Introduction

Cell plasticity is a crucial feature of cancer progression and metastasis. This plasticity enables tumour cells to adapt to the ever-changing conditions occurring throughout the disease trajectory, from the initiation of a primary tumour followed by dissemination to and colonization of distant organs and the outgrowth of distant metastases. Research over the past 20–30 years has revealed that dynamic activation of the epithelial–mesenchymal transition (EMT) programme is one of the major processes underlying cancer cell plasticity1.

EMT originally referred to the transdifferentiation of locally confined epithelial cells into a mesenchymal, motile phenotype, initially observed during early embryonic development2, in which EMT contributes to processes such as gastrulation, neural crest formation and heart development3. This programme is also essential for various physiological processes in adult tissues, for example, wound healing and tissue homeostasis3. Pathological reactivation of EMT has a pivotal role in several pathological processes, including fibrosis and the development and progression of cancer.

The EMT programme is primarily exerted by a core set of EMT-activating transcription factors (EMT-TFs), including SNAIL (SNAI1), SLUG (SNAI2), the basic helix–loop–helix factors TWIST1 (TWIST) and TWIST2, and the zinc finger E-box-binding homeobox factors ZEB1 and ZEB2 (ref. 3) (Fig. 1a). In cancers of epithelial cell origins (carcinomas), ‘classical’ EMT can often be observed in the invasive regions of primary tumours. Activation of EMT is reflected in a loss of epithelial integrity via the downregulation of genes responsible for cell–cell adhesion, apical–basal polarity and basement membrane formation. Notably, cancer cells rarely acquire a fully mesenchymal state but often undergo partial and transient EMT, resulting in a hybrid epithelial/mesenchymal (E/M) phenotype, which maintains both epithelial and mesenchymal properties and expression of established markers associated with both of these phenotypes. Throughout this Review, we refer to this state as the hybrid E/M state. The EMT programme also activates genes that enable cancer cell motility and invasion of the surrounding tissues as well as dissemination through (intravasation into and extravasation out of) blood and lymphatic vessels. On reaching distant anatomical locations, cancer cells can dynamically adapt to new environmental conditions. In this phase, the EMT programme is often reversed, and the underlying mesenchymal–epithelial transition (MET) promotes the proliferation and outgrowth of metastases.

Fig. 1: Central mediators, properties and functions in EMP.
figure 1

a, Schematic showing crucial effectors and functions of cancer cell plasticity (epithelial–mesenchymal plasticity (EMP)) exerted by the epithelial–mesenchymal transition (EMT) programme and the reverse mesenchymal–epithelial transition (MET) programme. Epithelial cancer cells rarely transition to a fully mesenchymal phenotype but rather exist in dynamic intermediate states characterized by a mixture of certain mesenchymal as well as epithelial features (referred to as a hybrid epithelial/mesenchymal (E/M) cell state), which is often also characterized by the highest levels of cellular plasticity and stemness. b, Overview summarizing the multiple oncogenic functions of enhanced EMP over the course of tumour progression. Classical EMT-associated properties enable cancer cells to migrate, invade, and intravasate and extravasate from blood vessels. MET enables the colonization of anatomical locations distant from the primary tumour as well as the outgrowth of metastatic lesions. Non-classical EMT traits can also support tumour initiation and support cancer cells in adapting to the changing conditions created by metabolic reprogramming, enhance their survival via upregulation of DNA repair and prevention of cell death, and promote immune evasion and resistance to chemotherapy and radiotherapy. Importantly, EMP is regulated by environmental signals arising from several sources, including cancer-associated fibroblasts or immune cells, or via effects of stress factors such as hypoxia or therapeutic intervention. CTCs, circulating tumour cells; DTCs, disseminated tumour cells.

Full size image

This dynamic and reversible nature of the EMT programme, integrating both EMT and the reverse process of MET, is detectable in many tumours, in which the hybrid E/M state is activated only transiently. This scenario is best described by the term epithelial–mesenchymal plasticity (EMP)1,4 and provides a driving force for cancer evolution. Notably, EMT-TFs have crucial roles not only in classical EMT but also in many additional processes, including neural, neuroendocrine, haematopoietic, and mesenchymal cellular differentiation and homeostasis1,5,6. Thus, the potency of EMT-TFs in regulating crucial cellular processes goes beyond EMT, including additional roles in supporting cancer progression. EMT-TFs already contribute to the early stages of tumour initiation by conferring stemness properties that later also facilitate metastatic colonization. Throughout tumour progression, EMT-TFs enable cancer cells to survive and adapt to fluctuating conditions via metabolic reprogramming, altered DNA repair, immune evasion and enhanced survival capabilities, all of which can also have a role in treatment resistance. Notably, data published in March 2025 demonstrate that EMT also favours the generation of genetic alterations, resulting in the emergence of genomically unstable cancer cells7. Thus, we consider cancer cell plasticity to be the best term to comprehensively refer to all EMT-TF-mediated traits and their overall effects1.

Crucially, the EMT programme not only enables cancer cells to dynamically adapt to environmental factors but can also be induced by various environmental signals, such as those arising from cancer-associated fibroblasts (CAFs) or immune cells, and therapeutic interventions5,8,9,10,11,12 (Fig. 1b). Prominent environmental factors capable of activating EMT include the growth factors TGFβ, PDGF and HGF as well as hypoxia and many other cellular stress factors3,5,8,9,13,14,15. Environmental factors capable of inducing a reversion of the process (MET) in cancer cells are less well understood, with members of the bone morphogenetic protein family (such as BMP7) reported to be one of the most prominent mediators16 (Fig. 1a).

In summary, in addition to accumulating genetic alterations, an aberrantly high level of cancer cell plasticity, which often reflects transient and partial activation of EMT, is considered a second major driving force of cancer evolution from tumour initiation to metastatic progression. However, in contrast to irreversible genetic changes, those associated with EMT are plastic and are therefore potentially reversible. In this Review, we summarize the current clinical evidence supporting a role of EMT and EMP in cancer and discuss subsequent clinical opportunities and challenges.

