A developmental constraint model of cancer cell states and tumor heterogeneity

Summary

Cancer is a disease that stems from a fundamental liability inherent to multicellular life forms in which an individual cell is capable of reneging on the interest of the collective organism. While cancer is commonly described as an evolutionary process, a less appreciated aspect of tumorigenesis may be the constraints imposed by the organism’s developmental programs. Recent work from single-cell transcriptomic analyses across 24 cancer types has revealed the recurrence, plasticity, and co-option of distinct cellular states among cancer cell populations. Here, we note that across diverse cancer types, the observed cell states are proximate within the developmental hierarchy of the cell-of-origin. We thus posit a model by which cancer cell states are directly constrained by the organisms’ ‘developmental map’. According to this model, as a population of cancer cells traverses the developmental map, it generates a heterogeneous set of states whose interactions underpin emergent tumor behavior.

In Brief

Single-cell transcriptomic analyses in diverse cancer types have revealed the recurrence, plasticity and co-option of distinct cellular states, leading to a “developmental constraint” model whereby cancer cell state plasticity is restricted by the organism’s developmental map to access progenitor-like states and differentiated-like states of adjacent lineages.

The tumor is a highly complex organ composed of myriad cell types, each in a range of cellular states¹. The advent of single-cell RNA-seq has permitted a global view of the transcriptomes of cancer and microenvironment cells across individual tumors^2,3. Such transcriptomic analyses have enabled the identification of resident cell types as discontinuous clusters, as well as a continuum within an individual cluster indicating diverse cellular states^4,5. Over the past decade, transcriptional variation has been identified within the malignant cell compartment across 24 cancer types in over 80 different studies^6–8. Importantly, this variation within the malignant cell population comes in the form of functionally-related sets of differentially expressed genes. For example, in cancer cells of lung adenocarcinoma tumors, one set of differentially expressed genes – a gene module – includes CAV1, AGER and LMO7; genes which were previously described as important markers of the alveolar type I differentiation cell state⁸. Cancer cells expressing this alveolar differentiation gene module were found to be associated with better survival in lung cancer patients with residual disease⁹. This relationship between the expression of gene modules in cancer cells and the fate of the patient underscores the importance of understanding their emergence and role in the tumor.

Notably, an individual cell is not necessarily restricted to the expression of a single gene module but rather several modules may be expressed in conjunction to achieve a range of phenotypic properties². Thus, a cancer cell’s state appears to be assembled by the heterogeneous expression of gene modules^8,10. From single cell transcriptomic analyses of an individual patient’s tumor we generally find, among the cancer cells, overwhelming evidence for a range of states², pointing to transcriptional heterogeneity as a critical characteristic of tumor biology.

In this Perspective, we provide a conceptual framework to synthesize the findings from published single-cell transcriptomic atlases and propose a unified model on the causes and constraints of tumor heterogeneity. This view provides a pathway from the descriptive characterization of cancer cell states across distinct cancer types towards a more mechanistic understanding of the developmental pressures acting on a tumor’s composition of cell states. Interrogating the developmental constraints inherent to a tumor’s cellular states offers a research program poised to lead to new therapeutic strategies and improved clinical outcomes.

Cell state recurrence, co-option of programs and cellular plasticity

Extensive work on the transcriptional variation among cancer cells has revealed three important properties. First, the gene modules underpinning states recur across tumors of the same cancer type¹¹, and even across cancer types⁸. In lung adenocarcinoma, tumors with distinct driver mutations (EGFR, KRAS and MET) transcriptionally converge to a population of cells in the alveolar, ciliated-like and proliferative states¹². Similarly, across the three genetic subtypes of breast cancer (ER+, HER2+ and triple-negative) malignant cells share common cancer cell states^13,14. Moreover, in glioblastoma¹⁰, prostate cancer^15,16, ovarian¹⁷ and colorectal cancer¹⁸, analyses of tumors with distinct driver mutations have revealed a convergence of transcriptional heterogeneity, indicating that the repertoire of cancer cell states is a fundamental property of tumorigenesis. This, in turn, implies conservation of a range of functional properties that may be required to sustain tumors and fuel progression.

