Kidney Cancer Journal 99
subtype.53 The e.2 subgroup, on the other hand, was found
to demonstrate significant overlap with the previously described
ccA subtype,53,61 which is distinguished by overexpression
of angiogenesis-related genes such as such as FLT4,
FLT1, and VEGFB.61,62,64 Thus, given that the most widely
utilized treatments in metastatic RCC are immune checkpoint
inhibitors, angiogenesis inhibitors, or combinations
of the two,4,7,65,66 transcriptomic information regarding
inflammation or angiogenesis may also have predictive
potential.
Unlike prognostic biomarkers which inform on the
spontaneous trajectory of the disease, predictive biomarkers
have the potential to inform on treatment selection, which
may be tailored to the particular patient. Beyond the obvious
benefit of selecting the most efficacious treatment,
predictive biomarkers have the potential to reduce adverse
events as well as cost from exposures to non-beneficial
drugs. The use of gene expression profiles to predict treatment
response was considered in an exploratory analysis of
the IMmotion150 trial, a phase 2 study comparing the combination
of atezolizumab (a PD-L1 inhibitor) and bevacizumab
(a neutralizing VEGF antibody) (atezo+bev) to
atezolizumab (atezo) or sunitinib monotherapy.67 McDermott
and colleagues showed that predefined gene signatures
of angiogenesis (Angio), immune T cell infiltration (Teff),
and myeloid inflammation (Myeloid) may have predictive
potential. They found that patients with the AngioHigh gene
signature had an improved progression-free survival (PFS)
compared to those with AngioLow with the angiogenesis inhibitor,
sunitinib.67 Interestingly, PBRM1 mutations were
found to be enriched in the AngioHigh subgroup. In line with
these findings, a similar analysis of the COMPARZ trial evaluating
two anti-angiogenic agents, sunitinib and pazopanib,
found that high expression of angiogenesis genes was
associated with improved response rates.64 Furthermore, tumors
with PBRM1 mutations demonstrated significantly
higher angiogenesis gene expression scores than those with
BAP1 mutations.64 In the IMmotion150 trial, patients with
the Teff
High gene signature had improved PFS with atezo+bev
compared to the sunitinib and the atezo arms. Intriguingly,
when examining the effect of myeloid signature on outcomes
within the Teff
High population, a PFS advantage of
atezo+bev over atezo was observed in the Teff
High /Myeloid-
High /MyeloidLow subgroup
High subgroup (n=66) but not in the Teff
(n=66). This finding may identify a particular
population of patients which would benefit from combination
ICI and TKI therapy. However, this is a relatively small
study and has not yet been reproduced in larger datasets.
These studies highlight the potential role of the tumor
microenvironment (TME) on clinical outcomes. Bioinformatics
techniques can be leveraged to gather information
about the TME from RNA extracted from tumors. One such
approach is to utilize single sample gene set enrichment
analysis (ssGSEA) in which signature gene panels attributed
to particular cell types are utilized to disentangle a heterogeneous
tumor.68 Our group expanded upon this method
by exploiting patient-derived tumorgrafts (patient tumors
implanted in mice), where the human TME is ultimately replaced
by the host.69 By focusing on human RNA and subtracting
from the patient tumor the transcriptome of the
tumorgraft, one is left with the human TME transcriptome.
Utilizing this dissection algorithm and a cutoff of 20-fold
to distinguish TME vs tumor genes, an empirically derived
TME (eTME) signature was obtained. Clustering analyses of
the ccRCC TCGA (KIRC) tumors according to the eTME revealed
two subgroups, an inflamed subtype (IS) and a noninflamed
subtype (NIS). Interestingly, the IS was enriched
for BAP1 mutations (P = 7.7E-5) and demonstrated a worse
prognosis compared to NIS. Furthermore, the eTME-IS subtype
correlated with systemic inflammatory markers such
as elevated platelet counts and decreased hemoglobin levels,
as well as worse prognosis in three distinct cohorts totaling
approximately 1,000 patients. This correlation draws a link
between inflammatory subtypes of RCC and key prognostic
variables in IMDC or MSKCC models. Notably, the presence
of such variables is associated with intermediate/poor risk
disease, which suggests that inflammatory tumors are particularly
aggressive in patients. Interestingly, in the Checkmate
214 trial comparing the combination of nivolumab
and ipilimumab to sunitinib, the intermediate/poor risk
groups appeared to derive the most benefit from ICI,65
which is consistent with the predictive potential of the IS
subtype. Thus, the combined use of NGS and bioinformatics
has potential to predict responses to therapy and is an active
area of investigation.70
Intratumoral Heterogeneity
One important disadvantage of the aforementioned studies
is the use of a limited number of tumor samples per patient
(typically just one), as this may fail to capture intratumoral
heterogeneity (ITH). In one of the earliest attempts to measure
ITH in ccRCC, Gerlinger and colleagues performed
multiregional WES of two primary tumors and demonstrated
that only 31% of the somatic mutations were ubiquitous
amongst all sampled regions.71 Furthermore, the previously
described “cc-A” and “cc-B” gene expression patterns could
be identified in spatially distinct areas of an individual patient’s
tumor, highlighting how ITH can confound efforts
to establish effective prognostic models based on analyses
of a single sample.71,72 This laid the groundwork for the
TRAcking Cancer Evolution through therapy (Rx) (TRACERx)
consortium, which prospectively collects tumor samples
and performs multiregional sequencing, when possible,
over time.73 In a recently published report, Turajlic and colleagues
sequenced 1,206 primary tumor samples from 101
patients with the use of a 110 gene panel, allowing for an
unprecedented view of the molecular diversity within a single
tumor.74 Multiregional sampling allowed for the detection
of clonal and subclonal somatic mutations. Thus,
the prevalence of PBRM1 (55%), SETD2 (25%), BAP1 (19%)
and other driver mutations could be more accurately calculated.
74 Interestingly, while BAP1 and PBRM1 mutations
could be identified in the same tumor, they were typically
located in spatially distinct regions, consistent with previous
reports that these mutations anticorrelate with one
another and are found in different areas of the same
tumors.29,32,34,74
Genomic data obtained from spatially distinct regions
provided the ability to assess the timing of mutations, and
thus patterns of tumor evolution could be inferred (Figure
2). In the TRACERx studies, seven distinct patterns could
be identified utilizing rule-based clustering. However, 36.6%
(37/101) of cases could not be assigned an evolutionary subtype.
Subtypes were assessed for nuclear grade, stage, microvascular
invasion, genomic instability, and degree of ITH.
The most aggressive subtype, based on the aforementioned
parameters, was the “multiple clonal drivers” subtype,
which contained truncal aberrations in two or more of the
following: BAP1, PBRM1, SETD2, or PTEN.74 It is important
to note that in the “multiple driver mutations” subtype, the
temporal relationships of driver mutations were indistinguishable,
which separates this group from subsets defined