The challenges of tumor genetic diversity

Abstract

We review and discuss the implications of genomic analyses documenting the diversity of tumors, both among patients and within individual tumors. Genetic diversity among solid tumors limits targeted therapies, as few mutations that drive tumors are both targetable and at high prevalence. Many more driver mutations and how they affect cellular signaling pathways must be identified if targeted therapy is to become widely useful. Genetic diversity within a tumor, intra-tumor genetic heterogeneity, makes the tumor a collection of subclones, related yet distinct cancers. Selection for pre-existing resistant subclones by conventional or targeted therapies may explain many treatment failures. Immune therapy faces the same fundamental challenges. Nevertheless, the processes that generate and maintain heterogeneity might provide novel therapeutic targets. Addressing both types of diversity requires genomic tumor analyses linked to detailed clinical data. The trend toward sequencing restricted cancer gene panels, however, limits the ability to discover new driver mutations and assess intra-tumor heterogeneity. Clinical data presently collected with genomic analyses often lack critical information, substantially limiting their use in understanding tumor diversity. Now that diversity among and within tumors can no longer be ignored, research and clinical practice must adapt to take diversity into account.

Introduction

Genomic tumor analyses over the past decade, exemplified by The Cancer Genome Atlas (TCGA),¹ have greatly enhanced our understanding of the diversity among different types of cancer and among different patients' tumors of the same type. This work has also made it increasingly apparent that cancers are diseases of genetic diversity both among patients and among cancer cells within a single patient's tumor.² If a patient's tumor consists of multiple related yet distinct cancers, then we need to reconsider many present approaches in cancer research and therapy.

Here we review the evidence for clinically important diversity both among patients' tumors and within individual tumors, discussing their implications for cancer research and clinical practice.

Diversity among tumors

Personalized therapy and the search for common, targetable driver mutations

Much motivation for large-scale genomic tumor analysis was to identify “driver” mutations that might provide therapeutic targets. Driver mutations activate oncogenic signaling or inactivate tumor suppressor pathways, with few drivers expected in any single tumor.³ The remarkable success of imatinib in treating CML and promising early results with other targeted therapies^{4, 5} suggested that identifying new driver genes could allow development of novel therapies targeting drivers or the pathways they alter, potentially allowing the same agent to be used in different tumor types.^{6, 7}

In early examples, the driver mutation itself (ABL kinase activated by genomic translocation in CML, mutated BRAF in melanoma, EGFR mutation or amplification in lung cancer) was the target of therapy.⁴ A driver mutation may instead make a tumor highly sensitive to inhibition of a particular cellular function, providing a therapeutic window for tumor-specific attack. For example, DNA repair deficiencies of BRCA-mutant tumors can make them highly sensitive to PARP inhibitors.^{8, 9} Combining genomic analyses of thousands of tumors and cell lines with exhaustive cell-line testing can provide a comprehensive catalog matching known mutations with drug sensitivity.¹⁰ In principle, if the mutations driving a tumor and how they affect cellular pathways were known then a rational choice of patient-specific therapy or combinations of therapies might be possible.^{7, 11, 12} Identifying most clinically relevant driver mutations is thus critical for such personalized cancer therapy.

Evaluating the “landscape” of mutations among tumors distinguishes driver mutations from “passenger” mutations passively accumulated during tumorigenesis.¹³ Genes mutated more frequently than by chance among tumors of a particular type, forming peaks in the landscape, are driver-mutation candidates.^{3, 14}

Challenges to personalized therapy

Although many driver mutations have been identified, genomic analysis of solid tumors has mostly found foothills in the landscapes of mutations, with few large, high-prevalence, targetable peaks. Few genes are altered in more than 10% of cases for any type of cancer, and fewer pass that threshold in multiple types of cancer (Table 1). Some of the most frequent alterations (TP53, PTEN, CDKN2A) are losses of tumor suppressor genes due to mutation or genomic deletion, whose effects cannot yet be readily targeted with drugs.

Table 1. Mutations at more than 10% prevalence.

