The process of oncogenesis begins with the transformation of a single, founding neoplastic cell - a consequence of cell cycle dysregulation in conjunction with abrogated apoptotic signaling. This results in uncontrolled proliferation, tempered by resource limitation, overcrowding, and eventual toxic substrate accumulation - shaping new and local environmental conditions to be overcome ( Figure 1A ). Throughout tumor progression, individual cancer cells undergo a number of heritable molecular alterations that fuel evolution and heterogeneity. Clones with alterations that enhance cellular survival and proliferation experience increased fitness and undergo positive selection. Likewise, deleterious alterations result in decreased fitness, such that clones undergo negative selection and possible eradication from the tumor population. The resultant tumor population then is comprised of numerous subpopulations, each distinct in their abilities to access a range of advantageous cell states.

Several tumor progression models to date have been described ( Figure 1B ), including linear evolution, where mutations are acquired in a stepwise fashion and driver mutations fuel selective sweeps of clonal dominance throughout tumor evolution (20); branching evolution, where clones evolve simultaneously, resulting in multiple subclonal branches that demonstrate selection for clones with increased fitness over time within the tumor; neutral evolution, where expansion in the absence of stringent selection leads to passive accumulation of genomic alterations (e.g., the Big Bang model of colorectal tumor growth) (21); and punctuated evolution, depicted by mutation bursts or cataclysmic genomic rearrangements resulting in the sudden accrual of genomic changes (20). Irrespective of the mode of evolution in treatment-naive tumors, therapy of all types can either contribute to increased intratumoral heterogeneity or impose a selective pressure that results in the expansion of a resistant subclone (22).

While cancer has long been considered a genetic disease, where heritable DNA alterations serve as substrates for evolution (gene-centric model of evolution), it is now evident that non-genetic sources of phenotypic variation play a critical role in tumor development, progression and therapeutic resistance (23–25). These include epigenetic alterations (e.g., DNA methylation, histone modifications) as well as transcriptomic variation – both of which operate at much faster rates than does the acquisition of genetic mutations, thus serving as substrates for evolution even in the absence of any genetic events (26, 27). Variation at the level of the genome, transcriptome, or epigenome can also contribute to tumor plasticity, i.e., the degree to which a tumor population can flexibly and reversibly transition cell states to respond to stress (28, 29) ( Figure 1C ). The presence and integration of both heterogeneity and plasticity within a tumor results in many possible evolutionary avenues for tumor growth and survival. Indeed, there is growing evidence of ‘dynamic phenotypic heterogeneity’, where cancer cells can be phenotypically ‘re-trained’ by chemotherapy, resulting in the induction of drug-tolerant states (25, 30–34).

It is increasingly evident that reducing evolution to an allele-centric framework incompletely captures the context in which evolution occurs (35). From the early stages of transformation to metastasis, cancer cells are exposed to an array of niche microenvironmental and therapeutic pressures - which in total impart selective forces that shape their genetic and phenotypic profiles. This results in billions of cancer cells that are locally optimized to have their own distinct cell states conducive to their survival. Further, cells can fluctuate among different metastable cell states, broadening the population’s effective phenotypic landscape, thereby increasing adaptive capacity. With this breathtaking diversity comes clear implications for how to improve upon current therapies, which typically manage to target only a proportion of all cell states within a tumor, resulting in the outgrowth of clones that circumvent therapy by retaining or adopting non-targeted cell states.

Lack of tumor response or relapse has been noted in response to single agents (targeted therapy, immunotherapy) as well as combination chemotherapy, resulting in the ongoing search for second, third, and fourth-line agents in many cancers, despite their relative ineffectiveness (33–44). In contrast to the ‘scorched earth’ approach of chemotherapy, targeted therapies aim to spare normal cells by targeting specific cancer cell dependencies, driven by a single molecule or reliance on a certain cellular pathway (42). Similarly, ‘precision medicine’ seeks to rationally target individual branches, with their respective dependencies, within the evolutionary tree of a patient’s tumor. These approaches all rely on genetic characterization of tumors to identify therapeutic targets or biomarkers that predict tumor response to existing targeted therapy options (43, 44). Such agents include BRAF inhibitors, which selectively eliminate or inhibit the growth of cells that harbor BRAF mutations and imatinib, which specifically inhibits the aberrant tyrosine kinase produced by the BCR-ABL gene fusion in Philadelphia chromosome-positive chronic myelogenous leukemia (CML) (45).

