Precision Oncology: Between Vaguely Right and Precisely Wrong

Current Public Perception and Cracks in the Paradigm

Clinical studies have failed to demonstrate that systematic cancer genome sequencing results in significant benefit to patients. One study showed that in only 5% of cases does genome sequencing result in “actionable” information (7), detection of a mutated driver oncogene for which there is a target-selective drug available and accessible (8). Often, well-established activating mutations are detected that cannot yet be pharmacologically blocked. For instance, one third of all human tumors are driven by mutations in RAS genes (9), yet development of drugs to block KRAS gene function has been challenging (10). One may ask: Given that Ras is so widely mutated and its intracellular pathways fan out to regulate thousands of genes (11), would targeting Ras indeed represent a “precision” intervention? Is precision the “right” strategy at all?

Despite limited clinical impact, Precision Oncology, glorified by a handful of highly publicized anecdotal success stories continues its ascendance in public awareness. The U.S. governments' new Cancer Moonshot Program launched in 2016 with great fanfare has Precision Oncology as one of its central pillars (12). Questions of cost effectiveness are drowned out by ubiquitous direct-to-consumer advertisements for Precision Oncology by cancer care centers (13). And the skepticism of some investigators regarding the very concept of “driver mutations” (14, 15) is overpowered by the gratifying logics of targeting a molecular root cause.

Ironically, the same genome sequencing technologies that have brought us Precision Oncology have uncovered a series of inconvenient facts that explain why Precision Oncology cannot be the universal solution. Deep sequencing with next-generation sequencing and single-cell omics technologies have revealed inconsistencies in the paradigm of “driver mutations” (16) and the somatic mutation theory of cancer, warranting a more critical stance toward Precision Oncology (15, 17).

Genetic Heterogeneity in Tumor Specimens

The notion of massive cell heterogeneity in tumors is now broadly acknowledged. But the minds of investigators still operate largely in the category of genetic as opposed to nongenetic heterogeneity (18–20). Before clinical researchers recognized tumor cell heterogeneity as a challenge, cartoons of mutated oncogenic pathways (21) were used as a scheme of causality. These were mentally projected onto every cell of the tumor under the tacit assumption that the tumor is a homogenous mass of identical cells. This idea goes back to the textbook concept of clonal expansion of a dominant cell clone carrying the oncogenic mutation that “takes over” the entire tumor. Reality departs from such cartoonish pathway models in two ways: (i) Every individual cancer cell carries a distinct set of mutations, hardly surprising given the genomic instability of cancer cells; (ii) there is little evidence of a “purifying selective sweep” by the most malignant clone carrying “driver mutations” that would “homogenize” the tumor.

Although one finds signatures of clonal selection in some tumors (22–24), clonal diversity dominates. Genetic heterogeneity is now readily recognized by deep sequencing and genomic analysis of multiple biopsies from different regions of the same tumor (25). Analysis of regional diversity of genomes in colorectal and liver tumors (26, 27), with 100 million distinct coding region mutations found in a single tumor, illustrates this genetic diversity (26). A systematic analysis of the frequency distributions of mutations in bulk tumor genomes suggests that neutral evolution is the rule rather than exception (28). Single-cell studies confirm multiple, branching evolutionary trajectories proceeding concomitantly in leukemia populations (29).

With such genetic diversity within a single tumor, we must reexamine what exactly needs to be measured to identify “oncogenic mutations” in the diagnostic step of Precision Oncology. There is no single tumor genome to sequence. Genomic DNA extracted from a bulk tumor sample may capture the most dominant clone. The “other” tumor cells that do not carry the targeted “driver mutation” may yet survive and be the source of tumor recurrence.

In contrary to the regional genetic diversity uncovered by sequencing a tumor following surgery, temporal changes in genetic diversity have been less studied. Analyses of individual cell genomes in serial samples extracted at multiple time points indicate complex subclone dynamics that reflect independent evolutionary paths and copresence of distinct stages, far from the standard model of multistep cumulative acquisition of mutations (22, 29). Circulating tumor cells (30) or even tumor (ct) DNA (31) now afford a noninvasive window to the tumor, enabling longitudinal monitoring without repeated biopsies. But given the cellular heterogeneity of tumors, do these “liquid biopsies” offer representative samples of entire tumor cell populations? For example, ctDNA analysis of patients with colorectal cancer who had developed resistance to anti-EGFR therapy revealed up to 12 distinct subclones in one patient, each harboring a different mutation in RAS or BRAF genes (32).

