Cancers are diseases caused by changes in DNA sequences. Some changes occur in the protein-coding part of the DNA sequence, in other words, in the stretches of DNA that include the instructions to make a given protein. Other changes occur in the remaining parts of the DNA that do not code for proteins, which accounts for about 98% of the human genome. Modern technologies allow us to identify these DNA changes, but, up until recently, this has only been possible for the protein-coding part of the DNA. Many studies have thus analyzed DNA changes in the protein-coding part of the human genome, while the larger, non-coding part remains rather unexplored.

Advances in technology means that large datasets are becoming available where changes in DNA sequences are identified across the entire genomes of a collection of cancer patients. However, it is not clear which of these DNA changes play a role in the development of cancer and which are neutral with no effect on cancer.

Now, Juul et al. have developed a new method, named 'ncdDetect', to search the human genome and identify stretches of DNA that when changed give cancer cells an advantage and allow them to grow. Juul et al. refer to these DNA stretches as 'driver elements', and, after analyzing the genomes from 505 patients with cancer, they identified some known driver elements and some potentially new ones. For example, possible driver elements were found in non-coding parts of the DNA that regulate genes called SMUG1 and CD1A. Both of these genes encode proteins that had been linked to cancer in the past, but driver elements had not previously been described in the nearby non-coding regions. Juul et al. also found a number of possible driver elements that might be important to consider in the treatment of cancers.

Importantly, not all the candidate driver elements identified with ncdDetect are true drivers. The changes in DNA vary greatly between different types of cancer and even between different cases of a single type of cancer. Understanding and describing this variation continues to be a challenge in identifying driver elements, and so Juul et al. plan to keep improving the method to make sure that the driver elements it identifies are all trustworthy.

Cancer is caused by somatically acquired changes in the DNA sequence of genomes (Stratton et al., 2009). Recently, large-scale sequencing of cancer-genomes coordinated by the International Cancer Genome Consortium (ICGC), The Cancer Genome Atlas (TCGA), and others has catalogued the molecular changes across hundreds of cancer samples (Hudson et al., 2010; Weinstein et al., 2013). The quest is now to analyze and understand the role of these changes in cancer development. The aberrations in non-coding regions are of particular interest as they have only become evident with the advent of whole cancer-genomes. Here, we develop the method ncdDetect for non-coding cancer driver detection. The method captures the heterogeneities of the mutational processes in cancer and aggregates signals of mutational burden as well as functional impact in the significance evaluation of a candidate driver element. We apply the method to 505 TCGA whole-genomes (Fredriksson et al., 2014).

Cancer arises by an evolutionary process where natural selection operates on genetic variation stemming from randomly occurring somatic mutations. Thousands of somatic mutations distinguish tumor tissue from healthy tissue, as a result of the mutational processes that cancer cells go through during the lifetime of a cancer patient. The somatic mutations are identified through Next-Generation Sequencing (NGS) and commonly labeled according to their effect on cancer development: Driver mutations are subject to positive selection during the evolutionary process of cancer, as they offer the cell a growth advantage and contribute to the expansion of tumors. By definition, driver genes contain one or more driver mutations. Passenger mutations, on the other hand, have a neutral, or perhaps slightly negative, fitness contribution to the cell, and accumulate as passive passengers during the evolutionary process of cancer (Stratton et al., 2009; Pon and Marra, 2015). Many more passenger than driver mutations exist in cancer cells, and distinguishing the two is challenging (Marx, 2014). Typically, the strategy is to search for signs of recurrent positive selection across a set of cancer genomes.