Clinical evidence for EMT in cancer

Substantial evidence that activation of EMT correlates with disease progression and adverse clinical outcomes in patients with cancer has accumulated over the past two decades. EMT is a multifaceted process and is often only transiently and partially active, as reflected in context-dependent individual changes in gene expression. Hence, the identification of EMT cannot rely only on a few biomarkers, especially in clinical settings1. Several EMT signatures and scoring systems have been developed to overcome this challenge17,18. Novel single-cell analysis techniques, combined with spatial information (spatial transcriptomics and/or proteomics), have become extremely useful as methods of confirming the presence, prevalence and clinical relevance of EMT. This utility is evident in pan-cancer single-cell RNA sequencing analyses capable of quantifying the activity of all genes in thousands of individual cells from the same specimen. These studies have included samples obtained from primary tumours and metastatic lesions and have revealed direct associations between the activation of common hybrid E/M and cellular plasticity programmes19,20,21,22. A detailed 3D spatial transcriptomics approach has combined spatial information with single-cell transcriptomics to identify cellular transition states corresponding to classical morphologies, such as the transitions from glandular to mucinous or mucinous to solid, involving EMT in samples from patients with colorectal cancer (CRC)23. Nonetheless, the transient and hybrid nature of EMT in cancer remains a challenge when attempting to standardize clinical assessments24.

Looking beyond simple survival correlations, data from manifold experimental studies involving different model systems and from clinical studies have demonstrated the role of EMT activation in all stages of cancer progression, including initiation, the formation of precursor lesions, invasion, dissemination, metastasis and, importantly, treatment resistance. Seminal experimental studies, including mouse models of breast25,26,27,28,29, pancreatic30,31,32,33, skin34,35,36 and lung cancer37, were the basis for demonstrating the role of EMT and EMP in various human cancer types (Table 1).

Table 1 Overview of clinical evidence supporting a role of epithelial–mesenchymal transition in specific tumour entities
Full size table

Tumour initiation and precursor lesions

Abundant experimental evidence exists that expression of EMT-TFs is associated with both the onset of tumorigenesis and the early stages of tumour progression. This tumour-initiating capacity most probably reflects the stemness properties associated with partial EMT, thereby indicating a role in cancer cell stemness3,5,38,39,40,41,42. Moreover, data from a mouse model of pancreatic cancer indicate that EMT favours the acquisition of both mutations and genomic instability and that this feature might also support all stages of tumour progression7. Precursor lesions are often characterized by increased cancer cell proliferation and consequent tumour growth. In this context, activation of EMT is not always linked with proliferation arrest (such as in dormancy states), with distinct forms of EMT also being associated with hyperproliferation (as reviewed in detail elsewhere43). Verifying the participation of EMT in the earliest stages of the initiation of human tumours is often difficult owing to a lack of tumour tissue samples obtained specifically at this stage. However, multiple examples indicate that EMT can also participate in the formation of benign precursor lesions and the extent of activation increases during malignant progression (Table 1). In pancreatic cancer, both pancreas intraepithelial neoplasia and intraductal papillary mucinous neoplasm precursor lesions (associated with KRAS alterations) can show evidence of partial EMT with attendant expression of EMT-TFs44,45,46,47,48,49. Similarly, an analysis of premalignant CRC lesions (associated with APC mutations) found increased expression of EMT mRNA signatures, including expression of vimentin, E-cadherin and TGFβ in colonic mucosal samples obtained from patients with ulcerative colitis50, in which areas of active inflammation, which carry the highest risk of transformation, had higher levels of EMT activation51. Elsewhere, expression of the EMT-associated proteins SNAIL and TWIST has been detected in adenomas but not in the non-malignant colonic mucosa52. Notably, initial mutations found in precursor lesions (including KRAS and APC mutations) can support EMT activation, even during these very early stages of tumour development. For example, higher expression of TWIST, SNAIL and SLUG has been detected in tissue samples obtained from women with invasive cervical intraepithelial neoplasia grade 3 (CIN3) versus those with CIN2 and CIN1 lesions53,54,55. Additionally, expression of HPV16-related oncogenes has been shown to induce TWIST2 and vimentin in cervical cells as well as to promote invasion54. Similarly, a stepwise reduction in levels of the EMT-suppressing microRNA (miR)-200 family has been observed moving from benign oesophageal epithelium, through to benign Barrett epithelium (reflecting inflammation typically arising from gastroesophageal reflux) to high-grade dysplastic Barrett lesions56. Mechanistically, this observation might reflect upregulation of FOXF1, which induces EMT and a Barrett-like phenotype in non-neoplastic human squamous oesophageal cell lines57. A similar gradual increase in EMT marker expression and SNAIL activation has also been detected during transition from benign endometrial epithelium through to endometrial glandular dysplasia and serous endometrial intraepithelial carcinoma in samples obtained from women with endometrial cancer58,59. EMT is also prevalent in fibrosis, a precursor to cancer in various organs, including the liver, kidneys, lungs and breasts14. Targeting these relationships could ultimately provide opportunities for cancer prevention.

Invasion

EMT has been implicated in the transition from pre-invasive disease to invasive malignancy in several cancer types. Various stimuli, including hypoxia, have been reported to activate EMT in cancer cells, particularly at the invasive front of tumours60. Hypoxia in the invading primary tumour can even influence the fate of subsequent disseminated tumour cells (DTCs), which acquire an EMT-associated, therapy-resistant dormancy phenotype, as demonstrated in mouse models of breast cancer61. Initial observations in samples obtained from patients with CRC indicate morphological changes reminiscent of EMT and correlating with invasion and metastasis. Here, E-cadherin-low, mesenchymal-like budding cancer cells were identified at the invasive front, whereas cancer cells present in distant metastases tended to have an epithelial morphology similar to that of the centre of the primary tumour, suggesting a high level of plasticity as tumours progress towards metastasis62. A later study confirmed that such budding cancer cells often have a hybrid E/M state63. Evidence from patients with pancreatic ductal adenocarcinoma and those with oesophageal cancer demonstrates hybrid E/M induction at the tumour–host interface, with E-cadherin loss and high levels of TGFβ expression at the invasive edges of tumour buds64,65. Further examples of a role of EMT in tumour progression include progression from ductal carcinoma in situ to invasive breast cancer via TBX3-mediated upregulation of SLUG and TWIST66. Moreover, limited expression of microRNAs associated with suppression of EMT, such as miR-126 and miR-218 (ref. 67), and expression of EMT markers such as SNAIL and SPARC68 were predictive of disease progression from ductal carcinoma in situ to invasive breast cancer69,70. Similarly, in an analysis of samples from patients with carcinoma in situ of the bladder, invasiveness was associated with the loss of membranous E-cadherin and increased expression of SNAIL, SLUG and ZEB1 (ref. 71). In an analysis of circulating tumour cells (CTCs) from patients with non-muscle-invasive bladder cancer, ZEB1, TWIST1, TIMP2 and VIM (encoding vimentin) expression increased from stage Ta to T1. Furthermore, higher levels of TWIST1 expression correlated with inferior recurrence-free survival72. An analysis of samples from patients with oesophageal cancer revealed upregulation of SLUG in the transition from Barrett oesophagus to carcinoma73, with downregulation of miR-200 family members reported elsewhere56. Conversely, activation of EMT was not detected in premalignant cutaneous actinic keratoses, neither in individuals without cancer74 nor in transplant recipients75, but was detected in skin SCC samples in both populations. These observations might reflect loss of ZEB1 suppression by the EMT suppressors OVOL1 and OVOL2, which has been shown to lead to disease progression from actinic keratoses to invasive skin SCC76.