Neither Mutations nor Epimutations Are Strictly the “Cause” of Cancer

Nowhere is the disconnect between nominal “oncogenic mutation” and cancer phenotype as lucidly manifested as in precancerous lesions that only rarely become cancer. With next-generation sequencing of somatic tissues, a large number of mutations that affect oncogenes have been detected in healthy skin cells. Potentially oncogenic mutations (in BRAF, TP56, PIKC3, etc.) are found in normal skin and at high frequency in nevi, the vast majority of which never develop into malignant melanoma (33–36). As in tumors, these mutations may even display signs of clonal selection and expansion. Similarly, a large fraction of epithelial cells in Barrett's esophagus also carry oncogenic mutations. But again, only <1% of Barrett's esophagus progresses to esophageal adenocarcinoma (37, 38). For a systematic review of driver mutations found in noncancer tissues, see Kato and colleagues' work (39). A logical interpretation is that so-called “driver mutations” are not actual omnipotent “drivers” but represent, in the simplest scheme, molecular lesions that are causatively necessary but not sufficient for pathogenesis. But the common notion that “additional hits” are needed may also be too simplistic.

If genomic mutations are not sufficient, one should also consider the logical complement that accumulation of specific mutations may not be necessary in the first place. Although absence of evidence is never evidence of absence, whole-genome sequencing of ependymoma has failed to identify recurrent somatic mutations (40), and several childhood cancers, notably leukemia and rhabdomyosarcoma, may be caused by just a single mutation (41). This is not as outlandish as one may think if one overcomes the tacit habit of reducing complex phenotypes to genotypes, a common epistemic practice in which each new “hallmark of cancer” (42), such as invasion or angiogenesis, is mapped to a plausible mutation. In reality, a single genetic alteration in a master regulatory gene may profoundly alter the developmental trajectory of cells on Waddington's canonical “epigenetic landscape” of cell development. In this (mathematically permissive) abstraction, every point on the landscape is a cell phenotype, and the major valleys into which cell naturally roll represent stable normal differentiated cell types (43). Mutations may push cells into “side-valleys” where further descent (=differentiation) is blocked, causing a maturation arrest that is a fundamental feature of all neoplasia (19). Once in the wrong developmental path, the derailing mutation may not be necessary anymore to maintain the malignant state. Such hysteresis in which activity of an oncogenic protein is required only in a transient developmental window (44) obviously defies the therapeutic targeting of “causal” mutations because dependency on the (tumor initiating) oncoprotein is lost.

A prosaic manifestation of “phenotype innovation without mutation” is the diversity of cell types in the metazoan body, each of which constitutes a distinct, stable, and inheritable cell phenotype that is not produced by genomic alterations. It is then only a small step to explain minimally mutated or even mutation-less cancer if one considers the old idea that cancer cells are pathologic (immature) cell types not meant to be implemented but once realized, are trapped in developmental dead ends (45).

When no plausible mutation is found, it is common practice to implicate “epimutations,” that is, changes in DNA methylation or covalent histone modifications (ref. 46; such epigenetic alterations have little to do with Waddington's original “epigenetic landscape” despite the use of the same term; refs. 47, 48). These reversible changes of the “epigenetic state” of gene loci must be programmed by specific transcriptional regulation via recruitment of chromatin-modifying enzymes by site-specific transcription factors (49). Thus, chromatin changes reflect the altered gene expression patterns that implement the pathologic cell phenotype and are the result rather than cause of the coordinated regulation of genes by the transcriptional network that integrates inputs from a variety of signaling pathways (48, 50). Therefore, when invoking “epigenetic” causation, one must consider that epigenetic modifications are fundamentally distinct from genetic alterations, belonging to a separate explanatory category: A chromatin modification is not a random event (the defining property of mutations) that occurs in a single cell and must be selected for to spread through the cell population. Instead, epigenetic modifications merely reflect the coordinated gene expression patterns, which in turn are controlled by regulatory signals in the tissue that act in parallel on large numbers of cells to modulate gene expression. These tissue-level signals have much more leverage than genetic mutations because they affect a large proportion of the cells in parallel. Perhaps, such events of abnormal signaling are the rate-limiting factors in progression of precancerous lesions, which may already carry oncogenic mutations. Then mutations would not be the “cause” of macroscopic (clinical) cancer.

This view applies to epigenetics in both senses and has fundamental epistemic and practical consequences: The development of clinical cancer in premalignant or dormant lesions may not be driven solely by “additional mutational hits.” Nongenotoxic tissue-level stress, for example, chronic inflammation and toxic stress, alters the behavior of entire cell populations through signaling. Such nongenetic perturbations may be a much more potent trigger of growth in a dormant tumor and thus may be considered the “effective cause” of a clinical tumor. Such directed and concerted cell population level dynamics would explain rapid recurrence but defy the picture of clonal evolution that underlies Precision Oncology.