Second, the recurrent gene modules are mostly not uniquely expressed in cancer, but rather shared with gene modules induced in normal physiology and tissue homeostasis. One class of these shared – or co-opted – gene modules relate to biological processes such as stress¹⁹, the interferon response²⁰, and hypoxia programs²¹. The stress-like cell state occurs at a higher frequency among cancer cell populations than non-cancer ones¹⁹. Pan-cancer studies defined a catalog of recurrent gene modules, including one with the genes ISG15, ISG20, IFI6, HLA-DR; which are collectively expressed in the interferon response^7,8. It has also been shown that cancer cell-derived interferons regulate functional polarization of tumor-associated monocytes, predicting a better response to immunotherapy²⁰.

Beyond these biological process-related gene modules, a second class of gene modules corresponds to cell identity. For example, the pigmented cell state in melanoma expresses melanocytic gene programs of the skin²², and mature luminal-like cell states in breast cancer express programs of luminal cells in the breast¹³. Collectively then, the gene modules underlying the cancer cell states appear to be co-opted from developmental and differentiation programs. The cancer cell-of-origin is co-opted in many cancers, in addition to the co-option of states reminiscent of other tissue specific cell types, as we discuss below.

A third crucial aspect of cancer cell states is their potential to transition between states by invoking cellular plasticity. Using single-cell and lineage tracing technologies, studies in a number of cancer types have revealed the occurrence of remarkable cellular plasticity during tumor evolution²³. Murine lung adenocarcinoma comprises a transcriptionally distinct subpopulation of ‘highly plastic cancer cells’, which when transplanted into a syngeneic mouse are able to generate a tumor that reconstitutes the full transcriptional heterogeneity of the original tumor²⁴. In vivo studies have also revealed the switching between the invasive and proliferative subtypes of melanoma, supporting plasticity between melanoma cell states²⁵. In rhabdomyosarcomas, the mesenchymal cancer cell state has been described as having tumor initiating capacity and inherent plasticity to recapitulate all the states of the initial rhabdomyosarcoma tumor²⁶. In glioblastoma, Neftel et al. isolated two of the four dominant glioblastoma states – mesenchymal-like and neural-progenitor like – and implanted subpopulations of three of the four dominant glioblastoma states – astrocyte-like, mesenchymal-like, and neural-progenitor like – revealing that these individual cell identity cancer states have tumor-initiating capacity and lead to tumors with a reconstitution of all four neuronal states¹⁰. Furthermore, using an in vivo barcoding experiment the authors suggest that all four glioblastoma cell states have the capacity to differentiate to other states.

The ability for individual cancer cells to reconstitute a tumor with the transcriptional diversity of the original demonstrates the highly plastic nature of cancer cell states, as well as their general non-genetic dependencies. In the same glioblastoma study¹⁰, tumors with distinct genetic alterations (EGFR, CDK4, NF1, and PDGFRA) were generally composed of the same four neuronal cell identity cancer cell states, though at different frequencies. For example, in tumors with EGFR amplifications, the frequency of cells in the astrocyte-like cell state was higher than that in tumors with other genetic alterations. Furthermore, genetic alterations in basal cells or luminal cells induce cell state changes, as has been shown in basal cell carcinoma^27,28. The effect on cancer cell state frequencies as a function of the genetic alterations suggests an interaction between these two basic mechanisms. We may refer to such a process as ‘plastigenic’, as it involves both genetic and non-genetic contributions. This is observed in gliomas, where IDH-mutant and IDH-wildtype gliomas harbor differential capacities for cellular plasticity²⁹. Interestingly, a plastigenic process reliant upon both genetic changes and the cell’s basic abilities for plasticity may be required for tumorigenesis.

