The Transcriptional and Epigenetic Landscape of Cancer Cell Lineage Plasticity

During normal development, cells undergo specific differentiation and organization programs to produce tissues with defined functions. More specifically, multipotent embryonic progenitor cells undergo nongenetic phenotypic changes in response to environmental cues to yield lineage-specific progenitors and, eventually, terminally differentiated cells (1). It was originally postulated by Conrad Waddington that mature, differentiated cells were unable to produce progeny distinct from their committed lineage (2). Although some cell types and tissues exhibit a restricted capability for phenotypic plasticity, later studies revealed a remarkable level of cell state flexibility in many adult organs (3), particularly in the context of tissue repair and pathologic stress (4). Lineage plasticity can manifest via dedifferentiation, the reversal of a differentiated phenotype back to a progenitor-like state that may then redifferentiate to another cell lineage, or transdifferentiation, whereby cells switch from one lineage to another without transit via an intermediate pluripotent state.

The unlocking of phenotypic plasticity during tumor evolution is now recognized as a critical hallmark of cancer (5). By exploiting dedifferentiation and transdifferentiation, cancer cells can acquire new molecular features to facilitate metastasis and evade systemic therapies. Although genetic alterations can promote cancer cell lineage plasticity, this phenomenon is primarily driven by epigenetic and transcriptional reprogramming. Epigenetic/transcriptional mechanisms are advantageous over genetic mechanisms, since they enable tumors to respond more rapidly to volatile microenvironmental pressures and therapy (6) and are more permissive of reversible phenotype switching, whereas genetic changes can lock cells into an evolutionary trajectory that may become disadvantageous in the future. This review describes different types of cancer cell lineage plasticity, the epigenetic and transcriptional underpinnings of such plasticity, and how cancer plasticity might be leveraged therapeutically.

Cells exist in a certain configuration—termed a cell state—based on specific molecular (e.g., gene expression) and functional (e.g., self-renewal capacity) features. Multiple regulatory levels, including epigenetic and transcriptional, are fundamental for determining and maintaining cells in a given state. However, it is becoming increasingly clear that cells are not locked in a particular state but rather can be rewired and transition along a continuum of phenotypes. A key feature of this rewiring is its complexity, with transitions being frequently characterized by uncertain outcomes and ambiguity in cancer cell identity. Nevertheless, common manifestations of lineage plasticity observed in solid tumors and with documented roles in cancer growth, metastasis, and therapy resistance are summarized in Fig. 1A and described in more detail below.

The loss of epithelial phenotype and acquisition of mesenchymal properties, a process referred to as epithelial–mesenchymal transition (EMT), was first described in the context of developmental genetics by Elizabeth Hay and colleagues when studying chick embryogenesis (7, 8). In the intervening time, compounding evidence has pointed to EMT as a developmental program ubiquitously co-opted by cancer cells to shed epithelial cell polarity and manifest a mesenchymal, motile phenotype (refs. 9–11; Fig. 1B). This process is bidirectional; mesenchymal-like cells are capable of gaining epithelial features through a mesenchymal–epithelial transition (MET). The dynamic flexibility of these processes is collectively termed “epithelial–mesenchymal plasticity.”

The regulatory framework controlling epithelial–mesenchymal plasticity is well defined, incorporating both core intrinsic elements, namely, the EMT-transcription factors (TF) SLUG, SNAIL, TWIST, and ZEB1/2, as well as context-dependent extrinsic factors such as extracellular matrix stiffness and heterotypic signals originating from tumor-associated stromal cells (ref. 9; see “Interplay between the Tumor Microenvironment and Lineage Plasticity”). A plethora of studies have demonstrated that EMT is associated with enhanced metastatic capacity, therapy resistance, and stem-like features (11). Consistent with these findings, tumors derived from tissues of mesenchymal origin are frequently highly invasive (e.g., glioblastoma) or characterized by early metastasis (e.g., melanoma; ref. 12). However, these observations contrast with a parallel body of work showing that stable mesenchymal features can abrogate metastatic outgrowth in preclinical models along with reports that tumor cells with an epithelial phenotype retain complete metastatic capacity. For example, recent findings from the Brisken lab and others have demonstrated that, in mouse models of estrogen receptor–positive breast cancer, cells in an epithelial differentiation state exhibit increased fitness, antiapoptotic signaling, and metastatic potential (13, 14).

The divergent and somewhat contradictory findings surrounding epithelial–mesenchymal plasticity in cancer progression may be a result of studying extreme epithelial and mesenchymal states in a process that is likely transient and episodic. More recent evidence has shown that cells existing in a hybrid EMT (hEMT) state—that is, coexpression of epithelial and mesenchymal markers in the same cells—have the greatest malignant and metastatic potential (15, 16). More­over, isolation and sequencing of circulating tumor cells (CTC) from patients with breast cancer revealed that the hEMT state is linked to elevated expression of genes that promote a stem cell–like state (17). Notably, this hybrid phenotype is very common, with a large-scale transcriptomic profiling effort of over 7,100 human epithelial tumors uncovering a hEMT state in ∼40% of tumors (18). Whether the hEMT state represents a stable phenotype or is a snapshot of the fluctuating EMT status of individual cells remains an open question, although recent single-cell data from multiple tumor types suggest the existence of a continuum of cell states (19).

Histologic transitions from adenocarcinoma to neuroendocrine carcinoma are frequently observed in prostate, lung, and bladder tumors (Fig. 1B). However, a neuroendocrine gene signature was found to be associated with metastasis and poor outcome across 21 epithelial malignancies, suggesting that this phenomenon may be underappreciated (20). The concept that the neuroendocrine cell state arises via lineage plasticity rather than expansion of preexisting alternative lineage clones is supported by the observation that prostate tumors frequently retain the genetic hallmarks of their precursor adenocarcinoma histology (21). Validating this hypothesis, murine lineage tracing studies revealed that neuroendocrine cancer cells can arise from alveolar type 2 and tracheobronchial basal cells in the lung and luminal epithelial cells in the prostate (22–24). Moreover, in a series of elegant multiomics and spatial profiling experiments of human lung adenocarcinomas, the Ciriello lab provided additional evidence for epithelial–neuroendocrine transdifferentiation as a result of epigenetic and transcriptional reprogramming (25). Epithelial-to-neuroendocrine lineage plasticity frequently manifests following therapies targeted against driver oncogenes: anti-EGFR therapy in EGFR-mutant lung adenocarcinoma, ALK inhibition in ALK-rearranged lung cancer, and anti–androgen receptor (AR) pathway therapy in prostate adenocarcinoma (26–30). However, evidence for such plasticity in untreated tumors suggests that it can be induced by other factors, such as microenvironmental stressors, hypoxia, or oncogene dysregulation. For example, in genetically engineered mouse models (GEMM) of pancreatic ductal adenocarcinoma, ductal–neuroendocrine lineage plasticity is governed by deregulated MYC expression in the context of KRAS mutation (31). Similarly, normal epithelial cells of prostate and lung origin can be reprogrammed into neuroendocrine-like cancer cells by ectopic upregulation of AKT, MYC, and BCL-2 concomitant with loss of the tumor suppressors TP53 and RB1 (32).