Signs of recurrent positive selection across cancer genomes can be detected by comparing the somatic mutation frequency to an estimated background mutation rate (Pon and Marra, 2015). However, modeling the background mutation rate is complicated as it varies along the genome with a large degree of heterogeneity (Lawrence et al., 2013; Polak et al., 2015; Alexandrov et al., 2013). Not only does the mutation rate in cancer exhibit high variation between different cancer types; this is also the case between different samples of the same cancer type. Furthermore, the mutational processes are affected by various genomic features, primarily replication timing, expression, and the position-specific sequence context (Lawrence et al., 2013; Bertl et al., 2017). It is thus crucial to take these features into account when estimating the background mutation rate in cancer. Another strategy for detecting signs of positive selection in cancer is to rank mutations according to their impact on protein function (Marx, 2014). In particular, point mutations might introduce alterations in the amino acid sequence of a protein, and thereby change its function (Reva et al., 2011).

Systematic mutational screens of cancer exomes have expanded the set of known cancer driver genes over the past decade (Forbes et al., 2015). Many tools exist for the identification of such genes (Dees et al., 2012; Lawrence et al., 2013; Gonzalez-Perez and Lopez-Bigas, 2012; Tamborero et al., 2013; Reimand and Bader, 2013), and at present, 616 cancer driver genes are catalogued as causally implicated in cancer (Cosmic, 2016). However, less than 2% of the genome codes for protein. While it is established that non-coding elements play diverse roles in regulating the expression of protein-coding genes, only few studies systematically explored the role of non-coding somatic mutations in cancer development (Weinhold et al., 2014; Lochovsky et al., 2015; Melton et al., 2015; Fredriksson et al., 2014; Mularoni et al., 2016; Lanzós et al., 2017). The first identified non-coding driver element was the TERT promoter with highly recurrent mutations across several cancer types (Huang et al., 2013; Horn et al., 2013). In general, the functional understanding of non-coding regions is poor compared to protein-coding regions, challenging the interpretation of non-coding mutations (Khurana et al., 2016).

We develop the method ncdDetect for detection of non-coding cancer driver elements. With this method, we consider the frequency of mutations alongside their functional impact to reveal signs of recurrent positive selection across cancer genomes. In particular, the observed mutation frequency is compared to a sample- and position-specific background mutation rate, which is estimated based on various genomic annotations. A scoring scheme (e.g. position-specific evolutionary-conservation scores) is applied to further account for functional impact in the significance evaluation of a candidate cancer driver element.

To strengthen our conclusions regarding the driver potential of candidate elements, we draw on additional data sources. Non-coding mutations may perturb gene expression patterns, and we thus correlate their presence with expression levels in an independent data set (Ding et al., 2015). Likewise, we correlate mutation status with survival information for these candidates.

What sets ncdDetect aside from other non-coding driver detection methods is the position-specificity, and the derived ability to include genomic annotations of varying resolution down to the level of individual positions. In one existing non-coding driver detection method, the position- and sample-specific probabilities of mutation are derived, much like in ncdDetect but are then aggregated across a candidate element during significance evaluation (Melton et al., 2015). This means that knowledge about the exact position and probability of a mutation is not fully utilized. In another method, the genome is divided into bins according to the average value of replication timing (Lochovsky et al., 2015), and in a recent method, the significance evaluation is performed by locally conditioning on the number of observed mutations within a candidate element (Mularoni et al., 2016). To our knowledge, no existing non-coding driver detection method derive and apply position- and sample-specific probabilities of mutation in the significance evaluation of a candidate driver element, and allows the use of position-specific scores and accurate evaluation of their expectation across a candidate element. This unique feature of ncdDetect means that candidate elements of arbitrary size and location can be analyzed, and that the potential large variation of mutational probabilities within a candidate element is handled.

With ncdDetect, we model the different levels of heterogeneity in the somatic mutation rate known to be at play in cancer and evaluate the relative merit of different position-specific scoring-schemes. The result is a driver detection method tailored for the non-coding part of the genome, and with it we aim to contribute to the understanding of non-coding cancer driver elements.

ncdDetect evaluates if a given non-coding element is under recurrent positive selection across cancer samples. The method takes as input (a) a candidate genomic region of interest, (b) position- and sample-specific probabilities of mutation, and (c) position- and sample-specific scores measuring mutational burden or impact.

