An analysis of genetic heterogeneity in untreated cancers

Abstract

Genetic intratumoral heterogeneity is a natural consequence of imperfect DNA replication. Any two randomly selected cells, whether normal or cancerous, are therefore genetically different. We re-analyzed the extent of genetic heterogeneity within untreated cancers with particular regard to its clinical relevance. We found that homogeneity of predicted functional mutations in driver genes was the rule rather than the exception. In primary tumors with multiple samples, 97% of driver gene mutations in 38 patients were homogeneous. Moreover, among metastases from the same primary tumor, 100% of driver mutations in 17 patients were homogeneous. With a single biopsy of a primary tumor in 14 patients, the likelihood of missing a functional driver gene mutation that was present in all metastases was 2.6%. Furthermore, all functional driver gene mutations detected in the primary tumor were present among all metastases. Last, we found that individual metastatic lesions responded concordantly to targeted therapies in 91% of 44 patients. These data indicate that the cells within the primary tumors that gave rise to metastases are genetically homogeneous with respect to functional driver gene mutations and suggest that future efforts to develop combination therapies have the potential to be curative.

Introduction

Cancer is an evolutionary process spanning multiple decades. During the expansion of cell populations, intratumoral heterogeneity (ITH) arises as a natural consequence of imperfect DNA replication^1–4. Whenever a cell divides, a few mutations across the whole genome are acquired. In a tumor comprised of billions of cells, every conceivable point mutation is expected to be present in at least a few cells. At the genetic level, not only is every cancer type different, but also every tumor of the same type and every cell of the same tumor are different. This extensive heterogeneity has been considered a major barrier to drug development and long-term disease control^5–10. However, the success (even if short-lived) of several forms of targeted therapies suggests that intratumoral heterogeneity does not preclude initial therapeutic response. For example, in patients with metastases – who represent the majority of patients treated with therapeutic agents – it would be difficult to observe an objective response if some metastatic lesions did not harbor the targeted driver gene mutation in the vast majority of their cells. How can the successful responses to targeted therapies be reconciled with the intratumoral heterogeneity that has been observed in next generation sequencing studies?

Here we re-evaluate sequencing data in the literature with particular regard to the clinical significance of intratumoral heterogeneity. As a result of the different forms of tumor heterogeneity and the recent focus on subclonal heterogeneity, some discrepancies have arisen between the interpretations of observed heterogeneity and its clinical implications^2,11. Other discrepancies arise from loose distinctions between functional driver gene mutations and passenger mutations because not every mutation within a bona fide driver gene actually drives tumorigenesis^2,12–14. When these factors are taken into account, the sequencing data are in harmony with clinical experience. Homogeneity of true driver gene mutations emerges as the rule rather than the exception in treatment-naïve cancers.

Driver gene mutations and their role in tumor evolution

Before we begin with a quantitative description of tumor heterogeneity at the genetic level, we review some of the basic principles underlying the genetic determinants of cancer. Solid tumors typically require alterations of three driver genes to convert a normal cell to a cancer cell^2,15–20. This number can vary among cancer types and individual patients. Each of these alterations promotes tumorigenesis by providing a selective growth advantage to the cells within their microenvironment. In other words, driver gene mutations result in an increase in cell division or a decrease in cell death, resulting in a net cell gain overall. Relatively small changes in the cell birth rate, b, or in the death rate, d, can dramatically alter the net growth rate, given by r = b − d^21–26. For example, assume a tumor grows exponentially with a volume doubling time of 150 days²⁷. The growth rate is then r = ln(2)/150 ≈ 0.5% per day. If the cells within the tumor divide every 4 days^28,29, then b = 1/4 = 0.25 per day and the death rate is d = b − r = 0.245 per day according to the above given formula. Suppose a driver gene mutation then occurs in one cell of this tumor. A driver gene mutation causes an increase in the birth rate of on average 0.4%²¹, though some driver gene mutations can confer much stronger or weaker selective advantages^30,31. A typical new birth rate is then b′ = b(1 + 0.004) = 0.251 per day. If the death rate is unchanged, then the new growth rate of this cell is r′ = b′ − d = 0.6% per day. The new mutation therefore increases the net growth rate by 20% per day (= 0.6%/0.5%). The number of these mutated cells will then double every ≈ 120 days (= ln(2)/0.006) as opposed to 150 days of the cells without this additional driver gene mutation. Over many months to years, this difference is sufficient for cells with driver gene mutations to progressively outgrow the cells without this new mutation in the tumor^21,32–35.

