the head and neck squamous cell carcinoma
showing the highest median value (2.63%), a
subset of melanoma samples could reach a
frequency of 12.5%. Further, although the
CD8+Temracell was underrepresented in the
tumor, a few tumor samples of lung cancer
and melanoma and a large fraction of renal
tumors showed much higher frequencies of
such cells (fig. S30). Thus, therapeutic strat-
egies targeting the above cell subtypes should
consider the variability across cancer types.
The tumor mutation burden (TMB) has
been associated with the efficacy of ICB ( 40 ).
By partitioning tumors into TMB-high and
TMB-low groups (fig. S31A), we found that
only the frequency of CD4+TFH/TH1 cells
showed a strong correlation with TMB (FDR <
0.001, PVE > 26.3%) (figs. S28F and S32A).
This association could be validated in pan-
cancer and multiple individual cancer types
using the bulk TCGA (The Cancer Genome
Atlas) data (fig. S32B). We also identified a
positive association betweenFAT1(fatty acid
translucase 1) mutations andTNFRSF9+Treg
cells (Fig. 4B and figs. S31B and S32C). Thus,
the T cell composition in the tumor could be
affected not only by the number of potential
neoantigens, reflected by TMB but also by
specific somatic mutations of cancer cells.
To reveal the overall pattern of T cell com-
positions across cancer types, we inspected
the frequency correlations among metaclus-
ters in the tumor and identified several highly
correlated modules of metaclusters (fig. S33A).
One module consisted of the three ISG+meta-
clusters, whereas the four CD8+Texcell popu-
lations, TFH/TH1cells,andTNFRSF9+Tregcells
formed another module. Metaclusters with sim-
ilar gene signatures but from different compart-
ments (CD8+Tc, CD4+TH, and CD4+Tregcells)
tended to cluster together (for example, CD8+
Tc17 and CD4+TH17 cells). The potentially tumor-
reactive metaclusters were negatively correlated
with certain metaclusters, which could be ex-
plained by several mechanisms, including the
dynamic state transitions between metaclusters.
For example, CD8+terminal Texmetacluster
was negatively correlated with CD8+ZNF683+
Trmand CD8+GZMK+Temmetaclusters (fig.
S33B) and showed an aforementioned state
transition relationship. The positively correlated
metaclusters usually had significant overlap be-
tween signature genes (P< 0.01, hypergeometric
tests)(Fig.4Candfig.S34),suggestingthat
the same regulators induced similar transcrip-
tional programs in different T cell populations.
Using those overlapped signature genes and
the NicheNet algorithm ( 41 ) to find shared
ligands, we found thatRORCand other type-
17 response-related genes were potentially
induced in both TH17 and Tc17 by ligands such
asIL23AandIL15(fig. S35A). We identified
shared ligands for the potentially tumor-
reactive metaclusters. For the ISG+metaclusters,
the type I interferon-encoding genes were
expectedly ranked as the top ligands by the
NicheNet analysis (fig. S35B). For terminal
Tex,TFH/TH1, andTNFRSF9+Tregcells, a total
of 325 shared signature genes were identified,
18 of which were TFs (for example,TOX,TOX2,
VDR,ZBED2,ETV7,ZNF282, andHIVEP1) (Fig.