Malignant progression requires the activation of distinct EMT programmes. Data from studies involving the MMTV–PyMT mouse model of metastatic breast cancer demonstrate that one type of EMT-dedifferentiation programme facilitates cancer cell invasion in a PRRX1-dependent manner, whereas the activation of an adult, inflammatory EMT programme requiring SNAIL activity but not PRRX1 can contribute to inflammation and fibrosis77. Similarly, two individual EMT trajectories can be regulated by two distinct epigenetic chromatin modifiers, PRC2 and KMT2D–COMPASS. Separate inhibition of each modifier unlocks two distinct EMT phenotypes: a highly metastatic quasi-mesenchymal phenotype associated with markers of stemness and an inferior prognosis, and a fully transitioned mesenchymal phenotype associated with an improved prognosis78.

Circulating tumour cells

Invasive cancer cells can intravasate into the blood or lymphatic vessels and disseminate as CTCs, either as single cells or in small clusters3,8,79,80,81. Increasing evidence indicates that EMT activation in subgroups of CTCs facilitates successful dissemination82,83. This observation was originally demonstrated in breast cancer, in which CTCs in blood express EMT markers84, with higher levels of expression of TWIST and vimentin found in CTCs obtained from patients with metastatic disease relative to those with early-stage breast cancer85. Data from other studies indicate that CTCs from patients with breast or prostate cancer co-express both mesenchymal and epithelial genes, indicating a hybrid E/M state86,87,88,89. Similar findings have been described for lymphatic vessel CTCs, which also demonstrate a stem-like tumour-initiating capacity, in a mouse model of melanoma90.

Further research demonstrates that epithelial CTC clusters in cooperation with single hybrid E/M state CTCs confer the highest risk of metastatic dissemination79,91. Enhanced phenotypic plasticity of CTCs between epithelial and mesenchymal states is predictive of an inferior response to therapy and an inferior prognosis in patients with breast cancer79,91,92. This observation has subsequently been confirmed in CTCs obtained from patients with pancreatic cancer93, hepatocellular carcinoma94, CRC95,96, prostate cancer97, bladder cancer72,98 and small cell lung cancer99 as well as in a pan-cancer study, in which the presence of stem-like, EMT marker-positive CTCs was associated with an increased risk of metastatic disease and inferior survival100.

DTCs, latency, colonization and metastasis

CTCs leave the bloodstream and settle as DTCs in niches within organs, particularly the bone marrow9,101,102. DTCs can persist in a dormant state for many years, explaining the long latency between colonization and clinically evident metastases that is often seen in patients with prostate cancer and certain subtypes of breast cancer. Metastatic colonization from dormant DTCs requires an incompletely understood switch back to an active growth mode, often coupled with epithelial re-differentiation (MET)35,103, enabling the subsequent outgrowth of macrometastases. By contrast, the dormant DTC state is coupled with a hybrid E/M phenotype9,34,101,104, as already proposed on the basis of initial findings in CRC. Thereby, activation of certain EMT-associated factors, such as ZFP281, maintains the growth-arrested dormancy state105. The underlying stimuli that orchestrate DTC reactivation are an area of ongoing research interest. For example, in oestrogen receptor-positive breast cancer xenograft mice, a dormancy state accompanied by downregulation of E-cadherin and activation of ZEB1 and ZEB2 is initiated upon dissemination. Reactivation of E-cadherin expression can force these DTCs out of dormancy, resulting in the development of lung metastases104. In another study involving mouse xenograft models, luminal breast cancer cells obtained from organs with a high metastatic burden were shown to derive from DTCs with a stemness and EMT phenotype after epithelial re-differentiation coupled with an increased proliferative capacity106. Similar findings have been described for other cancer types. For example, dormancy in CRC is regulated by ZEB2 and is strongly associated with the mesenchymal CMS4 The Cancer Genome Atlas subtype, which has the worst prognosis of the four CRC subtypes107,108.

Substantial data on the role of EMT in DTC dormancy, re-differentiation and metastatic outgrowth have been obtained from animal models using genetic and/or molecular tracing techniques. However, obtaining direct evidence of this activity in patients with cancer is extremely difficult owing to the extensive phenotypic plasticity of cancer cells as well as the lack of reliable methods of detecting dormant DTCs in tissues. A number of studies have demonstrated the presence of DTCs in seemingly unaffected liver tissue samples obtained from patients with primary CRC or pancreatic cancer109. Bone marrow provides a more readily accessible source of potential DTCs110.

Clearly, an intervention with the ability to control or inhibit the switch from a dormant to a proliferative state could have substantial therapeutic benefit by blocking disease relapse. It is highly possible that the well-described mechanistic links between hybrid EMT, stemness and tumorigenic capacity associated with high levels of cellular plasticity are necessary for inducing dormancy but also to awaken DTCs as metastasis-initiating cancer cells.