	Tumor type
Gene	Head and neck squamous	Uterine endometrioid	Lung adeno	Lung squamous*	Stomach adeno	Colorectal (non-hypermutated)*	Renal clear cell	Breast	Prostate adeno	Urothelial	Renal chromophobe	Ovarian serous*	Thyroid papillary
TP53	73%	28%	46%	81%	48%	60%		38%		53%	31%	96%
PIK3CA	37%	57%		16%	24%	18%		39%
KRAS		22%	36%		16%	43%
PTEN		67%			11%				17%		11%
FBXW7		16%				11%				12%
ARID1A		34%			33%
BRAF			11%										62%
CDKN2A	49%			15%
KEAP1			19%	12%
NFE2L2	12%			15%
AGTR1	16%
APC						81%
ARID5B		12%
BAP1							13%
CASP8	11%
CDH10	11%
CHD4		17%
CTCF		19%
CTNNB1		30%
EGFR			17%
FAT1	29%
FGFR2		14%
GATA3								13%
HRNR			27%
KIF26B		14%
KMT2D	18%
MLL2				20%
NKX3-1									17%
NOTCH1	20%
NSD1	12%
PBRM1							39%
PIK3R1		33%
PLSCR1	14%
PPP2R1A		11%
SEMA5A	14%
SETD2							16%
SPOP									12%
STK11			19%
VHL							54%
ZNF804B	11%

For each tumor type, the percentage of tumors having a sequence mutation or copy-number change in a gene is specified for each gene having more than 10% mutation prevalence in that tumor type. Based on published TCGA data obtained via cBioPortal.⁷⁵

^*Not including tumors with only copy-number changes.

Diversity of driver mutations among tumors and low prevalence of targetable drivers in solid tumors pose major obstacles to personalized therapy. For example, the NCI-MATCH trial uses mutations in solid tumors or lymphomas to assigns patients to particular targeted therapies. Of the first 500 patients enrolled, however, only 9% were found to have targetable mutations, and only 33 patients were successfully assigned to a treatment arm; information on responses of these patients is not yet available.¹⁵ Many more driver mutations and the ways that they affect cellular pathways will need to be identified before personalized cancer therapy is widely available.

Immunotherapy in principle provides a way to overcome the challenges from genetic diversity among patients' tumors. Unleashing a patient's immune system to destroy a tumor via patient-specific neoantigens exploits that diversity among tumors: differences from other patients' tumors should not matter. Even immunotherapy, however, can be limited by the second type of tumor genetic diversity, the diversity that occurs within individual patients' tumors.

Diversity within tumors

The challenges posed by diversity within an individual patient's tumor may dwarf those posed by diversity among tumors. Although geographic, phenotypic and epigenetic diversity also occur,^{16, 17} we focus on heritable differences in DNA, intra-tumor genetic heterogeneity, as the quantifiable type of within-tumor diversity that poses the greatest challenge to therapy.

Subclonal evolution

Forty years ago, Nowell noted that intra-tumor evolution via mutation and selection continues after tumor initiation.¹⁸ Cancer cells were already known to have deficiencies in DNA repair, providing higher mutation rates than in normal cells.¹⁹ By the time a tumor is clinically detectable it can be a genetically diverse collection of subclones.

Figure 1 illustrates this process. Mutations present in the cell that initiates the tumor clone are “truncal”²⁰ and shared by all progeny unless some cells lose that genomic locus. Truncal mutations include drivers initiating the tumor and passengers mutated in the initiating cell but not responsible for tumor initiation.³ Mutations in later generations of cells that provide selective survival advantage to a subclone are also driver mutations, while other later mutations are passengers. The tumor at presentation thus may contain multiple subclones differing in their complements of the later mutations.

Figure 1.