The recent observation of tumors with mixed adenocarcinoma/neuroendocrine features following targeted therapy is suggestive of epithelial-to-neuroendocrine plasticity occurring via a transitory, metastable phenotypic state(s). Spatial profiling of a prostate tumor with distinct adenocarcinoma and neuroendocrine foci revealed cells along a continuum of plastic gene expression signatures between the two extreme differentiation states (33), a concept reinforced by the frequency of therapy-resistant prostate tumors exhibiting amphicrine phenotypes (i.e., coexpression of both AR and neuroendocrine markers; refs. 30, 34, 35). Similarly, mixed/transitory high-plasticity cell states emerge during lung cancer progression in GEMMs and human patient-derived xenografts (PDX) and are postulated to underlie chemoresistance (36).

Squamous transdifferentiation represents an alternative lineage plasticity mechanism (Fig. 1A) that often develops as a subpopulation within another primary histology. Although adenosquamous histology is most commonly observed to arise de novo in pancreatic ductal adenocarcinomas (37), squamous differentiation in the lung has been reported in the context of acquired resistance to EGFR tyrosine kinase inhibitor therapy as well as ALK inhibitors in the setting of EML4–ALK fusion–positive non–small cell lung cancer (NSCLC; refs. 38, 39). Although limited in number, cases of squamous transformation with molecular profiling were found to retain the characteristic genetic features of the adenocarcinoma (e.g., EGFR mutation), supporting the hypothesis of lineage plasticity. This is reinforced by mouse models of KRAS-driven lung adenocarcinoma whereby loss of a kinase, LKB1, mediates epigenetic reprogramming that drives tumor cells into a plastic state to enable squamous transformation (40). More recently, squamous differentiation has been reported in human prostate tumors following AR pathway antagonism (35) and in a prostate cancer GEMM driven by MYCN overexpression and RB1 loss (41). The observation that prostate, lung, and pancreatic adenocarcinomas with squamous differentiation are enriched for mutations in oncogenes such as MYC and KRAS (KRAS^G12D) suggests that certain genotypes may be more permissive for squamous-like identity (42–44).

The concept of epithelial-to-endothelial transition (Fig. 1A), more commonly referred to as vascular or vasculogenic mimicry, was first described by Maniotis and colleagues in 1999, who observed a neoteric blood supply in malignant melanoma formed by stem-like cancer cells transdifferentiating into an endothelial-like phenotype and forming vascular-like networks (45). Neovascularization via endothelial transdifferentiation has subsequently been reported in various other solid tumors, including breast, prostate, lung, and glioblastoma (ref. 46; Fig. 1B). Conceptually, the concept of vascular mimicry is akin to vasculogenesis during embryonic development, providing a conduit for increased plasma and blood cells to meet the metabolic demands necessitated by rapid growth. Not surprisingly, therefore, cancer cells undergoing vascular mimicry have been reported to hijack embryonic vasculogenesis programs including activation of Nodal signaling (47) and upregulation of vascular endothelial-cadherin (VE-cadherin; ref. 48), which functions as an “adhesion zipper” between the endothelial-like cancer cells. Interestingly, the development of Kaposi sarcoma is characterized by the transdifferentiation of endothelial cells into epithelial tumor cells (49), hinting that the epithelial–endothelial transdifferentiation axis is bidirectional.

Melanomas that are resistant to angiogenesis inhibitors display greater levels of vascular mimicry (46). This could be the result of the inability of these inhibitors to effectively suppress cancer cell–derived neovascular networks, a hypothesis supported by in vitro studies using human melanoma cell lines (50). Although definitive evidence of treatment-induced vascular mimicry has yet to be observed, it is tempting to speculate that the poor clinical efficacy of angiogenesis inhibitors could be due, at least in part, to the phenomenon of epithelial-to-endothelial transdifferentiation.

Although there are clear distinctions between epithelial–mesenchymal, epithelial–neuroendocrine, epithelial–squamous, and epithelial–endothelial plasticity, it must be noted that there is also substantial overlap between the features and mediators of these phenomena (Fig. 1 and Box 1). An important concept is that these various transitions are generally associated with hybrid metastable states (Fig. 1A) characterized by elevated plasticity and stem-like features. Notably, breakdown of lineage commitment and transition into a hyperplastic state is not unique to cancer but is also observed during wound healing when stress-responsive enhancers become activated and override homeostatic enhancers that govern lineage specificity (51).

Another important concept when considering lineage plasticity is that a population of cells within a heterogeneous tumor can be endowed with heightened stemness, plasticity, and drug resistance, which is frequently referred to as the “cancer stem cell” or “therapy-tolerant persister” paradigm. When modeling acute responses to anticancer therapies across drug-sensitive human cancer cell lines, Settleman and colleagues made the initial discovery that a small population of “cancer stem–like cells (CSC)” maintain viability (52) and serve as a reservoir from which heterogeneous drug-resistance mechanisms can evolve to enable tumor reestablishment. Notably, the persister state was found to be transiently acquired and relinquished by individual cells within a population (52), suggesting that tumor cells are constantly in flux along a continuum of “stem-ness.” The intimate relationship between stem-like states and lineage plasticity is an important principle when considering transcriptional and epigenetic mechanisms governing cancer cell state transitions.

The coordinated actions of lineage-specific TFs are essential for normal development. In contrast, this tightly coordinated process goes awry in cancer, rerouting transcriptional pathways to enhance cell plasticity and drive disease progression. In this context, a major sequela of the malignant state is activation of aberrant lineage-specific TF activities. For example, therapy-mediated epithelial–neuroendocrine transdifferentiation in prostate cancer involves activation of BRN2, ASCL1, FOXA2, and N-MYC, TFs that mediate neural differentiation programs in normal development (53–56), whereas squamous transdifferentiation of pancreatic and lung adenocarcinomas is mediated by p63, a master transcriptional regulator of the normal squamous lineage (57, 58).