Driver genes can be classified into well-defined signaling pathways and their effects depend on the tissue origin of the cells. A few dozens of driver genes are recurrently mutated across many cancer types. Most driver genes are recurrently mutated only in a few tissues and cancer types. Functional consequences of mutated driver genes complement each other, resulting in patterns of co-occurrence and mutual exclusivity among driver gene mutations^36,37. In the case of oncogenes, a single missense mutation generally represents the genetic alteration responsible for activating it. In the case of tumor suppressor genes, inactivation typically requires two separate mutations. One of these mutations is usually intragenic (producing a stop codon, for example), while the other is often a large deletion that inactivates the other allele^38,39. It is important to note that driver genes (e.g., NOTCH1 or CDH1) can act as oncogenes in one cancer type but as tumor suppressor genes in other cancer types^2,40, reflecting the different signaling circuitries that define organogenesis.

The first driver gene mutation allows the formation of a small clonal expansion, creating a benign lesion^17,41,42. These lesions typically grow to a size of a few million cells and are usually undetectable clinically. The second driver gene mutation results in a second wave of clonal local expansion, often leading to a clinically detectable, though still benign, tumor^19,43–45. The third mutation endows the tumor cell not only with a further selective growth advantage but also with the ability to expand its environment by invading through the basement membrane^19,45, thereby defining malignancy (a.k.a. cancer). Advanced tumors typically contain frequent gains and losses of focal genomic regions, chromosome-arms, and whole chromosomes^39,46–48. Depending on the cancer type, whole-genome duplication (WGD) occurs in 10–80% of cancers which could lead to subsequent chromosomal alterations^38,49. To date, it has been impossible to determine whether the rate of chromosome gains and losses (chromosome instability) increases during tumor progression. However, a new approach employing organoids should make this possible in the future⁵⁰. Despite intense efforts, no genetic alterations have been identified that unambiguously endow the cell with the ability to metastasize⁵¹. The process of metastasis seems stochastic; once a cancer has developed (i.e., acquired invasive growth capability), it may only be a matter of time before a cell invades a vessel and seeds a distant metastasis¹³.

Driver gene mutations are “clonal” if they are present in virtually all cells of the cancer. Clonal mutations are also called “truncal” because they are in the trunk of the tumor evolutionary tree. Subclonal mutations represent those that are present in only a subset of the cancer’s cells. Subclonal mutations are sometimes described as “branched” because they occur on a branch of the tree when the evolutionary trajectory of the tumor can be assessed. Though three driver gene mutations appear to be sufficient for the development of a malignant solid tumor, more than three driver genes can be observed in cancers, because the evolutionary process of tumors never stops. These additional mutations can be clonal, but are more likely to be subclonal compared to the first three mutations driving the disease.

Heterogeneity among driver gene mutations in benign tumors and expanded clones occurs frequently^{18–20,45,52–54}. However, this heterogeneity can essentially be erased in the primary tumor by a mutation that endows a strong growth advantage (e.g., the third mutation in the chain resulting in the advent of malignancy). In the parlance of population genetics, such a mutation results in a clonal sweep: the vast majority of the cells within the cancer descend from this mutant cell, which outcompeted other clones in the developing tumor (Fig. 1).

Fig. 1: Clonal sweeps give rise to driver gene mutation homogeneity.

Subclonal cells containing different driver genes emerge over time. Subclones of cells with different driver gene mutations are colored yellow, orange, red, or green. a | Driver gene mutation heterogeneity in a small lesion. b | Lesion grows with the expansion of both the yellow and the red subclones. Some subclones may progress, others remain stable or regress. c | The red subclone sweeps through the lesion and eradicates the preexisting driver gene heterogeneity harbored by the yellow subclone. New driver gene mutations in another subclone (green) may be acquired during the growth of the lesion.

Relatively few mutations result in a gain-of-function or loss-of-function of a driver gene to confer a selective growth advantage. These driver gene mutations are functional because they increase the rate of cell division or decrease the rate of cell death. Just because a mutation occurs in a driver gene does not mean that it drives tumorigenesis. Mutations in driver genes that are not functional should be considered passenger mutations because of their effectively neutral consequences on selection^13,23,40,55. With sufficiently deep sequencing, essentially every possible point mutation in every gene will be observed. Many candidate driver gene lists already contain more than 1000 genes, and driver gene mutations are becoming increasingly likely to be false-positives. Each mutation in a driver gene needs to be carefully assessed before functional consequences should be indicated. In this review, we have attempted to rigorously distinguish mutations that are likely to be functional from those that are not.

A final and often unappreciated point about genetic heterogeneity is that it is not confined to tumors. Any two randomly selected normal cells from a healthy adult contain hundreds to thousands of genetic alterations that distinguish the two cells^3,48,56–60. One can precisely quantify heterogeneity and evolutionary relationships through various metrics such as the Simpson index, the Shannon index, or the Jaccard similarity index^26,61–66. Applied to normal cells and cancer cells, these metrics reveal that cancer cells are more similar to each other than normal cells are to each other. For example, the fraction of distinct genetic alterations between any two random cancer cells from a single cancer is much smaller than between any two random normal cells from a normal organ^56,64. The reduced heterogeneity in cancer cells is a consequence of the clonal sweeps described above, wiping out all prior heterogeneity among the other clones (Fig. 1). Viewed in such a quantitative context and in comparison to normal cells, the extent of intratumoral genetic heterogeneity does not emerge as a distinctive feature of cancer. Nevertheless and independent of the degree of heterogeneity, the key questions addressed in this review remain the same: in which situations is genetic heterogeneity clinically important and why?