Correspondence between cancer cell states and the developmental map

The recurrence, co-option and plasticity of cancer cell states has led us to note an apparently general pattern relating to their occurrence: within a given cancer type, the observed cell states are related along the developmental hierarchy. Consider melanoma, where it has been long known that gene expression programs can distinguish particular subtypes^30,31. Melanoma tumors have been classified into either the invasive/undifferentiated subtype, marked by the expression of AXL³², or the differentiated/proliferative subtype marked by the expression of melanocyte lineage factor MITF³³. Beyond the expression of marker genes, single-cell analyses have shown that melanoma cancer cells express cell type specific gene programs and are present in entirely distinct cell identity cellular states. Among these states are (1) a differentiated melanocyte cell state, (2) a neural crest-like cell state, and (3) an invasive/mesenchymal-like cell state^22,34–36. Importantly, the normal counterparts of these cell identity states are directly adjacent in the developmental map, according to which the cell types originate. Melanoblasts – thought to be the cell-of-origin for melanoma – originate developmentally from the ectoderm³⁷, which gives rise to neural crest cells³⁸. The states that recur in melanoma cancer cells thus correspond precisely to the developmentally adjacent cell states of the melanoblasts cell-type of origin.

Lung adenocarcinoma tumors originating from alveolar type II cells exhibit 7 cancer cell states^12,24,39,40: the (1) alveolar type I-like cell state enriched with the expression of alveolar type-I specific genes such as Ager, Cav1, and Hopx, (2) the alveolar type II-like marked by the expression of surfactant proteins Sftpc, Sftpb and Napsa, among others, (3) the mixed-alveolar-type I/II cell state expressing mixed lineage genes such as Muc5b, Nkx2–1, and Pdpn, (4) a club-like cell state that expresses secretary genes such as Scgb3a1 and Scgb3a2, (5) an embryonic-liver-like cell state marked by the expression of genes including Cldn2, and Dlk1, (6) a gastric- or gut-like cell state expressing genes such as Dgat2 and Itih2, and (7) an EMT cell state expressing the canonical markers such as Vim and Zeb1. Applications of single-cell transcriptomic and multiplex immunofluorescence have identified cell identity states along the developmental trajectory specifically in primary human lung adenocarcinoma tumors^12,39 as well as in pan-cancer studies^7,8. Five of these states are marked by the expression of gene expression programs from cell types within the lung tissue: the alveolar type I-like, alveolar type II-like, mixed-alveolar-type I/II, club and ciliated cell states, that map developmentally to cells derived from a common lung progenitor^41,42 (Figure 1). The gastric-like or gut cell-like state expressed gene modules of a developmentally primitive cell state – the foregut progenitor cell state. Finally, the embryonic-liver gene program^24,43 is enriched for the expression of a more distant cell-type that is developmentally accessible through the primitive gut progenitor state. These observations suggest that malignant cells in lung adenocarcinoma navigate both in the ‘reverse’ direction to de-differentiate towards progenitor cell states and ‘forward’ direction along the developmental map to differentiate to mature cell states.

Figure 1. Correspondence between lung cell types in development and cancer cell states in lung adenocarcinoma.

a, An illustration of dimensionality reduction projections of single-cell transcriptomic data from normal lung¹¹² and mouse embryonic developmental atlases¹¹³, which were used to generate the developmental hierarchy of cell types shown in c. b, Same as a, for cell identity-related cancer cell states in lung adenocarcinoma, as curated from single-cell transcriptomic studies^12,24,39. c, (Top) A simplified developmental hierarchy of cell types identified using the approach diagrammed in a, within the foregut branch. Marker genes are indicated for relevant developmental cell types. (Bottom) Lung adenocarcinoma cancer cell states identified using the approach diagrammed in b, are arranged according to the developmental hierarchy, matched by corresponding gene expression programs.

Recent work from Desai and colleagues has revealed that transgene activation in alveolar type I cells also leads to adenomatous transformation with a more indolent histology⁴⁴. At early stages of tumorigenesis, the tumors originating from alveolar type I cells transition through an intermediate state expressing both alveolar type I and alveolar type II marker genes that are normally mutually exclusive⁴⁴. During later stages of tumor development, the tumors harbor states expressing lung progenitor cell states and gastric cell states, reminiscent of cell states derived from alveolar type II cells⁴⁴. Thus, irrespective of the cell-of-origin, lung adenocarcinoma cancer cells appear able to traverse the foregut developmental branch to assume distinct states through developmentally restricted cellular plasticity.