Perhaps the most prominent example of lineage-specific TFs that are aberrantly activated in cancer to drive plasticity are the core EMT-TFs of the SNAIL, TWIST, and ZEB families (59). These TFs enable epithelial cancer cells to acquire features associated with a mesenchymal lineage by directly repressing epithelial genes (e.g., CDH1, encoding E-cadherin) and activating mesenchymal genes (e.g., CDH2, encoding N-cadherin, and VIM, encoding Vimentin). However, the terminology “EMT-TF” is perhaps unfortunate, since this class of factors can also mediate epithelial–neuroendocrine (60–62), epithelial–endothelial (63), and endothelial–mesenchymal (64) cell state transitions as well as elicit dedifferentiation, in some cases independent of their EMT-inducing capacity (65–67). Thus, EMT-TFs could be regarded as “pan-plasticity” factors, at least in some cancer contexts (Box 1).

Acquisition of new lineage-specific TF activity in cancer cells often occurs concomitantly with the activation of one or more of the core embryonic stem cell pluripotency TFs, OCT4, SOX2, KLF4, and NANOG. These factors not only promote cancer cell dedifferentiation to a more stem-like state (68–71), as expected, but can also promote transdifferentiation, as illustrated by SOX2-mediated epithelial–neuroendocrine plasticity and EMT (70, 72, 73) and KLF4-mediated acinar-to-ductal phenotype switching in pancreatic cancer (74). Thus, pluripotency TFs can enable complex transdifferentiation/dedifferentiation coalescence that engenders highly plastic cell states.

Loss of developmental TFs that specify the original tumor lineage is another feature of cancer cell lineage plasticity, with some key examples being loss of PDX1 in pancreatic cancer, which enables tumors to suppress the epithelial state and gain mesenchymal/squamous (pancreatic) features (75); loss of the luminal epithelial lineage factor AR in prostate cancer, which can lead to neuroendocrine transdifferentiation and/or dedifferentiation to a more stem-like state (35, 76); loss of GATA3 in breast cancer, resulting in the acquisition of basal/mesenchymal phenotypes (77); loss of the lung epithelial lineage factor NKX2-1 in lung cancer, which is associated with switching to squamous cell states and various gut lineages (78); loss of HOXA5 and SMAD4 in colon cancer, which enables dedifferentiation to a progenitor-like state (79, 80); and loss of the master regulator of melanocyte differentiation, MITF, in melanoma, leading to reactivation of a neural crest progenitor state (81). A number of noteworthy elements related to lineage plasticity arising from reduced TF expression/activity are apparent. First, in some cases, the TFs that are lost were the initial primary drivers of tumorigenesis and growth—such as AR in prostate cancer—highlighting the extreme adaptability of tumor progression. Second, targeted therapies are a significant selection pressure leading to TF loss and/or inactivation. Treatment of a GEMM of melanoma with the BRAF inhibitor vemurafenib causes downregulation of SOX10, a neural crest lineage–specifying TF essential for melanocyte development, which provokes tumor dedifferentiation and therapy resistance (82). Other prominent examples are loss of the sex steroid hormone receptors AR and ERα in prostate and breast cancers, respectively, in response to hormonal therapies that directly inhibit their activity (35, 83). Interestingly, the development of more potent hormonal therapies for prostate cancer has led to a substantial increase in the frequency of tumors exhibiting highly plastic phenotypes (84), a cautionary tale when applying therapies that inhibit lineage-specific TFs. Finally, paralleling other aspects of lineage plasticity (see below), epigenetic mechanisms frequently underpin loss of lineage-specific TFs. For example, hormone therapy–induced lineage switching in breast and prostate cancers has been linked to epigenetic silencing of the ESR1 and AR genes (85–87). Such epigenetic mechanisms could enable oscillating loss/gain of key epithelial TFs, in turn driving complex rounds of cell state switching, in response to microenvironmental cues and therapy.

The above examples imply scenarios whereby concomitant gain of TFs that enhance cell plasticity and loss of TFs that maintain differentiated epithelial states converge to yield unequivocal binary lineage switching. However, as mentioned above, such binary phenomena are likely to be rare and associated with reduced cellular flexibility. Instead, selection pressures enable the acquisition of high-plasticity hybrid states; interplay between, and opposing actions of, TFs are important mediators of such states. In the hybrid EMT setting, TFs that counteract EMT-TFs to prevent full transition have been referred to as phenotypic stability factors (59); one such factor is GRHL2, which supports the proliferation of epithelial cancer cell growth (88, 89). Similarly, hybrid states are also observed in cell line models of prostate cancer and patient tumors undergoing epithelial–neuroendocrine transdifferentiation, whereby TFs mediating the luminal epithelial state (i.e., AR, FOXA1) can be coexpressed with TFs driving neuronal and stemness transcriptional circuitry (i.e., N-MYC, SOX2, ASCL1, BRN2, etc.; refs. 34, 35).

In addition to gain and loss of TF function, modulation of the activity of cancer regulatory TFs to support cancer cell lineage plasticity is frequently observed, a process referred to as TF “reprogramming” or “rewiring.” One notable example is the altered function of two key prostate cancer–associated TFs, AR and FOXA1, in response to AR-targeted therapies. In the normal prostate and in early-stage prostate cancer, FOXA1 and AR play complementary and critical roles in maintaining the luminal epithelial lineage (90, 91). However, both chronic and acute AR-targeted therapies can cause redistribution of AR/FOXA1 cistromes from epithelial-associated to neuroendocrine-associated regulatory elements in cell lines, PDXs, and patient tumors (34, 92, 93), illustrating a paradigm whereby TF reprogramming is a driver of cell state switching. The mechanisms underlying TF reprogramming in cancer cell lineage plasticity are complex and multifactorial but include altered expression/activity of other TFs and chromatin-modifying factors, often in a combinatorial manner. Indeed, reprogramming of FOXA1 and AR activities in prostate cancer can be mediated via interplay with pluripotency TFs (OCT4 and NANOG) and the histone methyltransferase EZH2, leading to altered cistromes, interacting protein partners, and gene targets (34, 68, 69).