Forms of tumor heterogeneity

Genetic differences between cancers of two patients, each with a different tumor type (intertype heterogeneity) are well-known (e.g., mutations present in prostate cancer vs. in pancreatic cancer). Even cancers of the same type in two different individuals are genetically very different and may share very few or no somatic mutations (intratype heterogeneity). These differences are the basis for precision medicine: patients are treated with drugs that target the genetic alterations that are present in their particular tumor. This contrasts with conventional chemotherapeutics, in which all patients with a given tumor type are treated identically. Perhaps the epitome of personalized medicine is illustrated by the “tumor type-agnostic” approval of immune checkpoint inhibitors for cancers⁶⁷. The drug pembrolizumab is now recommended for treatment of patients whose tumors are mismatch repair deficient, regardless of the tumor type. A similar tumor type-agnostic indication for patients with tumors bearing TRK mutations was recently approved⁶⁸.

This review focuses on three forms of heterogeneity that affect the same cancer in a single induvidual^2,69. Intraprimary heterogeneity refers to the genetic heterogeneity between two cells of the same primary tumor; intermetastatic heterogeneity refers to the genetic heterogeneity between the cells that seed distant metastasis, and intrametastatic heterogeneity refers to the genetic heterogeneity between two cells of the same metastasis (Fig. 2).

Fig. 2: Three forms of heterogeneity within a single patient.

Subclonal cells containing different driver genes emerge over time. Subclones of cells with different driver gene mutations are colored yellow, orange, or green. a | Intraprimary heterogeneity: Subclones containing different driver gene mutations expand in parallel. b | Intermetastatic heterogeneity: Cells with different driver gene mutations disseminate and colonize distant sites, leading to driver gene heterogeneity among the founding cells of different metastases. c | Intrametastatic heterogeneity: Mutations in the founding cells of a metastasis clonally expand so that they are present in all cells of the metastasis. However, additional driver gene mutations can be acquired during the growth process of the metastatic lesion. Whether intrametastatic heterogeneity can arise from the dissemination of new clones from one metastatic lesion to another is the subject of ongoing research^80,140,141.

Intraprimary heterogeneity

Intraprimary heterogeneity can directly affect patient outcomes only when the primary tumor cannot be excised. As depicted in Fig. 3, in ~57% of all newly diagnosed cancers in the US, the primary tumor is surgically resectable. For selected common solid cancers (representing 81% of all new cancer cases in the US), ~70% of primary tumors are surgically resectable and intraprimary heterogeneity is clinically irrelevant⁷⁰. Nonetheless, intraprimary heterogeneity can sometimes provide important prognostic information^61,66,71,72. Moreover, when the primary tumor cannot be completely resected, such as is nearly always the case in glioblastomas, intraprimary heterogeneity among driver genes can limit the response to therapies that target such driver genes. Such driver heterogeneity explains, for example, why agents that target the EGFR variant translocation have not achieved notable clinical success; not all cells within most glioblastomas contain this translocation⁷³.

Fig. 3: Majority of primary tumors are surgically resectable at the time of diagnosis.

a | Estimated incidence of selected solid cancers in the United States in 2018⁷⁰. *Solid cancer types with more than 10,000 estimated new cases per year were selected. Selected cancer types represent approximately 81% of all new cancer cases in the US. Hematological cancers, cancer types with less than 10,000 estimated new cases per year, and cancers for which surgery is not routinely recommended (i.e. small-cell lung cancers), or for which the primary tumor often cannot be completely resected (i.e. glioblastoma) were excluded. b | Fraction of resectable primary tumors across cancer types in the US. Approximately 70% (984,506/1,400,960) of newly diagnosed cases of these solid cancer types (panel a) and approximately 57% (984,506/1,735,350) of all newly diagnosed cancer cases are resectable.

