Trends in

OPEN ACCESS Cancer

Review

Transcriptomic subtyping of gastrointestinal

malignancies
Tim R. de Back 1,2,3, Sander R. van Hooff 1,2,3, Dirkje W. Sommeijer 4,5, and Louis Vermeulen 1,2,3,5,*

Gastrointestinal (GI) cancers are highly heterogeneous at multiple levels. Tumor Highlights
heterogeneity can be captured by molecular profiling, such as genetic, epige- Epithelial, metabolic, immune, and mes-
netic, proteomic, and transcriptomic classification. Transcriptomic subtyping enchymal transcriptomic subtypes from
unsupervised clustering appear to be
has the advantage of combining genetic and epigenetic information, cancer
universal across gastrointestinal malig-
cell-intrinsic properties, and the tumor microenvironment (TME). Unsupervised nancies and often bear comparable
transcriptomic subtyping systems of different GI malignancies have gained inter- prognostic and predictive information.
est because they reveal shared biological features across cancers and bear
Single cell sequencing has demonstrated
prognostic and predictive value. Importantly, transcriptomic subtypes accurately the biological and clinical relevance of the
reflect complex phenotypic states varying not only per tumor region, but also cancer cell-intrinsic compartment.
throughout disease progression, with consequences for clinical management.
Here, we discuss methodologies of transcriptomic subtyping, proposed taxon- Combining gene- and pathway-level,
bulk and single cell transcriptomic classi-
omies for GI malignancies, and the challenges posed to clinical implementation, fications appears to provide most infor-
highlighting opportunities for future transcriptomic profiling efforts to optimize mative patient stratification, rather than
clinical impact. using a single (consensus) taxonomy.

Transcriptomic subtypes reflect com-

plex phenotypic states that vary within
The value of transcriptomic cancer subtyping a tumor and change during disease
Tumor heterogeneity of GI malignancies poses considerable challenges to the clinical manage- progression, which affects clinical
management.
ment of patients and determines their outcome [1]. Tumor heterogeneity arises from a multiface-
ted process, driven by (epi-)genetic events, immune response, composition of the TME, and cell International prospective clinical trials
of origin of the tumor [2]. Heterogeneity occurs both between two primary tumors from the same stratified for molecular subtypes, and
histology (intertumor heterogeneity) and within a single tumor (intratumor heterogeneity) [1,2]. research into non-invasive subtype
detection methods, could enhance
Given that distinctive tumor molecular traits require tailored management, efforts have been clinical implementation of transcriptomic
made to develop prognostic and predictive patient stratification systems based on tumor biology. taxonomies.
Patient stratification through panel-based genetics has been studied extensively and is widely
adopted in clinical practice, because it is relatively straightforward, cost-effective [3–5], and infor-
mative for prognosis, selection of targeted therapies, and response prediction [5–9]. However,
these strategies are inherently limited by the number of genes included, and their predictive
1
value is mostly restricted to a limited number of systemic therapies, such as anti-EGFR therapy Cancer Center Amsterdam, Laboratory
in RAS/BRAF wild-type colorectal cancer (CRC) [10]. Moreover, genetic tumor characterization for Experimental Oncology and
Radiobiology, Center for Experimental
only partially reflects tumor heterogeneity, because it omits DNA translation and epigenetic regu- and Molecular Medicine, Meibergdreef 9,
lation of gene expression. As an illustration, similar genetic alterations may lead to highly dissimilar 1105 AZ, Amsterdam, The Netherlands
2
cancer phenotypes [11] and, conversely, different mutations may lead to comparable biological Amsterdam Gastroenterology
Endocrinology Metabolism, Laboratory
and clinical tumor behavior [12], underlining the need for a multidimensional approach. for Experimental Oncology and
Radiobiology, Center for Experimental
Transcriptomic cancer subtyping provides essential information, because it captures not only and Molecular Medicine, Meibergdreef 9,
1105 AZ, Amsterdam, The Netherlands
gene mutations, but also epigenetics, tumor cell properties, and the TME [13,14], reflecting key 3
Oncode Institute, Meibergdreef 9,
elements of tumor heterogeneity. Numerous transcriptomic taxonomies have emerged for GI 1105 AZ, Amsterdam, The Netherlands
4
cancers [15–17], with demonstrated clinical relevance. Although advances in accessibility are Flevohospital, Department of Internal
Medicine, Hospitaalweg 1, 1315 RA,
still required, RNA sequencing (RNA-seq) has become more affordable, faster [18,19], and rela- Almere, The Netherlands
tively easily interpretable through well-established analytical tools [20–22]. This is in contrast with 5
Shared senior authorship.

842 Trends in Cancer, September 2024, Vol. 10, No. 9 https://doi.org/10.1016/j.trecan.2024.06.007

© 2024 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Trends in Cancer
OPEN ACCESS

taxonomies based on proteomics or multi-omics data, which accurately reflect cancer heteroge- *Correspondence:
[email protected]
neity, but are costly, require strong expertise, and are not yet ready for clinical implementation (L. Vermeulen).
[23,24]. Currently, cancer subtyping using RNA-seq may provide the optimal tradeoff between
feasibility and level of acquired information (Figure 1), perhaps tipping the balance toward imple-
mentation of precision medicine for patients with GI cancers. Here, we review the methodologies,
clinical applicability, barriers, and future directions of transcriptomic subtyping in a rapidly evolving
field of research. Our focus is on RNA seq-based transcriptomic subtyping, but important alter-
native gene expression-based taxonomies are also discussed. We discuss these developments
in three core GI cancer types: CRC, gastric cancer (GC), and pancreatic ductal adenocarcinomas
(PDAC).

Methodology of transcriptomic subtyping

Practical aspects of RNA-sequencing
Tissue preservation and type of RNA impact RNA-seq experiments (Box 1). RNA-seq is often car-
ried out on bulk RNA from whole tumor or macrodissected slides, representing cancer cell-
intrinsic and TME properties. Given that the TME is an important determinant of tumor behavior
[25,26], transcriptomic profiling of bulk RNA appears to be clinically meaningful. Alternatively,
low tumor cellularity in bulk RNA samples is common, especially in heavily fibrotic tumors
[27,28]. This complicates the detection of tumor cell-specific gene expression profiles (GEPs), be-
cause these samples are rich in RNA from the TME, thus diluting the GEPs of tumor cells [28,29].
To address this issue, studies have performed microdissection of tissue slides to separately
characterize GEPs of the tumor cell versus the stroma compartments [27,30,31], minimizing
TME dilution, yet preserving its information. Recently, single cell-sequencing techniques have
further improved the appreciation of cell types and cell states within a tumor [32,33]. Single cell

Immuno- Phospho-
histochemistry proteomics

FFPE-based FF-based
sequencing sequencing
DNA panel Single-cell
sequencing sequencing
RNA
sequencing
Unsupervised
Level of information

Supervised
Clinical feasibility

clustering clustering
WGS

Bulk Microdissected
RNA RNA

Method
Trends in Cancer

Figure 1. Molecular profiling in the clinical setting. Graphical representation of the balance between the clinical feasibility
of molecular profiling methods and the acquired level of information from the respective methods. Generally, complex
molecular profiling methods, such as single cell sequencing or phosphoproteomics, deliver a wealth of molecular
information, but are compromised by clinical feasibility. By contrast, clinically feasible techniques, such as
immunohistochemistry (IHC) or DNA panel sequencing, offer a restricted amount of information. In this context, RNA-
sequencing appears to provide the optimal tradeoff between clinical feasibility and acquired level of information.
Abbreviations: FF, fresh-frozen; FFPE, formalin-fixed paraffin-embedded; WGS, whole-genome sequencing.

Trends in Cancer, September 2024, Vol. 10, No. 9 843

 Trends in Cancer
OPEN ACCESS

Box 1. Sample and RNA type matter for RNA sequencing

Tumor RNA for RNA-seq is generally obtained from fresh-frozen (FF) tissue specimens. FF preservation of tumor tissue yields
high-quality RNA suitable for in-depth transcriptomic tumor characterization (e.g., novel transcripts, gene fusions, or gene iso-
forms). However, FF-preserved tissues are not standardly available in the clinic, and collection and storage of cryopreserved
samples are tedious. Alternatively, RNA can be obtained from archival formalin-fixed paraffin-embedded (FFPE) tissue samples,
which are widely available in the clinic and, therefore, highly suitable for clinical translation of molecular subtyping strategies. Due
to delays between the acquisition of tumor tissue and subsequent FFPE preservation, RNA stemming from FFPE samples is
highly degraded, complicating downstream RNA-data analyses [102,103]. With the development of novel library preparation
protocols, incorporating exome capture and ribosomal RNA depletion, these hurdles have been largely overcome, and studies
are showing high concordance of gene expression profiles between matched FF and FFPE tissue specimens [104,105].

Generally, protein-coding mRNA is preferred as the substrate for transcriptomic profiling of tumors, because it contains
direct information on up- or downregulated cellular processes, providing clear-cut therapeutic targets. However, mRNA
easily degrades, especially after FFPE preservation, and is regularly nontumor cell specific [106]. Therefore, transcriptomic
profiling using small RNAs, such as miRNAs, or long noncoding RNAs (lncRNAs), has gained interest [106–108].