In tumors of lung squamous cell carcinoma – another histological subtype of lung cancer – cancer cell states recapitulate the cell identities derived from the proximal lung progenitor, including the basal-like cell state, the ciliated-like cell state and the club-like state⁴⁵. Notably both lung histology tumors share the club-like and ciliated-like cell state (Figure 2).

Figure 2. The developmental map constrains the occurrence of cancer cell states in seven cancer types.

(Top) A simplified developmental lineage of cell types based on a singlecell transcriptomic atlas of mouse embryonic development⁴⁵ and tissue-specific reviews for pancreas,⁴⁸ lung,^41,42 skin,^37,38 brain,⁴⁹ muscles, and blood lineages.⁵⁰ (Bottom) A synthesis of cancer cell states as detected in 7 cancer types. To generate this map, we first curated a list of single-cell transcriptomic atlases for each cancer type: pancreatic ductal adenocarcinoma,^51,52 lung adenocarcinoma,^12,24,39 lung squamous cell carcinoma,⁴⁷ melanoma,^22,34,35,36 glioblastoma,¹⁰ acute myeloid leukemia,53,54,55 and rhabdomyosarcomas.²⁶ Second, we curated lists of the identified cancer cell states that are related to cell identity and excluded those related to biological processes (such as cycling and stress). Finally, we arranged the cancer cell states in accordance with the developmental map. Arrows indicate evidenced cell state transitions. Stars indicate the purported cell of origin. Developmental cell type and cancer cell states are colored by cancer type.

In glioblastoma tumors, Neftel et al. provided evidence that cancer cells assumed either an astrocyte progenitor-like state (APC), a neural progenitor-like state (NPC), an oligodendrocyte progenitor-like state (OPC) or a mesenchymal like cancer cell state (MES)¹⁰. Other studies have revealed a more granular classification of these cell states^46,47. The APC-like and OPC-like states are transcriptomically similar to the astrocyte progenitor cell and oligodendrocyte progenitor cell, respectively, and these are developmentally adjacent, deriving from a common glial progenitor cell⁴⁸. The NPC state is marked by gene expression programs similar to the neural progenitor cell that is developmentally upstream on the developmental map⁴⁸ (Figure 2). While in glioblastoma the cell-of-origin remains to be one of the major unresolved questions^49,50, the putative cell types are the oligodendrocyte progenitor cell^47,51 and differentiated neurons⁵². Thus, the glioblastoma cell state plasticity programs apparently enable malignant cells to not only assume upstream stem-like states, but also those differentiated states accessible through multiple common upstream progenitor states. If we assume that the oligodendrocyte progenitor cell is the cell-of-origin, the observed cell state heterogeneity suggests that the malignant OPC-like cells are constrained to traverse the map to exploit accessible states both in ‘reverse’ mode to de-differentiate into the NPC-like state as well as in ‘forward’ mode to differentiate into the APC-like state. Alternatively, if a differentiated neuron is the cell-of-origin for a glioblastoma, the malignant cell states would still traverse the developmental map in the ‘reverse’ mode to de-differentiate into the NPC-like state and the ‘forward’ mode to differentiate into the APC-like and OPC-like state.

Pancreatic ductal adenocarcinoma has been historically defined transcriptionally as either ‘classical’ – tumors with pancreatic progenitor gene expression programs – or ‘basal-like’ – tumors with basaloid or mesenchymal-like transcriptional programs^53–55. Single-cell studies have added insight into this classification by revealing that individual tumors simultaneously harbor subpopulations of both states⁵⁶. Specifically, these studies have identified (1) a cancer cell state co-opting the acinar cell program of the pancreas (acinar-like), (2) a classical cancer cell state expressing gene programs of the pancreatic progenitor cell and (3) a neuroendocrine-progenitor like cell state that shares transcriptional similarities to the endocrine progenitor in the pancreas^57,58 (Figure 2). The authors also identified two additional intriguing cell states: the basaloid and the squamoid⁵⁷ (Figure 2). Recent work has elucidated that large pancreatic ducts comprise dNp63 expressing basal cell-types⁵⁹, however, the developmental trajectory of these cells and the trajectory that tumor cells take to acquire these states are still being explored⁶⁰. Examining the developmental map, the basal and squamous cells of the esophagus are derived within the same foregut developmental path as the pancreas⁶¹. One prediction is thus that pancreatic cells assume the basal and squamous states by traversing the map through early pancreatic-esophageal progenitor states.