Intraprimary heterogeneity can be assessed through multi-region or single cell sequencing of primary tumors. Several studies of this sort have revealed driver gene mutations present only in a subset of the evaluated regions of some tumors^74–81. Some of this heterogeneity can be explained by sequencing noise, low neoplastic cell content, or low sequencing depth in individual samples^81–84. For example, if the depth of sequencing for a specific genomic position is only n = 10 reads, and the neoplastic content of the tumor DNA used for sequencing is 50%, then the probability of completely missing that mutation (k = 0) when it occurs in 100% of the cancer cells is $\left(\begin{array}{c}n\\ k\end{array}\right)\cdot f^k\cdot {\left(1-f\right)}^{n-k}=5.6\%$ (assuming that one of two alleles in the tumor cells is mutated, f = 0.5/2). The probability of missing at least one of three such mutations is $1-{\left(1-{\sum }_{k=0}^1\left(\begin{array}{c}n\\ k\end{array}\right)\cdot f^k\cdot {\left(1-f\right)}^{n-k}\right)}^3=57\%$ if it is assumed that the mutation needs to be observed twice (k < 2) to ensure that it is real rather than an artifact of sequencing. Multi-sample analysis accentuates this problem because the probability that sequencing depth or neoplastic cell content is low in at least one sample strongly increases with the number of analyzed samples²³. Most sequencing data analysis methods have been optimized for single-sample analysis, and few methods have been described to minimize artifacts in the context of multi-sample analysis^82,85. Some of the newer methods to control for multi-sample analysis artifacts have been applied to the sequencing data described in this study (Supplementary Methods S1).

Minimal functional consequences of subclonal driver gene mutations

Intraprimary heterogeneity is conferred by subclonal mutations. But are the subclonal mutations that occur in driver genes functional? As noted above, just because a mutation occurs in a driver gene does not mean that the particular mutation actually drives tumorigenesis. To address this question, we reanalyzed data from 38 untreated primary epithelial tumors derived from six cancer types in which multi-region sequencing had been performed (13 ovarian⁸⁶, 10 colorectal^23,79,87,88, 9 breast⁷⁸, 4 pancreatic⁶⁴, 1 gastric⁸⁹, and 1 endometrial cancer¹³). In each region, we classified mutations as present if their “present probability” was at least 80% according to the Bayesian inference model of Treeomics⁸² (Supplementary Methods S1). Mutations were classified as absent if their “absent probability” was at least 80%. Mutations that were present in at least one sample but not in all samples were classified as subclonal. Ambiguous mutations that did not reach these presence or absence probability thresholds in all samples of a patient were excluded from the analysis to minimize effects of low sequencing depth or low neoplastic cell content. We thereby identified 19 subclonal, non-synonymous mutations within the 299 driver genes listed in the TCGA consensus list⁴⁰ (Supplementary Table S1; Fig. 4a). The number of subclonal mutations was considerably less than the number of clonal mutations (19 subclonal vs. 143 clonal in the same 38 cancers).

Fig. 4: Intratumoral heterogeneity in untreated primary tumors and among metastases.

Intraprimary heterogeneity analysis based on 96 samples from 38 subjects (13 ovarian⁸⁶, 10 colorectal^23,79,87,88, 9 breast⁷⁸, 4 pancreatic⁶⁴, 1 gastric⁸⁹, and 1 endometrial cancers¹³; Supplementary Methods S1). Intermetastatic heterogeneity analysis based on 67 metastases samples of 17 subjects (6 pancreatic⁶⁴, 4 endometrial^13,104, 3 colorectal⁷⁹, 2 breast¹⁰³, 1 gastric⁸⁹, 1 prostate¹⁰⁵ cancers). a | Driver gene mutations present in all samples from a single primary tumor were more frequently predicted to be functional than those present in only a subset of the samples from a primary tumor (54% vs. 11%, P < 0.001). The fraction of subclonal functional driver gene mutations (11%) was not significantly different from the fraction of clonal or subclonal functional passenger gene mutations in the same tumor (3.3% and 2.3%). b | Mutations in driver genes that were present among all metastases samples of a subject were more frequently predicted to be functional than those present only in a subset of metastases samples (65% vs. 0%, P < 0.001). The fraction of subclonal functional driver gene mutations (0%) was not significantly different from the fraction of clonal and subclonal functional passenger gene mutations in the same samples (4.1% and 6.6%). c | On average 69% and 66% of the mutations per patient were clonal (homogeneous) among primary tumor samples and among metastases, respectively. Mutations in putative driver genes were significantly more homogeneous among primary tumor samples (90%, P < 0.001) and among metastases (84%, P < 0.0048) than mutations in all genes (sum of passenger genes and driver genes). Likely functional driver gene mutations were even more homogeneous among primary tumor samples (98%, P < 0.0042) and among metastases (100%, P < 0.0018) than other categories of mutations. Two-sided Fisher’s exact tests were used in panels a and b. Two-sided Wilcoxon rank-sum tests were used in panel c. Thick black bars denote 90% confidence interval. Numbers in brackets denote number of variants in each group. ** P < 0.01; *** P < 0.001.