Prognosis and survival

Statistically significant associations have been demonstrated between a wide range of EMT features and overall survival (OS), relapse-free survival, disease-free survival (DFS) and/or progression-free survival (PFS) (Supplementary Table 1). Notably, of the 99 outcome measures, 92 indicate an adverse effect of EMT on survival, of which 84 are statistically significant. The data show notable directional consistency with higher expression of EMT-TFs and markers of hybrid E/M or mesenchymal states correlating with inferior outcomes, and higher expression of inhibitors of EMT as well as markers of epithelial states associated with improved outcomes. These robust associations between an EMT gene expression signature and an increased risk of metastatic disease extend to CRC111, pancreatic cancer112, SCC113, prostate cancer114 and breast cancer115. The prognostic value of these signatures for inferior survival has also been proven for a large variety of cancers, including breast, hepatic, CRC and gastric cancers111,116,117,118 as well as in a pan-cancer retrospective analysis of The Cancer Genome Atlas data119. In rectal cancer, EMT-like changes at the invasive front are associated with both inferior OS and an increased risk of metastatic dissemination, with ZEB1 implicated as the controlling EMT-TF24,112,120,121,122,123,124,125,126. In oesophageal cancer, patients with reduced E-cadherin expression or positivity for SNAIL have inferior clinical outcomes127. Likewise, in patients with pancreatic cancer, loss of membranous E-cadherin expression in tumour buds located at the invasive front correlates with inferior 30-month survival64. In patients with pT1 non-muscle-invasive bladder cancer, reduced E-cadherin expression is associated with significantly worse 10-year PFS128,129,130. Similarly, SNAIL is an independent predictor of disease recurrence and progression in patients with non-muscle-invasive bladder cancer131, in which a correlation between increased TWIST1 expression and reduced relapse-free survival has been established72.

Interestingly, novel molecular classifications of various cancer types have revealed correlations between specific subtypes characterized by EMT signatures and inferior clinical outcomes. These include claudin-low breast cancer132,133, CMS4 CRC134, and the quasi-mesenchymal or basal subtype of pancreatic cancer135,136. Clinical implementation of these robust relationships with prognosis might lead to clinical benefit in terms of tailoring therapy and, ultimately, in guiding the use of possible EMT-targeted therapies.

Therapy resistance

Resistance to standard-of-care therapies provides another EMT-associated barrier to effective cancer treatment. Mechanistically, EMT-mediated treatment resistance can arise from a number of features of the hybrid E/M or mesenchymal states. These features include elevated expression of drug efflux transporters11,13,137, activation of alternative signalling pathways, such as AXL kinase in the context of resistance to EGFR inhibitors12,138, and activation of immune escape mechanisms5,9,11,13. Conversely, most chemotherapeutic agents as well as radiotherapy can promote EMT, which could result in resistance to subsequent lines of treatment. By contrast, using agents with the capacity to reverse EMT could leave tumours in a more therapy-sensitive state, leading to more favourable responses to subsequent lines of therapy and improved outcomes. In keeping with this concept, a review comparing therapies used in clinical trials with the capacity to either induce or (more rarely) inhibit EMT respectively found accelerated or repressed disease progression and mortality with these two therapeutic categories12.

Disease relapse owing to treatment resistance is strongly correlated with activation of EMT in many cancer types, including breast92,132, colorectal95,107,139, pancreatic140 and prostate cancers, in which an EMT-associated neuroendocrine subtype can emerge141,142. A link between EMT activation and acquired resistance is evident for most non-surgical treatment modalities, including chemotherapy, radiotherapy, endocrine therapy, targeted therapy and immunotherapy (Box 1).

Cancer cells that can survive the effects of these various treatment modalities by transitioning to a hybrid E/M state provide the basis for minimal residual disease and subsequent disease recurrence. These highly resistant cancer cell types are therefore the ultimate targets for novel therapies.

Clinical manifestations of EMT in non-epithelial tumours

EMT-TFs also have important roles in the development and homeostasis of non-epithelial tissues3,6,143,144. Consequently, tumours of non-epithelial origins, such as glioblastoma, sarcoma and leukaemia, do not harbour the classical characteristics of EMT, particularly the loss of expression of epithelial markers and concurrent gain of mesenchymal markers145,146. Nevertheless, EMT-TFs such as ZEB1 or SNAIL can still have unfavourable effects on cancer biology and their presence is associated with inferior clinical outcomes145. These observations demonstrate that these factors not only activate classical EMT but also have a general role in regulating cellular plasticity in most tissue types. Glioblastoma is characterized by extensive local and regional invasion and spreading, and subtypes with EMT-related gene expression, particularly upregulation of ZEB1, confer inferior responses to treatment and a worse overall prognosis147,148,149. Degradation of EMT-TFs can suppress this invasive behaviour in mouse models of glioblastoma150. In Ewing sarcoma, the YB1–HIF1α axis, which is known to activate EMT-TFs, is associated with an increased risk of metastases and inferior survival outcomes151. Similarly, in osteosarcoma, ZEB1 expression is associated with a higher tumour grade, an increased risk of metastases and inferior clinical outcomes144,152. Other examples in which EMT activation correlates with survival, treatment resistance and/or metastatic dissemination include liposarcoma153, carcinosarcomas in various organs154, uterine and ovarian cancers155,156, and mesothelioma157.

Furthermore, EMT-TFs (ZEB1, ZEB2, TWIST and SNAIL) have important regulatory roles in haematopoiesis158,159,160,161,162. Consequently, dysregulated expression of these proteins is also associated with disease progression and inferior outcomes in patients with various haematological malignancies. These malignancies include B cell lymphoma163, T cell lymphoma164, multiple myeloma165, and chronic166 and acute158,167 myeloid leukaemia. These observations further demonstrate that EMT-TFs not only activate classical EMT features but also have a general role in regulating cellular plasticity in most tissues and, consequently, in non-epithelial malignancies.

Clinical opportunities and challenges

Clinically relevant EMT markers

EMT clearly involves a huge array of different drivers and many different mediators resulting in substantial cellular heterogeneity along the EMT–MET continuum (Fig. 1a). EMT biomarkers can be grouped into various categories and have been described in detail elsewhere1,3,5,8,9,10,11,15,24,80,101,168. These biomarkers include established immunohistochemical markers used routinely in pathology labs (such as vimentin, E-cadherin, pan-cytokeratin) to elaborate EMT signatures and, potentially, various multimodal measures, including phenotypic protein markers and signatures; genomic and/or transcriptomic signatures (including 3D genome organization169); activation of driver or suppressor pathways directly or indirectly affecting EMT; transcription factors and related epigenomic regulators for different EMT stages; microRNAs and long non-coding RNAs; and exosomes and CTCs.

