Somatic evolutionary timings of driver mutations

Tumor cells accumulate numerous somatic mutations. Some of these mutations directly contribute to tumor growth and progression and are commonly referred to as driver mutations. Knowledge of the relative timing of driver mutations is essential for understanding cancer progression as a whole and for optimizing treatment for individual patients [1‐4]. For this reason, much attention has been paid to identifying the distribution and frequency of driver mutations among tumor cell populations [5‐7].

Mutations found in most cells of a tumor can be detected by estimating the fraction of cancer cells harboring a particular variant, or cancer cell fraction (CCF) [8]. This approach has been utilized to analyze the extensive data available through the Cancer Genome Atlas (TCGA), a comprehensive database that contains genomic changes in hundreds of thousands of tumors (single tumor genome sequencing results) from 33 types of cancers [9]. When CCFs were estimated for driver mutations found in tumors from TCGA, most driver mutations were present at high CCF, meaning most driver mutations were found in the majority of cells in a tumor [10]. Mutations found in a large proportion of tumor cells are likely to have occurred at the earlier stages of tumor growth, because such early-arising mutations are inherited by all cells in a tumor following clonal evolution [2, 11]. Therefore, it was inferred that the majority (>70%) of driver mutations occur early in tumor growth for all types of cancer surveyed [12].

However, tumors consist of heterogeneous cell lineages, each of which may be driven by different driver mutations [1‐3]. For example, a study of renal cell carcinoma reported 73–75% of driver mutations to have likely occurred at a later time in tumor evolution [13]. This study employed a different methodology to detect early mutations, where variants from multiple regions of a given tumor (M-seq) were analyzed. Using an M-seq approach, variants found in a majority of sampled regions are classified as early (ubiquitous) mutations [6]. Additional studies have employed M-seq methodology to study different types of cancer, and these studies have identified many mutations that are private to one or only a few regions of a tumor [14‐21], or mutations that are present at different points in time [22, 23]. This identification of many private mutations indicates that single-tumor profiles, such as those in the TCGA database, do not completely capture the spectrum of late-arising driver mutations. Therefore, conclusions based on analyses of TCGA data [12] may not apply to all the cancers and/or patients.

Now that M-seq data from various cancers are available, it is possible to comprehensively explore the relative preponderance of driver mutations arising in early and late stages of tumor growth. Here we perform a meta-analysis using samples from 101 individuals representing various cancer types [13, 14, 24‐32], which revealed that the fraction of driver mutations occurring early in tumor growth varies extensively among cancers as well as among individuals. We evaluate the frequency of late-arising driver mutations in primary and metastatic tumors.

We obtained sequencing read counts of mutant and wild type alleles, and their chromosomal positions, for 101 tumor data sets from 11 published studies that contained at least three tumor samples per patient (Table 1) [13, 14, 24‐32]. Our analysis focused on single nucleotide variants (SNVs) and insertions and deletions (indels) that arose somatically in the tumors of individual patients. To identify mutations that likely affected cancer progression or development (driver mutations), we first extracted mutations in coding regions of the genome by mapping the chromosomal position of each variant onto the reference human genome (hg19) from the Ensembl database [33]. We excluded mutations in intergenic regions, because their functional effects on cancer development are rarely known. We also excluded synonymous mutations, because these mutations are not expected to significantly affect protein function.

To ensure that our findings regarding the distribution of drivers are robust to methods of driver determination, we used five schemes to determine mutations that are cancer drivers. We first determined driver mutations based on whether the affected gene has been previously implicated in cancer. We annotated every mutation occurring in a cancer-associated gene listed in the COSMIC cancer gene census [34, 35] as a driver (Driver annotation type I). Among these genes, those without functional annotation in cancer (oncogene or tumor suppressor gene) could be false-positives. Therefore, we applied a second, more stringent approach, in which we used only known oncogenes and tumor suppressor genes listed in COSMIC (Driver annotation type II).

Some sites in the genome are more frequently mutated in cancer than others, and these somatic variant hot spots are believed to play a role in cancer [36]. In our third approach, we annotated driver mutations when mutations were located within 15 nucleotides of somatic variant hot spot (Driver annotation type III), because mutations at hot spots and neighboring regions may be cancer drivers. Variant hot spots were those identified as individual substitution hot spots as presented in Chang et al. [36], those mutated >10 individuals in the COSMIC database, or those mutated in at least two individuals in our 101 datasets.

Furthermore, many computational methods have been created to determine variants with functional roles in cancer [37]. In the fourth approach, we used one such driver mutation prediction tool, IntOGen [38] (http://​www.​intogen.​org/​analysis). The IntOGen pipeline examines genes that frequently have mutations with high functional impact and regions of the protein sequences where mutations frequently occur (Driver annotation type IV). To predict driver mutations with IntOGen, we input all mutations (point mutations and indels) with chromosomal positions, wild type and mutant nucleotides, and strand information obtained from Ensembl database.

