of AMPK and ETC genes, increased ribose sugar metabolism,
and demonstrated a profoundly worse outcome than metabolically
traditional chRCC.24 This metabolically divergent
subtype of chRCC lacked the typical chromosomal aberrations
of chRCC, and four of the six displayed sarcomatoid
de-differentiation. Taken as a whole, these landmark studies
contributed tremendously to our understanding of the molecular
biology of RCC, and the findings are driving the development
of molecularly-based prognostic models across
histological subtypes in RCC
Identification of Prognostic and Predictive Biomarkers
Successful prognostication of patients is critical to both the
clinical practice of oncology and research, as it influences
treatment approaches and may be used to stratify patients
in clinical trials. Classically, the most robust prognostic variables
were related to histology (subtype, grade, tumor
size), clinical features (performance status and pace of
disease), or laboratory parameters. For localized
disease treated with curative nephrectomy,
the two main prognostic models
are the UCLA Integrated Scoring System
(UISS) and the Mayo Clinic Stage, Size,
Grade, and Necrosis (SSIGN) score.54,55
The UISS score integrates pathological
(tumor size and nuclear grade) with clinical
(ECOG performance status) variables,
whereas the SSIGN score utilizes strictly
pathological variables. Despite the frequent
use of the aforementioned models
in patient prognostication, a recent analysis
has called into question the predictive
power of these nomograms.56
Somatic mutations in either
PBRM1 or BAP1 tend to be
mutually exclusive and activate
distinct gene ex pression
programs in tumors leading
to differentiated pa thological
features, ultimately causing
divergent clinical outcomes.
As such, these discoveries
resulted in the first mo lecular
classification of ccRCC.
Utilizing prospectively collected data from
the ASSURE trial,57 a randomized phase III
placebo-controlled trial of sunitinib and sorafenib in the adjuvant
setting, Correa and colleagues analyzed 8 recurrence
prediction models including the SSIGN and UISS tools.
They reported that the predictive power for each model’s
predefined outcome, as measured by the co-occurrence
index (C-index), fell considerably for all models in the dataset.
However, among these models, the SSIGN score performed
best with a C-index of 0.688 (95% CI: 0.686 –
0.689). Most other nomograms provided only marginal improvement
relative to the TNM staging system.56 Integrating
underlying genomic information with clinical and
histological variables is one strategy to improve these models.
We previously demonstrated that BAP1 and PBRM1 expression
by immunohistochemistry (IHC) are highly
correlated with mutation status (P = 3E-58 and 4E-23 respectively34
and that BAP1/PBRM1 expression by IHC are
independent predictors of both disease-free survival and
overall survival in the localized disease setting.31,34 PBRM1
and BAP1 expression status did not, however, add independent
prognostic information to the SSIGN score in a multivariate
analysis of nearly 1,500 cases of localized RCC.34
One possible explanation is that BAP1-mutant tumors are
significantly more likely to demonstrate higher nuclear
grade, stage, and necrosis than PBRM1-mutant tumors.34-37
Thus, despite progress, additional strategies to stratify
patients in the localized disease setting remain needed.
For advanced disease, the most commonly used prognostic
models are the Memorial Sloan Kettering Cancer
Center (MSKCC) risk model58 and the International Metastatic
RCC Database Consortium (IMDC) risk model.59 Both
98 Kidney Cancer Journal
models integrate clinical variables (time from diagnosis to
systemic treatment and performance status) and laboratory
variables including calcium and hemoglobin (as well as
lactate dehydrogenase in the MSKCC model vs platelet and
neutrophil counts in the IMDC model) but are agnostic to
the molecular biology of the underlying disease. Many of
the key driver mutations in RCC have been shown to have
negative (BAP1, SETD2, PTEN, and TP53) and positive
(PBRM1) prognostic significance,24,34,53,60 although it is important
to interpret the prognostic implications of these
mutations in the context of histology. For example, PBRM1
mutations which tend to associate with favorable prognosis
in ccRCC were strongly associated with reduced survival in
type I pRCC (p<0.0001).24 Regardless, retrospective analysis
of two phase III trials in metastatic RCC, COMPARZ (firstline
sunitinib vs pazopanib) and RECORD-3 (first-line sunitinib
vs everolimus) identified PBRM1, BAP1, and TP53 as
having independent prognostic value.40 Their prognostic
significance persisted after multivariate
analysis with the traditional variables in
the original MSKCC model. This resulted
in the development of a genomically-annotated
MSKCC risk model which stratifies
patients into four risk groups (favorable,
good, intermediate, and poor) versus three
in the original. The genomically-annotated
model had a more balanced distribution
across these groups and an improved ability
to predict overall survival. The C-index
of the original model was 0.567 (95% CI:
0.529 – 0.604) vs 0.637 (95% CI: 0.529 –
0.604) with the new model.40 While promising,
this model will need to be validated
in the prospective setting.
Prognostic models have also been developed
using gene expression data. One of the first such attempts
in ccRCC was by Brannon and colleagues who
collected transcriptomic data in 47 ccRCC primary tumor
samples. Unsupervised clustering analyses revealed two subgroups,
referred to as clear cell types A and B (ccA/B), which
had distinctly different clinical outcomes with a median
survival of 59 vs 36 months respectively in a validation cohort
of 177 patients (P=0.004).61 A 34 gene panel (Clearcode34),
which could delineate ccA and ccB subtypes, was
found to better predict relapse-free survival and cancer-specific
survival than both the UISS and SSIGN models in a cohort
of nearly 500 patients with localized RCC.62 Similar
findings were seen in a smaller metastatic cohort (54 patients),
however, only ccB classification added prognostic
value when incorporating the IMDC and MSKCC risk models.
63
More recently, cluster of cluster analysis including DNA
copy number, mRNA, microRNA, DNA methylation, and
protein expression data from the TCGA cohort revealed 9
unique clusters across all histological subtypes,53 three of
which were enriched in histologically defined ccRCC (e.1,
e.2, and e.3). These demonstrated significant differences in
survival, with e.3 demonstrating the poorest prognosis and
e.2 demonstrating the most favorable. Tumors in the e.3
cluster tended to have frequent loss of the CDKN2A gene,
frequent BAP1 mutations, and overexpression of cell cycle
and hypoxia-related genes.53 Additionally, e.3 was enriched
in inflammatory gene expression relative to e.2, and expression
of PDCD1 (encoding for PD1) and CTLA4 were independent
predictors of poor prognosis across the ccRCC