Subclonal evolution of a tumor. Left, colored regions represent cancer cells; white background, cells with normal DNA. Starting from the tumor-initiating clone (clone 0), subclones are born and expand (numbered triangles) or die (diamonds) over time. Each subclone contains all mutations present in its progenitors, plus its subclone-specific mutations. Right, mutant-allele fractions (MAF), the fraction of DNA in a sample that shows a tumor-specific mutation, at time of tumor sampling. Examples shown for (sub)clone-specific mutations in 2 tumor samples and for the whole tumor, based on heterozygous mutations at normal copy number in cancer cells and 20% of all cells in the sample having normal DNA (80% tumor “purity”). For a heterozygous mutation and 80% purity, the fraction of cells with the mutation, its cancer cell fraction (CCF), is 2.5 times its MAF. Mutations in the originating clone have higher MAF/CCF values than those in subclones. Mutations in subclones can be missed either due to sampling (e.g., subclones 3 and 6 in Sample 1; subclones 4 and 5 in Sample 2) or due to having MAF values too low to be detected (e.g., if the detection limit is 0.05: subclone 2 in Sample 2; subclones 5 and 6 in the whole tumor). Adapted from Bozic et al, PLOS Comp Biol 12: e1004731. doi:10.1371/journal.pcbi.1004731. Used under the Creative Commons Attribution License 4.0, https://creativecommons.org/licenses/by/4.0/.

With subclonal evolution, context can also turn a former passenger mutation into a driver. Some early studies of mouse cancer models were best explained by differences among cancer cells in a tumor.^21-24 Under Nowell's model, a subclonal mutation might provide resistance to a drug, loss of dependence on a truncal driver mutation, or a tendency to metastasize, favoring some subclones. A “passenger” mutation thus can become a “driver” in the context of tumor progression or therapy.

Measuring intra-tumor genetic heterogeneity

Next-generation sequencing (NGS)²⁵ of genomic DNA allows detailed investigation of intra-tumor genetic heterogeneity. NGS examines millions of individual nucleic acid pieces in parallel, with each piece assessed for mutations at the genomic loci it covers.²⁶ NGS provides counts of the pieces of DNA covering each genomic locus (sequencing depth) and how many of those pieces show the mutant sequence. The fraction of pieces showing the mutation at a locus is the mutant allele fraction (MAF). Truncal mutations will typically have higher MAF values in a tumor sample than mutations restricted to later-developing subclones (Fig. 1).

NGS of multiple regions of a tumor at adequate sequencing depth to detect subclonal mutations is one straightforward approach. Gerlinger et al²⁷ combined this approach with other analytical methods to establish dramatic differences among geographically distinct portions of renal cell carcinomas. About two-thirds of mutations were not detected across all regions of individual tumors, including some mutations that were potential candidates for targeted therapy.

A single tumor sample can provide information on heterogeneity if the different contributions to MAF values in the sample are taken into account. Besides different MAF values for truncal versus subclonal mutations (Figure 1), changes of ploidy from whole-genome duplications and local genomic copy-number aberrations (CNA) alter MAF values. Furthermore, normal DNA from stromal or immune cells provides an “impurity” that lowers all MAF below their values in cancer cells per se (Figure 1).²⁸ Several methods to correct for purity, ploidy, and CNA have been proposed to unravel the subclonal composition of tumors,²⁹ providing estimates of the fraction of cancer cells that contain each mutation, the cancer-cell fraction (CCF), of major interest in terms of subclonal tumor evolution and potential response to therapy.

To estimate genetic heterogeneity from tumor samples simply without determining subclonal relationships, we developed a measure called mutant allele tumor heterogeneity (MATH).³⁰ (Figure 2) MATH incorporates subclonality of point mutations and CNA at mutated loci, while including a first-order correction for tumor “impurity.” Heterogeneity measured by MATH may be characteristic of a tumor as a whole, as multiple samples of the same tumor can have similar MATH values³¹ and MATH has been used to demonstrate maintenance of intra-tumor heterogeneity during passage of patient-derived xenografts in mice.³²

Figure 2.