In mesoderm-originating tumors, a similar correspondence is found between the cell states of the tumor and those within the developmental lineage. For example, in rhabdomyosarcomas, the cell-of-origin is striated muscle and its tumors harbor (1) a subpopulation of cells in the differentiated muscle cell state (expressing genes such as like MYLPF, ACTC1, and TSPAN3), (2) a mesenchymal-enriched population (expressing extracellular matrix and mesenchymal genes including MMP2, CD44, PTN, POSTN and THY1), and (3) ‘ground state’-like cells reflecting a similarity to a muscle progenitor cell, which express muscle-lineage specific genes such as MYOD, MYOG and Desmin (indicating a commitment to the muscle lineage), but not specifically inducing myosin²⁶. Again, it is evident that the corresponding developmental stages are adjacent along the developmental map (Figure 2).

Landmark studies in acute myeloid leukemia (AML) have demonstrated the heterogeneity in this cancer type by identifying populations of cells with distinct cell surface marker expression⁶². Two main subpopulations have been recognized: (1) primitive AML cells or leukemic stem cells that sustain the tumor and promote therapeutic resistance and (2) differentiated AML cells that may retain hematological functions to create pathological interactions with the tumor microenvironment⁶³. While it was initially believed that transcription factor alterations in AML led to ‘differentiation blocks’⁶⁴, whereby progenitor states are locked-in, the application of single-cell RNA-seq has now revealed that malignant AML cells transcriptionally mimic features of cell states along the myeloid development hierarchy⁶⁵. During normal hematopoiesis, hematopoietic stem cells give rise to lineage-committed progenitor states, including the common lymphoid progenitor and common myeloid progenitor that ultimately give rise to mature cell types⁶⁶. AML tumors encompass cell states that recapitulate both progenitor states such as the hematopoietic stem-like state, the progenitor-like state and differentiated cell types such as the cDC-like cell state and the monocyte-like cell state^65,67,68. Consistently, cell states in AML tumors map to the developmental stages of hematopoiesis.

Notably, the ‘mesenchymal-like’ cell state appears to recur across different cancer types⁶⁹. Epithelial malignant cells often deploy a transition to a mesenchymal-like state to acquire motility and invasive properties by invoking the developmental program for an epithelial-to-mesenchymal transition (EMT)⁷⁰. While the mesenchymal-state recurs across cancer types, a recent pan-cancer study described five non-overlapping gene expression programs that independently describe a mesenchymal state. These EMT or mesenchymal-like gene expression programs vary with specificity to distinct cancer types⁷. This suggests that while there may be shared features of the mesenchymal state across tumor types, such as the expression of canonical marker Vimentin⁷¹, it is possible the cells retain a memory of its developmental history. Supporting the notion that the mesenchymal states encode tissue specific memory, are observations from reconstitution experiments, where the mesenchymal state from rhabdomyosarcoma and glioblastoma tumors is isolated and injected into a mouse the resulting tumors recapitulate the histology and cell state diversity of the original tumor, indicating the retention of elements from the original cancer initiating cell type^10,26. Furthermore, comparative analyses of the transcriptional dynamics during EMT have revealed that the transition to the mesenchymal cell state is largely context-dependent⁶⁹, and indeed the context-specificity of ‘EMT genes’ was also observed in primary tumors⁷². Another study developed a mesenchymal-stromal decoupling approach that revealed three distinct patterns of mesenchymal gene expression programs in squamous, gynecological and gastrointestinal tumors⁷³. Collectively, it appears that the mesenchymal state may be best interpreted as a common cancer motif⁷¹ with context-specific features involving contributions from developmental, epigenetic and genetic factors.