Both miRNAs and lncRNAs have regulatory roles in gene expression, either directly via binding to DNA promotor regions or
indirectly via post-transcriptional gene regulation [107,108]. The advantage of using miRNAs for transcriptomic profiling
lies in the fact that they are typically smaller and more stable, facilitating accurate sequence detection [107]. Furthermore,
miRNAs, but especially lncRNAs, can be exceptionally tumor specific, making them attractive therapeutic targets [108]. To
date, several miRNA- and lncRNA-based profiling efforts have been carried out successfully, although the clinical
relevance of these strategies needs to be further elucidated [107,109–113].

transcriptomic subtyping efforts clearly demonstrate the intratumor heterogeneity of GI cancers,

which could aid personalized medicine [34,35].

Subtype discovery: supervised versus unsupervised clustering

After RNA-seq data acquisition and preprocessing, transcriptomic class discovery is carried out
(Box 2). An important step involves feature selection, which can take place in either a supervised
or unsupervised manner [13]. Initial transcriptomic classification strategies mainly performed
supervised clustering, using predefined genes with proven prognostic value or associations

Box 2. Class discovery using whole-transcriptome data

The discovery of transcriptomic subtypes generally follows a common workflow, in which feature selection and clustering
are essential steps for class discovery [14].

Feature selection refers to the identification of informative genes for clustering, in a (biology) supervised or unsupervised fashion
(Figure I). Many transcriptomic subtyping studies use most variably expressed genes [17,27,114], but components determined
by dimensionality reduction can also be used [77,115,116]. Dimensionality reduction is a method to transform high-dimensional
data into a few low-dimensional components that explain the variance in the data. Several techniques are available, widely clas-
sified as linear or nonlinear strategies [117]. Linear techniques decompose the data as a linear combination of the original var-
iables while maintaining pairwise distances between data points. Non-negative matrix factorization (NMF) [77,116] and principal
component analysis (PCA) [70] are most standardly used for feature selection. NMF provides a small number of metagenes
based on the total gene list and creates a matrix of the expression level of each metagene per sample [118]. PCA transforms
complex gene expression data into principal components that explain data variance, ordered from the largest to the smallest
explained variance [119]. For clustering purposes, the components (i.e., features) that capture most of the variance are typically
selected. Nonlinear dimensionality reduction techniques transform high-dimensional data into nonlinear low-dimensional
components while preserving the local structure of the data [117]. Examples are t-distributed stochastic neighbor embedding
(t-SNE) and uniform manifold approximation and projection (UMAP). A comparison between PCA, t-SNE, and UMAP in
analyzing large bulk transcriptomic data sets suggested the superiority of UMAP over other techniques, because it effectively
identified batch effects, predefined biological subgroups, and new clusters with biological and clinical significance [120].

Feature selection is followed by clustering, which can be either hierarchical or nonhierarchical, and iterative or non-iterative
[14]. Hierarchical clustering assumes a developmental relationship between subtypes, whereas nonhierarchical clustering
identifies subtypes as separate entities and, therefore, may be more suitable for transcriptomic class discovery (e.g., k-means
clustering and NMF). Iterative approaches perform repetitive clustering (generally 500–1000 runs) and determine the definitive
clusters based on a consensus score (consensus clustering), whereas non-iterative approaches perform a single clustering
analysis. Most commonly used class discovery approaches are iterative nonhierarchical clustering techniques, either with
(NMF [118]) or without dimensionality reduction (original consensus clustering [121]).

844 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

Feature selection Clustering

Unsupervised Hierarchical
Genes/components

Dimension B
Patients Dimension A

Supervised Nonhierarchical
Genes/components

Dimension B

Patients Dimension A

Most variably expressed genes

Classical supervised feature selection Optional: iterative
Biology-supervised feature selection

Trends in Cancer

Figure I. Class discovery in transcriptomic clustering efforts. Feature selection refers to the process of
selecting genes or components of interest for clustering. Unsupervised efforts generally use the most
variably expressed genes, whereas supervised efforts use prognosis- or therapy-informed genes or
components, or a biological subset of genes or components. Clustering is either hierarchical, assuming a
developmental relation between the clusters, or nonhierarchical, and is mostly iterative, although non-
iterative clustering is an option.

with therapeutic response [13,14]. Although these approaches can be clinically meaningful [36–38],
they are biased due to upfront feature selection and, thus, are unable to fully capture underlying
biological variation, resulting in limited generalizability [38,39]. To resolve this limitation, an unbi-
ased, unsupervised clustering approach using whole-transcriptome gene expression data without
any prespecified selection criteria is regularly utilized [15–17]. The resulting taxonomies have been
recapitulated in various clinical settings, capturing biological heterogeneity and providing prognos-
tic and predictive information [40–42].

Alternatively, semisupervised clustering is common practice [43–45]. This method considers a

few essential clinical, histopathological, or genetic variables for clustering based on gene expres-
sion data [13]. Notably, with the former two selection classes, the data are filtered on sample level
(columns in a clustering matrix), whereas with the latter data are filtered on gene level (rows
in a clustering matrix) (Box 2). Selecting specific genes for clustering potentially limits the extent
of biological variation that can be detected and, therefore, might be perceived as a biology-
supervised manner of clustering, with similar limitations as other supervised clustering approaches.

Alternative approaches for transcriptomic profiling

Although RNA-seq is becoming more accessible, the required amenities are difficult to implement
in the clinic. Hence, several alternative classification strategies have been developed. An example

Trends in Cancer, September 2024, Vol. 10, No. 9 845

 Trends in Cancer
OPEN ACCESS

is gene panel-based classification of existing or novel transcriptomic taxonomies, using targeted

RNA-seq or NanoString technologies [46–49]. This strategy reduces all the aforementioned
hurdles, while maintaining the clinical impact. The same goes for immunohistochemistry (IHC)-
based detection of previously or newly established transcriptomic subclasses [50–52]. These
panels generally include between two and four markers, making them easy to implement. However,
IHC-based detection of transcriptomic subtypes remains inferior to RNA-seq; for instance, not every
RNA seq-derived subtype can be distinguished with IHC [50–52]. An upcoming alternative is image-
based classification of tumor tissue slides, using deep-learning algorithms [53–55]. Although this
approach is cost-effective and time efficient, it requires expertise in using deep-learning algorithms
and might be challenging to implement in a clinical setting.

These alternative approaches facilitate clinical implementation of transcriptomic subtyping by

compromising on the acquisition of transcriptome-wide data. This is an important disadvantage,
because it reduces biological information, limits incorporation of novel developments, and mini-
mizes opportunities for personalized medicine.

Unsupervised taxonomies
Colorectal cancer
CRC is a heterogeneous disease, with a worldwide incidence of 1.15 million cases and over 570
000 annual cancer-related deaths [56]. As such, many studies have proposed supervised (Box 3)
and unsupervised gene expression-based taxonomies to capture its intrinsic heterogeneity and

Box 3. Supervised transcriptomic taxonomies of GI malignancies

Using differentially expressed genes between early and late-stage CRC [38,122], or genes related to recurrence risk or
poor prognosis [36,37,123], patients with high-risk CRC in need of adjuvant chemotherapies could be identified among
patients with localized disease. Although one signature correlated with benefit of oxaliplatin treatment [37], most CRC
signatures were not associated with adjuvant therapy success. For GC, histology-supervised transcriptomic subtyping
successfully distinguished diffuse from intestinal (proximal or distal nondiffuse) GC, bearing prognostic and potential
predictive value [124,125]. For PDAC, several prognosis-supervised taxonomies were established [126–128], of which
one bore in vitro predictive value for gemcitabine sensitivity [128].

Biology-supervised stratification studies for GI malignancies are numerous. For CRC, well-studied examples are the IHC-
based Immunoscore [129,130], and the Complexity INdex in SARComas (CINSARC) signature, comprising 67 chromo-
somal instability-related genes [131]. Whereas CINSARC is a prognostic signature, Immunoscore correlates with immuno-
therapy benefit [130]. Taxonomies based on RNA-seq identified spectra of biological subgroups for CRC, such as
secretory cell-related, necroptosis-related and immunometabolism-related subtypes, each showing distinct prognostic
value, but limited predictive value [45,132,133].

For GC, biology-supervised clustering by the Asian Cancer Research Group (ACRG) identified a poor-prognostic stromal
MSS/EMT subtype, alongside an MSI, an MSS/TP53+, and an MSS/TP53– subtype [70]. These subtypes hold potential for
predicting response to immunotherapy, and ERBB2- and EGFR-targeting therapies. More recently, four novel GC sub-
types were identified based on a 32-gene signature, among which was a TGF-β-overexpressing mesenchymal subtype
with poor prognosis [44]. No clear correlation with the ACRG subtypes was found. These subtypes were predictive for re-
sponse to adjuvant 5-fluorouracil (5-FU) and platinum chemotherapy and immune checkpoint inhibitors. Additionally, peri-
toneal metastases of GC also contained a poor-prognostic mesenchymal-like subtype, potentially correlating with
immunotherapy response [134]. Several other taxonomies also demonstrated prognostic value and potential to predict im-
munotherapy response [115,135–138].

For PDAC, several biology-supervised clustering efforts focused on predicting immunotherapy response [139–141].
Recently, a TME classification was established through clustering using 25 functional gene signatures, identifying an
immune-enriched (IE), immune-enriched fibrotic (IE/F), fibrotic (F), and immune-depleted (D) subtype [142]. IE and IE/F cor-
related with favorable prognosis, potential immunotherapy response, and previously defined aberrantly differentiated en-
docrine exocrine (ADEX), and immunogenic subtypes [16], and were enriched in lung metastases. F and D conferred poor
prognosis, correlated with basal-like and squamous subtypes [16,27], and D was enriched in liver metastases. This tax-
onomy adequately addressed the contribution of the TME in defining prognosis and therapeutic vulnerability of PDAC.