Numerous examples of prognostic and predictive factors relating to the role of EMT in cancer progression exist (Supplementary Table 1 and Box 1) and can potentially also be used and/or adapted for monitoring of treatment responses. Many could potentially also be used for the selection of patients at high risk who might require intensified adjuvant therapy, for treatment selection potentially favouring treatments effective in epithelial–mesenchymal transitioned cells, and for the monitoring of post-treatment phenotypes and likely subsequent responses to later-line therapy. For example, investigators identified different human triple-negative breast cancer molecular subtypes including EMT-enriched signatures and used these to effectively predict the responses of cell lines to therapies targeting specific signalling pathways that are active during EMT170. Subsequently, research by another group demonstrated that patients without an initial EMT phenotype who did not have a complete response to standard chemotherapy often have evidence of a transition to an EMT tumour phenotype with consequences for treatment selection171. In an attempt to predict the outcomes of patients with oligometastatic CRC, investigators conducted a molecular analysis of samples from 121 patients and established three different molecular subtypes of liver metastases with distinct patterns of OS. The stromal subtype, characterized by high levels of an EMT gene expression signature, was associated with the least favourable prognosis172. Histological features have also been used to monitor the extent of EMT. For example, tumour buds, which can be identified in histopathological sections of subsets of various cancer types, including CRC, pancreatic ductal adenocarcinoma and oral squamous cell carcinoma (SCC), are strongly associated with EMT173, and tumour budding has emerged as a prognostic biomarker in clinical assessments of samples from patients with high-risk CRC. Intriguingly, these and other EMT-related features might be detectable in haematoxylin and eosin images using computational pathology and/or machine learning analysis methods174.

These and many other studies (summarized in Supplementary Table 1 and Box 1), if successfully implemented clinically, have the potential to identify patients that might benefit from an altered (more-aggressive or targeted) treatment approach. However, no consensus EMT-related biomarkers have currently been validated for clinical use, and the heterogeneity associated with EMT suggests that such biomarkers might need to be personalized or, as a minimum, tailored to very specific clinical settings.

In this regard, an appreciation exists that very few complex biomarkers (such as gene expression signatures) are currently reimbursed for use in oncology. Those that are available internationally, such as Oncotype Dx for patients with breast cancer, have required large-cohort prospective trials involving thousands of patients to demonstrate clinical utility. The majority of companion biomarkers that are approved and reimbursed alongside specific interventions require generic reagents and do not involve proprietary, patented, or protected scoring assays and systems. They can be carried out by local laboratory personnel and involve the simple enumeration of a single moiety such as HER2 immunohistochemistry (IHC) and in situ hybridization, IHC-based quantification of PD-L1, or assessments of BRAFV600E mutations using reverse transcription-PCR or next-generation sequencing. EMT-based biomarkers will probably take some time to enter clinical practice, possibly in conjunction with EMT-targeted therapies. Simple, single-moiety assays are the most straightforward to validate; nonetheless, the technologies underlying multiomics, digital pathology and spatial biology-based analyses are all improving exponentially and might provide solutions that can overcome these limitations in the near future. Clinical assessments of EMT therefore remain a challenge that has yet to be addressed (Box 2).

Therapeutic strategies

EMT and associated processes, such as EMP, are correlated with tumour latency and recurrence, subtype switching, stemness, minimal residual disease, resistance to therapy, and metastatic dissemination in many cancer types, and thus provide an important clinical opportunity. In some scenarios, the roles of EMT in malignancy oppose each other, for example, the apparent requirement of reverse transition (MET) for metastatic competence after EMT has promoted initial invasion, therapy resistance and immune suppression1,3,5. These dichotomies indicate a need for extremely precise analysis to define the various scenarios in which therapeutic exploitation of EMT might be effective. Methods of sampling and analysing individual cancers can be married with state-of-the-art technologies for cancer-predominant and even patient-specific assessments of EMT processes for possible clinical targeting.

Strategies for targeting EMP exist across the spectrum of cancer pathogenesis and progression. These strategies include aiming towards a differentiated phenotype by blocking (initial) EMT to avoid invasive escape, intravasation, and extravasation or by driving EMT-positive dormant cells through MET to make them responsive to standard-of-care therapies and immunotherapy, or aiming towards an undifferentiated mesenchymal phenotype by blocking MET and thus supressing metastatic outgrowth at the target site and/or by driving cells towards extreme EMT to retain them in a dormant state with the opportunity to directly target the mesenchymal phenotype based on identified unique vulnerabilities such as in their metabolism. Additional strategies aim at entirely blocking plasticity to freeze cells in the current state (plastistatic therapy) or at inducing the transdifferentiation of cancer cells into non-proliferating cells (such as adipocytes) (Fig. 2).

Fig. 2: Strategies to target EMP in cancer.
figure 2

Various strategies for targeting cancer cell epithelial–mesenchymal plasticity (EMP) exist across the spectrum of cancer pathogenesis and progression. These include various types of action: aiming towards a differentiated phenotype (yellow) either by blocking (initial) epithelial–mesenchymal transition (EMT), thus avoiding invasive escape, intravasation, and extravasation (1), or by driving EMT-positive dormant cells through mesenchymal–epithelial transition (MET) to make them responsive to standard-of-care therapies and/or immunotherapies (2), or aiming towards an undifferentiated mesenchymal phenotype (red) by blocking MET, which is required for metastatic outgrowth at the target site (3), or by driving cells towards extreme EMT to maintain them in a dormant state (4), with the opportunity to directly target the hybrid epithelial/mesenchymal (E/M) or mesenchymal (M) phenotypes based on the identified unique vulnerabilities, for example, in their metabolism (5). Additional strategies include blocking plasticity entirely to freeze cells in the current state (plastistatic therapy) (6) and inducing the transdifferentiation of cancer cells into non-proliferating cells such as adipocytes (7; depicted in green).

Full size image

Although various manipulations of cellular positioning on the E–M axis could generate therapeutic effects as described above, inhibiting or reversing the occurrence of EMT in cancer cells is currently the most clinically advanced strategy. A vast wealth of underlying preclinical data has led to the development of agents targeting a range of aspects of the EMT process. Various EMT triggering, initiating and effector mechanisms are being targeted in ongoing trials, some of which now have results available. These initiatives can be classified based on the type of EMT pathway component they target, as reviewed in detail elsewhere10 (Table 2 and Supplementary Table 2).