The developmental constraint model

As discussed above, support for the notion that cancer cell states co-opt developmental states comes from studies across several cancer types that have identified similar cell identity related states through independent analytical methods and datasets. Based upon the composition of the malignant cellular states in the seven cancer types described in Figure 2, it seems evident that the appearance of cancer cell states is constrained by the underlying developmental map. We thus propose a ‘developmental constraint’ model, whereby the malignant population – originating from the initial state of the cell-of-origin – traverses the states accessible by their developmental trajectory, both toward progenitor-like states and differentiated-like states of adjacent lineages, collectively forming a heterogeneous population. Specifically, the developmental constraint model restricts cell states transitions to occur only among states directly related in the developmental map.

The developmental constraint model indicates the importance of the cancer cell-of-origin in determining the cell state content of the heterogeneous malignant population. While for a number of cancer types the cell-of-origin remains unknown⁷⁴, in a few cancer types the cell-of-origin is known from studies of genetically engineered mouse models with oncogene activation. In lung adenocarcinoma, for example, where Kras activation in alveolar type II cells initiates tumors⁷⁵, the observed states are consistent with the notion of a cancer population traversing the developmental map to spawn the related cellular states (Figure 1).

Further support for the model derives from the evidence that tumors in a given tissue give rise to distinct cell states depending upon the precise cell-of-origin. Lung adenocarcinomas and lung squamous cell carcinoma tumors, for example, originate from developmentally distant cell types – alveolar type I/II and basal cells, respectively. Consistently, lung squamous cell carcinoma cell tumors co-opt the states from the proximal lung cell types while lung adenocarcinoma co-opt the states lung alveolar cell types. We note that in mouse models of lung squamous cell carcinoma Sox2 activation in combination with loss of Cdkn2ab and Pten in both alveolar type II or basal cells generate tumors with lung squamous cell carcinoma histology, however their specific patterns of cell state heterogeneity remain unknown⁷⁶. Therefore, notwithstanding the potency of driver mutations to determine tumor histology, the specific pattern of cell state heterogeneity may be determined by the tumor’s cell-of-origin.

Molecular encodings of developmental constraints

The developmental constraint model may enable insight into the function of the co-opted states in tumorigenesis, by studying the underlying molecular mechanisms and by reference to their roles within developmental pathways. The developmental constraints that restrict cell state transitions along the developmental map likely include aspects of chromatin remodeling required for cell state transitions. This follows from the observation that lineage-commitment programs in embryonic development and differentiation are typically accompanied by highly dynamic chromatin changes⁷⁷. For example, in the hematopoietic system, a population of hematopoietic stem cells gives rise to a variety of cell types by a specific sets of transcription factors and chromatin remodelers that establish regulatory chromatin states and subsequently lead to lineage differentiated cells^78,79. Analogously, during tumorigenesis, as an oncogene-transformed cancer cell-of-origin establishes a tumor, it is likely that chromatin changes enable the generation of tumor heterogeneity⁸⁰. We speculate that cancer cells rewire their chromatin state to a plasticity-permissive state that is epigenetically reminiscent of a developmental cell type.

Chromatin remodelers such as histone-lysine demethylases are frequently mutated in cancer and can promote activation of gene expression programs allowing bidirectional cell state transitions⁸¹. Furthermore, gain-of-function EZH2 mutations in lymphomas and myelomas alter the regulatory genome by expansive Histone3-Lys27 trimethylation that restrict the cells from specific cell states^82,83. More recently, Loukas et al. found that disrupting the epigenetic regulatory network led to a competitive advantage for cancer cells in different tumor microenvironments⁸⁴. Additionally, a number of studies have revealed the role of chromatin modifiers in establishing and regulating cancer cell states. For example, the histone demethylase PHF8 establishes a metastatic and invasive melanoma cell state⁸⁵ and inhibition of KDM1A, another histone demethylase, suppresses a neuroendocrine cell state in small cell lung cancer⁸⁶. We thus speculate that transcription factors and chromatin remodelers enable co-option of developmental epigenetic networks to establish specific regulatory elements such as bivalent promoters^87,88 and chromatin loops to reprogram enhancer-promoter contacts⁸⁹. Consequently, identifying chromatin accessibility-imposed constraints on a cell’s identity may reveal the principles by which the developmental map is encoded and exploited by the malignant cell population.