846 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

improve the course of the disease. Given that it is extensively studied, CRC could serve as a case
example for other GI malignancies.

To capture the full spectrum of CRC heterogeneity and strengthen the clinical relevance of strat-
ification systems, unsupervised clustering efforts have been undertaken. In 2015, six taxonomies
based on bulk RNA were integrated into four consensus molecular subtypes (CMSs) [15]. These
subtypes have been extensively studied for relevant biological and clinical characteristics (Box 4)
[25,40,57]. Notably, the poor-prognostic subtypes (CMS4, stem-like, CCS3, and C4) appear to
be largely determined by stromal tumor content [25,26]. Additionally, differences in GEPs within
a tumor can lead to the assignment of varying subtypes to the same patient, raising the possibility
of patient misclassification [25,58]. When using cancer cell-intrinsic GEPs, tumors can be more
robustly classified by tumor sample, rather than tumor region, hence minimizing confounding
by intratumor heterogeneity [59,60].

To exploit this notion, unsupervised clustering was performed on murine patient-derived xeno-
grafts (PDXs), devoid of human TME components, eliminating stromal bias by using human-
specific microarray probes, leading to the establishment of five CRC cancer Intrinsic Subtypes
(CRIS-A–D) [61]. These showed prognostic value and comprised a cancer-cell intrinsic subtype
with high TGF-β signaling, not specific for CMS4. This subtype defines a previously undetected
patient group that might benefit from adjuvant chemotherapy and novel agents targeting TGF-
β signaling. Furthermore, CRIS-C detects patients with CRC sensitive to anti-EGFR treatment,
whereas CRIS-E defines patients with anti-EGFR refractory CRC. However, the CRISs are disad-
vantaged by possible xenotransplantation-related artefacts, such as selection drifts after PDX
propagation and engraftment.

Box 4. The consensus molecular subtypes of colorectal cancer

The CMSs were founded on six unsupervised taxonomies established with bulk RNA data [114,116,143–146]. Through
network analyses, four recurrent consensus subtypes were identified [15]: CMS1, immunogenic, BRAF-mutated, and
MSI, with good prognosis in localized disease; CMS2, epithelial with activation of MYC and WNT target genes and good
prognosis in localized and metastatic disease; CMS3, metabolic with goblet cell enrichment and KRAS mutations, and
poor prognosis in metastatic disease; and CMS4, mesenchymal, angiogenic, and rapidly disseminating, with TGF-β acti-
vation and poor prognosis. To enable classification of both patient cohorts and individual tumor samples, a random forest
classifier and single-sample predictor were developed [15]. Recently, a classification was established with genes related to
low intratumor heterogeneity, leading to more CMS congruence between multiregional samples and primary–metastasis
pairs (congruent CMSs) [60]. Furthermore, a cell-line, patient-derived xenograft [147,148], and mouse-model classifier
were established [149] to facilitate fundamental research.

The CMSs are informative for all stages of the adenoma–carcinoma sequence [150–153]. Notably, the CMSs do not rep-
resent mutually exclusive, static entities, but are rather plastic and reflective of molecular states with potential overlap
[90,154]. Adenomas mainly belong to CMS3 and, upon progression, subtype switching toward CMS4 regularly occurs,
whereas the CMS2 subtype appears to be rather stable during disease progression [150,151]. Once metastasized, liver
metastases are often CMS2- or CMS4-like and peritoneal metastases are predominantly CMS4-like [152,153]. To elimi-
nate the liver tissue background of liver metastases, which can bias CMS classification, a liver-metastasis specific CMS
classifier has been introduced [155]. Additionally, for both liver and peritoneal metastases, additional subdivision into three
groups has been proposed, with distinct biological and clinical features [152,153].

Importantly, the CMSs hold predictive value. CMS1, enriched for MSI tumors, generally responds well to immunotherapy
and appears to be sensitive to bevacizumab treatment; CMS2 benefits from oxaliplatin-based chemotherapy and, in cases
of RAS wild-type status, EGFR inhibition; and CMS4 is relatively resistant to adjuvant chemotherapy and benefits from
irinotecan-based regimens in the metastatic setting [40]. Despite their therapeutic relevance, clinical implementation of
the CMSs is challenging. To facilitate this process, alternative detection methods have been developed. The CMSs can
either be detected via IHC [51,156–158], targeted RNA-seq using gene panels [46–49], or classification of stained tumor
slides using deep-learning algorithms [55,159,160]. To circumvent RNA-degradation issues, CMS classification on miRNA
has been proposed [107], as well as FFPE-curated classification pipelines [48,92].

Trends in Cancer, September 2024, Vol. 10, No. 9 847

 Trends in Cancer
OPEN ACCESS

Given the limitations of bulk RNA transcriptomic subtyping and the advent of single cell-sequenc-
ing techniques, a novel cancer-cell intrinsic CRC taxonomy was established, termed the intrinsic
Consensus Molecular Subtypes (iCMSs) [35]. Two epithelial cell states were identified, iCMS2 (i2)
and iCMS3 (i3), with CMS2 predominantly being classified as i2, and CMS1 and CMS3 as i3.
CMS4 was classified as both i2 and i3. The subgroup of CMS4/i3 tumors had a significantly
worse prognosis compared with CMS4/i2 tumors. Given that i3 tumors could be subdivided
into two groups based on mismatch repair (MMR) status, and fibrosis identified CMS4 tumors
within i2 and i3, a composite classification was proposed based on intrinsic subtype, MMR sta-
tus, and fibrosis, comprising five classes: i2_MSS_NF, i2_MSS_F, i3_MSS_NF, i3_MSS_F, and
i3_MSI. Among these subtypes, i3_MSS_F had the worst prognosis. i3_MSI correlated with re-
sponse to immunotherapy, and drug response signatures for folinic acid, fluorouracil, and oxali-
platin (FOLFOX), 5-fluorouracil (5-FU), and cetuximab were upregulated in i2 tumors, whereas
response signatures for FOLFIRI and gefitinib were upregulated in i3 tumors. These results sug-
gest prognostic and predictive value of the iCMSs.

Recently, a pathway-level bulk transcriptomic classification for CRC was proposed [62]. This
classification identified three pathway-derived subtypes, PDS1–3, dependent on cell differentia-
tion state, distinguishable by a stem maturation index, as a gradient from the expression of
MYC targets (stem-like, PDS1) toward the expression of Polycomb Repressive Complex (PRC)
targets (differentiation-like, PDS3). The PDS3 subtype represents a previously overlooked, highly
differentiated, slow-cycling subset of CMS2 tumors, thus uncovering heterogeneity within CMS2.
Notably, PDS3 was associated with the worst prognosis in localized disease, in contrast to clas-
sic CMS2 survival patterns. CMS1 and CMS4 tumors were predominantly classified as PDS2,
whereas PDS1 mainly comprised CMS2 and CMS3 tumors. Although the prognostic relevance
of the PDSs was evident, their predictive value remains to be demonstrated.

Ultimately, parallel classification using abovementioned transcriptomic taxonomies may provide

the most unbiased and informative stratification of patients with CRC that considers cancer
cell-intrinsic features, the TME, and gene-level and pathway-level information (Figure 2A).

Gastric cancer
GC is the fifth most-common cancer worldwide, and the fourth leading cause of cancer-related
death [56]. Chronic infection with Helicobacter pylori is an important risk factor for its develop-
ment [63]. Interpatient differences in histopathology [64], genetic tumor characteristics [17],
and therapy response [65] indicate tumor heterogeneity; hence, transcriptomic stratification
efforts have been pursued (Box 3).

Based on unsupervised cell line clustering, two cancer cell-intrinsic subtypes of GC were identi-
fied: G-INT (metabolic/proliferative) and G-DIF (mesenchymal/proliferative), partially related to
Lauren’s histopathological classification (Figure 2B) [66]. In validation cohorts, G-INT had a signif-
icantly better prognosis compared with G-DIF and correlated with benefit from adjuvant 5-FU
chemotherapy, which is mainly administered in Asia. To translate these findings to the clinic,
human GC samples were clustered, and three subtypes were identified: metabolic, proliferative,
and mesenchymal [67]. The metabolic subtype was predictive for 5-FU chemotherapy efficacy,
whereas in vitro PI3K inhibitors were effective for the mesenchymal subtype. Recently, a two-
subtype taxonomy was proposed, comprising a mesenchymal phenotype (MP), with activated
TGF-β signaling, versus an epithelial phenotype (EP) [68]. The MP subtype associated with
poor prognosis and lacked benefit from adjuvant chemotherapy. A recent five-subtype system
identified, among others, a mesenchymal subtype (MSC) and an intestinal subtype with stem-
like features (INT/S). Exclusively the MSC subtype correlated with stromal infiltration, but both

848 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

(A) Epithelial Immune-stromal Ref. Figure 2. Molecular stratification of patients with

gastrointestinal (GI) cancer using multiple transcriptomic
taxonomies. Parallel transcriptomic classification of patients
Gene

CMS2 CMS3 CMS1 CMS4 15

with GI cancer using bulk, single cell, gene-level and pathway-
level taxonomies, could refine molecular patient stratification. (A)
Bulk