A critical epigenetic mechanism underlying cancer cell lineage plasticity is the addition and removal of posttranslational modifications of histones, the so-called histone code. In particular, a histone code state characterized by the simultaneous enrichment of transcriptionally active [e.g., histone H3 lysine 4 monomethylation (H3K4me1) or trimethylation (H3K4me3)] and transcriptionally repressive [e.g., H3K27 trimethylation (H3K27me3)] posttranslational modifications, termed bivalent chromatin, plays a major role in determining the expression of plasticity-associated genes. Bivalently marked genes tend to exhibit a low level of transcription that can be activated by the loss of H3K27me3 or repressed by the loss of H3K4me3, a status referred to as transcriptionally “poised” (94). In normal development, chromatin bivalency is enriched at elements regulating the expression of lineage-specific TFs and other developmental factors, enabling dynamic and rapid control of cell state (95, 96). A general feature of lineage plasticity is hijacked control of the bivalent chromatin state, resulting in the activation of genes that promote plasticity and repression of genes that block plasticity, which collectively enable dedifferentiation and transdifferentiation processes (Box 1). For example, in prostate cancer, bivalent chromatin controls genes involved in epithelial and neuronal identity; stimuli that enhance cell plasticity (i.e., androgen withdrawal or overexpression of N-MYC) cause an increased H3K27me3:H3K4me3 ratio at epithelial genes and a decreased ratio at neuronal genes, leading to acquisition of neuronal features (97). Moreover, genes encoding EMT-TFs and other plasticity-associated TFs such as N-MYC and SOX9 are dynamically regulated by bivalent chromatin during the metastatic progression of melanoma, enabling dynamic control of mesenchymal cell states (98).

Enzymes responsible for controlling histone posttranslational modifications, and in particular the bivalent chromatin state, are key regulators of lineage plasticity. EZH2 is the enzymatic subunit of the Polycomb-repressive complex 2 (PRC2), which catalyzes the deposition of H3K27me3. Inhibition of EZH2 activity can block and, in some cases, reverse epithelial–neuroendocrine transdifferentiation in human and mouse models of prostate cancer and small cell lung cancer (SCLC; refs. 34, 56, 72, 99–102). By regulating the deposition of H3K27me3, EZH2 also promotes EMT in pancreatic (103) and prostate cancers (104), the latter via direct repression of the CDH1 promoter. Another key mediator of chromatin bivalency and the histone code more generally is LSD1, an enzyme that demethylates H3K4me1, H3K4me2, H3K9me1, and H3K9me2. In embryonic stem cells, LSD1 regulates pluripotency and stemness by controlling chromatin bivalency at key developmental genes, including lineage-specific TFs such as FOXA2 (105). Given its overarching role in pluripotency, it is not surprising that dysregulation of LSD1 in cancer cells has been linked to EMT (106, 107), epithelial–neuroendocrine plasticity (108, 109), and maintenance of cancer stem cell states (110). EZH2 and LSD1 are increasingly recognized as therapeutic targets, a concept explored in more detail below.

Another key epigenetic factor regulating cancer cell lineage plasticity is DNA methylation, a modification that is deposited by DNA methyltransferases (DNMT1, DNMT3A, and DNMT3B) and removed by ten-eleven translocation proteins (TET1, TET2, and TET3). DNMT1, in particular, has been identified as a major mediator of plasticity. In prostate cancer, upregulation of DNMT1 in neuroendocrine prostate cancer (NEPC) tumors is associated with dramatically altered DNA methylation patterns; neuronal, cell–cell adhesion, developmental, EMT, and stem cell transcriptional programs are enriched within genes exhibiting altered methylation, providing evidence that DNMT1 is a direct mediator of plasticity in this disease context (21, 111). In support of this theory, DNMT1 inhibition dampens the neuroendocrine phenotype in cell line, organoid, and xenograft models of prostate cancer (112). Paralleling these observations, SCLC is characterized by elevated DNMT expression and exhibits similar DNA methylation profiles to NEPC, including hypomethylation of neuronal TFs ASCL1, HES6, and ONECUT2 (111). Elevated DNMT1 activity and specific DNA methylation profiles have also been linked to EMT and stemness in breast cancer (113, 114) and pancreatic cancers (115). Although the above examples reveal a robust link between elevated DNMT1 and lineage plasticity, this principle does not appear to apply to other DNA methylation–modifying enzymes. As examples, DNMT3A loss-of-function mutations can drive stemness and malignant progression of leukemia cells (116, 117), while TET1 can enhance self-renewal and expansion of CSCs in triple-negative breast cancer cell lines (118). The precise mechanisms underlying these complex relationships between DNA methylation and cell plasticity states remain to be elucidated.

The unprecedented resolution afforded by single-cell profiling has deepened our understanding of the origins and transitions between transcriptional states as well as the evolutionary paths that tumors follow. Indeed, such profiling has demonstrated that cells can display both developmental and unique evolutionary trajectories, which often involve a transient increase in cellular plasticity followed by the adoption of increasingly metastatic states (119).

Pseudotemporal approaches to evaluate single-cell data have reinforced the notion that perturbation of tumor suppressors or therapeutic pressure accelerates plasticity-mediated progression by creating novel cellular trajectories as well as illustrating the complexity and heterogeneity of cell state changes. For example, single-cell transcriptomic/epigenomic analysis of tumor progression in a GEMM expressing an oncogenic Kras-G12D mutant in combination with Tp53 deletion, which recapitulates human lung adenocarcinoma progression, revealed a loss of fidelity of the lung lineage coincident with the acquisition of transcriptional features associated with lung progenitors, embryonic ectoderm, and EMT (36, 120). In contrast to embryogenesis, tumor progression in this model was characterized by the persistence of earlier cell states coincident with the acquisition of new cellular states, giving rise to intratumoral diversity. A similar phenomenon has been described in SCLC at the single-cell level, with multiple, concurrent plasticity-related resistance mechanisms and cell states (e.g., EMT, neuroendocrine-like phenotypes) coexisting within the same tumor (121, 122). Likewise, in prostate cancer, single-cell RNA and chromatin accessibility profiling of patient tumors, organoids, and cell line models has demonstrated that AR-targeted hormonal therapies cause the onset of multilineage phenotypes comprised of epithelial, neuroendocrine, mesenchymal, and tuft-like cell states, among others (41, 123, 124).

Single-cell profiling has also provided new insight into the mediators of cancer cell lineage plasticity, uncovering evolutionary bottlenecks and vulnerabilities that could pave the way for new antiplasticity therapies. Use of single-cell chromatin accessibility profiling detected heterogeneous self-renewing populations of stem-like cells in human glioblastoma multiforme (GBM) tumors (125). Further work revealed that YAP/TAZ signaling is responsible for instilling stem-like properties; inhibition of this pathway irreversibly locked differentiated GBM cells in a nontumorigenic state, preventing plasticity and regeneration of the stem-like cell compartment (126). Likewise, intratumoral subtype switching in SCLC has been reported as a mechanism of acquired resistance to platinum-based chemotherapy (127). Longitudinal single-cell transcriptome analysis of SCLC tumors uncovered a role for MYC-mediated NOTCH signaling activation in promoting a temporal shift from ASCL1-positive SCLC to a nonneuroendocrine YAP1-positive state, a subtype switch that could be blocked with a NOTCH inhibitor (128). Finally, single-cell studies in prostate and breast cancer have implicated JAK–STAT and FGFR signaling in lineage plasticity from luminal epithelial to neuroendocrine and mesenchymal states, revealing pharmacologic inhibition of JAK–STAT as a viable therapeutic strategy (123, 124, 129).