Table 2 Selected trials testing EMT-targeting agents or agents with indirect effects on EMT
Full size table

Prominent examples of EMT-targeting agents in clinical trials

Eribulin

In contrast to most chemotherapies, which tend to induce EMT12, the microtubule inhibitor eribulin, which is approved as second-line or third-line therapy for metastatic breast cancer and as a first-line therapy for advanced-stage liposarcoma by several regulatory agencies, can reverse TGFβ-induced and/or chemotherapy-induced EMT in breast cancer cell lines both in vitro and in vivo175, which might explain the efficacy of eribulin in comparison with other chemotherapies176. Interestingly, eribulin seemed to prolong OS to a greater extent than PFS in patients with breast cancer in a pooled analysis of data from two phase III trials (HR for PFS 0.9, HR for OS 0.85)176,177 and to a much more limited extent in patients with advanced-stage liposarcoma in a phase III trial (HR for PFS 0.52, HR for OS 0.51)178. These differences might reflect a reduction in the emergence of new metastases (owing to indirect suppression or reversal of EMT) rather than effects on established lesions176. Eribulin has also been shown to sensitize oral SCC cell lines to the anti-EGFR antibody cetuximab via induction of MET179 and to reverse EMT in triple-negative breast cancer cells in vitro, thus sensitizing them to other chemotherapies175. Combinations that enhance sensitivity to other agents can potentially be administered at lower doses and/or are more likely to be appropriate for incorporation into combination regimens. Moreover, emphasizing the use of chemotherapies and other agents that might prevent EMT activation in settings in which, potentially, several agents or regimens are available might be a prudent approach. Thus far, eribulin provides the best example of a chemotherapy that does not induce but rather might reverse EMT and could provide a prototype for the development of other novel drugs with similar mechanisms of action.

A number of signalling pathways have an established role in EMT induction, including TGFβ, AXL kinase, netrin 1, NOTCH and hypoxia-inducible factors. Many of these pathways have multiple other effects beyond control of EMT such that clinical benefit arising from targeted inhibition of these pathways might not necessarily be only the result of EMT inhibition. Here, we focus on the inhibition of netrin 1 and AXL kinase in clinical trials.

Netrin 1 inhibition

Netrin 1 has a role in both embryogenesis and cancer progression. Stimulation of this signalling pathway activates EMT, angiogenesis, migration and invasion. The mechanism of netrin 1-induced EMT activation might be indirectly mediated via the PI3K and ERK signalling pathways as blocking these can inhibit netrin 1-induced EMT180. In a phase I trial, monotherapy with the anti-netrin 1 antibody NP137 resulted in one partial response (response rate 7.1%) and stable disease in a further eight patients (clinical benefit rate 57.1%) with advanced-stage endometrial cancer, with evidence of suppression of EMT181. NP137 was well tolerated, with mild-to-moderate infusion-related reactions as the most frequent treatment-related adverse event182. Clinical testing of NP137 has been extended to LiverNET1 (NCT05546879), adding NP137 to bevacizumab and atezolizumab in patients with advanced-stage hepatocellular carcinoma, and ImmunoNET1 (NCT05605496), a phase II trial designed to assess NP137 as an add-on therapy to standard immunotherapies in patients with advanced or metastatic solid tumours. Two further studies are combining NP137 with FOLFIRINOX (folinic acid, 5-fluorouracil, irinotecan and oxaliplatin) chemotherapy in patients with resectable (NCT06203821) or locally advanced (NCT05546853) pancreatic cancer. The FDA designated this agent as an orphan drug for patients with pancreatic cancer in November 2024.

AXL kinase inhibition

AXL, a member of the TAM (TYRO3, AXL and MERTK) family of receptor tyrosine kinases (RTKs) is commonly associated with both EMT induction and treatment resistance183. In breast cancer cells, AXL is associated with a more plastic phenotype and, when suppressed, can reverse the SLUG-dependent mesenchymal phenotype without affecting SLUG levels184. Upregulation of AXL expression is evident in various cancer cells of mesenchymal phenotypes, in metastatic lesions and on acquisition of resistance to agents targeting other RTKs such as EGFR, HER2 and ALK185. This often reflects rewiring of the AXL signalling pathway in the mesenchymal state, which can be achieved via cross-activation of AXL and other kinases, such as c-MET or TANK-binding kinase 1 (TBK1), to trigger sustained downstream phosphorylation of ERK or AKT3 (refs. 186,187). Owing to the immunomodulatory functions of TAM receptors, activation of AXL signalling is also associated with immune evasion and resistance to anti-PD-(L)1 antibodies188. Overcoming this signalling crosstalk is therefore crucial to reversing resistance and effectively targeting EMT. A number of inhibitors of varying levels of selectivity have been developed, including the AXL-specific inhibitors DS-1205c and bemcentinib, the AXL and MET inhibitors glesatinib, S49076 and BPI-9016M, the pan-TAM inhibitor sitravatinib, and agents that inhibit TAM RTKs and other kinase moieties such as foretinib, gilteritinib, ningetinib and AL2846 (Table 2 and Supplementary Table 2). Trial designs thus far have not allowed assessments of the value of adding AXL inhibitors to existing treatments, being either single-arm trials or comparing arms with very different standard-of-care treatments integrated with AXL inhibition. The most promising single-agent response rates have been 26% with glesatinib in patients with c-MET-altered or AXL-altered non-small-cell lung cancer189 and 23% with foretinib in those with advanced-stage hepatocellular carcinoma190. In a phase II trial in which 54 patients with non-small-cell lung cancer received bemcentinib plus pembrolizumab, including 3 patients with STK11-mutant disease, which is typically associated with resistance to anti-PD-(L)1 antibodies, all had objective responses191. However, no particularly notable response rates, PFS or OS outcomes have been reported from other trials testing these agents in combination regimens (Supplementary Table 2).