Using mutant-alllele fraction (MAF) values to assess intra-tumor heterogeneity. Left, a low-heterogeneity tumor; right, a high-heterogeneity tumor. Circles represent MAF values of individual tumor-specific mutated loci; the curves are density plots (smoothed histograms) of the distributions. Mutant-allele tumor heterogeneity (MATH) for each tumor is the percentage ratio of the width to the center of its MAF-value distribution, with the median taken as the center and the median absolute deviation (MAD) taken as the width. From Mroz et al, PLOS Medicine 12: e1001786. doi:10.1371/journal.pmed.1001786, 2015, under the Creative Commons Attribution License, https://creativecommons.org/licenses/by/4.0/.

Critically, intra-tumor genetic heterogeneity is not the same as mutation frequency assessed from the number of mutated genomic loci. A tumor with many truncal mutations might not yet have developed subclones. Alternatively, a tumor-initiating clone with few mutations could have evolved into a genetically diverse tumor even with few late-developing mutations. Indeed, heterogeneity measured by MATH bears little relation to the number of mutated loci in a tumor sample (Figure 3).^{30, 33} This easily overlooked distinction between mutation numbers and genetic heterogeneity may be particularly important for immunotherapy.

Figure 3.

Intra-tumor genetic heterogeneity is not the same as mutation frequency. Scatter plot of MATH values versus number of tumor-specific mutations for 74 head and neck squamous cell carcinomas (p = 0.21, Kendall rank correlation). From Mroz and Rocco, Oral Oncology 49:211-215, 2013, with permission.

Challenges to conventional and targeted therapy

The clinical risks of high intra-tumor genetic heterogeneity are no longer just theoretical or restricted to case studies. In head and neck squamous cell carcinoma (HNSCC), high MATH was related to higher mortality in two independent data sets totaling nearly 400 patients.^{33, 34} (Figure 4) Among 2433 breast cancer patients, those whose tumors were in the top quartile of MATH values had shorter breast-cancer-specific survival than those with tumors in the bottom quartile.³⁵ For simplicity we discuss below the ways that intra-tumor heterogeneity can affect targeted therapy, based on the assumption that a suitable target has been found. The same principles, however, apply to conventional therapies and should help explain the documented relation of high intra-tumor heterogeneity to worse outcome in conventional therapy.^33-35

Figure 4.

High intra-tumor heterogeneity is related to higher mortality in HNSCC. Kaplan-Meier curves for 305 HNSCC patients in TCGA grouped by high (>32) versus low MATH values. (Hazard ratio, 2.18; 95% CI, 1.44 to 3.30; p < 0.001). A significant relation of high MATH to high mortality was verified in a Cox multiple regression accounting for 9 other covariates (including smoking history, age, and HPV status). From Mroz et al, PLOS Medicine 12: e1001786. doi:10.1371/journal.pmed.1001786, 2015, under the Creative Commons Attribution License, https://creativecommons.org/licenses/by/4.0/.

For targeted therapy to cure a cancer in the context of subclonal evolution (Figure 1), all subclones need to contain the target (a truncal driver mutation with CCF=1) and no subclone can contain additional mutations or mechanisms that negate the therapy. Even if a target is found for a patient, either a low CCF for the target or a resistance mutation, both likely consequences of intra-tumor heterogeneity, would pose additional problems. Intra-tumor heterogeneity will almost always limit the effectiveness of targeted therapy.

If heterogeneity means that the target is not a truncal mutation, then targeted therapy will have no direct effect on cancer cells that lack it. A low MAF or CCF for a driver mutation can identify targets that are likely to fail, yet it is almost impossible, from a single pre-treatment tumor sample, to assure a target is truncal with CCF=1.²⁷

Having a target with CCF=1 is insufficient for targeted therapy success. By the time that a tumor has reached the 10⁹ cells to be clinically detectable, some subclone will almost certainly contain a resistance mutation. Bozic and Nowak³⁶ estimate that a radiographically detectable tumor may have at least 10 resistant subclones, each providing a different resistance mechanism. So even if targeted therapy initially leads to clinical remission, resistant subclones will likely result in recurrence. This initial response followed by regrowth is common with targeted therapy.^{6, 37, 38} Supporting this failure mechanism, earlier melanoma progression times following targeted therapy tended to occur in patients whose resistance mutations had sufficiently high MAF to be detected in pretreatment tumors.³⁹

