Not All Predictors Of RCC Clusters Are The Same: An individual sample predictive model to classify patients with metastatic renal cell carcinoma
Clear cell renal cell carcinoma (RCC) patients have multiple approved therapies, including anti-angiogenesis tyrosine kinase inhibitors (TKIs) and immuno-oncology therapies (IO), but lack a clinically validated biomarker. Using RNAseq data from the IMmotion 151 clinical trial (IM151)1-3, 7 RCC biologic clusters have been defined3. Several groups have attempted to predict these clusters on various RCC datasets4,5 with negative results, suggesting that the association with therapy response from IM151 could not be reproduced. We hypothesized that the specific approaches used to generate cluster predictions led to the misinterpretation of findings. Both published models used standardization (z-scores) to normalize the data within their patient cohorts, imposing an expected gene expression distribution in which [~]50% of patients have apparently higher-than-average expression, artificially impacting the proportion of cluster assignments and leading to potential misclassification. We developed a machine learning (ML) model, IRIS-RCC (Individual RNA-seq Intrinsic Subtyping for RCC), to predict treatment (TKI vs. IO) for patients using an individual-sample model using the IM151 trial (N=823) and validated the model on the JAVELIN Renal 101 trial (JR101; N=726)3,6. Our method normalizes gene expression within a given sample using ratiometric expression. This method results in different cluster assignments for individual tumors and distinct clinical correlations. An additional advantage of individual-sample predictions is that they can be readily applied in a prospective setting, where patients must be classified one at a time. IRIS-RCC is currently being validated in a prospective biomarker-driven Phase II clinical trial (OPTIC RCC).