Relationship between the developmental constraint model and existing models

One long-discussed notion describes cancer as a de-differentiation process, suggesting ‘development in reverse’⁷⁰. While a transition to stem-like states is common between this and the developmental constraint model, the de-differentiation model does not account for the possibility of also assuming more differentiated states that are distinct from the cell-of-origin. As we have noted, in glioblastomas and lung adenocarcinoma the cancer cell population invokes several states within the accessible developmental map and acquires states that require ‘forward differentiation’.

The cancer stem cell (CSC) concept is another prominent model for tumor growth^90,91. Under this model, tumors are composed of sub-populations that are related by a hierarchy⁹⁰. At its apex, is a stem cell or stem cell-like population, whose proliferation leads to cellular states that ‘roll down’ to assume differentiated states⁹². A crucial aspect of the CSC model is that the differentiated states have a limited propagation potential, such that only the CSCs are able to initiate new tumors. Indeed, Tirosh et al. in their study of oligodendroglioma identified a stem-like state and described it as evidence for the CSC model¹¹.

Both the CSC and developmental constraint models agree on the occurrence of cell state heterogeneity and its maintenance. According to the CSC model, cellular heterogeneity results from stem-like cancer cells differentiating into non-CSC malignant cells. In contrast, in the developmental constraint model, heterogeneity occurs as the malignant cell population traverses in both directions: forward towards differentiation and in reverse, de-differentiating towards stem-like states. This is supported by studies in colorectal cancer that revealed the Lgr5+ stem cell population is replenished by Lgr5-cancer cells⁹³ and in melanoma, revealing that cancer cell states transition between identities to enable metastasis and tumor growth³⁵. Conceptually, it suggests that cancer cell states – rather than being organized hierarchically – can each access their inherent plasticity to traverse the available states within the constraints of the developmental map.

As we have reviewed above, recent work has revealed that the bulk of the tumor consists of rapidly proliferating and differentiated cell states⁹⁴. Moreover, there is evidence that distinct states – including stem-like states – are able to reconstitute the tumor by cellular plasticity. Essentially then, the developmental constraint model provides a contrasting model centered around ‘plasticity’ by which no population holds a privileged propagating potential, and operating within the contexts of developmental constraints that act to limit the range of possible cellular states.

Another related concept is lineage infidelity – or trans-differentiation – whereby a tumor converts from one histological subtype to another. Trans-differentiation has been most notably observed in lung and prostate cancer that enables resistance to therapy⁹⁵. Histological trans-differentiation of lung adenocarcinomas to aggressive neuroendocrine derivative (small-cell lung cancer) frequently occurs in patients with EGFR mutant lung adenocarcinoma treated with targeted therapy against mutant EGFR^96,97. Most recently, clinical and retrospective studies have reported that lung adenocarcinoma patients treated with targeted therapy against EGFR and KRAS mutations have demonstrated squamous transformation^98,99. During the lung adenocarcinoma to lung squamous cell carcinoma histological transition, cancer cell heterogeneity transitions from cells in ‘alveolar’ cell states to cancer cells expressing basal-cell related gene modules through a multi-lineage cell state¹⁰⁰. This can be readily explained by the developmental constraint model, as alveolar cell states are separated from basal cell states by multiple upstream progenitor states, though still within the same lung progenitor developmental branch (Figure 2). Similarly, in prostate cancer, under therapeutic pressure, such as anti-androgen therapy, prostate adenocarcinomas trans-differentiate to neuroendocrine cancers¹⁰¹; acquiring states within the same developmental lineages¹⁰².

Our model posits a strong constraint on the establishment of new states, even under intense evolutionary pressures such as therapeutic intervention. The mechanisms underlying trans-differentiation are an active area of research, and it will be interesting to study how therapeutic intervention remains constrained by the developmental lineage.