Despite their promise, targeted approaches have their limitations. First, a single biopsy produces a restricted representation of the various niches occupied by a tumor and is unlikely to resolve the complete genomic landscape, with a recent study showing >100 million coding region mutations existing within a single tumor (33, 46–48). Given this, even if the predominant subclones harboring the detected molecular phenotype are targeted effectively, other subclones that are below the limit of detection and harbor different cellular dependencies may still survive and expand; indeed, this has been demonstrated by the suboptimal outcomes of patients treated with agents targeting sub-clonal driver mutations or copy number gains (18, 19). Second, sensitive subclones may also acquire therapeutic tolerance or resistance, subverting the effect of therapy and contributing to relapse and re-emergence of heterogeneity ( Figure 1D ). CML illustrates both of these points, as imatinib resistance has been shown to be due to pre-existing or acquired resistance (8, 9).

Further, focusing solely on genomic alterations neglects the contributions of non-genetic mechanisms of tumor resistance, which can be induced by therapy (29, 49). Recent studies demonstrate that in several cancer types, therapeutic intervention initiates cellular reprogramming that induces a drug-tolerant phenotype in the absence of a pre-existing resistant clone (13); in this case, continued targeted therapy may accelerate tumor progression. For instance, continued treatment with a BRAF inhibitor can cause metastasis of RAS/BRAF-mutant melanoma (48, 50). Further, treatment with EGFR and BRAF inhibitors in colorectal cancers can increase overall mutability and the likelihood of resistance, demonstrating that targeted therapy can transiently enhance evolutionary potential by accelerating genetic diversity (51).

The dominant framework for cancer resistance studies for the last 15 years has consisted of high-throughput sequencing analyses of pre- and/or post-treatment tumor biopsies. While highly informative, this approach is limited in its capacity to provide comprehensive understandings of the longitudinal evolutionary process in a tumor, specifically due to resolution limitations that preclude capture and tracking of rare clones over time. As a result, only sporadic snapshots of a cancer cell’s journey are captured following therapeutic exposure. Moreover, the design and development of targeted agents to date have largely been informed by genomic alteration measurements following therapy. However, the vast heterogeneity inherent to tumors and their unique evolutionary trajectories as they adapt to an assortment of microenvironments and respond to various stimuli are such that many evolutionary outcomes are possible for a given tumor. Indeed, recent single-cell studies have shown that multiple cell states are often present within a tumor and that different cell states can have different sensitivities to therapy (52). Therefore, therapeutic strategies to date may eliminate the majority of a tumor population, but certain subpopulations can survive and drive relapse ( Figure 1D ). Regardless of mechanism of action, it is unlikely that any single or combination of therapeutic agents can adequately address the large range of present and potential phenotypes (i.e., cellular states and dependencies) that can emerge across clones within a tumor. Thus, the rational next step will be to also reduce the total number of potential cell states and associated dependencies within a tumor.