Despite the increasing catalog of single-cell data, findings in relation to cellular trajectories inferred by pseudotemporal approaches should be taken with caution, as they make the assumption that transcriptional similarity denotes a developmental relationship. The next wave of single-cell technologies that enable lineage barcoding (for review, see ref. 130) will provide a more robust capture of plasticity, cell state transitions, and clonal mosaicism. Indeed, application of a lineage recorder system that exploits CRISPR–Cas9-induced insertions and deletions to a mouse xenograft model of metastatic pancreatic cancer reinforced the notion—originally proposed from bulk transcriptomic data—that cells occupy a continuum of EMT states and that “late-hybrid” cells have the highest metastatic potential (131). More recently, an equivalent CRISPR–Cas9-based lineage tracing system was incorporated into the Kras-G12D mutant/TP53 deletion lung cancer GEMM, yielding an autochthonous mouse model that provided important insights into tumor evolution, plasticity, and subclonal origins (119).

The intimate relationship between metabolism and tumor cell plasticity is evident by the distinct metabolic profile between stem-like and differentiated cell populations within a heterogeneous tumor. In general, stem-like cells exhibit suppression of oxidative phosphorylation (OXPHOS), resulting in redox niches with low reactive oxygen species (ROS) levels. Multiple epigenetic and transcriptional drivers of lineage plasticity have been demonstrated to control this metabolic reprogramming (Fig. 2). For example, in mouse models of liver cancer, NANOG directly represses OXPHOS genes and concomitantly activates fatty acid oxidation to support cancer cell self-renewal and drug resistance (132). Similarly, JAK–STAT pathway activation (described above as a feature of lineage plasticity) promotes fatty acid β-oxidation activity to sustain breast cancer stem-like cells (133). A switch from OXPHOS to glycolysis is also a common feature of lineage plasticity in tumors. The SNAIL–G9a–DNMT1 complex mediates promoter methylation and repression of fructose-1,6-biphosphatase (FBP1) in breast cancer cell lines, leading to a switch from mitochondrial OXPHOS to glycolysis that supports EMT, stemness, and tumorigenicity (134). Acquisition of enhanced glycolysis is also associated with epithelial–neuroendocrine plasticity in prostate cancer, mediated, at least in part, by aberrant epigenetic/transcriptional activities of the histone lysine demethylase KDM8 and HIF1α (135).

The above examples describe how epigenetic enzymes and TFs mediate metabolic reprogramming concomitant with increased cancer cell plasticity. However, the interplay between these processes is very much a two-way street, perhaps best epitomized by the fact that the lineage-shaping activities of chromatin-modifying enzymes are dependent on metabolic pathways and oncometabolites that provide the “fuel” for epigenetic modifications, particularly DNA and histone methylation (Fig. 2). For example, one-carbon metabolism is central to lineage plasticity due to its role in the synthesis of substrates for methylation reactions. The methionine cycle controls the conversion of methionine into S-Adenosyl methionine (SAM), the universal methyl donor for histone and DNA methyltransferase enzymes, such as EZH2. Accordingly, regulation of SAM metabolism dictates a cell's ability to impart histone and DNA modifications to support lineage reprogramming (136, 137). Indeed, methionine has been reported to be required for the maintenance of human embryonic stem cells (138). Similarly, in malignancy, stem-like cells rely on methionine and SAM biosynthesis for their self-renewal and tumor-forming capacity (139, 140). In triple-negative breast cancer, methionine deprivation is associated with H3K4me3 demethylation and suppression of the proplasticity/developmental TF SOX9, effects that were partially rescued by SAM supplementation (140). Increased SAM levels have also been reported during prostate cancer epithelial–neuroendocrine transdifferentiation (112). Beyond SAM, other oncometabolites that support lineage plasticity via dysregulation of DNA and histone methylation include fumarate and 2-hydroxyglutarate, both of which can inhibit TET DNA demethylases and histone demethylases: by suppressing TET-mediated DNA demethylation of miR-200 promoters, fumarate was found to stabilize EMT-related TFs (141), whereas 2-hydroxyglutarate was recently demonstrated to mediate an altered chromatin landscape that drives breast cancer cell state transitions and heterogeneity (142).

The microenvironment in which a tumor grows can have a profound impact on cancer cell lineage plasticity. These external cues have been referred to as extrinsic regulators of plasticity (143), as opposed to cell-intrinsic factors that are predominantly described above. Although in many cases the precise mechanisms underlying the cross-talk between the microenvironment and cancer cell plasticity remain to be uncovered, an emerging body of work is elucidating epigenetic and transcriptional mechanisms underlying this phenomenon.

Growth factor secretion by cancer-associated fibroblasts (CAF) can substantially influence the transcriptional and epigenetic regulators of cancer cell plasticity. For example, Lau and colleagues demonstrated that hepatocyte growth factor release by CAFs induced the AP-1 TF FRA1 (encoded by FOSL1) in liver cancer cells; subsequent transcriptional changes mediated by FRA1, including induction of the NOTCH signaling factor HEY1, resulted in enhanced plasticity, stemness, and cancer-initiating capacity (144). Acidic fibroblast growth factor (FGF1) derived from pancreatic cancer CAFs can signal through AKT and GSK-3β to increase MYC half-life in tumor cells, which increases MYC promoter occupancy and transcriptional activity (145). Another study found that NSCLC CAFs release various growth factors, most notably insulin-like growth factor 2, to induce the stem TF NANOG in lung cancer, which drives transcriptional programs associated with stemness and plasticity (146). Collectively, this body of evidence supports the concept that paracrine signals play a major role in establishing transcriptional and epigenetic programs required for cancer cell plasticity. Importantly, the composition of such signals and the subsequent impact on the cancer cell state will depend on the specific microenvironmental niche; lung cancer brain metastases exhibited epithelial–neuroendocrine lineage switching, but this was reversed when the tumor cells were separated from brain stroma (147).