For colorectal cancer (CRC), each consensus molecular

Pathway

PDS1 PDS3 PDS1 PDS3 PDS2 PDS2 62

subtype (CMS) could be further divided into a stem-like,
S-L D S-L D R differentiated, or regenerative subtype, with an intrinsic indolent
(i2) or aggressive tumor cell compartment (i3), and a
i2 i3 i2 i3 i3 i2 microsatellite stable nonfibrotic, microsatellite stable fibrotic, or
Single cell

microsatellite unstable phenotype, resulting in refined subtypes

Gene

35
MSS MSS MSS MSS MSS MSI MSS MSS with distinct clinical value [15,32,65]. (B) For gastric cancer
NF NF NF NF NF F F
(GC), bulk transcriptomic and single cell classifications identified
Survival intestinal and mesenchymal cancers, among which the
Adj CTx intestinal cancers could be further divided into a microsatellite
ICB unstable subgroup, and TP53-mutated microsatellite stable
Anti-EGFR subgroups, with distinct clinical value [34,66–70]. Additionally,
organ-specific subgroups appear to exist. (C) For pancreatic
Favorable Unfavorable
cancer, bulk transcriptomic, cancer cell-intrinsic, and single cell
classifications consistently define classical and basal-like
(B) Overlapping subtypes Other Ref. (mesenchymal) subtypes [16,27,30,77–85]. The classical
subtype also harbors a metabolic and an immune subgroup,
G-INT/EP G-DIF/MP 66/68
and the latter also has features from basal-like tumors.
MES MET Furthermore, a pancreas-specific secretory subgroup has been
PRO 67
identified. Some taxonomies described subtypes that did
Bulk

not overlap with other classifications, as displayed under

MSI MSS/TP53- MSS/TP53+ MSS/EMT 70
‘Other’. Of note, clinical associations of the subtypes are not
based on randomized clinical trials, but deduced from
INF
INF INT GST INT INT-S GST INT INT-S MSC 69 preclinical experiments or from gene expression signatures of
EBV+
the (overlapping) subtypes. Abbreviations: ADEX, aberrantly
Differentiated Undifferentiated
Single cell

differentiated endocrine and exocrine; Adj CTx, adjuvant

chemotherapy; CI, cancer-intrinsic; D, differentiated; Desmopl.
C2 C3 C5 C3 C1 C4 34
desmoplastic; EBV, Epstein–Barrvirus; EMT, epithelial to
mesenchymal transition; EP, epithelial phenotype; F, fibrotic; G-
Survival
DIF, genomic diffuse; G-INT, genomic intestinal; GST, gastric;
Adj CTx
i2, iCMS2; i3, iCMS3; IC, intermediate co-expressor; ICB,
ICB
immune checkpoint blockade; INF, inflammatory; INT-S,
Favorable Unfavorable intestinal with stem-like features; INT, intestinal; MES,
mesenchymal; MET, metabolic; MP, mesenchymal phenotype;
(C) Overlapping subtypes Other Ref. MSC, mesenchymal; MSI, microsatellite unstable; MSS,
microsatellite stable; NF, nonfibrotic; PRO, proliferative; QM,
Classical QM Exocrine 77
quasi-mesenchymal; R, regenerative; Ref. reference; S-L, stem-
Classical Basal 27 like; SC, single cell.
Classical-AB Basal-AB 79
Bulk

Classical Immune Stroma Basal Desmopl. 81

Progenitor Immune Squamous ADEX 16

Epithelial Mesenchymal Secretory Compound 80

Progenitor Classical Basal Glycomet 84

CI

C3 C1 C4 C2 30

Classical Squamoid-basaloid Treatment enriched 85

SC

scClassical scBasal IC 86

Survival
Adj CTx
ICB

Favorable Unfavorable

Epithelial Metabolic Immune Mesenchymal Organ-specific

Trends in Cancer

Trends in Cancer, September 2024, Vol. 10, No. 9 849

 Trends in Cancer
OPEN ACCESS

subtypes showed poor prognosis, reduced response to chemotherapy, and attenuated dissem-
ination and drug resistance upon TGF-β inhibition in preclinical models [69]. A comparison with
previous taxonomies showed correlation between the MP/MSC subtypes and G-DIF, the micro-
satellite stable (MSS)/epithelial–mesenchymal transition (EMT) subtype, and the genome stable
subtype [17,66,70], indicating validity of this subgroup. Notably, microsatellite unstable (MSI)
and Epstein–Barr virus (EBV)-positive tumors were classified as either inflammatory (INF) or intes-
tinal (INT). Similarly, peritoneal metastases of GC could be divided into an EMT versus a non-EMT
transcriptomic subtype [71]. The EMT subtype was characterized by increased Hippo pathway
signaling, and TEAD1 inhibition resulted in significant tumor suppression in vivo.

At the single cell level, five intrinsic subtypes of GC have been identified [34]. The first three sub-
types correlated with Lauren’s histopathological diffuse, intestinal, and mixed subtypes, with poor
prognosis of diffuse cancers. The fourth subgroup expressed chief cell markers and RNF43, with
activated Wnt/β-catenin signaling, consistent with the rare fundic gland-type GC [72]. The fifth
subtype expressed immune-related genes and was associated with EBV infection. These sub-
types not only uncovered clear tumor cell-intrinsic heterogeneity of GC, matching previous bulk
transcriptomic subtypes, but also highlighted additional rare subgroups.

Pancreatic ductal adenocarcinoma

PDAC is the most common histology of pancreatic cancer, characterized by a dismal prognosis
with high lethality, with little improvement over the past few decades [73]. The poor outcome for pa-
tients with PDAC partly results from dense stroma, hindering drug delivery [74]. Additionally, tumor
heterogeneity is widespread, with only a few commonly altered genes (KRAS, TP53, SMAD4, and
CDKN2A), although a long tail of infrequently mutated genes, stressing the need for individualized
therapy [75,76]. Thus, transcriptomic subtyping efforts have been undertaken (Box 3).

One of the first unsupervised classifications divided patients into the classical, quasi-
mesenchymal, and exocrine-like subtypes, with overexpression of epithelial, mesenchymal,
and digestive enzyme genes, respectively (Figure 2C) [77]. The quasi-mesenchymal subtype
conferred the worst prognosis and predicted response to gemcitabine, whereas the classical
subtype predicted response to erlotinib. Using tissue microdissection, two cancer-intrinsic sub-
types were identified, termed classical and basal-like subtypes, closely related to the established
classical and quasi-mesenchymal subtypes [27]. The basal-like subtype correlated with poor
prognosis and benefited from adjuvant chemotherapy, similar to the quasi-mesenchymal sub-
type. However, basal-like tumors responded poorly to FOLFIRINOX compared with classical
PDACs [78]. Several other studies identified similar subtypes [79–81]. Additionally, another
transcriptomic taxonomy described four subtypes associated with tumor differentiation: a poor
prognostic (adeno-)squamous subtype with enrichment for TP53 and KDM6A mutations, a
pancreatic progenitor subtype with expression of genes involved in early pancreatic development,
an immunogenic subtype with upregulation of pathways involved in acquired immune suppression,
and an aberrantly differentiated endocrine exocrine (ADEX) subtype with upregulation of genes
involved in KRAS activation, and endocrine and exocrine differentiation [16]. (Adeno-)squamous
tumors had the worst prognosis and immunogenic tumors were potentially targetable with
immunotherapy.

A consensus taxonomy proposed three immunophenotypic subtypes, comprising a poor-

prognostic innate immune subtype (representative of the quasi-mesenchymal, basal-like, and
squamous subtypes), a T cell-dominant subtype (representing the exocrine-like and ADEX sub-
types), and a tumor-dominant subtype (reflecting the classical pancreatic progenitor subtypes)
[82,83]. Although the previously defined immunogenic subtype showed similarities with the T

850 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

cell-dominant subtype, it was not integrated in this consensus taxonomy. Clinical value of these
consensus subtypes remains to be demonstrated.

Recently, four novel cancer-intrinsic subtypes (C1–C4) were established, also showing parallels
with previous classical (C1 and C3), quasi-mesenchymal (C4), and exocrine-like (C2) subtypes,
with C4 showing the worst disease-free survival [30]. In this classification, the classical subtype
was divided into a worse prognostic, immunogenic, protein-folding subgroup (C1) and a better
prognostic, metabolic, protein-translating subgroup (C3). Using patient-derived organoids, a pre-
viously undiscovered subtype was identified, which was enriched for glycolysis, carbohydrate,
and glucose metabolism, termed the glycomet subtype, and was associated with poor prognosis
[84]. Two single cell-clustering efforts identified a classical and squamoid-basaloid/basal sub-
type, showing strong concordance with previous subtypes [85,86]. Furthermore, a treatment-
enriched subtype, similar to the exocrine/ADEX subtypes, and an intermediate subtype were de-
fined, corroborating previous taxonomies.