Several classes of EMP-targeting agents have also been developed in model systems or can be repurposed from existing compounds and might provide new treatment approaches10,192,193,194,195,196 (Table 3). These classes include repurposed drugs with co-incidental EMT-reversing or targeting properties such as salinomycin, metformin, glitazones (also known as thiazolidinediones) and MEK inhibitors194; agents targeting established EMT-activating signalling pathway components such as TGFβ, NOTCH, ERK1, ERK2, WNT, FGF, IGF and RTKs (reviewed in detail elsewhere197); direct targeting and/or downregulation of EMT-TFs via epigenetic reprogramming198; applying non-coding microRNAs, which are a central part of physiological control in EMT199,200, and inducing the transdifferentiation of cancer cells into non-proliferating adipocytes or basal–luminal cancer subtype transition27,201. Moreover, directly targeting aggressive cancer cells in a hybrid E/M or mesenchymal state based on unique vulnerabilities will provide novel opportunities; one example is ferroptosis, a newly discovered iron-dependent cell death pathway that is strongly dependent on EMT activation and ZEB1 expression. Research in this area could provide direct therapeutic options against up-until-now untargetable, highly resistant cancer cells in a hybrid E/M or mesenchymal state202,203,204,205. Programmes aiming to identify novel drugs using high-throughput screens are ongoing206 and are expected to offer further therapeutic opportunities, as summarized in detail elsewhere10. The immunosuppressive effects of EMT are well established168,207. Consequently, inhibition of EMT or activation of MET might immediately enhance endogenous immune-mediated clearance of cancer cells and/or improve the efficacy of immune-checkpoint inhibitors193.

Table 3 Candidate agents and antibodies with potential to target epithelial–mesenchymal plasticity
Full size table

Challenges in targeting or modulating EMT in the clinic

Trial design and end points

Developing clinical interventions specifically targeting EMT will require consideration of the histology-specific EMT phenotype as well as trial designs, readouts and end points that are optimized for detection of the clinical activity of such interventions (Table 4). Window-of-opportunity trials provide an opportunity to test early, short-term interventions predominantly in patients with localized disease. This type of trial design has led to paradigm shifts in the treatment of early-stage breast cancer (such as in the I-SPY trials)208. In these trials, treatment regimens are usually administered in the neoadjuvant setting for a defined period between initial diagnosis and primary tumour resection. However, these trials rely on the supposition that short-term efficacy end points (most commonly pathological complete response in neoadjuvant studies) are reliable surrogates for the longer-term end points that reflect benefit for patients (generally DFS and OS). A surrogate relationship between pathological complete response and survival outcomes might prove to be the case with directly cytotoxic therapies, for which shrinkage of the primary tumour and eradication of micrometastases tend to correlate. However, the same might not hold true for EMT-directed therapies for several reasons.

Table 4 Clinical scenarios and challenges relating to clinical targeting of EMT in specific tumour types
Full size table

First, agents that successfully modulate EMT and thus sensitize either early-stage or metastatic tumours to other treatments might not themselves greatly affect cancer cell proliferation or survival. In certain scenarios, inhibiting EMT could lead to increased tumour cell proliferation, with the goal instead being to render the malignant cells more sensitive to co-administered cytotoxic agents. Consequently, cytoreduction as a study end point should be reserved for trials testing such combinations, with biomarker measurements of successful EMT modulation as well as safety and tolerability being the most suitable end points for studies testing single agents with the goal of sensitization.

Second, the core goal of adjuvant and neoadjuvant therapies is to eradicate micrometastases, the main drivers of cancer relapse and subsequent mortality. EMT-targeted therapies hold genuine promise to achieve this goal even where other therapies might not, owing to the role of EMT in cancer cell dormancy and the survival of micrometastases. However, a finding of primary tumour shrinkage is unlikely to reflect dormancy reversal or eradication of micrometastatic disease. Therefore, trials in this setting should include DFS as a primary or co-primary end point to avoid missing important positive efficacy signals.

Finally, a core goal of EMT-directed therapy, particularly in patients with metastatic disease, is to avoid inducing resistance both to ongoing treatment and to subsequent lines of therapy. Therefore, trial designs that only record response depth and the duration of treatment with the current therapy risk missing possible downstream improvements in efficacy. Consequently, trials should include sufficient follow-up to also record the durations of subsequent treatment responses and survival outcomes. Taking eribulin as an example of a treatment with an established ability to reverse EMT, in both breast cancer177,209 and liposarcoma178, the median OS benefit is substantially longer than median PFS on the drug (fourfold and sixfold longer, respectively), implying that eribulin either delays tumour progression or sensitizes tumour cells to subsequent therapies. Eribulin would be unlikely to be approved or reimbursed for use on the basis of the modest PFS benefits alone.

Potential for situation-specific contrary outcomes

As described previously, cellular plasticity, exerted by the dynamic and reversible nature of the EMT programme, is a major driving force towards metastasis. Therefore, despite the promise of EMT targeting across the cancer management continuum, care will be required moving forward as EMT can have dual and opposing roles in certain stages of disease progression, which can exist simultaneously in the same cancer patient. For example, disseminated cancer cells appear to need to undergo reverse transition (MET) to develop metastatic competence after EMT has promoted initial invasion, treatment resistance and immune suppression35,103. When attempting to target dormancy, a solution might therefore be to reverse the dormant state to render dormant cells more sensitive to standard-of-care treatments. However, such reversal of dormancy could also lead to accelerated proliferation106 and increased metastatic potential104. However, a precedent exists suggesting that induction of MET is beneficial, even when it comes at the cost of increased cancer cell proliferation210. An alternative approach would be to develop plastistatic therapies that retain cells in the current state and thus preclude phenotype switching10. For example, RepSox, a selective inhibitor of the TGFβ type I receptor, effectively blocks EMT induction and stabilizes hybrid E/M cells in a state that is more closely aligned with epithelial differentiation, resulting in reduced metastatic potential in mouse models of breast cancer and CRC206. Modulation of epigenetic regulators, such as chromatin remodelling and DNA methylation enzymes, might offer a means of constraining cells within the hybrid E/M state. Regardless of the type of agent, plastistatic therapies aim to retain cancer cells in a dormant but non-toxic state and are therefore unlikely to have direct cytotoxic effects. As a result, such agents will probably require sustained administration or need to be administered in combination with agents of other therapeutic classes to achieve robust efficacy. These dichotomies highlight the need for precise analysis to define the various scenarios in which therapeutic exploitation of EMT might be effective. Methods of sampling and analysing individual cancers might need to be combined with state-of-the-art technologies for cancer-specific and even patient-specific assessments of EMT processes for sufficiently precise clinical targeting.