To determine whether these 19 subclonal mutations were likely to be functional, we pooled information from various databases and used bioinformatic methods to create a two-phase algorithm, called LiFD (Likely Functional Driver; Supplementary Methods S1). In the first phase of LiFD, mutations that are annotated in OncoKB⁹⁰, the catalog of validated oncogenic mutations (CGI, Cancer Genome Interpreter¹²), known cancer hotspots⁹¹, or present at least 4 times in COSMIC⁹² (Catalogue of Somatic Mutations in Cancer) were predicted to be functional. If a mutation was not annotated as functional in the first phase, we used CHASMplus^14,93, FATHMM⁹⁴, CanDrA⁹⁵, CGI¹², and VEP⁹⁶ to predict the functional consequences of this mutation in the second phase of LiFD. If the majority of the methods that produced a result predicted functionality, we classified the mutation as likely functional (Supplementary Methods S1; Supplementary Fig. S1). This method was lenient in that it allowed mutations scored as significant as judged by only a subset of tools to be considered functional in LiFD.

Through LiFD, we found that clonal mutations in putative driver genes were significantly more likely to have functional consequences than subclonal driver gene mutations in the same tumors (Fig. 4a; Supplementary Table S1). Only two (11%) of the 19 subclonal mutations compared to 77 (54%) of the 143 clonal mutations were predicted to be functional (P < 0.001, two-sided Fisher’s exact test; Fig. 4a). The two likely functional subclonal driver gene mutations occurred in PTEN in the same patient, one in each of two regions of the cancer. When evaluated at the individual tumor level, 97% (37/38) of the tumors evaluated had no functional subclonal driver gene mutations (Fig. 4). On average, we found 2.1 likely functional driver gene mutations per primary tumor.

Survival analysis of patients with subclonal driver gene mutations

Another clinically important question is whether patients whose tumors have subclonal mutations in driver genes and are thus heterogeneous with respect to driver gene mutations have a worse prognosis than patients without such heterogeneity^{61,71,76,97–99}. To address this question, we reanalyzed data from 100 early-stage non-small-cell lung cancers where 62 subjects were reported to have at least one subclonal driver gene mutation⁷² (either point mutation or short insertion or deletion). We did not find a statistically significant difference in disease-free survival between patients that exhibited subclonal driver gene mutations (n = 62) and those that did not (n = 38) based on the originally reported heterogeneity and driver classification (Fig. 5a). When the heterogeneity classification and the driver gene mutation classification described above were used (Supplementary Methods S1), the number of cancers harboring subclonal driver gene mutations decreased from 62 to 32. Nevertheless, no significant difference in patient outcomes was observed (Fig. 5b). Though it would be reasonable to expect that tumors that had acquired additional driver gene mutations would be more aggressive, allowing escape from host control and conferring worse survival, this was not the case. We again found that clonal mutations in driver genes were significantly more likely to be functional than subclonal ones although the high mutation rate in lung cancers complicates the driver functionality prediction (33% vs 20%, P < 0.001, two-sided Fisher’s exact test).

Fig. 5: Subclonal driver gene mutations did not lead to worse patient outcomes in patients with non-small-cell lung carcinomas.

Analysis based on data of Jamal et al.⁷². a | No statistically significant difference in disease-free survival between patients that harbored subclonal driver mutations (n = 62) and those that did not harbor any subclonal driver gene mutations (n = 38), according to the originally provided driver and heterogeneity classification. Shaded areas denote 90% confidence interval. The hazard ratio of subjects with subclonal driver gene mutations was 0.51 (95% CI: 0.24 − 1.1; P = 0.088, likelihood ratio test). b | When the LiFD algorithm for identifying functional driver gene mutations was applied, the number of patients that harbored subclonal driver gene mutations was 32 and the number of patients that did not harbor any functional driver gene mutations was 68. No statistically significant difference in disease-free survival between patients that harbored subclonal functional driver gene mutations and those that did not harbor subclonal functional driver gene mutations was observed. The hazard ratio of subjects with subclonal functional driver gene mutations was 1.4 (95% CI: 0.61 − 3.0; P = 0.46, likelihood ratio test). In panel b, a different heterogeneity classification was performed than in panel a (Supplementary Methods S1).

Single biopsies generally provide adequate information for precision medicine

Intraprimary tumor heterogeneity also informs the number of biopsies required for choosing the optimal targeted therapies for metastatic lesions. For example, if only a single region of a primary tumor is biopsied, what would be the probability of selecting a functional (and perhaps targetable) driver gene mutation that was not present in all metastases? Conversely, what would be the probability of missing a functional driver gene mutation that was present in all metastases if only a single biopsy were used for sequencing analysis? To address these questions, we reanalyzed data from 14 treatment-naïve subjects in whom at least one sample of the primary tumor and at least two distinct metastases were sequenced¹³. First, any detected functional driver gene mutations present in a primary tumor biopsy were also present among all metastases of that patient. Second, we found that the proportion of functional driver gene mutations present in all metastases but missing from a primary tumor biopsy was on average 2.6% (Supplementary Methods S1; Supplementary Table S2). These data support the conclusion that in most patients, a single biopsy of a primary tumor captures the information necessary for therapeutic choices about the treatment of extant or presumptive metastases. Because untreated samples of the primary tumor and of multiple metastases are rarely available, these analysis results are based on a relatively small cohort, representing only five cancer types. Further research will be required to determine the clinical scenarios in which multiple rather than single biopsies are advantageous. For example, when lesions are multi-focal, such as in the esophagus, evaluation of biopsies from several sites clearly provides useful information^66,71,72,100.