Although combination therapies provide hope that a subclone evading one therapy is unlikely to evade another, tumor genetic diversity also limits their effectiveness, as combinations face the same challenges as individual therapies. Diversity among tumors in terms of driver mutations is a major obstacle. If, as noted above, not even 10% of patients can be matched with a single targeted therapy based on the present catalog of driver mutations, then only a few percent are likely to have more than 1 identified target. Diversity within a tumor will limit combination therapy if some subclones lack the targets or any subclone has mutations providing resistance to the combination. Combined resistance might be a significant problem with a large tumor burden in metastatic disease.⁴⁰ Combination therapies can also fail due to improper timing of administration. For example, if a single agent is used until recurrence is observed and then replaced with a second agent, progeny of a subclone with resistance to the first agent might have time to develop resistance to the second agent before it is used.⁴⁰

Challenges to immunotherapy

Large numbers of neoantigens might be thought to help overcome intra-tumor heterogeneity. Indeed, the best therapeutic responses to inhibition of immune checkpoints have been in tumors with high mutation frequencies and high neoantigen burdens.^41-43

Mutation frequencies and intra-tumor heterogeneity are not the same,^{30, 33} however. A recent report distinguished neoantigen numbers from heterogeneity in responses to immune checkpoint blockade.⁴³ Neoantigens shared among all cancer cells, truncal mutations, provided the best targets. The authors could not detect immune responses to subclonal antigens, noting that immune responses to subclonal antigens also would not target all tumor cells. Furthermore, subclonal mutations have now been shown to alter antigen presentation, making a subclone resistant to immune attack.⁴⁴ Thus intra-tumor heterogeneity likely poses challenges for immunotherapy similar to those seen for targeted therapy

Meeting the challenges of tumor genetic diversity

Now that the challenges posed by genetic diversity among and within tumors are clear, both types of diversity must thoroughly inform oncologic research and practice. To account for and understand diversity, clinical trials, genomic studies, and projects that combine clinical and genomic data need to examine diversity directly. Fundamental research on the mechanisms that generate and maintain intra-tumor diversity may point to novel therapeutic approaches that exploit diversity rather than are limited by it.

Clinical trials

The NCI-MATCH trial, other basket trials, and trials of new agents do take diversity among tumors into account, as they use patient-specific driver mutations to match patients with agents. Clinical trials have not yet, however, incorporated diversity within tumors into trial design and interpretation.

Intra-tumor heterogeneity complicates the interpretation of responses to therapy. The initial response of a tumor to an agent may roughly represent the fraction of cancer cells in the tumor that both contain the intended target and are not resistant. The 30% decrease in lesion diameters for a partial response (PR) under RECIST criteria⁴⁵ is approximately a 66% decrease in tumor volume. Thus a tumor with CCF=0.5 for the target might not show PR even if the drug were completely active on the cells that did contain the target, or if there was a high CCF for resistance mutations. On the other hand, the dramatic and maintained effects of drugs in “exceptional responders”⁴⁶ may represent not only the unique oncogene addictions of their tumors but also a lack of subclones with resistance. Similarly, responders to induction chemotherapy, who can have better survival than non-responders,^{47, 48} may be patients whose tumors have lower intra-tumor heterogeneity. Time to progression may be a measure of the size of resistant subclones rather than any inherent property of the drug-target combination, as suggested both by theoretical⁴⁰ and observational³⁹ studies. Without knowing the underlying intra-tumor genetic heterogeneity, it will be difficult to assign the mechanism of failure.

As both initial responses to therapy and times to progression may represent heterogeneity rather than the intrinsic efficacy of therapy, it makes little sense to ignore heterogeneity in clinical trials. Studies of new agents might best be performed on patients having homogeneous tumors with high target CCFs, based on pre-treatment genomic analysis; alternatively, patients might be stratified by pre-treatment tumor heterogeneity or target CCF value. Interpretation of clinical responses should incorporate information about heterogeneity, to tease out responses related to the therapeutic target from those representing a consequence of heterogeneity.