4C and fig. S34). Thus, similar transcriptional
machinerieswerelikelyusedbythethreedif-
ferent cell populations to respond to TME
stimuli. Further, both terminal Texand TFH/
TH1 cells producedCXCL13and type 1 response
cytokines, andTGFB1wasinferredasoneof
the top potential ligands inducing their
shared signature genes (fig. S35C). Meanwhile,
TNFRSF9+Tregcells highly express TGF-b–
induced TFs, includingSOX4andTGIF2(table
S3). Moreover,IFNB1was inferred as one of
the top potential ligands inducing the shared
signature genes betweenTNFRSF9+Tregcells
and terminal Texor TFH/TH1 cells (fig. S35, D
and E). These observations suggested that
TGF-band interferons may affect the tran-
scriptional program and abundance of the
potentially tumor-reactive T cells.Immune types of pan-cancer defined by the
composition of T cells
Next, using the frequencies of these correlated
metaclusters, we found that tumor samples
could be clustered into eight groups (C1 to
C8) (Fig. 5A). The C1 and C2 harbored high
frequencies of terminal Texcells, and C1 also
had the highest frequency ofTNFRSF9+Treg
cells. Tumors of C3 to C8 harbored a low fre-
quency of terminal Texcells and high fre-
quency of CD8+ZNF683+CXCR6+Trmcells and
could be further divided into groups domi-
nated by naïve T cell (C7), enriched naïve T cell
(C8), enriched Temracell (C6), enriched Tc17 or
TH17 cell (C4), and with a low frequency of
TNFRSF9+Tregcell (C5), respectively. On the
basis of the linear model analysis, we found
that our grouping could explain more varia-
bilities of the T cell composition of the tumor
than other factors (figs. S28 and S36). Although
each immune type included mixed cancer types,
certain cancer types exhibited clear preferences
(fig. S37). For example, nearly half of esophageal
and nasopharyngeal carcinoma tumors were of
C1. By contrast, thyroid carcinoma and uterine
corpus endometrial carcinoma were enriched
in C3, suggesting that a large proportion of
these two cancer types with high T cell sup-
pression might still benefit from immuno-
therapybecauseofthepresenceofISG+
activating T cells and the low abundance of
terminal Texcells. More than half of mela-
nomas were of C2, which has high Texcell but
lowerTNFRSF9+Tregcell frequency, which is
consistent with their tendency to respond to
ICB. Both basal cell carcinoma and hepato-
cellular carcinoma were enriched in C4, indi-
cating that their tendencies are inflammatorythrough IL17-producing T cells. These T cell–
based immune types provide a reference to
understand the overall tumor-infiltrating T cell
properties, which may help guide the devel-
opment of newer therapies and patient strat-
ification instead of the conventional cancer
type metrics.
Such immune type classification may have
clinical implications. Using immune type sig-
natures to stratify the TCGA cancer patients,
we found that the TexloTrmhi(C3 to C8) tumors
had better overall survival than that of
TexhiTrmlotumors (C1 and C2) across cancer
types or in multiple individual cancer types,
including lung adenocarcinoma, hepatocellular
carcinoma, and renal papillary cell carcinoma
(Fig. 5B and fig. S38A). Because T cells are
the direct target for many immunotherapies,
the T cell–based immune types could logically
be associated with the treatment efficacies.
Reanalysis of published data of PD-1 anti-
body treatment for melanoma ( 42 ) indicated
that responsive tumors had a lower frequency
of terminal Texcells and a higher frequency
of naïve T cells (Fig. 5C and fig. S38B). The Tex
cell connection was reproduced in another
dataset ( 43 ), showing that the responder group
was enriched with more TexloTrmhitumors
than that of the nonresponders (Fisher’s exact
test,P= 0.025) (fig. S38C). Pretreatment tumors
in responders also had a higher frequency of
Tc17 (Fig. 5C), implicating an important role of
Tc17 in ICB treatment. Further investigation is
needed to reveal how this finding is tied to the
notion that Tc17 could also go into exhaustion
in the tumor.Discussion
We systematically characterized the T cells
from various human cancers, investigating
different aspects from gene expression signa-
ture and heterogeneity to state transitions
and regulations. Multiple tumor-enriched
metaclusters—including Tex,TFH/TH1, and
TNFRSF9+Tregcells—deserve particular atten-
tion because of their potential tumor reactivities.
Our analyses revealed diverse paths to T cell
exhaustion and the cancer type preference of
those paths (fig. S39). Such landscape depiction
deepens our understanding of cancer immu-
nity and will facilitate therapeutic development.
The T cell states and infiltration in tumors
are affected by multifaceted elements, such as
tumor-intrinsic and metabolic factors ( 8 ). In
our data, the TMB shows a positive association
with TFH/TH1 cells, whereas the BMI exhibits
a positive association with TFHcells. Because
both TMB ( 40 )andBMI( 44 ) have been
previously linked to ICB responses, our find-
ings highlight the importance of TFH-related
cell populations in the antitumor response.
Additionally, specific mutations could affect
T cell compositions in the TME.FAT1mutations
are positively correlated withTNFRSF9+TregcellZhenget al.,Science 374 , eabe6474 (2021) 17 December 2021 8 of 11
RESEARCH | RESEARCH ARTICLE