.
An intriguing finding emerged from analysis of CpG island methylator phenotype (CIMP) tumors within the pRCC cohort. CIMP-positive tumors, defined by widespread promoter hypermethylation across multiple gene loci, represent a molecularly distinct subset of pRCC with particularly aggressive biology and poor clinical outcomes compared to non-CIMP pRCC.
All CIMP-positive pRCC tumors in the cohort were assigned ultra-high multi-classifier risk scores exceeding 14.0, placing them uniformly in the highest risk tier of the scoring system. This finding demonstrates that the multi-classifier system captures a biologically meaningful dimension of tumor aggressiveness that correlates with epigenetic alterations, suggesting the lncRNA and image-based classifiers encode information correlated with CIMP status.
The consistent identification of CIMP tumors as ultra-high risk by the multi-classifier system provides a biologically grounded validation of the scoring approach and suggests potential utility in identifying this molecularly aggressive pRCC subgroup for whom standard surveillance and treatment protocols may be insufficient without augmented management strategies.
The multi-classifier system was independently validated in two separate patient cohorts distinct from the training dataset, addressing the critical limitation of single-cohort model development. Performance metrics in the independent validation cohort and the TCGA cohort confirmed maintenance of discriminatory ability consistent with training cohort results, supporting the generalizability of the approach to patients from different institutions and treatment eras.
The four-lncRNA molecular classifier alone achieved C-index values ranging from 0.700 to 0.777, representing meaningful prognostic information from RNA expression data alone. When combined with MobileNetV3 digital pathology features and clinicopathological variables, the incremental gain in C-index confirms that each data modality contributes non-redundant prognostic information to the composite score.
Univariate and multivariate Cox regression confirmed that the multi-classifier risk score was an independent predictor of recurrence-free survival after adjustment for tumor stage, grade, patient age, and sex. This independence from conventional prognosticators is a prerequisite for clinical utility and distinguishes the multi-classifier from tools that merely recapitulate established staging information.
This study establishes proof of concept that multi-modal prognostic classifiers integrating molecular, digital pathology, and clinical data can substantially improve recurrence prediction in papillary RCC compared to single-modality approaches. The multi-classifier C-index of up to 0.858 represents clinically meaningful discriminatory performance that could inform post-surgical surveillance intensity and adjuvant therapy decisions.
The identification of four prognostic lncRNAs as components of the molecular classifier adds to the growing body of evidence supporting non-coding RNA biology as a source of clinically actionable biomarkers in renal malignancies. The specific lncRNAs CYTOR and LUCAT1 have previously been implicated in cancer biology, providing biological plausibility for their inclusion in the prognostic model.
Prospective clinical implementation of the multi-classifier system would require development of standardized RNA extraction and expression profiling protocols compatible with formalin-fixed paraffin-embedded surgical specimens, as well as validation of the MobileNetV3 digital pathology model across diverse scanning platforms and staining protocols. These translational steps are essential before the tool could be deployed in routine clinical practice.