Rapid tumor growth frequently outstrips the process of vascularization, resulting in a hypoxic microenvironment. An emerging knowledge base has linked hypoxia to epigenetic and transcriptional reprogramming of plasticity in cancer cells. Arguably the most well-studied mechanism underlying this relationship is induction of the TF HIF1α, which occurs in response to hypoxia and can directly activate genes encoding mediators of EMT, such as TWIST, ZEB1, SNAIL, lysyl oxidase, and TGFβ (148–151). HIF1α can also regulate other forms of cancer cell lineage plasticity: in hypoxic GEMM prostate tumors, HIF1α protein is stabilized and can interact with the TF FOXA2 to drive expression of a gene expression program that mediates epithelial–neuroendocrine plasticity and metastasis (152). Beyond cell-autonomous mechanisms, hypoxia can also modulate stroma to influence the plasticity of cancer cells. For example, hypoxia induces expression of VEGF by stromal cells, which can activate stem-related transcriptional programs in breast and lung cancer cell lines via MYC and SOX2 (153).

Inflammation is another important microenvironmental factor that is intimately connected to lineage plasticity. Early lineage tracing studies were key in establishing this relationship. For example, a murine model of bacteria-induced prostatic inflammation resulted in basal–luminal transdifferentiation that was linked to carcinogenesis (154). Although the epigenetic and transcriptional bases that link inflammation and cancer cell plasticity remain to be fully established, one emerging theme is the activation of transcriptional regulators such as NF-κB, JAK–STAT3, and NOTCH by proinflammatory chemokines and cytokines (e.g., TNFα, IL1β, IL6, IL8, and CXCL8). Indeed, IL6 can elicit stemness by activating JAK–STAT3 and NOTCH in breast cancer models in vitro and in vivo (155, 156). More recent studies have also highlighted the essential role of JAK–STAT inflammatory signaling in prostate cancer lineage plasticity, whereby this pathway is essential for establishing complex admixed cellular phenotypes (including neuroendocrine-like and stem-like populations) and resistance to AR-targeted therapies (123, 124). Soluble proinflammatory factors are frequently produced at high levels in cancer cells but can also be derived from other tumor microenvironmental components, most notably myeloid cells (156, 157) and CAFs (158).

Mechanical features of the TME can also regulate cancer cell transcriptional and epigenetic plasticity. Integrins in cancer cells sense TME stiffness and mediate activation of key plasticity TFs, such as TWIST and YAP/TAZ, leading to EMT transcriptional programs (159, 160). YAP/TAZ-mediated mechanotransduction was also reported to enhance stem-like traits in breast and pancreatic cancer cell lines via modulating autophagy (161).

Two other aspects of the tumor:microenvironment interplay in relation to cancer cell plasticity are worth noting. First, although our discussion has had a unidirectional focus—that is, microenvironmental cues that influence transcriptional and epigenetic regulators of plasticity—this is, of course, a bidirectional relationship. An elegant example of such bidirectionality was provided by Somerville and colleagues, who demonstrated that squamous transdifferentiation of human and mouse pancreatic cancer cells results in the conversion of pancreatic stellate cells into inflammatory CAFs; mechanistically, this was mediated by p63-mediated transcription of genes encoding proinflammatory secreted factors in the cancer cells (37). Thus, p63 regulates a transcriptional program that simultaneously drives squamous transdifferentiation (57, 58) and fibroblast reprogramming (37). Second, experimental conditions can dramatically influence cancer cell plasticity, which has significant implications for studying the influence of the microenvironment on this process. Indeed, cell states of pancreatic adenocarcinoma cell lines and organoids were highly sensitive to culture medium, and artificial states altered anticancer drug responses (162). Additionally, nonmalignant and neoplastic mammary epithelial cells can gain stem-like characteristics via epigenetic mechanisms when grown in two-dimensional culture, but this phenomenon is mitigated if cells are cultured within three-dimensional matrices (163). Thus, approaches to study the impact of the tumor microenvironment on transcriptional and epigenetic mechanisms underlying cancer cell plasticity should use models and optimize culture conditions that faithfully recapitulate patient-relevant cell states (162, 164).

The ability of cancer cells to evade attack and elimination by immune cells is a hallmark of cancer and emerging evidence points to a critical role for nongenetic cell plasticity in promoting nonimmunogenic cell states (Fig. 3 and Box 1). This concept was first established via the identification of a relationship between epithelial–mesenchymal plasticity and an immunosuppressive phenotype (165, 166). Surveying clinical data from lung cancer, a strong correlation between “tumor EMT score” (based on expression of 76 canonical EMT genes) and expression of the immune checkpoint PD-L1 was uncovered (165). This finding was validated in a subsequent large-scale analysis of 11 cancer types, which demonstrated that tumors with a high EMT score exhibited elevated expression of immune-checkpoint proteins, including PD-1, PD-L1, PD-L2, and CTLA-4 (166). Subsequent work has begun to provide mechanistic insight into the link between EMT and immune evasion. For example, an elegant set of experiments from the Weinberg lab using a SNAIL-driven fluorescent reporter in the MMTV-PyMT breast cancer mouse model demonstrated the relevance of this EMT-TF in directly mediating resistance to immune-checkpoint inhibitors (ICI; refs. 167, 168). Although SNAIL^lo epithelial cells were immunogenic, SNAIL^hi mesenchymal-like cells promoted an immunosuppressive tumor microenvironment characterized by cytotoxic T lymphocyte exclusion, M2 protumor macrophage polarization, and decreased response to anti–CTLA-4 therapy. Interestingly, in mixed SNAIL^lo/SNAIL^hi tumors, only a small proportion (<10%) of mesenchymal-like cells were required to shield the surrounding epithelial cells from immune attack (167), likely due to secretion of immunosuppressive factors, such as TGFβ, by mesenchymal-like cells (169). SNAIL directly binds to the promoters of key immunosuppressive paracrine factors, including Cd73 (Nt5e), macrophage colony-stimulating factor (Csf1), and secreted phosphoprotein 1 (Spp1), identifying a putative mechanism by which it inhibits antitumor immune activity (168). A similar immunosuppressive mechanism was identified in ovarian cancer whereby SNAIL directly binds to and upregulates the expression of genes encoding the CXCL1 and CXCL2 chemokines, leading to infiltration of myeloid-derived suppressor cells (MDSC; ref. 170). Direct transcriptional regulation by ZEB1 can activate expression of immunosuppressive factors (PD-L1 and CD47; ref. 171) and repress expression of T cell–attracting chemokines, such as CXCL10 (172). Thus, direct transcriptional control of immunoregulatory factors by EMT-TFs could be a key contributor to tumor immune evasion and response to immunotherapy. Interestingly, positive feedback loops between tumor cells and immune cell populations can enhance this phenomenon: for example, MDSCs can reinforce an EMT/stemness phenotype and facilitate metastasis in syngeneic mouse models of breast cancer (173).