Genomic studies

Large-scale genomic studies to discover more drivers, combined with studies on the cellular processes those drivers affect, would help meet the challenge of genetic diversity among tumors. This means distinguishing driver mutations at very low prevalence, down to just a few percent among patients' tumors, from passenger mutations. Lawrence et al⁴⁹ estimate that up to 5000 tumors of each type, a scale 10 times that attempted by TCGA,¹ are required before most driver mutations present down to 2% of tumors can be identified. Further basic-science and translational work will then be needed to match these low-prevalence drivers to targetable effects on cellular pathways. Genomic analysis of recurrent and metastatic tumors^{44, 50} or of circulating cell-free tumor DNA^{51, 52} provides a similar opportunity to identify resistance mutations. Even if resistance mutations are undetectable in a primary tumor due to low CCFs, subclone selection through therapy will likely lead to detectable, high CCFs in recurrences. Genomic analysis also can assess intra-tumor diversity, providing critical information generally unavailable thus far for translational or clinical work.

Unfortunately, some recent genomic analyses use methods that hinder all these approaches to addressing tumor diversity, pre-selecting extremely restricted portions of the genome through hybridization prior to NGS. Judicious pre-selection allows many more samples to be sequenced, or at higher sequencing depth, at the same cost as whole-genome sequencing.⁵³ For example, whole-exome sequencing (WES), pre-selecting about 30 million bases (MB) covering nearly 20,000 protein-coding genes, has allowed remarkable advances in genomics at reasonable cost.⁵⁴ Many recent studies, however, restricted genomic coverage much further, to just a few hundred known cancer genes.^{35, 50, 55-58}

Whatever the economic advantages of sequencing a small panel of cancer genes, it does not foster discovery of unknown driver mutations or resistance mechanisms. It also makes assessment of intra-tumor heterogeneity less reliable. A tumor will only have mutations in a handful of known cancer genes, while assessing its heterogeneity requires MAF values for many mutated loci. Illustrating this difficulty, when Pereira et al calculated tumor MATH values from sequencing 173 genes, about one-third of tumors did not even meet their low threshold of having 5 mutated loci to be eligible for MATH analysis.³⁵

Nevertheless, many data in the new NCI Genomic Data Commons (GDC)⁵⁹ are now from a highly restricted gene panel, via Foundation Medicine.⁶⁰ While having such genomic data readily available is an advance in many respects, it does little to address issues central to tumor genetic diversity. Furthermore, current GDC specifications do not require reporting the MAF values needed to assess heterogeneity.⁶¹ Genomic data need both wide genome coverage and MAF values to support progress against the challenges of tumor genetic diversity.

Repositories combining clinical and genomic data

Repositories that link patients' clinical and genomic data, like that begun by TCGA and continued in the NCI GDC, could be extremely valuable for understanding tumor genetic diversity as well as other issues in oncology. Such repositories will be even more useful if the clinical and genomic data are further linked to biobanks of associated tissue specimens, as in the multi-institutional ORIEN network.⁶² Genomic and clinical data from controlled trials provide advantages of uniform treatment and patient populations, clinical record quality, and follow up. Comprehensive repositories of representative samples of patients not in trials could provide a solid basis for developing and testing new hypotheses relating to the standard-of-care cancer therapy that most patients continue to receive as research on new therapies continues. Design and oversight of these repositories, however, must take tumor genetic diversity into account, as both clinical characteristics of patients and the choice of therapy would be expected to influence the relations of particular driver mutations⁷ and of intra-tumor heterogeneity^{33, 34} to outcome.