Pan-gastrointestinal subtyping
Most GI taxonomies have defined cancer subgroups with shared molecular features, such as
epithelial (Wnt upregulation and TP53 loss), metabolic (KRAS mutated), immunogenic (MSI and
BRAF mutated) and mesenchymal tumors (TGF-β overexpression) (Figure 2), posing the question
whether a pan-GI taxonomy would be feasible. In this respect, the tissue-of-origin of cancers can
dominate pan-cancer unsupervised clustering, since tumors of specific organ systems tend to
cluster together [87]. GI malignancies were separated into three clusters, defined by EBV positivity,
MSI, and colorectal origin, indicating that other molecular features beyond cell of origin are of im-
portance in clustering efforts. A dedicated attempt to perform transcriptomic clustering of GI ma-
lignancies (excluding PDAC) failed to define molecular subgroups that spanned anatomic
boundaries, because clustering was strongly influenced by tissue type [88]. When using genomic
and methylation data, not only shared molecular features across GI malignancies were found, but
also considerable differences in DNA methylation and CIN depending on upper or lower GI tract
localization. In an organoid model, regional differences in response to Wnt signaling were observed
[89]. Altogether, these data suggest comparable molecular subgroups across GI malignancies, al-
though regional differences in cell of origin are important and may lead to the identification of organ-
specific subtypes.

Clinical implementation
Transcriptomic taxonomies across GI malignancies have captured tumor heterogeneity with clin-
ical impact. However, clinical implementation of these taxonomies has been hindered by technical
aspects, such as a long turnaround time and complicated subtype detection using low-quality
RNA from formalin-fixed paraffin-embedded (FFPE) samples. Furthermore, their clinical utility
needs further addressing, such as the exact predictive value for therapy response and the neces-
sity for repeated tissue sampling, especially when monitoring cancer progression and therapy re-
sistance over time, which are affected by subtype switching [90]. In addition, practical aspects,
such as high costs and required expertise, have been an issue.

Although technical issues remain, critical advances have been made. As an example of reduced
sequencing turnaround time, real-time DNA methylation subtyping during surgery has been suc-
cessfully performed for central nervous system tumors [91]. In addition, optimization of subtype
detection on low-quality RNA-seq data stemming from FFPE tissues has been carried out
[81,92]. This approach has an added benefit, because alternative strategies lack acquisition of
transcriptome-wide gene expression data for detailed exploration of tumor characteristics and
treatment vulnerabilities, reducing clinical impact.

Trends in Cancer, September 2024, Vol. 10, No. 9 851

 Trends in Cancer
OPEN ACCESS

Although the clinical value of the CMSs for CRC have been studied thoroughly in retrospective co- Outstanding questions
horts, prospective randomized clinical trials stratified for molecular subtypes are largely lacking for What is the optimal transcriptomic
GI malignancies [40]. For GC and PDAC, a consensus subtyping taxonomy remains to be estab- classification method: supervised or
unsupervised, gene or pathway level,
lished and clinically validated, respectively [83]. Another clinical challenge is subtype switching
and bulk or single cell RNA?
during disease progression, which might affect disease management [90]. Recently, CMS
subtyping of metastatic CRC samples was performed using liquid biopsies, with RNA extraction What is the clinical value of transcriptomic
from extracellular tumor vesicles. The use of liquid biopsies allowed for non-invasive and repeated subtypes in GI malignancies?
subtyping during disease progression, offering a patient-friendly solution for monitoring cancer How could the universally identified
progression, subtype evolution, therapy resistance, and candidate systemic therapies [93]. poor-prognostic mesenchymal sub-
type of GI tumors be most effectively
Practical hurdles have been largely overcome by reduction of RNA-seq costs and alternative clas- treated?

sification strategies (Box 4). Similarly, for both GC and PDAC, IHC-based [52,94–96], gene panel- Should bulk and single cell transcriptomic
based [97,98], or image-based subtyping strategies [53,54,99,100] have been developed. taxonomies be combined into a refined
patient stratification tool for optimal
prognostic and predictive value?
Concluding remarks
With the advancement of RNA-seq and bioinformatics techniques, transcriptomic subtyping of Would it be feasible to design a pan-GI
cancers has become a prime research focus, capturing inter- and intratumor heterogeneity with cancer transcriptomic taxonomy instead
relevance for clinical practice. Supervised transcriptomic clustering based on preconceived clinical of individual cancer-type classifications?

parameters is informative for detecting poor-prognostic gene signatures or gene signatures predic- How could multilabel transcriptomic
tive for systemic therapies, and could complement unsupervised taxonomies. These signatures classification of a single tumor
generally capture biological features that overlap with unsupervised classification approaches, comprehensively guide clinical
but are restricted in generalizability by the selections made. Unsupervised clustering has proven management?

valuable in this respect through unbiased subtype detection, reflecting the full spectrum of molec- How could multiregion subtyping
ular tumor heterogeneity. Recently, cancer cell-intrinsic classification efforts have further advanced of a single tumor using spatial
the field, providing novel insight into transcriptomic traits driving tumor cell growth and heterogene- transcriptomics improve insight into
the complex phenotypic states (i.e.,
ity within the cancer cell compartment. These classification efforts have improved subtype consis-
subtypes) within a tumor?
tency, reduced misclassification through stromal bias, and added an extra layer of prognostic and
predictive value to former classifications schemes, as observed for CRC. What are the clinical feasibility and value
of liquid-biopsy based transcriptomic
subtyping of different GI malignancies?
The clinical utility of the multitude of GI transcriptomic stratification systems needs further ad-
dressing (see Outstanding questions). Strikingly, most independent subtyping efforts identify sim-
ilar tumor classes, as observed not only for CRC, but also for GC, PDAC, and cross-cancer
(Figure 2), perhaps allowing for a pan-GI transcriptomic taxonomy. This would dramatically in-
crease the number of eligible patients in subtype discovery efforts and clinical trials. However,
the cell of origin of cancers might complicate the detection of molecular subgroups spanning an-
atomic boundaries. Additionally, alternative subtyping approaches, for instance clustering based
on single cell RNA-seq data or pathway-level clustering, appear to subdivide established sub-
types into refined classes, with additional clinical relevance. Future tumor classifications should
include the interconnection with established subtyping systems to refine and combine taxon-
omies. These taxonomies should be investigated in international prospective clinical trials to eval-
uate their therapeutic implications.

Importantly, molecular subtypes do not appear to represent mutually exclusive entities, but rather
complex parallel phenotypic states varying per tumor region, which change during disease pro-
gression and under therapeutic pressure [57,93]. Multilabel transcriptomic classification of CRC
was recently shown to improve clinical value, and deserves further investigation [101]. Subtype
switching is a clinical reality, and would require repeated tissue-based subtyping during the
course of the disease. Therefore, low-cost and less-invasive subtype detection methods deserve
further attention, such as repeated liquid biopsy-based molecular classification during disease
progression [93].

852 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

Acknowledgments
This work was supported by The New York Stem Cell Foundation (NYSCF-I-R43), the European Research Council (ERC-
CoG 101045612 - NIMICRY), ZonMw (Vici 09-15018-21-10029), the Dutch Cancer Society (KWF 14182), and The Oncode
Institute Technology Development Fund (P2021-0011). L.V. is a New York Stem Cell Foundation —Robertson Investigator.
The funder had no role in the study design or manuscript. The authors wish to thank Sarah Derks and Maarten F. Bijlsma for
helpful discussions.

Declaration of interests
L.V. received consultancy fees from Bayer, MSD, Genentech, Servier, Roche, Novartis, and Pierre Fabre, but these had
no relation to the content of this publication. L.V. is an employee of Genentech. The other authors declare no competing
interests.