Intermetastatic heterogeneity

Intermetastatic heterogeneity is the most important form of heterogeneity for patients with primary tumors that can be completely excised^2,13,101. Intermetastatic heterogeneity of driver gene mutations determines whether all lesions have the capacity to respond to a given targeted therapeutic agent. If even a single lesion lacks the driver gene mutation being targeted, and therefore continues to grow following the initiation of therapy, it is much less likely that an objective response will be observed than if all lesions harbor the mutation¹⁰².

Most studies of intratumoral heterogeneity have focused on primary tumors although metastases are responsible for most cancer-related deaths. Moreover, intermetastatic heterogeneity provides a uniquely informative view of intraprimary heterogeneity. If intraprimary heterogeneity were important for treatment response of metastatic disease, some metastases would be derived from the subclones in the primary tumor that define this heterogeneity. In prior studies addressing this issue, patients have often been treated with toxic or mutagenic agents which complicate the interpretation of mutations observed in metastases. We therefore surveyed the literature for patients in which at least two distant treatment-naïve metastases underwent genome/exome-wide sequencing¹³. Across all cancer types, and among tens of thousands of patients in whom genome-wide sequencing was performed, only 17 subjects were found to fulfill these requirements^{13,64,79,89,103–107} (6 pancreatic, 4 endometrial, 3 colorectal, 2 breast, 1 gastric, 1 prostate cancer; Supplementary Methods S1). Using the LiFD classification framework, we found that 65% (44/68) of all clonal non-synonymous mutations in driver genes were predicted to be functional while no (0/14) subclonal mutations were predicted to be functional (Fig. 4b; Supplementary Table S3). Hence, all of the predicted functional driver gene mutations were present in all metastatic lesions of individual patients (Fig. 4c). The fraction of subclonal driver gene mutations predicted to be functional (0%) was not significantly different from the fraction of clonal or subclonal passenger gene mutations predicted to be functional (4.1% and 6.6%, respectively). We repeated this functional analysis with a more expansive driver gene list and obtained similar results (Supplementary Fig. S2, Supplementary Methods S1).

For previously treated metastases, varying degrees of intermetastatic heterogeneity of driver gene mutations have been reported within and across cancer types^{64,74,101,108–112}. This is not unexpected because therapies create selective bottlenecks that unmask additional mutations. Which of these additional mutations in driver or passenger genes actively contribute to progression and resistance is often unclear, particularly when no functional analysis was performed. Moreover, the selective bottleneck enforced by a therapeutic agent can be very different from the selective bottlenecks operating during cancer initiation and progression: a potent driver mutation of cancer initiation may not contribute to resistance and a potent resistance mutation may not drive carcinogenesis. Some of the additionally observed driver gene mutations can be explained by their role in conferring resistance. For example, KRAS mutations following treatment with EGFR inhibitors, loss of PTEN following treatment with PIK3CA inhibitors, or FGFR2 mutations following treatment of cholangiocarcinoma patients with FGFR inhibitors^{6,9,27,113–117}. The mutations that confer resistance to targeted agents such as these are lesion-specific, and can differ considerably among the metastases of a single patient. This type of intrametastatic heterogeneity is very important for selecting second line therapies. However, it is typically not relevant for selecting the initial therapies for newly diagnosed patients because it is usually only present in a tiny fraction of metastatic cells prior to therapy^8,23.

Clinical correlates of intermetastatic heterogeneity

The success of targeted therapies for most cancers is dependent on the homogeneity of the targeted driver gene mutations among metastases. At present, all targeted therapies are based on oncogene alterations. Tumor suppressor gene alterations cannot yet be targeted by drugs because there is currently no way to restore the function of an inactivated protein. The conclusions described above, based on genome-wide sequencing of metastatic lesions in untreated patients, are strongly supported by sequencing studies of individual oncogenes in metastases from the same patients. For example, it has been shown that mutations in KRAS, NRAS, BRAF, and PIK3CA are nearly always concordant across all metastatic lesions in colorectal cancer patients^118,119. The same EGFR mutations are similarly almost always found in all metastatic lesions in lung adenocarcinoma patients¹²⁰, and the same BRAF and NRAS mutations are found in metastatic lesions of patients with melanomas^121,122.