Much about repository planning can be learned from experience with TCGA. Its emphasis on acquiring many tumor samples⁶³ evidently led to a tradeoff so that only minimal accompanying clinical data were required. Missing clinical data from TCGA HNSCC cases made us unable to address critical issues like the relation of heterogeneity to outcome in patients whose tumors were excised surgically without adjuvant therapy.³³ Also, the multiple types of analyses envisioned by TCGA placed a premium on having large frozen tumor samples. TCGA cases thus under-represented lower stages of HNSCC, so we were unable to resolve whether heterogeneity increases with T classification.³³

As tumors collected more recently are presumably from well-designed genomic studies having extensive clinical data, there is no reason for continued incomplete clinical annotation. Furthermore, with many genomic analyses now possible on standard formalin-fixed, paraffin-embedded pathology specimens,⁶⁴ under-representation of lower-stage disease can likely be rectified. Repositories might also greatly accelerate progress if their clinical data included information difficult or impossible to obtain from present databases,⁶⁵ on outpatient therapy, recurrences, local versus distant failures, and disease-specific survival.

Unfortunately, the structure of the new GDC has not yet moved far enough toward comprehensiveness in either case representation or clinical data. The open invitation to researchers to submit data to the GDC⁶⁶ may result in samples of convenience, with their inherent difficulties in interpretation, rather than representative sets of linked genomic and clinical data. More distressing, the current version of the GDC clinical data dictionary does not require information on therapy,⁶⁷ without which few clinically reliable conclusions can be drawn regardless of how much genomic data are provided.

The centralized oversight of the new US Moonshot could make a dramatic difference in clinical-genomic repositories. It could seek out representative samples of tumors, providing support for genomic analysis and clinical data acquisition, and extra support for efforts directly related to assessing heterogeneity, such as collection and analysis of multiple portions of individual tumors. It could condition funding for genomic analyses on sharing extensive, regularly updated and audited clinical data. Similar to how the early TCGA commitment to genomic analysis⁶⁸ fostered genomic technology, a large-scale commitment by the Moonshot to automated parsing of clinical records could foster development of improved technologies to turn clinical electronic records into useful research data, amplifying the value of associated genomic data.

Targeting mechanisms underlying intra-tumor heterogeneity

Even though intra-tumor genetic heterogeneity poses challenges to current therapies, it might provide novel therapeutic targets. There are already intriguing hints about how this heterogeneity could be exploited once further research documents how it is established and maintained.

First, a tumor's heterogeneity represents diversity produced by mutations followed by the selection of clones with high fitness. This requires just enough, but not too much, generation of diversity.⁶⁹ It might be possible to tip that balance in a heterogeneous tumor toward genetic instability that diminishes cancer-cell survival.⁷⁰

Second, selection could be exploited with a view toward making a tumor more homogeneous and treatable. If therapy selects a particular subclone, the post-treatment tumor should initially become more homogeneous than the pre-treatment primary.⁷¹ The data of Gerlinger et al^{27, 72} suggest that tumors from patients treated with everolimus were more homogeneous than those of untreated patients. Such genetic consolidation of tumors might help identify subclone-specific targets that were not seen in pre-treatment biopsies, providing consolidation therapy for pharmaceutical rather than surgical follow-up.

Third, continuing evolution within a tumor represents subclones both competing and cooperating; either process might be targeted if the mechanisms were better understood. For example, a model of three-dimensional tumor growth suggests that subclones with greater cellular migratory capacity outcompete other subclones by being better able to reach tumor regions that favor cell division over cell death.⁷³ Migratory activity might thus provide a therapeutic target. Cooperation among subclones might also provide therapeutic targets. In mouse models a small subclone can enhance growth and invasiveness of other cancer cells via the microenvironment.⁷⁴ Targeting those subclones or inhibiting their influence could provide a novel therapeutic approach. Research on competition and cooperation among subclones, presently in early stages, might provide many other ways to take advantage of the mechanisms underlying intra-tumor heterogeneity.

Conclusion

Although genetic diversity both among and within tumors poses substantial challenges to personalized medicine, including targeted and immunotherapy, these challenges may be met if diversity is addressed directly. Identifying more driver mutations and how they affect cellular processes will address challenges from diversity among tumors. Assessing diversity within tumors will allow evaluation of its effects. Better understanding of intra-tumor heterogeneity should provide new therapeutic targets. Expanded and improved combinations of genomic and clinical data will facilitate this and many other types of cancer research.