36 Clear Cell Renal Cell Carcinoma Consensus Transcriptomic Programs Reveal Converging Trajectories Towards Aggressive Disease

Abstract

Abstract Background Clear cell renal cell carcinoma (ccRCC) is characterized by a branching genomic architecture in which biallelic VHL inactivation is followed by divergence into PBRM1- or BAP1-mutant lineages, with subsequent acquisition of additional alterations contributing to progression and aggressive behavior. However, driver mutations alone incompletely explain the molecular and phenotypic heterogeneity of ccRCC. Approximately 40% of cases lack a detectable BAP1 or PBRM1 mutation, and tumors with similar mutation profiles can exhibit substantial variation in microenvironment composition and clinical behavior. Transcriptomic profiling offers a complementary lens for capturing tumor phenotype, but existing ccRCC subtyping frameworks are often confounded by tumor microenvironment (TME) composition and applied as discrete classifiers despite substantial intratumoral heterogeneity. We sought to develop a framework that defines reproducible ccRCC transcriptomic programs, separates tumor-intrinsic from extrinsic signals, and quantifies their continuous usage across tumors to relate gene expression to genotype, histology, and clinical behavior. Methods We developed a multi-cohort, multi-rank consensus workflow that identifies recurrent non-negative matrix factorization (NMF) factors, termed consensus transcriptomic programs (CTPs), across independent ccRCC bulk RNA-seq datasets and across a range of factorization dimensionalities. CTP usage in individual tumors was scored against a gene-wise permuted null distribution to enable cross-dataset comparison and statistical testing. The workflow was applied to three ccRCC datasets (IMmotion151, n = 823; JAVELIN Renal 101, n = 726; TCGA, n = 614) and validated in two additional cohorts (TRACERx Renal, CheckMate 025/010/009). CTPs were mapped onto single-cell RNAseq of human ccRCC tumors and matched normal kidney, and onto 131 RCC patient-derived tumorgraft lines, enabling separation of tumor-cell-intrinsic, tumor-cell-extrinsic, and mixed programs. Trajectory inference was performed by applying diffusion mapping to tumor-intrinsic CTPscores, yielding a continuous pseudotime axis. Trajectory and transcriptomic stage (TS) were assigned to each tumor and tested for associations with driver alterations, histopathology, and clinical outcomes. Intra-tumoral validation was performed using Visium spatial transcriptomics of paired conventional clear cell and sarcomatoid regions, and multiregional DNA/RNA sequencing from TRACERx Renal. Results We defined 17 CTPs, distinguishing tumor-cell-intrinsic (n = 6), tumor-cell-extrinsic (n = 6), and mixed (n = 5) programs. Tumor-cell-intrinsic CTPs were associated with canonical ccRCC genetic drivers, including VHL (R1), PBRM1/KDM5C (R2), BAP1 (R4), PTEN/TSC1 (R3), and CDKN2A/TP53 (MP-Prolif), and two programs associated with non-clear cell histologies (R5: TFE3/TFEB fusions; R6: NF2 mutations). Diffusion mapping of tumor-intrinsic CTPscores revealed two branching trajectories, PBRM1-like (R2 > R4) and BAP1-like (R4 > R2), that diverged at an Intermediate state and converged on a shared aggressive Late state characterized by R3 and MP-Prolif utilization. The pseudotime axis defined a transcriptomic stage (TS) associated with stepwise increases in Fuhrman nuclear grade, driver alteration burden, whole-genome instability, myeloid and stromal infiltration, and poor clinical outcomes. Spatial transcriptomic analysis of paired conventional clear cell and sarcomatoid regions revealed a shift from predominantly Intermediate TS in conventional regions to Late TS in sarcomatoid regions (p < 0.001), with trajectory assignments largely homogeneous within tumors. Multiregional sequencing in TRACERx demonstrated concordant trajectory across regions in 66% of patients and stepwise increases in driver burden and genomic instability across TS, in some cases aligning with acquisition of private alterations such as 9p (CDKN2A) deletion or TSC1 mutations. TS remained independently prognostic in TCGA after adjustment for stage, grade, and BAP1/PBRM1 status (HR 3.47, 95% CI 1.87–6.43 for Late vs. Early). In IMmotion151 and JAVELIN Renal 101, R1 and MP-Prolif utilization interacted with treatment arm: R1-utilizing tumors derived limited benefit from immune checkpoint inhibitor (ICI)/VEGF combinations relative to sunitinib monotherapy, whereas MP-Prolif-utilizing tumors exhibited greater relative benefit from combination therapy. Conclusions We present an atlas of recurring transcriptomic programs in ccRCC and a framework that bridges genotype, tumor-cell-intrinsic gene expression, microenvironment remodeling, and clinical outcome. Two findings have particular translational relevance. First, TS provides a quantitative axis that may refine risk stratification in localized disease beyond grade and stage, with potential application in adjuvant treatment decisions. Second, individual tumor-intrinsic CTPs (R1, MP-Prolif) interact with treatment arm in two phase III trials, identifying candidate predictive biomarkers that may have been obscured within composite transcriptomic subtypes. Beyond these clinical applications, the framework also offers a parsimonious explanation for discrepant reports linking PBRM1 mutations to either angiogenic or inflamed microenvironments by positioning TS as a hidden stratifier of genotype-TME associations. The atlas and analytical tools are disseminated as the rC3TP R package, enabling reproducible CTP scoring, trajectory assignment, and TS calling in user-supplied datasets.