Perhaps most importantly, are the actual responses observed in patients treated with targeted therapies consistent with the predicted homogeneity among driver genes emphasized in this review? Very few published studies provide detailed data on the response of individual lesions to targeted therapies. In general, most studies report only the data required to meet RECIST criteria for response, i.e., whether the sum of the diameters of all measured lesions decreases or increases¹⁰². An objective response is reported if the sum of those diameters decreases by more than 30%, and tumor progression is reported if the sum increases by more than 20%. However, we gathered data from two clinical trials that more directly addressed the question considered here as a proof of concept, i.e., if one metastatic lesion responds to a targeted therapy, do all index metastatic lesions respond too, or do some lesions continue to grow?

In the first of the two trials, 33 patients with melanoma and at least 2 index lesions were treated with targeted therapy^123,124 (dabrafenib, trametinib, GSK2141795). All patients had a V600E mutation in BRAF in their primary tumors. Of these 33 patients, a decrease of ≥30% in diameter of at least one index lesion was observed in 27 patients (Fig. 6a; Supplementary Table 4). We then determined how often one of the other index lesions in these 27 patients grew during the initial treatment period, generally 8 to 16 weeks and prior to the emergence of resistance. We found that in 23 of the 27 patients, none of the other index lesions grew by ≥10% (representative examples in Fig. 6b).

Fig. 6: Lesions of individual patients respond similarly to targeted therapy.

Patients are represented by humanoid cartoons. Circles within the humanoids represent responding, stable, or growing lesions (green, blue, and red, respectively). A lesion was considered to respond if it shrank by at least 30% in diameter; a lesion was considered to grow if its diameter increased by at least 10%; and a lesion was considered to be stable if it did not grow by at least 10% or shrink by at least 30%. a | At least one lesion responded in 27 of 33 melanoma patients^123,124. In 23 patients (gray humanoids), no lesion grew. In four patients (yellow humanoids), one of the lesions grew while the others responded, i.e., a heterogeneous response was observed. In six patients (red humanoids), no lesion responded. b | Examples of different types of responses to targeted therapy. All lesions responded in patient M37. One lesion responded less well than three other lesions in patient M40. None of the lesions responded in patient M29. One lesion responded, two lesions remained stable, and a fourth lesion grew in patient M08. c | At least on lesion responded in 8 of 11 thyroid cancer patients¹²⁵. In eight patients (gray humanoids), no lesions grew. In three patients (red humanoids), no lesion responded. Additional information about these patients’ responses are provided in (Supplementary Tables S4–S5). In 91% (40/44) of the patients analyzed (with melanomas or thyroid cancers), all lesions responded similarly to targeted therapy.

In the second trial, 11 patients with metastatic thyroid cancers were treated with pazopanib (VEGFR/PDGFR/RAF inhibitor) and trametinib (MEK inhibitor)¹²⁵. Of these 11 patients, a decrease of ≥30% in diameter of at least one index lesion was observed in 8 patients (Fig. 6c; Supplementary Table 5). In all 8 patients, none of the other index lesions grew by ≥10% during the first 6 months after the start of therapy or until end of treatment (whichever occurred first).

The conclusions from both trials are thereby similar. When an objective tumor response from a targeted therapy is observed in one metastatic lesion, it is common (89% of 35 patients) for all lesions in that patient to respond to the therapy. If we include cases where no lesion achieved an objective response, similar responses (regression, growth, or no stable diameters) of all index lesions were observed in 91% (40/44) of the patients. Note that this does not mean that all index lesions respond in an identical fashion. In some cases, all lesions regressed to the same extent (Fig. 6b). In other cases, the timing and degree of response varied. The timing and degree of the response is dependent on a host of factors other than the presence of the targeted driver gene mutation in the metastatic lesion. In particular, the timing and extent of a response depends on the vascularity in each tumor because this determines the dose actually delivered to the lesion. The microenvironment can also impact drug delivery and local immunity might play a role¹²⁶. Moreover, the degree of a response depends on the number of cells in the lesion that contain a mutation that can confer resistance (intrametastatic heterogeneity). These additional factors are currently beyond the control of the oncologist. But unless all metastatic lesions contain the targeted mutation, a targeted therapy will usually not be very useful. Fortunately, these clinical results confirm the above noted sequencing studies which demonstrated in all 17 patients evaluated (and in all 67 distinct metastases from those patients), if one metastasis contained a predicted functional driver gene mutation, all the other metastases of the same patient contained the identical mutation.