References
1. Punt, C.J. et al. (2017) From tumour heterogeneity to advances 20. Ahmaderaghi, B. et al. (2022) Molecular subtyping resource: a
in precision treatment of colorectal cancer. Nat. Rev. Clin. user-friendly tool for rapid biological discovery from transcrip-
Oncol. 14, 235–246 tional data. Dis. Model. Mech. 15, dmm049257
2. De Sousa, E.M.F. et al. (2013) Cancer heterogeneity–a multi- 21. Hait, T.A. et al. (2019) The EXPANDER integrated platform for
faceted view. EMBO Rep. 14, 686–695 transcriptome analysis. J. Mol. Biol. 431, 2398–2406
3. Arriola, E. et al. (2023) Cost-effectiveness of next-generation 22. Reyes, A.L.P. et al. (2019) GENAVi: a shiny web application for
sequencing versus single-gene testing for the molecular diag- gene expression normalization, analysis and visualization. BMC
nosis of patients with metastatic non-small-cell lung cancer Genomics 20, 745
from the perspective of Spanish reference centers. JCO Precis. 23. Li, L. et al. (2023) Comprehensive proteogenomic characteriza-
Oncol. 7, e2200546 tion of early duodenal cancer reveals the carcinogenesis tracks
4. Lotan, Y. et al. (2018) Modelling cost-effectiveness of a of different subtypes. Nat. Commun. 14, 1751
biomarker-based approach to neoadjuvant chemotherapy for 24. Li, S. et al. (2023) Integrative proteomic characterization of
muscle-invasive bladder cancer. BJU Int. 122, 434–440 adenocarcinoma of esophagogastric junction. Nat. Commun.
5. Pataky, R.E. et al. (2023) Real-world cost effectiveness of a 14, 778
policy of KRAS testing to inform cetuximab or panitumumab 25. Isella, C. et al. (2015) Stromal contribution to the colorectal
for third-line therapy of metastatic colorectal CANCER in British cancer transcriptome. Nat. Genet. 47, 312–319
Columbia, Canada. Pharmacoecon. Open 7, 997–1006 26. Calon, A. et al. (2015) Stromal gene expression defines poor-
6. Alwers, E. et al. (2019) External validation of molecular subtype prognosis subtypes in colorectal cancer. Nat. Genet. 47, 320–329
classifications of colorectal cancer based on microsatellite 27. Moffitt, R.A. et al. (2015) Virtual microdissection identifies dis-
instability, CIMP, BRAF and KRAS. BMC Cancer 19, 681 tinct tumor- and stroma-specific subtypes of pancreatic ductal
7. Berg, M. et al. (2015) Molecular subtypes in stage II-III colon adenocarcinoma. Nat. Genet. 47, 1168–1178
cancer defined by genomic instability: early recurrence-risk 28. Biankin, A.V. et al. (2012) Pancreatic cancer genomes reveal
associated with a high copy-number variation and loss of aberrations in axon guidance pathway genes. Nature 491,
RUNX3 and CDKN2A. PLoS One 10, e0122391 399–405
8. Zhao, Q. et al. (2022) Comprehensive profiling of 1015 patients' 29. Iacobuzio-Donahue, C.A. et al. (2003) Exploration of global
exomes reveals genomic-clinical associations in colorectal gene expression patterns in pancreatic adenocarcinoma using
cancer. Nat. Commun. 13, 2342 cDNA microarrays. Am. J. Pathol. 162, 1151–1162
9. Cancer Genome Atlas Network (2012) Comprehensive molecu- 30. Birnbaum, D.J. et al. (2021) Transcriptomic analysis of laser
lar characterization of human colon and rectal cancer. Nature capture microdissected tumors reveals cancer- and stromal-
487, 330–337 specific molecular subtypes of pancreatic ductal adenocarcinoma.
10. Kopetz, S. et al. (2019) Encorafenib, binimetinib, and cetuximab Clin. Cancer Res. 27, 2314–2325
in BRAF V600E-mutated colorectal cancer. N. Engl. J. Med. 31. Maurer, C. et al. (2019) Experimental microdissection enables
381, 1632–1643 functional harmonisation of pancreatic cancer subtypes. Gut
11. Prahallad, A. et al. (2012) Unresponsiveness of colon cancer to 68, 1034–1043
BRAF(V600E) inhibition through feedback activation of EGFR. 32. Suvà, M.L. and Tirosh, I. (2019) Single-cell RNA sequencing in
Nature 483, 100–103 cancer: lessons learned and emerging challenges. Mol. Cell
12. Corless, C.L. et al. (2011) Gastrointestinal stromal tumours: 75, 7–12
origin and molecular oncology. Nat. Rev. Cancer 11, 865–878 33. Zhang, Y. et al. (2021) Single-cell RNA sequencing in cancer
13. Bijlsma, M.F. et al. (2017) Molecular subtypes in cancers of the research. J. Exp. Clin. Cancer Res. 40, 81
gastrointestinal tract. Nat. Rev. Gastroenterol. Hepatol. 14, 34. Zhang, M. et al. (2021) Dissecting transcriptional heterogeneity
333–342 in primary gastric adenocarcinoma by single cell RNA sequencing.
14. Wang, X. et al. (2013) Dissecting cancer heterogeneity–an Gut 70, 464–475
unsupervised classification approach. Int. J. Biochem. Cell 35. Joanito, I. et al. (2022) Single-cell and bulk transcriptome
Biol. 45, 2574–2579 sequencing identifies two epithelial tumor cell states and refines
15. Guinney, J. et al. (2015) The consensus molecular subtypes of the consensus molecular classification of colorectal cancer.
colorectal cancer. Nat. Med. 21, 1350–1356 Nat. Genet. 54, 963–975
16. Bailey, P. et al. (2016) Genomic analyses identify molecular 36. Salazar, R. et al. (2011) Gene expression signature to improve
subtypes of pancreatic cancer. Nature 531, 47–52 prognosis prediction of stage II and III colorectal cancer.
17. Cancer Genome Atlas Research Network (2014) Comprehensive J. Clin. Oncol. 29, 17–24
molecular characterization of gastric adenocarcinoma. Nature 37. Yothers, G. et al. (2013) Validation of the 12-gene colon cancer
513, 202–209 recurrence score in NSABP C-07 as a predictor of recurrence in
18. Goodwin, S. et al. (2016) Coming of age: ten years of next- patients with stage II and III colon cancer treated with fluoroura-
generation sequencing technologies. Nat. Rev. Genet. 17, cil and leucovorin (FU/LV) and FU/LV plus oxaliplatin. J. Clin.
333–351 Oncol. 31, 4512–4519
19. Satam, H. et al. (2023) Next-generation sequencing technology: 38. Jorissen, R.N. et al. (2009) Metastasis-associated gene
current trends and advancements. Biology (Basel) 12, 997 expression changes predict poor outcomes in patients with

Trends in Cancer, September 2024, Vol. 10, No. 9 853

 Trends in Cancer
OPEN ACCESS

Dukes Stage B and C colorectal cancer. Clin. Cancer Res. tumor heterogeneity of colorectal cancer. Nat. Commun. 15,
15, 7642–7651 4342
39. Li, A. et al. (2009) Unsupervised analysis of transcriptomic pro- 61. Isella, C. et al. (2017) Selective analysis of cancer-cell intrinsic
files reveals six glioma subtypes. Cancer Res. 69, 2091–2099 transcriptional traits defines novel clinically relevant subtypes
40. Ten Hoorn, S. et al. (2022) Clinical value of consensus molecu- of colorectal cancer. Nat. Commun. 8, 15107
lar subtypes in colorectal cancer: a systematic review and 62. Malla, S.B. et al. (2024) Pathway level subtyping identifies a
meta-analysis. J. Natl. Cancer Inst. 114, 503–516 slow-cycling biological phenotype associated with poor clinical
41. Van Cutsem, E. and Ducreux, M. (2016) Colorectal and gastric outcomes in colorectal cancer. Nat. Genet. 56, 458–472
cancer in 2015: the development of new agents and molecular 63. Higashi, H. et al. (2002) Biological activity of the Helicobacter
classifications. Nat. Rev. Clin. Oncol. 13, 69–70 pylori virulence factor CagA is determined by variation in the ty-
42. Takahashi, S. et al. (2021) Advanced colorectal cancer sub- rosine phosphorylation sites. Proc. Natl. Acad. Sci. U. S. A. 99,
types (aCRCS) help select oxaliplatin-based or irinotecan- 14428–14433
based therapy for colorectal cancer. Cancer Sci. 112, 1567–1578 64. Lauren, P. (1965) The two histological main types of gastric
43. Zheng, Y. et al. (2023) CSTF2 mediated mRNA N(6)- carcinoma: diffuse and so-called intestinal-type carcinoma. An
methyladenosine modification drives pancreatic ductal adeno- attempt at a histo-clinical classification. Acta Pathol. Microbiol.
carcinoma m(6)A subtypes. Nat. Commun. 14, 6334 Scand. 64, 31–49
44. Cheong, J.H. et al. (2022) Development and validation of a 65. Wagner, A.D. et al. (2006) Chemotherapy in advanced gastric
prognostic and predictive 32-gene signature for gastric cancer. cancer: a systematic review and meta-analysis based on aggre-
Nat. Commun. 13, 774 gate data. J. Clin. Oncol. 24, 2903–2909
45. Zhou, R. et al. (2023) Expression pattern of secretory-cell- 66. Tan, I.B. et al. (2011) Intrinsic subtypes of gastric cancer, based
related transcriptional signatures in colon adenocarcinomas de- on gene expression pattern, predict survival and respond differ-
fines tumor microenvironment characteristics and correlates ently to chemotherapy. Gastroenterology 141, 476–485
with clinical outcomes. Mol. Oncol. 17, 499–517 67. Lei, Z. et al. (2013) Identification of molecular subtypes of gas-
46. Buechler, S.A. et al. (2020) ColoType: a forty gene signature for tric cancer with different responses to PI3-kinase inhibitors
consensus molecular subtyping of colorectal cancer tumors and 5-fluorouracil. Gastroenterology 145, 554–565
using whole-genome assay or targeted RNA-sequencing. Sci. 68. Oh, S.C. et al. (2018) Clinical and genomic landscape of gastric
Rep. 10, 12123 cancer with a mesenchymal phenotype. Nat. Commun. 9, 1777
47. Shinto, E. et al. (2020) A validation study for recurrence risk 69. Jang, E. et al. (2023) Clinical molecular subtyping reveals intrin-
stratification of Stage II colon cancer using the 55-gene classi- sic mesenchymal reprogramming in gastric cancer cells. Exp.
fier. Oncology 98, 534–541 Mol. Med. 55, 974–986
48. Morris, J.S. et al. (2021) Development and validation of a gene 70. Cristescu, R. et al. (2015) Molecular analysis of gastric cancer
signature classifier for consensus molecular subtyping of colo- identifies subtypes associated with distinct clinical outcomes.
rectal carcinoma in a CLIA-certified setting. Clin. Cancer Res. Nat. Med. 21, 449–456
27, 120–130 71. Tanaka, Y. et al. (2021) Multi-omic profiling of peritoneal metas-
49. Piskol, R. et al. (2019) A clinically applicable gene-expression tases in gastric cancer identifies molecular subtypes and thera-
classifier reveals intrinsic and extrinsic contributions to consen- peutic vulnerabilities. Nat. Can. 2, 962–977
sus molecular subtypes in primary and metastatic colon cancer. 72. Miyazawa, M. et al. (2016) Gastric adenocarcinoma of the fun-
Clin. Cancer Res. 25, 4431–4442 dic gland (chief cell-predominant type): a review of endoscopic
50. Gonzalez, R.S. et al. (2016) Immunohistochemistry as a surro- and clinicopathological features. World J. Gastroenterol. 22,
gate for molecular subtyping of gastric adenocarcinoma. 10523–10531
Hum. Pathol. 56, 16–21 73. Mizrahi, J.D. et al. (2020) Pancreatic cancer. Lancet 395,
51. Trinh, A. et al. (2017) Practical and robust identification of mo- 2008–2020
lecular subtypes in colorectal cancer by immunohistochemistry. 74. Elahi-Gedwillo, K.Y. et al. (2019) Antifibrotic therapy disrupts
Clin. Cancer Res. 23, 387–398 stromal barriers and modulates the immune landscape in pan-
52. Muckenhuber, A. et al. (2018) Pancreatic ductal adenocarci- creatic ductal adenocarcinoma. Cancer Res. 79, 372–386
noma subtyping using the biomarkers hepatocyte nuclear 75. Waddell, N. et al. (2015) Whole genomes redefine the muta-
factor-1A and cytokeratin-81 correlates with outcome and tional landscape of pancreatic cancer. Nature 518, 495–501
treatment response. Clin. Cancer Res. 24, 351–359 76. Jones, S. et al. (2008) Core signaling pathways in human pan-
53. Saillard, C. et al. (2023) Pacpaint: a histology-based deep learn- creatic cancers revealed by global genomic analyses. Science
ing model uncovers the extensive intratumor molecular hetero- 321, 1801–1806
geneity of pancreatic adenocarcinoma. Nat. Commun. 14, 77. Collisson, E.A. et al. (2011) Subtypes of pancreatic ductal ade-
3459 nocarcinoma and their differing responses to therapy. Nat.
54. Flinner, N. et al. (2022) Deep learning based on hematoxylin- Med. 17, 500–503
eosin staining outperforms immunohistochemistry in predicting 78. Aung, K.L. et al. (2018) Genomics-driven precision medicine for
molecular subtypes of gastric adenocarcinoma. J. Pathol. 257, advanced pancreatic cancer: early results from the COMPASS
218–226 trial. Clin. Cancer Res. 24, 1344–1354
55. Sirinukunwattana, K. et al. (2021) Image-based consensus mo- 79. Chan-Seng-Yue, M. et al. (2020) Transcription phenotypes of
lecular subtype (imCMS) classification of colorectal cancer pancreatic cancer are driven by genomic events during tumor
using deep learning. Gut 70, 544–554 evolution. Nat. Genet. 52, 231–240
56. Sung, H. et al. (2021) Global Cancer Statistics 2020: 80. Dijk, F. et al. (2020) Unsupervised class discovery in pancreatic
GLOBOCAN estimates of incidence and mortality worldwide ductal adenocarcinoma reveals cell-intrinsic mesenchymal
for 36 cancers in 185 countries. CA Cancer J. Clin. 71, features and high concordance between existing classification
209–249 systems. Sci. Rep. 10, 337
57. Sveen, A. et al. (2019) Predictive modeling in colorectal cancer: 81. Puleo, F. et al. (2018) Stratification of pancreatic ductal adeno-
time to move beyond consensus molecular subtypes. Ann. carcinomas based on tumor and microenvironment features.
Oncol. 30, 1682–1685 Gastroenterology 155, 1999–2013 e1993
58. Dunne, P.D. et al. (2016) Challenging the cancer molecular 82. Sivakumar, S. et al. (2017) Master regulators of oncogenic
stratification dogma: intratumoral heterogeneity undermines KRAS response in pancreatic cancer: an integrative network
consensus molecular subtypes and potential diagnostic value biology analysis. PLoS Med. 14, e1002223
in colorectal cancer. Clin. Cancer Res. 22, 4095–4104 83. de Santiago, I. et al. (2019) Immunophenotypes of pancreatic
59. Dunne, P.D. et al. (2017) Cancer-cell intrinsic gene expression ductal adenocarcinoma: meta-analysis of transcriptional sub-
signatures overcome intratumoural heterogeneity bias in colo- types. Int. J. Cancer 145, 1125–1137
rectal cancer patient classification. Nat. Commun. 8, 15657 84. Shi, X. et al. (2022) Integrated profiling of human pancreatic
60. Langerud, J. et al. (2024) Multiregional transcriptomics identifies cancer organoids reveals chromatin accessibility features asso-
congruent consensus subtypes with prognostic value beyond ciated with drug sensitivity. Nat. Commun. 13, 2169