Predictions and limitations of the developmental constraint model

How the heterogeneity in the malignant population confers functional relationships at the tumor-system level still remains elusive. The developmental constraint model predicts that the cancer cell states may be co-opted for large-scale spatial organization to achieve the physiological requirements of the tumor. From developmental biology it is known that setting up the embryonic axes and cell type specification in epiblast derivatives during gastrulation occurs by signaling molecule gradients and transcription factor activity. A number of studies have shown that cells are allocated to specific fates according to their spatio-temporal position in the streak¹⁰³. Specifically, the nodal–SMAD2 and SMAD3 signaling pathway¹⁰⁴ and the canonical Wnt signaling pathway¹⁰⁵ are critical in the posterior-distal axis. We also have a detailed understanding of the mechanisms by which the “toolkit genes” are deployed to set up morphogenic signals (such as bicoid in Drosophila) and the gene circuits leading to the patterning¹⁰⁶. Cancer cells may co-opt such developmental pathways to set up and maintain spatial axes of orientation among the states accessible from the developmental map. This co-option has been observed in teratomas, malignant transformation of germ cells, where organogenesis and spatial patterning allows for the formation of entire organs such as teeth, bone and other structures¹⁰⁷. Morphogenic signals may also be co-opted by the cancer population to particular individual states or maintain the heterogeneity. Such collective behavior may function to allow the malignant population to set up evasion programs, whereby an axis through the tumor may allow cells facing the immune system to adopt immune-evasive cell states while tumor cells within the core adopt a more proliferative state^8,108. Spatial transcriptomics studies are uncovering structural relationships among the tumor’s cancer cell states, indicating plausibility of higher-order architectures¹⁰⁹. Additional overlapping codes may specify axes of stem-like states, migration, and interactions with elements of the tumor microenvironment such as fibroblasts¹¹⁰.

The relationship between cancer cell states and the developmental map may shed light on mechanisms of drug tolerance with direct impact on patient outcomes. A central aspect to the developmental constraint model is the transitions between cancer cell states. One application of this aspect as a new therapeutic direction would thus be to inhibit developmentally constrained cell state transitions. Inhibiting cellular plasticity would seek to reduce the heterogeneity of cell states present in the tumor compartment. Indeed, efforts to inhibit cell state transitions have revealed that inhibiting HDAC1 can suppress the acinar-to-ductal plasticity in pancreatic cancer, providing a therapeutic opportunity for pancreatic ductal adenocarcinoma patients¹¹¹. Furthermore, inhibiting developmentally constrained cell state transitions may provide an opportunity to restrict adaptation to drug treatment and mechanisms of acquired resistance.

Another prediction of the developmental constraint model would be to better assess the stage of the tumor. Progression along the map appears to be related to the aggressiveness of the tumor, as, for example, neuroblastoma is an undifferentiated and highly aggressive tumor, whereas ganglioneuroma is differentiated and essentially a benign version. Through a better understanding of the obstacles and assistance that cancer cells face we may be able to exploit a tumor’s cell state composition to direct and refine therapeutics for improved patient outcomes.

One limitation of the developmental constraint model is that it does not describe the extent to which the cancer cell states may travel along the developmental map. For example, the model does not explain the restraints on AML cancer cells that restrict differentiation to the states along the lineage’s states. It may thus be that there are additional lineage constraints at the chromatin level or cancer-specific features that prevent the accessing by the cancer of developmental states that are in principle available to it. It remains unknown, however, whether the malignant cells of a tumor cannot access such developmental states or whether these simply do not provide the cancer cell with a competitive advantage in the specific tumor environment.

Overall, the developmental constraint model provides insight into the unifying constraints and strategies that establish the composition of the malignant cell states across tumor types. Among its strengths are its ability to make testable predictions about the set of states that compose a tumor with a particular cell of origin, as well as a tumor’s stage of progression given only its state composition. This view of the constraints imposed on tumor heterogeneity is supported by observations regarding cellular states, however, further work is required to incorporate the mechanistic dependencies of the cell state transitions, their relationship to the tumor spatial architecture, and interactions with the microenvironment. It will be particularly interesting to reveal the adaptive pressures that maintain the heterogeneity of the distinct functional states and their role in tumor physiology.

Highlights.

Not relevant for the Perspective format.