Paradoxical Responses to Targeted Therapy

A glaring illustration that cancer treatment is far more complex than can be captured by the Precision Oncology scheme is provided by a series of paradoxical effects: Attempts to block oncogenic pathways with precision drugs can instead stimulate progression (51, 52). Clinically, this is well documented in the case of melanoma treatment with BRAF inhibitors, a targeted therapy that has seen spectacular successes in the treatment of melanoma for a subset of patients (53, 54). Such paradoxical effects have also been reported in treatments with checkpoint inhibitors in immuno-oncology, the youngest child of Precision Ontology (55, 56).

Following the logics of target-selective therapy, benign nevi, which harbor BRAF mutations, should shrink when a melanoma patient is treated with a BRAF inhibitor. Some do but others actually grow under BRAF therapy (57, 58). BRAF inhibitors can trigger other (nonmelanoma) skin cancers, perhaps due to the compensatory activation of mitogenic MAPK when BRAF is inhibited (54, 59).

The rationale of Precision Oncology has therefore been expanded to targeting multiple pathways by combining treatment with BRAF and MEK inhibitors. Although combination treatment can substantially extend survival in melanoma patients, resistance and recurrence are observed in less than a year (60). The question is: how many backdoor pathways do tumors have available for escaping single-point targeted therapy, which becomes a whack-a-mole game due to the unforeseeably immense repertoire of evasion pathways. Where is then precision? Are we not better off acknowledging an inherent vagueness in intervening with complex processes and using broadly acting drugs, such as the class of chromatin modulators that target epigenetic enzymes? (61).

Expanding Precision Oncology to Embrace Complexity

The most salient manifestation of the limits of a “precision intervention” is the rapid and near inevitable development of therapy resistance and recurrence. Logically, tumor recurrence is at odds with Precision Oncology's targeting of the molecular root cause that drives the tumor, unless, as is done by default, recurrence is explained by a somatic evolution in the tumor: genetic mutations in the cancer cells' genome that confer resistance are selected for (62). Indeed, genetic mutations in the targeted pocket of kinases or those that activate alternative pathways can be routinely observed (63). However, in this gene-centric perspective, the convoluted and inevitable path to the phenotype of resistance is again reduced to a simple genetic cause, which takes effect thanks to a somatic equivalent of Darwinian evolution. But this rationale misses the complexity of the tumor, which provides many more ways for evading treatment. We summarize two central but forgotten principles (64):

Outlook: “Precision” Must Not Mean “Narrow”

What can we do to elevate the bar for “precision” and embrace all the aforementioned complexities? First and foremost, we need to learn more about the interplay between genetic and nongenetic heterogeneity of tumor cells and their dynamics, the plasticity of individual cells and of cell populations. The very technology that gave rise to Precision Oncology now affords high-throughput single-cell analysis to study these features. Advanced methods for single-cell resolution molecular profiling, such as RNA-seq, CyTOF, and MIBI (72), open a new window to the molecular signatures of individual tumor cells and to single-cell resolution of cell population structures (73, 74).

But only if interpreted within the appropriate theoretical framework of tumors as highly evolved complex adaptive systems (64) that are stochastic and nonlinear, and by considering that any intervention in such self-sustaining reactive systems can backfire (75), can these new analytic tools guide us beyond the confines of linear–mechanistic schemes of current Precision Oncology.

The sentiment of “precision” in current Precision Oncology depended on a narrow view that ignores the grander scheme of complex dynamics. The associated naïve mental picture of the molecular network of tumor cells working like a clockwork makes us, to recall Carveth Read, “precisely wrong,” when we need to first be “vaguely right,” to see beyond the trees and acknowledge the forest, and understand how the latter as a whole relates to the finer details of its parts. With new technologies for studying genetic and nongenetic cell population heterogeneity, tumor cell plasticity, and intercell communication, we are only at the beginning of achieving a new “precision” at the more encompassing level of systems behaviors to better understand the inherent resilience of tumors and to learn how to conquer it.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Grant Support

The authors are grateful for support from the American Association for Cancer Research and Breast Cancer Research Foundation through the BCRF-AACR grants for Translational Breast Cancer Research (#14-60-26-BROC to A. Brock) and the National Institute of General Medical Sciences (NIGMS) Grant R01GM109964 and NIGMS National Centers for Systems Biology Grant 2P50GM076547-06A1 (to S. Huang).