Epigenetic programs associated with high-plasticity tumor states also play a critical role in regulating cancer cell immunogenicity. For example, bivalent chromatin or DNA methy­lation at the promoters of key immune regulators, including major histocompatibility complex (MHC) class I genes, the type III interferon (IFN) receptor IFNLR1, and T cell–attracting chemokines (CCL5, CXCL9, and CXCL10), are associated with transcriptional repression and evasion of T cell–mediated immunity in leukemia, SCLC, various types of squamous cell carcinoma and triple-negative breast cancer (99, 174–176). Collectively, these studies demonstrate that epigenetic reprogramming associated with lineage plasticity can mediate immune escape and reveal immunosensitization therapies, as discussed below.

Clinically, lineage plasticity in cancer is not always recognizable, particularly since changes may be dynamic and occur at the cellular level. The most obvious recognition of lineage plasticity is when a primary tumor or metastatic biopsy shows mixed histology or a change in phenotype that differs from the original tumor histology. In the case of histologic transformation (e.g., SCLC or NEPC conversion), clinical genomic assays such as targeted exome sequencing may be performed to exclude a second primary cancer, which confirms similarities with the original cancer's gene alterations (e.g., EGFR mutation in NSCLC) and sometimes acquisition of additional alterations associated with plasticity or treatment resistance (e.g., loss of RB1 and/or TP53; ref. 21, 177). Tumor protein assessment may be performed by IHC to confirm loss of target expression (e.g., loss of the luminal epithelial lineage factor AR in a subset of prostate tumors undergoing neuroendocrine transdifferentiation; ref. 30) and acquisition of alternative lineage markers (e.g., neuroendocrine markers such as chromogranin, synaptophysin), although mixed luminal/neuroendocrine features are frequently observed. Often, the clinical detection of transdifferentiation is after it has already occurred; early detection strategies are important to identify susceptible individuals and for the development of novel approaches to target the lineage plasticity process.

Serial biopsies are not always feasible to perform in patients, and therefore a comprehensive understanding of the timing of activation of transcriptomic and epigenetic drivers of lineage plasticity in individual patients has been a major challenge. Noninvasive approaches such as liquid biopsies and molecular imaging are in early development and hold promise as tools for the early detection and monitoring of lineage plasticity. For example, loss of luminal programs may be suspected in prostate cancer based on features such as progressive disease with low or nonrising serum prostate-specific antigen or loss of prostate-specific membrane antigen, the latter evaluated using PET/CT imaging (178, 179). CTC analyses across varied cancer types have revealed heterogeneity in epithelial marker expression, pointing to EMT changes that may contribute to metastasis (180). Cell-free DNA (cfDNA) also holds significant promise for detecting both genetic and epigenetic changes associated with lineage plasticity. Indeed, cfDNA methylation can distinguish NEPC from prostate adenocarcinoma (181), and other emerging approaches such as cfDNA nucleosome footprinting (182) and 5-hydroxy­methylcytosine (5hmC) sequencing (183) may more robustly identify lineage plasticity changes as well as other epigenetic drivers of disease progression. Additional studies applying liquid biopsy platforms to evaluate lineage plasticity, with the aim of improving early detection strategies, disease monitoring, and therapy selection, are warranted.

Distinguishing early facilitators, key drivers, and downstream consequences of lineage plasticity has important biomarker and therapeutic implications. For instance, an early facilitator such as RB1 loss may not be sufficient to drive a change in phenotype but might still be a useful biomarker for strategies geared toward the prevention of lineage plasticity (184). Identifying key drivers or markers of lineage plasticity, such as TF activity, stem cell features, transcriptome, and/or DNA methylation changes, may serve to guide strategies geared toward targeting or reversing the plasticity process. Detecting downstream consequences of lineage plasticity, such as INSM1 or DLL3 expression that only emerges when a neuroendocrine differentiated state occurs, may be leveraged to target the new emergent phenotype (185).

As lineage plasticity most often emerges as a mechanism for cancers to metastasize and/or bypass therapy, patients developing lineage plasticity often have a poor prognosis. There are no current strategies specifically tailored toward targeting the lineage plasticity process. As new therapies are developed in the context of lineage plasticity, the goal of treatment may be to prevent, reverse, or treat the new lineage. For example, based on the hypothesis that RB1 and TP53 alterations occur “before” the onset of SCLC lineage plasticity, an ongoing trial is using these markers to intervene early in patients with NSCLC with SCLC-chemotherapy regimens to potentially prevent cell-state conversions and improve outcomes (NCT03567642). Approaches to target EMT-TFs or other lineage determinant TFs have been particularly challenging but, with advances in drug development including protein degraders, are looking increasingly feasible (186, 187).

Using epigenetic therapies to reverse cancer cell lineage plasticity and induce “re-differentiation,” which in some cases can resensitize tumors to standard-of-care therapies, is a compelling concept. Notably, this strategy would only be effective in the context of cancer cell states that are capable of transitioning and not “locked in” to new lineages. One prominent example is prostate cancer, whereby acquired neuroendocrine-like states can be reversed by EZH2 inhibition, which in some cases results in reexpression of the AR and resensitization to AR-targeted therapies (34, 72). This effect is, in part, due to the catalytic function of EZH2 (i.e., H3K27me3) but also via a noncanonical role of EZH2 in driving lineage plasticity by forming transcriptional complexes with N-MYC and AR (34, 56). LSD1 inhibitors can reverse neuroendocrine phenotypes in prostate cancer and SCLC and are also emerging as a viable strategy to redifferentiate cancer cells for therapeutic bene­fit (109, 188). Various epigenetic therapies including EZH2 inhibitors (e.g., tazemetostat, valemetostat, PF-06821497, and CPI-0209) and LSD1 inhibitors (e.g., CC-90011, seclidemstat, and JBI-802) are in clinical development across solid tumors, but trials are currently in biomarker-unselected populations. As these epigenetic therapies target factors with pleiotropic functions, understanding how they modulate lineage plasti­city and/or act to “redifferentiate” tumors, and in which context, will be critical. Cotargeting epigenetic pathways and lineage-specific drivers in patients at high risk of developing lineage plasticity, or sequentially cycling between these distinct therapeutic approaches, are intriguing possibilities but are yet to be tested clinically.