Lastly, it has been observed that genes that are causal in one cancer type many not be causal in every cancer type. Therefore, in annotation type V, we obtained a list of cancer associated genes for each cancer type from IntoGen database [38] in order to designate driver mutations following the approach outlined in annotation type I above. For analysis of data from Zhao et al. [32], we used cancer-associated genes identified for each patient by Zhao et al. to maintain consistency with their analysis.

Before distinguishing between early vs. late-occurring mutations, we made use of the extra power conferred by the M-seq approach to detect potential sequencing errors. Using Treeomics [39], we estimated the posterior probability that a putative variant read is actually present in a tumor given the reference and mutant allele read counts for multiple tumor samples. For each variant within a tumor sector, we annotated a mutation as present if the Treeomics-inferred posterior probability was greater than 0.95. We removed variants with a posterior probability less than 0.95 in all tumor samples from a patient. Similarly, a mutation was annotated to be absent if the Treeomics-inferred posterior probability was less than 0.05.

To distinguish early from late-arising mutations, we defined early-arising mutations as those that are found in all samples (ubiquitous mutations), and late-arising mutations as those that are private to only one sample (region-specific or private mutations). This definition is more stringent than the definition used in the previous TCGA analysis [12], because it excludes mutations found in some, but not all tumors. Consequently, mutations annotated to be early-arising are expected to be found in the progenitor of all tumor cells in a patient, and those designated late-arising status are expected to have arisen in only one tumor in a patient. We refer to all other drivers to be of intermediate origin. Note that early and late designations refer to relative timing of occurrence of mutations, they are not meant to convey absolute times. Supplementary Additional file 1 shows designations of drivers using five annotation schemes, chromosomal positions, count of samples with mutant allele after Treeomics treatment, patient ID, and study information.

Our results establish that the fraction of driver mutations occurring in the earliest stages of cancer varies among patients as well as cancer types. We have shown that, overall, the number of late driver mutations are equal to or greater than early drivers in 44% of the patients with metastatic tumors. This conclusion differs from some previous reports arguing that the majority of driver mutations happen early in cancer progression [12] or that tumors follow a neutral pattern of evolution after initial growth propelled by the effects of early driver mutations, i.e., intratumor heterogeneity is caused by passenger mutations [46].

Our observation that the number of late driver mutations is similar to early driver mutations does not inform us about the rate of driver mutation occurrence per cell or about the relative degrees of selective advantage conferred by early and late drivers. The number of cells that arise late in cancer progression (subclonal cells that have subclonal variants) is expected to exceed the number of clonal cells, so the number of subclonal variants is expected to exceed the number of clonal variants, which would result in the increased preponderance of mutations. In fact, the number of driver mutations was linearly correlated with the number of passenger mutations for both early (Fig. 6a) and for late (Fig. 6b) mutations (see also [32]). Actually, the number of early driver mutations per all early mutations (fraction of driver mutations) was similar to the fraction of late driver mutations (Fig. 6c). This pattern was different from the fractions of early and late driver mutations (Fig. 1). Thus, even when the numbers of early and late driver mutations are similar, it will not mean that the rate of driver mutation occurrence or accumulation per cell is the same.

We found that subclones with late drivers occur with significant frequencies; the average observed mutant frequency of late-arising mutations was 19% (with a standard deviation of 14%). In fact, 34% of late mutations were present at frequencies greater than 20% (Fig. 7a). However, the relative degrees of selective advantage conferred by early and late drivers is complex to assess from such frequency data, as for example, subclonal expansions may be caused by spatial constrain without positive selection [46]. Furthermore, a comparison of the frequency of late driver and passenger mutations is not able to inform about positive selection, because passenger mutations hitchhike with driver mutations—which would result in similar observed mutant frequencies for both [47]. As expected, the distribution of the observed mutant frequencies of late drivers was similar to that of late passengers. This pattern was also observed when all the data from all late mutations was pooled together (Fig. 7a) and when the comparison was restricted to individual regions that contained at least 10 late driver mutations (>10 mutations; Fig. 7b, c). However, it does appear that higher intratumor heterogeneity in the late stages is a result of the continued occurrence of genuine driver mutations with functional effects on tumor growth, because recent studies have found subclone-specific driver mutations in tumors using single-cell sequencing techniques. For example, putative driver mutations were identified that are unique to a subset of the clones of an individual bladder tumor detected through single cell sequencing [48]. We expect more detailed studies in the future to test the patterns that we have observed in the meta-analysis presented here.

In a meta-analysis of genome variation data from multiple tumor in each patient, we find that the numbers of late driver mutations are substantial: they often exceed the number of early drivers. No previous study has conclusively demonstrated this pattern, even though they have indicated presence of driver mutations in tumors. These results implicate driver mutations in the continued development of aggressive tumor growth and in progression during later events such as recurrence, metastasis well beyond the initial founding of the tumor. Finally, these results highlight the importance of accounting for intratumor heterogeneity when evaluating the mutational histories of tumor cell populations.