Researchers at Vanderbilt University Medical Center and Vindhya Data Science have developed IRIS-RCC (Individual RNA-seq Intrinsic Subtyping for RCC), a machine learning model designed to classify patients with metastatic clear cell renal cell carcinoma (ccRCC) into molecularly distinct subgroups using RNA sequencing data. The model, described in a brief communication published in npj Precision Oncology, was built on data from the IMmotion151 phase 3 trial and validated in the JAVELIN Renal 101 (JR101) dataset.
In cross-validation, IRIS-RCC achieved 87% overall accuracy in classifying tumors into three functional cluster groups, and validation in the JR101 cohort demonstrated that predicted cluster assignments were associated with differential outcomes on targeted and immunotherapy (IO) regimens. The authors caution that the analysis is retrospective and that prospective validation remains necessary before the model could inform clinical treatment decisions.
Key finding
In a retrospective validation using the JAVELIN Renal 101 dataset, metastatic ccRCC patients classified by IRIS-RCC as Cluster IO had significantly improved progression-free survival (PFS) with avelumab plus axitinib compared with sunitinib (log-rank p = 0.0001), while patients in Cluster Angio showed similar PFS on either regimen, though prospective validation is still required.
Study Design
| Model development dataset | IMmotion151 (IM151) phase 3 trial — RNAseq data from metastatic ccRCC tumor samples |
|---|---|
| Validation dataset | JAVELIN Renal 101 (JR101) — metastatic ccRCC patients treated with avelumab plus axitinib or sunitinib |
| Population | Patients with previously untreated metastatic ccRCC |
| Model approach | Random forest classifier trained on intra-sample gene expression ratios derived from 188 genes; three functional cluster groups: Cluster Angio (Clusters 1+2), Cluster IO (Clusters 4+5), and Cluster Exclude (Clusters 3+6) |
| Primary performance metric | Cross-validation accuracy, sensitivity, and positive predictive value (PPV) across 50 repeated bootstrapped cross-validation runs |
The model was developed to address a methodological limitation identified in prior classification attempts, specifically, the use of z-score normalization, which centers gene expression values relative to a given cohort rather than within an individual sample. The authors contend that this cohort-dependency makes z-score-based classifiers unreliable for real-world clinical deployment, where the size and composition of a patient population cannot be controlled. IRIS-RCC instead normalizes gene expression using within-patient gene ratios, enabling deterministic classification of individual samples independent of cohort context. The tool is freely available for academic use at irisrcc.vindhyadatascience.com.
Results and Safety
IRIS-RCC demonstrated 87% overall cross-validation accuracy on the IM151 training data, with sensitivity and PPV exceeding 80% for all three cluster groups. Cluster Angio, which represented approximately two-thirds of the training dataset, yielded the highest individual metrics at 95% sensitivity and 91% precision. Validation in the JR101 cohort showed that IRIS-RCC cluster assignments reproduced expected associations with independent biomarkers: Cluster IO patients had significantly higher CD8+ and PD-L1+ immunohistochemistry levels than Cluster Angio and Cluster Exclude patients (Kruskal–Wallis p < 0.001 for both), and mutation patterns, including enrichment of PBRM1 and KDM5C mutations in Cluster Angio and TP53 and BAP1 mutations in Cluster IO, were consistent with those observed in the original IM151 study.
In the JR101 validation cohort, IRIS-RCC-predicted Cluster IO patients treated with avelumab plus axitinib had superior PFS compared with those treated with sunitinib (log-rank p = 0.0001), while Cluster Angio patients had similar PFS regardless of treatment arm (log-rank p = 0.00006 favoring sunitinib in that arm). The authors compared these associations with those reported by Saliby et al., who applied a z-score-based random forest model to the same JR101 cohort. Both approaches identified similar directional trends, but IRIS-RCC produced larger hazard ratio differences between cluster groups; the Saliby et al. model showed only a modest association between Cluster IO and improved PFS in the avelumab plus axitinib arm (p = 0.03), compared with p = 0.0001 for IRIS-RCC, despite similar predicted Cluster IO prevalence (~25%) in both models. Safety data were not reported in this analysis.
The authors also evaluated the template-based classification method used in the BIONIKK trial, which relies on z-scored expression data projected onto cluster templates. When recapitulated and tested against IM151 as a reference standard, the four-template method achieved only 52% overall accuracy, rising to 64% with a six-template approach, compared with 87% for IRIS-RCC. In a theoretical experiment testing classification when all samples belonged to a single cluster, a scenario that exposes the distributional assumptions of z-scoring, the template methods performed near chance (approximately 17–18% accuracy), while IRIS-RCC was unaffected.
Clinical Context
Molecular subtyping of metastatic ccRCC has been an area of active investigation since the IMmotion151 trial identified seven transcriptionally distinct tumor clusters with differential associations with anti-angiogenic and IO-based regimens. However, attempts to translate these cluster assignments to other trial cohorts, including JAVELIN Renal 101 and BIONIKK, have produced inconsistent or counterintuitive findings. The current analysis suggests that methodological choices in gene expression normalization, particularly the use of cohort-level z-scoring, may account for a meaningful portion of this variability. By demonstrating that IRIS-RCC produces stronger and more consistent treatment-outcome associations than previously published approaches, the authors argue that individual-sample normalization should be a prerequisite for any classifier intended for clinical or prospective research use.
The authors also highlight an underappreciated practical concern: the majority of RNAseq samples in both JAVELIN Renal 101 and BIONIKK were derived from primary tumor tissue rather than metastatic sites. Data from the phase II OPTIC RCC and HCRN GU16-260 studies suggest that 38–54% of patients show discordant cluster assignments between primary and metastatic tumors, raising questions about the validity of treatment-outcome correlations derived from primary tissue in a metastatic disease context. The researchers acknowledge that IRIS-RCC requires prospective validation, noting that the ongoing OPTIC RCC trial is designed to address this need. A randomized prospective study assigning treatment based on molecular cluster, rather than correlating cluster assignments retrospectively, would be required to confirm clinical utility.
Sources
- Reddy A, Brewer WJ, Haake S, Rini B. “An individual sample predictive model to classify patients with metastatic renal cell carcinoma.” npj Precision Oncology. Article in Press. Accepted 28 September 2026. View source.