Intrametastatic heterogeneity

Intrametastatic heterogeneity does not impact the initial response to therapy but is responsible for disease recurrence after a response^6,8,9,115. Such recurrences result from mutations present in a small fraction of the cells within each metastasis prior to treatment; the larger the lesion, the more likely that such resistant cells exist⁸. Thus, treatment of relatively early metastatic states – with conventional chemotherapeutic agents, with targeted agents, or with immunotherapeutic drugs – are much more likely to be successful than treatment of bulky metastatic disease. Although we did not formally analyze intrametastatic heterogeneity here, a recent study of 2,520 metastases in which deep whole genome sequencing was performed showed that 96% of all driver gene point mutations were clonal³⁸. Similarly, 95% of driver gene mutations were shared among the metastatic lesions of individual patients in a study of 100 clear-cell renal cell carcinoma patients¹²⁷.

Conclusions

The results described above lead to several important conclusions. First, tumors are heterogeneous, but the term “heterogeneity” needs to be used with caution and nuance, as normal cells are heterogeneous too. We show that the extent of intraprimary heterogeneity of functional driver gene mutations is relatively small (mean of 2% per patient; Fig. 4c). We point out that unless the primary tumor cannot be excised, the extent of this heterogeneity is of little clinical consequence. More importantly from an oncologist’s viewpoint, the extent of heterogeneity of functional driver gene mutations among metastases of the same patient is minimal (Fig. 4c). Moreover, the sequencing data amassed in the literature is highly concordant with clinical experience (Fig. 6).

Multi-region sequencing enables a more refined inference of the genealogy of tumors, offering key insights into the nature of the tumorigenic process^128,129. For example, such sequencing has allowed investigators to determine the time course of tumorigenesis, which is now generally acknowledged to take decades^19,29,39,130 (a result also consistent with clinical experience). A particularly informative example of this principle was recently published: the first genetic alteration in kidney cancers occurs during early adulthood, decades prior to the onset of malignancy¹³¹. The growth of a primary tumor and its subclonal diversification occur much after the first genetic alterations, typically a few years to a decade before diagnosis^39,131. Multi-region sequencing is also critical to evaluate the potential of targeted therapies to be effective in cancers that cannot be surgically excised in their entirety, such as brain tumors, or to forecast the future evolutionary trajectory of a tumor^23,24,61,132 (Fig. 3). But for tumors that can be completely excised, sequencing of a single region from the primary tumor is generally adequate to find the clonal mutations susceptible to targeted therapies (Fig. 4; Supplementary Table S2).

Our study has several limitations. One of them is that our analyses of heterogeneity was limited to intragenic mutations (single base substitutions and small insertions and deletions). Other types of genetic alterations, as well as epigenetic alterations, undoubtedly play a role in tumorigenesis. For example, copy number alterations occur nearly ubiquitously in cancers and can confer selective advantages¹³³. Unfortunately, the target genes selected for by such copy number alterations are notoriously difficult to identify^71,133–137. With dramatic changes in copy numbers (such as occurs with true amplifications in ERBB2 or EGFR), the target gene can be identified. In the much more usual case of small changes in copy number (2 to 3-fold imbalances), it is unknown whether such copy number changes reflect a single underlying culprit gene, the combined effect of many genes^131,133,135, or simply represent passenger alterations arising as a result of chromosomal instability^39,49,138. Mutations in non-coding regions of the genome also play a role in certain cancers. However, except for mutations in the TERT promoter, individual non-coding mutations that drive tumorigenesis are rarely recurrent¹³⁹. Moreover, similar to copy number changes, it is currently very challenging to determine whether a given non-coding mutation is functional; tools like those used here are not yet available for predicting the effects on tumorigenesis of non-coding mutations. The same challenges apply to the thousands of epigenetic changes that occur in every cancer. Unless these changes occur in the ~300 well-documented driver genes^2,40, it is currently unfeasible to reliably discern which of them are likely to drive tumorigenesis, i.e., to cause a selective growth advantage in the actual human tumor microenvironment in which they occurred. Another limitation involves the difficulty of identifying functional genetic mutations. The LiFD classification framework combines the evidence of various databases and algorithms to minimize false-positives and false-negatives but is still only predictive rather than definitive. Finally, these studies of untreated cancers with numerously sampled primary tumors and metastases by necessity involved only a small number of cases. As sequencing becomes routine in clinical and research studies, we expect it will be possible to extend our type of evaluation to many other cancer cases.

The results reviewed here provide optimism for future targeted combination therapies. If intermetastatic heterogeneity in driver genes was routinely found, there would be little hope of achieving meaningful responses in most patients. We find that such heterogeneity is rare (Fig. 4), and this is compatible with the clinical success of targeted therapies in patients with metastatic disease (Fig. 6). Though these targeted therapies are not curative because of evolving resistance within metastases (intrametastatic heterogeneity), there is no theoretical reason why combinations of targeted therapies could not be curative⁹. Indeed, it has been shown that treatment of metastases with just two drugs for which no single alteration can confer cross-resistance should, in theory, cure most cancers⁸. These results apply not only to those driver gene mutations that are currently targeted but also to future targeted therapies – so such cures appear to be possible and eminently worthy of further pursuit.