854 Trends in Cancer, September 2024, Vol. 10, No. 9

Trends in Cancer
OPEN ACCESS

85. Hwang, W.L. et al. (2022) Single-nucleus and spatial tran- 108. Schwarzmueller, L. et al. (2020) Emerging role and therapeutic
scriptome profiling of pancreatic cancer identifies multicellular potential of lncRNAs in colorectal cancer. Cancers (Basel) 12,
dynamics associated with neoadjuvant treatment. Nat. Genet. 3843
54, 1178–1191 109. Weng, S. et al. (2023) Epigenetically regulated gene expression
86. Raghavan, S. et al. (2021) Microenvironment drives cell state, profiles decipher four molecular subtypes with prognostic and
plasticity, and drug response in pancreatic cancer. Cell 184, therapeutic implications in gastric cancer. Clin. Epigenetics 15,
6119–6137 64
87. Hoadley, K.A. et al. (2018) Cell-of-origin patterns dominate 110. Namkung, J. et al. (2016) Molecular subtypes of pancreatic
the molecular classification of 10,000 tumors from 33 types of cancer based on miRNA expression profiles have independent
cancer. Cell 173, 291–304 prognostic value. J. Gastroenterol. Hepatol. 31, 1160–1167
88. Liu, Y. et al. (2018) Comparative molecular analysis of gastroin- 111. Chen, H. et al. (2014) Long noncoding RNA profiles identify
testinal adenocarcinomas. Cancer Cell 33, 721–735 five distinct molecular subtypes of colorectal cancer with clinical
89. Adam, R.S. et al. (2020) Intestinal region-specific Wnt signalling relevance. Mol. Oncol. 8, 1393–1403
profiles reveal interrelation between cell identity and oncogenic 112. Chen, Y. et al. (2021) Classifying gastric cancer using FLORA
pathway activity in cancer development. Cancer Cell Int. 20, reveals clinically relevant molecular subtypes and highlights
578 LINC01614 as a biomarker for patient prognosis. Oncogene
90. Woolston, A. et al. (2019) Genomic and transcriptomic determi- 40, 2898–2909
nants of therapy resistance and immune landscape evolution 113. Arnes, L. et al. (2019) Comprehensive characterisation of
during anti-EGFR treatment in colorectal cancer. Cancer Cell compartment-specific long non-coding RNAs associated with
36, 35–50 pancreatic ductal adenocarcinoma. Gut 68, 499–511
91. Vermeulen, C. et al. (2023) Ultra-fast deep-learned CNS tumour 114. De Sousa, E.M.F. et al. (2013) Poor-prognosis colon cancer is
classification during surgery. Nature 622, 842–849 defined by a molecularly distinct subtype and develops from
92. de Back, T.R. et al. (2024) A consensus molecular subtypes serrated precursor lesions. Nat. Med. 19, 614–618
classification strategy for clinical colorectal cancer tissues. Life 115. Zhang, H. et al. (2023) Comprehensive transcriptomic analyses
Sci. Alliance 7, e202402730 identify KDM genes-related subtypes with different TME infil-
93. Bahrambeigi, V. et al. (2024) Transcriptomic profiling of plasma trates in gastric cancer. BMC Cancer 23, 454
extracellular vesicles enables reliable annotation of the cancer- 116. Schlicker, A. et al. (2012) Subtypes of primary colorectal tumors
specific transcriptome and molecular subtype. Cancer Res. correlate with response to targeted treatment in colorectal cell
84, 1719–1732 lines. BMC Med. Genet. 5, 66
94. Setia, N. et al. (2016) A protein and mRNA expression-based 117. Sumithra, V. and Surendran, S. (2015) A review of various linear
classification of gastric cancer. Mod. Pathol. 29, 772–784 and non linear dimensionality reduction techniques. Int.
95. Koh, J. et al. (2019) Development and validation of an easy- J. Comput. Sci. Inf. Technol. 6, 2354–2360
to-implement, practical algorithm for the identification of 118. Brunet, J.P. et al. (2004) Metagenes and molecular pattern dis-
molecular subtypes of gastric cancer: prognostic and thera- covery using matrix factorization. Proc. Natl. Acad. Sci. U. S. A.
peutic implications. Oncologist 24, e1321–e1330 101, 4164–4169
96. Noll, E.M. et al. (2016) CYP3A5 mediates basal and acquired 119. Yeung, K.Y. and Ruzzo, W.L. (2001) Principal component
therapy resistance in different subtypes of pancreatic ductal analysis for clustering gene expression data. Bioinformatics
adenocarcinoma. Nat. Med. 22, 278–287 17, 763–774
97. Heo, Y.J. et al. (2019) Reproduction of molecular subtypes 120. Yang, Y. et al. (2021) Dimensionality reduction by UMAP rein-
of gastric adenocarcinoma by transcriptome sequencing of forces sample heterogeneity analysis in bulk transcriptomic
archival tissue. Sci. Rep. 9, 9675 data. Cell Rep. 36, 109442
98. Rashid, N.U. et al. (2020) Purity Independent Subtyping of 121. Monti, S. et al. (2003) Consensus clustering: a resampling-
Tumors (PurIST), a clinically robust, single-sample classifier for based method for class discovery and visualization of gene
tumor subtyping in pancreatic cancer. Clin. Cancer Res. 26, expression microarray data. Mach. Learn. 52, 91–118
82–92 122. Agesen, T.H. et al. (2012) ColoGuideEx: a robust gene classifier
99. Wang, Y. et al. (2022) DEMoS: a deep learning-based ensem- specific for stage II colorectal cancer prognosis. Gut 61,
ble approach for predicting the molecular subtypes of gastric 1560–1567
adenocarcinomas from histopathological images. Bioinformatics 123. Sveen, A. et al. (2012) ColoGuidePro: a prognostic 7-gene
38, 4206–4213 expression signature for stage III colorectal cancer patients.
100. Kalimuthu, S.N. et al. (2020) Morphological classification of Clin. Cancer Res. 18, 6001–6010
pancreatic ductal adenocarcinoma that predicts molecular 124. Boussioutas, A. et al. (2003) Distinctive patterns of gene
subtypes and correlates with clinical outcome. Gut 69, expression in premalignant gastric mucosa and gastric cancer.
317–328 Cancer Res. 63, 2569–2577
101. Cascianelli, S. et al. (2023) Multi-label transcriptional classifi- 125. Shah, M.A. et al. (2011) Molecular classification of gastric
cation of colorectal cancer reflects tumor cell population cancer: a new paradigm. Clin. Cancer Res. 17, 2693–2701
heterogeneity. Genome Med. 15, 37 126. Donahue, T.R. et al. (2012) Integrative survival-based molecular
102. Penland, S.K. et al. (2007) RNA expression analysis of profiling of human pancreatic cancer. Clin. Cancer Res. 18,
formalin-fixed paraffin-embedded tumors. Lab. Investig. 87, 1352–1363
383–391 127. Haider, S. et al. (2014) A multi-gene signature predicts outcome
103. Masuda, N. et al. (1999) Analysis of chemical modification of in patients with pancreatic ductal adenocarcinoma. Genome
RNA from formalin-fixed samples and optimization of molecular Med. 6, 105
biology applications for such samples. Nucleic Acids Res. 27, 128. Kirby, M.K. et al. (2016) RNA sequencing of pancreatic adeno-
4436–4443 carcinoma tumors yields novel expression patterns associated
104. Li, J. et al. (2018) Accurate RNA sequencing from formalin-fixed with long-term survival and reveals a role for ANGPTL4. Mol.
cancer tissue to represent high-quality transcriptome from Oncol. 10, 1169–1182
frozen tissue. JCO Precis. Oncol. 2018, PO.17.00091 129. Pages, F. et al. (2018) International validation of the consensus
105. Hedegaard, J. et al. (2014) Next-generation sequencing of RNA Immunoscore for the classification of colon cancer: a prognostic
and DNA isolated from paired fresh-frozen and formalin-fixed and accuracy study. Lancet 391, 2128–2139
paraffin-embedded samples of human cancer and normal 130. Antoniotti, C. et al. (2022) Upfront FOLFOXIRI plus bevacizu-
tissue. PLoS One 9, e98187 mab with or without atezolizumab in the treatment of patients
106. Lu, J. et al. (2005) MicroRNA expression profiles classify human with metastatic colorectal cancer (AtezoTRIBE): a multicentre,
cancers. Nature 435, 834–838 open-label, randomised, controlled, phase 2 trial. Lancet
107. Adam, R.S. et al. (2022) Development of a miRNA-based Oncol. 23, 876–887
classifier for detection of colorectal cancer molecular subtypes. 131. Brunac, A.C. et al. (2022) CINSARC signature outperforms
Mol. Oncol. 16, 2693–2709 gold-standard TNM staging and consensus molecular

