The Genomically Annotated MSKCC Model
The original MSKCC model was modified based on the mutational status of the 3 genes. To the original five clinical and laboratory data the researchers added two features that queried the genomics: one additional point was added for the presence of BAP1 or TP53 mutation or both. The mutational status of PBRM1 was then queried: if PBRM1 is wild-type or if the patient has concurrent mutations in all 3 genes, another point is added.3
With the genomically annotated model, the highest possible score is 7. In the new model, patients are grouped into 4 risk categories: favorable (score: 0); good risk (score: 1); intermediate risk (score: 2); and poor risk (score: 3 or higher).
Patients in COMPARZ were randomized to be treated with sunitinib or pazopanib and OS outcomes from the original study were available for review. Based on the new risk categories, median OS was not reached for patients in the favorable- and good-risk categories, was 30.6 months and 17.4 months for those in the intermediate- and poor-risk categories, respectively.3 On statistical modeling, the genomically annotated model showed a better fit for OS in the training cohort compared with the original MSKCC model.3
When the researchers compared the redistributed risk groups with the original risk model, they noted that of the 61% of patients in the original intermediate-risk group (median OS: 26.6 months), 18% were reclassified into the good-risk group in the genomically annotated model (median OS: 35.5 months) and 38% were reclassified into the poor-risk group (median OS: 16.5 months vs 18.1 months for the same category in the original model).3
The annotated genomic risk stratification model was validated using complete datasets from patients in the RECORD-3 study, during which, patients were treated with either sunitinib or everolimus.3
Robert Uzzo, MD, chair of the department of surgical oncology, Fox Chase Cancer Center, Philadelphia, Pennsylvania, told Cancer Therapy Advisor that being able to predict a patient’s response to therapy and the individual’s risk for progression or death are essential in cancer therapy. “For patients with advanced, stage IV kidney cancer, most models used to predict these outcomes are poor,” he said.
He noted that the MSKCC model — one of the most frequently used models — has a relatively poor predictive value and pointed out that even in this study , the concordance-index (C-index) for OS, a measure of how accurate the model can be — was 0.59. “It’s the same as the flip of a coin, and demonstrates that the currently used models to predict risks in advanced RCC are poor,” he said.
Dr Uzzo noted that a strength of the study is that the inclusion of mutational status of the 3 genes makes the model perform better (ie, the model has better prognostic value). “But even though it increases the performance of the model and is statistically significant, it is not really clinically meaningful,” he said and noted that the C-value for OS with the new model was only 0.63.
Dr Kohli agreed. “The marginal increase in the concordance index after incorporating genomic variables with the clinical factors tells us that the model can be improved further by including hitherto unknown variables,” he said.
Dr Voss concurred as well. Not having a high C-index speaks to the heterogeneity of the disease, he pointed out. “It is 1 aspect of the model that requires to be reported,” he said. But he explained that there are other aspects of this model that make it clinically relevant.
“The uneven distribution of patients across the 3 risk categories in the old model was its Achilles heel. This model provides a better and more accurate distribution of patients across 4 risk categories,” he said.