Identifying suitable surrogate biomarkers of efficacy

Given the need for surrogate assessments in trials testing the efficacy of single-agent EMT-targeting therapies, a clear molecular profile of therapeutic EMT suppression would be of great value. However, as demonstrated by the plethora of published biomarkers linked with EMT (Fig. 1 and Supplementary Table 1) as well as the development of more comprehensive mRNA signatures in response to this complexity211, a surfeit of choice of translational assays to assess the activity of EMT inhibitors exists. In addition to gene expression profiling, trials could also consider assessments of effects on specific EMT-related characteristics such as changes in the tumour vasculature and neural invasion, tumour budding, and subtype shifts from a basal to a luminal histology (such as in the NeoSolti trial212) as well as molecular signatures. Building on the ubiquitous association of EMT with acquired resistance to a wide range of treatment modalities will require novel screening approaches (Box 1). Looking beyond the assessment of interventions designed to target EMT, consensus profiles of EMT induction and suppression would be of considerable value in the wider trial context when screening for mechanisms of resistance to novel therapies to assess the role of EMT at disease progression.

Spatial and temporal heterogeneity

The influence of spatial intratumoural heterogeneity is a frequent source of uncertainty when administering any biologically directed therapy. Many molecular targets have variable levels of expression between different areas of the same primary tumour and between separate metastatic lesions. For example, heterogenous HER2 expression as measured using IHC has been shown to correlate with inferior responses to HER2-targeted therapies in patients with HER2-positive breast cancer213. Similarly, heterogeneous expression of prostate-specific membrane antigen (PSMA), with some areas of primary prostate cancers entirely lacking in PSMA expression, correlates with inferior responses to 177Lu-labelled PSMA214. Given that EMT is likely to be driven, at least in part, by local conditions in the tumour microenvironment, including hypoxia and glucose deprivation, the extent of EMT is likely to vary. Consequently, assessments based on a limited volume of biopsy material might not accurately reflect the characteristics of the whole tumour. Considering precedent, in patients with CRC, EMT is generally more prevalent at the invasive edge, where it correlates with an inferior prognosis62,215.

Temporal heterogeneity relates to differences in the extent and effects of EMT as tumours progress from early through locally advanced to metastatic disease. As the treatment setting shifts from early-stage to metastatic disease, the goal of EMT targeting shifts from one of dormant cell eradication and prevention of metastatic dissemination to resensitization of metastatic disease, such that the most effective EMT-modulating therapy might change with disease stage. Consequently, successful therapy in patients with advanced-stage disease might not translate to the early-stage setting. In this regard, data from genomic profiling studies suggest that dormant metastatic cells are seeded early in primary tumour development216, implying that inhibition of EMT early in the course of disease might provide durable suppression of tumour development.

The value of EMT-targeted therapies will probably be tumour specific, depending on the timing of dissemination of the specific tumour type. Targeting EMT alone in patients with skin SCC, for example, might enable local disease control, which is one of the major clinical challenges in the treatment of these patients. However, adenocarcinomas, such as ovarian cancers, disseminate much earlier in the course of the disease relative to skin SCC. Thus, combining EMT-targeted therapy with cytotoxic therapy might be more appropriate in patients with such tumours.

Specificity and toxicity

Given the paucity of clinical testing of bona fide EMT-targeted or EMP-targeted therapies, not much is known thus far about the therapeutic index and toxicities of such agents. Published data are available for the netrin 1 inhibitor NP137, which is well tolerated, with infusion-related reactions as the most frequent adverse event182. The consequences for non-malignant cellular functions and processes have to be considered when developing any manipulation of abnormal cellular behaviour designed to achieve a therapeutic effect. We would anticipate that, as we move from agents targeting pathways that modulate EMT as just one of their actions, such as those targeting NOTCH or HIF signalling, to drugs that disrupt core EMT functionality such as the EMT-TFs or vimentin and cadherins, the adverse effects observed are likely to become more specific to off-target inhibition of EMT such as impaired wound healing, fibrosis and inflammation. Some data regarding the loss of function of EMT-TFs might be inferred from rare familial mutations and knockout mouse models. For example, mutations in the EMT-TF gene TWIST1 are responsible for Saethre–Chotzen syndrome, which is characterized by facial dysmorphism and fusions of fingers and toes217. SLUG function appears to be more important to epidermal health, with Slug-null mice having impaired wound healing and skin ulceration218. Inherited defects in SLUG can cause Waardenburg syndrome with symptoms such as loss of hair and skin pigmentation and sensorineural deafness219. Considering ZEB proteins, Zeb1-null mice die perinatally with signs of impaired thymic development and skeletal abnormalities220. In humans, loss-of-function ZEB2 mutations underlie Hirschsprung disease–intellectual disability syndrome, which involves aganglionic megacolon, intellectual disability and facial dysmorphism221. Notably, all of these effects relate to systemic absent or reduced EMT-TF function during embryogenesis or in early postnatal life, whereas the effects of inhibition in adulthood, which are most relevant to cancer treatment, are largely unknown. Notably, inducible systemic Twist1 inactivation in adult mice does not appear to affect the health of these animals, suggesting that targeting adult cancers with such an approach might be safe and is reassuring from a drug development perspective222. Taken together, these data suggest that interventions targeting EMT-TFs might be well tolerated, although monitoring of potential adverse effects in trials testing such interventions should include impairment of wound healing and haematopoiesis.

Ferroptosis

Mesenchymal state-specific cytotoxicity mediated by ferroptosis induction202,223 might have off-target effects, for example, on non-malignant mesenchymal tissues. Indeed, various pathogenic conditions have been linked with increased ferroptosis, including neurodegeneration, liver and lung fibrosis, autoimmune diseases, and chronic obstructive pulmonary disease223, such that selective targeting of ferroptosis induction to cancer cells or their specific metabolic states (such as those in a hybrid E/M cellular state) will probably be essential.

Conclusions

Partial and often transient activation of EMT equips cancer cells with an extraordinary level of phenotypic plasticity (EMP) that supports many stages of tumour progression towards metastatic dissemination and treatment resistance. Abundant evidence indicates crucial roles of EMT and EMP in many human cancer types, which opens the opportunity for novel diagnostic, predictive and, particularly, therapeutic options. Substantial efforts are being devoted to exploiting this knowledge for improved treatment and/or clinical management. Future translational and clinical work should seek to take advantage of these opportunities and tackle the associated challenges.