Trends in Cancer, September 2024, Vol. 10, No. 9 855

 Trends in Cancer
OPEN ACCESS

subtypes for clinical outcome in stage II-III colorectal carcinoma. 145. Budinska, E. et al. (2013) Gene expression patterns unveil a
Mod. Pathol. 35, 2002–2010 new level of molecular heterogeneity in colorectal cancer.
132. Luo, W. et al. (2022) Bulk and single-cell transcriptome profiling J. Pathol. 231, 63–76
reveal necroptosis-based molecular classification, tumor micro- 146. Roepman, P. et al. (2014) Colorectal cancer intrinsic subtypes
environment infiltration characterization, and prognosis predic- predict chemotherapy benefit, deficient mismatch repair and
tion in colorectal cancer. J. Transl. Med. 20, 235 epithelial-to-mesenchymal transition. Int. J. Cancer 134,
133. Bao, X. et al. (2023) An immunometabolism subtyping system 552–562
identifies S100A9(+) macrophage as an immune therapeutic 147. Eide, P.W. et al. (2017) CMScaller: an R package for consensus
target in colorectal cancer based on multiomics analysis. Cell molecular subtyping of colorectal cancer pre-clinical models.
Rep. Med. 4, 100987 Sci. Rep. 7, 16618
134. Wang, R. et al. (2020) Multiplex profiling of peritoneal metasta- 148. Sveen, A. et al. (2018) Colorectal cancer consensus molecular
ses from gastric adenocarcinoma identified novel targets and subtypes translated to preclinical models uncover potentially
molecular subtypes that predict treatment response. Gut 69, targetable cancer cell dependencies. Clin. Cancer Res. 24,
18–31 794–806
135. Ma, S. et al. (2023) Integrated analysis identities Rho GTPases 149. Amirkhah, R. et al. (2023) MmCMS: mouse models' consensus
related molecular map in patients with gastric carcinoma. Sci. molecular subtypes of colorectal cancer. Br. J. Cancer 128,
Rep. 13, 21443 1333–1343
136. Zhou, Q. et al. (2023) Cancer functional states-based molecular 150. Komor, M.A. et al. (2018) Consensus molecular subtype classi-
subtypes of gastric cancer. J. Transl. Med. 21, 80 fication of colorectal adenomas. J. Pathol. 246, 266–276
137. Xu, X. et al. (2022) The ferroptosis molecular subtype reveals 151. van de Weerd, S. et al. (2023) Consensus molecular subtype
characteristics of the tumor microenvironment, immunothera- transition during progression of colorectal cancer. J. Pathol.
peutic response, and prognosis in gastric cancer. Int. J. Mol. 261, 298–308
Sci. 23, 9767 152. Pitroda, S.P. et al. (2018) Integrated molecular subtyping
138. Zhang, C. et al. (2022) Development of cancer-associated fibro- defines a curable oligometastatic state in colorectal liver
blasts subtype and prognostic model in gastric cancer and the metastasis. Nat. Commun. 9, 1793
landscape of tumor microenvironment. Int. J. Biochem. Cell 153. Lenos, K.J. et al. (2022) Molecular characterization of colorectal
Biol. 152, 106309 cancer related peritoneal metastatic disease. Nat. Commun.
139. Xu, D. et al. (2023) Identification of the molecular subtype and 13, 4443
prognostic characteristics of pancreatic cancer based on CD8 154. Marisa, L. et al. (2021) Intratumor CMS heterogeneity impacts
+ T cell-related genes. Cancer Immunol. Immunother. 72, patient prognosis in localized colon cancer. Clin. Cancer Res.
647–664 27, 4768–4780
140. Zheng, H. et al. (2023) Single-cell and bulk RNA sequencing 155. Eide, P.W. et al. (2021) Metastatic heterogeneity of the consen-
identifies T cell marker genes score to predict the prognosis sus molecular subtypes of colorectal cancer. NPJ Genom Med.
of pancreatic ductal adenocarcinoma. Sci. Rep. 13, 3684 6, 59
141. Du, Y. et al. (2023) Integration of single-cell RNA sequencing 156. Ten Hoorn, S. et al. (2018) Classification of colorectal cancer in
and bulk RNA sequencing reveals that TAM2-driven genes molecular subtypes by immunohistochemistry. Methods Mol.
affect immunotherapeutic response and prognosis in pancreatic Biol. 1765, 179–191
cancer. Int. J. Mol. Sci. 24, 12787 157. Li, X. et al. (2021) A modified protein marker panel to identify
142. George, B. et al. (2024) Transcriptomic-based microenviron- four consensus molecular subtypes in colorectal cancer using
ment classification reveals precision medicine strategies for immunohistochemistry. Pathol. Res. Pract. 220, 153379
pancreatic ductal adenocarcinoma. Gastroenterology 166, 158. Kasurinen, J. et al. (2023) Phenotypic subtypes predict out-
859–871 comes in colorectal cancer. Acta Oncol. 62, 245–252
143. Marisa, L. et al. (2013) Gene expression classification of colon 159. Yoo, S.Y. et al. (2020) Whole-slide image analysis reveals
cancer into molecular subtypes: characterization, validation, quantitative landscape of tumor-immune microenvironment in
and prognostic value. PLoS Med. 10, e1001453 colorectal cancers. Clin. Cancer Res. 26, 870–881
144. Sadanandam, A. et al. (2013) A colorectal cancer classification 160. Nguyen, H.G. et al. (2022) Image-based assessment of extra-
system that associates cellular phenotype and responses to cellular mucin-to-tumor area predicts consensus molecular
therapy. Nat. Med. 19, 619–625 subtypes (CMS) in colorectal cancer. Mod. Pathol. 35, 240–248

856 Trends in Cancer, September 2024, Vol. 10, No. 9