It is currently unknown what agents, targetable states, and to what extent homogenization is feasible. Conceptually, the implementation of tumor homogenization could be systematically pursued through a three step process ( Figures 2B, C ). First is homogenization: as described by Tong et al., a selective pressure (i.e., therapeutic agent or combination of agents) can be introduced, coercing all tumor cells to exhibit a common phenotype that is vulnerable to a second agent which would eliminate the entire tumor population (also termed ‘collateral sensitivity’) (76, 86). Indeed, a recent study demonstrated that development of resistance to dasatinib treatment induces collateral sensitivity to non-classical BCR-ABL inhibitors, cabozantinib and vandetanib, in a murine model of acute lymphoid leukemia (87–89). Second is characterization of the homogenized state: extent of phenotypic homogenization within a tumor can be assessed through methods such as single-cell RNA-sequencing. Third is targeting of the homogenized state: therapeutic agents should be identified which can either eliminate the homogenized population through targeting of the shared phenotypic state, or funnel the homogenized cancer populations further into a defined targetable or sensitized state for elimination. Testing and selection of existing therapeutic agents at approved physiological doses would enable ease of clinical implementation.

The feasibility of homogenization strategies was first successfully tested in yeast (90), demonstrating that evolutionary dynamics can be manipulated for homogenization therapy. In support of the efficacy of homogenization therapy, a case study of a patient with ALK-rearranged non-small-cell lung cancer (NSCLC) has been described, where the authors postulated that cells exhibited unique, temporally restricted collateral sensitivities during adaptation to ALK inhibition (49, 91). Additionally, prior patient studies similarly have shown that convergent evolution in response to therapy is possible (16, 92, 93). The concept that cancer cell populations can be therapeutically modulated to transform cellular plasticity into therapeutic opportunities has been recently described in practice (49, 93, 94). For example, Frede et al. found that myeloma cells can modulate lineage restriction, adapt their enhancer usage, and employ cell-intrinsic diversity for survival and treatment escape, resulting in the co-existence of numerous distinct transcriptional states (95). Further, they demonstrated that standard therapy promotes transcriptional reprogramming while simultaneously reducing developmental potential, resulting in actionable immunotherapy targets (e.g., CXCR4) that could be exploited to overcome resistance (95). In another study, Lin et al. demonstrated that drug-induced antagonistic pleiotropy, the concept that genes can induce opposite effects on fitness in response to different drugs, can be leveraged to identify evolutionary traps which selectively target therapeutic resistance (96).

While cancer cells often rely on multiple stress response pathways to evade apoptosis and survive harsh tumor environments (e.g., the integrated stress response, cytosolic heat shock response, and unfolded protein response mediated by organelles such as the endoplasmic reticulum and mitochondria), induction of these cellular processes have also been noted to contribute to drug sensitivity of cancer cells. Activation of the integrated stress response in HER2+ breast cancer predicts a better response to trastuzumab therapy (97). In cancers with high protein turnover (e.g., multiple myeloma), agents that induce the unfolded protein response increase sensitivity to treatment with proteasome inhibitors through likely synergistic mechanisms (98). Further, cellular stress responses orchestrate common, potent responses across cells (i.e., via sweeping changes in cell state). For instance, ER stress induces the unfolded protein response, transducing multi-axis signaling and causing transcriptional reprogramming via IRE1α and ATF6, major translation modulation through phosphorylation of EIF2α, and pro-survival/-apoptotic signals dependent on resolution of ER stress (99, 100). In agreement, a recent lineage tracing study using TraCe-seq identified that efficacy of EGFR-inhibitor response is, in part, dependent on induction of ER stress (75). Similarly, replication stress has recently been described to activate immune-stimulating pathways, resulting in increased immune response to immunotherapies like PD-1/PDL-1 inhibitors across numerous cancer types, and serving as a reliable biomarker/predictor of clinical response to immune checkpoint blockade in patients (101, 102). For these reasons, induction of cellular stress responses may have great potential in actualizing phenotypic homogenization efforts ( Figure 2D ).

Homogenization strategies further require the ability to characterize cancer systems as they respond to sub-cytotoxic stress to uncover the nature of their responses, including the extent of their phenotypic uniformity and the duration of homogenization upon application and removal of stimulus. With the ongoing rapid development of multi-modal single-cell technologies, these characterizations will be greatly augmented. ClonMapper, Watermelon, and other dynamic expressed barcode systems enable one to distinguish how diverse clones differentially respond to therapy at single-cell resolution.