Arguably the most exciting application of plasticity-suppressing epigenetic therapies is to prime tumors for immune rejection. As described above, key epigenetic mediators of lineage plasticity (i.e., EZH2, LSD1, and DNMT1) play a critical role in establishing chromatin states that result in the repression of immune-activating genes and activation of immune-inhibitory genes (Fig. 3). It is therefore unsurprising that targeting these enzymes potentiates response to ICIs (99, 174–176, 189, 190). Furthermore, epigenetic therapies can also activate immune signaling through the viral defense pathway, a process referred to as “viral mimicry.” More specifically, inhibitors of EZH2, DNMT1, and LSD1 lead to derepression of endogenous retroviruses, resulting in the accumulation of dsRNA and antitumor type-I IFN responses, a phenomenon that can be harnessed to improve tumor response to ICIs (191–197). Importantly, combining DNMT inhibitors with histone deacetylase inhibitors (HDACi) was found to enhance viral mimicry and immunotherapy responses in NSCLC and ovarian cancer models (193, 194). Collectively, these studies demonstrate that epigenetic reprogramming associated with lineage plasticity can mediate immune escape and uncover novel therapeutic strategies for immunosensitization.

Could lineage plasticity be mobilized to treat cancer? Histo­rical and emerging evidence suggests that highly plastic malignant cells may be susceptible to enforced transdifferentiation into postmitotic differentiated cells. This strategy has been applied for decades in acute promyelocytic leukemia by treatment with the retinoid all-trans retinoic acid, which induces terminal differentiation of leukemic cells into granulocytes via activation of retinoic acid receptor-α (198). The application of differentiation therapy in solid tumors has been long studied with several agents, including retinoids, HDACi, and PPARγ agonists, inducing differentiation in solid tumor models (198). Although this approach has been challenging to translate to clinical benefit, an increased understanding of differentiation and lineage plasticity processes has resulted in new opportunities. For example, in preclinical models of breast cancers developing EMT in vivo, forced transdifferentiation of cancer cells into postmitotic adipocytes via the combination of MEK inhibition and the PPARγ agonist rosiglitazone resulted in cell-cycle arrest and repression of tumor growth and metastasis (199). Another breast cancer study demonstrated that retinoids can reverse EMT (i.e., promote MET) and thereby block metastasis of breast cancer, which occurs, at least in part, by causing a metabolic switch whereby fatty acids are directed toward lipid storage rather than β-oxidation (200). In IDH1-mutant glioma organoids or models of IDH1-mutant biliary cancer that exhibit dedifferentiation due to accumulation of 2-hydroxyglutarate, inhibition of IDH1 or DNA methylation using azacytidine enables tumor control by promoting differentiation (201, 202). Chemotherapy-mediated lineage plasticity in urothelial carcinoma leads to semi-squamatization, and targeting of cathepsin H promoted terminal squamous differentiation and cell death via pyroptosis (203).

Since lineage plasticity is often detected late after changes in phenotype manifest clinically, patients are typically treated based on their new histology extrapolated from other cancer types. Nonetheless, they still might harbor features of their ancestral tumor of origin, which should also be considered for disease management. The prognosis of lineage plasticity tends to be poor regardless of tumor type or mechanism of plasticity. An increased understanding of the biology of lineage plasticity and collaborative efforts across cancer types will hopefully lead to better biomarkers and therapies for this growing subgroup of patients with cancer.

Lineage plasticity is a hallmark of cancer and a critical facilitator of other oncogenic features such as metastasis, therapy resistance, dysregulated metabolism, TME reprogramming, and immune evasion. The centrality of phenotypic plasticity in cancer progression necessitates dissection of the molecular mediators and regulators of this phenomenon and development of strategies to effectively target them. Herein, we have focused on the transcriptional and epigenetic mechanisms that mediate cell state transitions in cancer, in the process highlighting features of these phenomena that may be exploited for patient benefit.

Given the extreme epigenomic/transcriptomic heterogeneity in cell states that can be found in highly plastic tumors, the advent of high-throughput single-cell and spatial profiling techniques alongside tools to trace individual cells is anticipated to drive progress in this field. Similarly, tumor models that accurately recapitulate the various manifestations of plasticity and that can be sampled at relevant time points are critical. However, one major challenge will be to effectively unravel and exploit the cell atlases that will arise from these comprehensive large-scale approaches; high-throughput genetic perturbation strategies will assist in this respect.

Achieving a more complete understanding of the fundamental epigenetic/transcriptional mechanisms underlying cell state transitions in cancer will unlock new therapeutic strategies to target, and tools to monitor, lineage plasticity. In this respect, a key question that remains poorly understood is whether plasticity is a cause or consequence of therapy resistance. We propose that the evaluation of rational combinatorial strategies will provide insight into this question and are likely to be necessary to realize significant benefit for patients. Using epigenetic therapies to sensitize tumors to immunotherapies and/or resensitize tumors to drugs that block distinct oncogenic signaling pathways is a compelling possibility that is under clinical evaluation. Of course, context will be key for effective targeting of lineage plasticity for patient benefit. Although a goal of precision cancer medicine is to treat patients with the right treatment at the right time, excess plasticity enables cancer cells to effectively deal with environmental stressors, signals, and therapy by switching on the right phenotype(s) at the right time. Since epigenetics and TFs play a major role in this extreme adaptability, understanding their context-dependent activities will be crucial to improve precision cancer medicine.

A. Davies is an employee of and holds stock in Pfizer, Inc. A. Zoubeidi reports grants from the Terry Fox New Frontier Program Project, the Canadian Institutes of Health Research, the U.S. Department of Defense, and the Prostate Cancer Foundation during the conduct of the study. H. Beltran reports personal fees from Merck, Pfizer, Inc., Foundation Medicine, Blue Earth Diagnostics, Amgen, Bayer, Oncorus, LOXO, Sanofi, Curie Therapeutics, and AstraZeneca, grants and personal fees from Novartis, and grants from Bristol Myers Squibb and Circle Pharma outside the submitted work. L.A. Selth reports grants from the Cancer Council of South Australia, the Freemasons Centre for Male Health and Wellbeing, Cancer Council New South Wales, and the Flinders Foundation during the conduct of the study.

This study was supported by the Cancer Council of South Australia (2012127 to L.A. Selth); a joint grant from the Freemasons Centre for Male Health and Wellbeing and the Flinders Foundation (L.A. Selth); Cancer Council New South Wales (23-04 to L.A. Selth); the Terry Fox New Frontier Program Project (F22-00589 to A. Zoubeidi); the Canadian Institutes of Health Research (CIHR-169173 to A. Zoubeidi); the U.S. Department of Defense (W81XWH-17-1-0653 to H. Beltran; W81XWH-18-1-0690 to A. Zoubeidi); the Prostate Cancer Foundation; and the NIH/NCI (R37CA241486 to H. Beltran; P50-CA211024 to H. Beltran). L.A. Selth is supported by a Principal Cancer Research Fellowship awarded by Cancer Council's Beat Cancer project on behalf of its donors, the state government through the Department of Health, and the Australian Government through the Medical Research Future Fund. The authors apologize that, due to space limitations, many important studies in the field could not be cited.