Position- and sample-specific model of the background mutation rate

A key feature of ncdDetect is the application of position- and sample-specific probabilities of mutation. These are obtained by a statistical null model, inferred from somatic mutation calls of a collection of cancer samples (Material and methods: Statistical null model) (Bertl et al., 2017). The model predicts the mutation rate from a set of explanatory variables, that is genomic annotations (Figure 1A). In the present paper, the null model is trained on 505 whole genomes distributed across 14 different cancer types generated by TCGA (Fredriksson et al., 2014) (Figure 1B).

Figure 1 with 1 supplement see all

Download asset Open asset

Figure 1—source data 1

The number of mutations observed for each of the 505 samples.

This data set relates to Figure 1B.

https://doi.org/10.7554/eLife.21778.004

Download elife-21778-fig1-data1-v3.txt

As explanatory variables, the model includes genomic annotations known to correlate with the mutation rate in cancer, as well as annotations we have found to improve the model fit. It is well known that the mutation rate varies between samples (Lawrence et al., 2013; Alexandrov et al., 2013). Indeed, the mean and median number of mutations per sample is approximately 32 × 10³ and 8 × 10³, respectively, with a large degree of variation between and within cancer types (Figure 1B). For example, the average number of mutations per sample is 73 times higher for colorectal cancer than for thyroid cancer, and within melanoma cancer, the least mutated sample has 224 times fewer mutations compared to the highest mutated sample. The mutation rate depends on the position-specific sequence context (Alexandrov et al., 2013) and correlates with replication timing and gene expression level (Lawrence et al., 2013). The mutation rate also varies between different types of genomic regions (Weinhold et al., 2014). Finally, we find that the local mutation rate in a window around a given genomic position helps to capture unaccounted mutation rate variation and increases the goodness of fit. The complete model specification, including the definition of the local mutation rate, is given in Material and methods: Statistical null model. Consequently, genomic annotations considered as explanatory variables in the null model for each sample are replication timing, tissue-specific gene expression level, trinucleotides (the nucleotide under consideration and its left and right flanking bases, thus taking into account the sample-specific mutational signature), genomic segment (3’ and 5’ untranslated regions (UTRs), splice sites, promoter elements and protein-coding genes) and local mutation rate (Figure 1C). Given these explanatory variables for a specific genomic position, the model predicts the particular probability of observing a mutation of a given type for a specific sample at this particular position.

Strand symmetric model

The reference sequence is divided into weak (A and T) and strong (G and C) base pairs (bps) (Cornish-Bowden, 1985). As strands generally cannot be distinguished in non-coding regions, we handle them symmetrically, with weak bps denoted with an A and strong bps denoted with a G. The 12 different types of point mutations are thus collapsed into the six classes, ‘A→C’ (thus including both A→C and T→G mutations), ‘A→G’, ‘A→T’, ‘G→T’, ‘G→C’ and ‘G→A’.

Our model considers four possible outcomes for each position: Transitions (TS_{{A→G, G→A}}), two types of transversions (TV_{{A→T, G→T}} and TV_{{A→C, G→C}}) as well as the reference class of no mutation.

Mutation-rate predictions and position-specific scores

For a given non-coding element of interest, the null model ensures the availability of position- and sample-specific probabilities of each of the four possible outcomes (Figure 2A). Due to the rarity of observing a mutation, the predicted mutational probabilities are of small magnitude (Figure 2B–C, Figure 1—figure supplement 1). Additional to these probabilities, each position is associated with a score that may be sample-specific, and may depend on the outcome class (Figure 2D–E). The scoring scheme can be freely defined, and in the present paper, we illustrate three different choices of scores, namely the number of mutations, log-likelihoods and conservation scores (Figure 2E). A wide variety of other scoring schemes can be considered. In particular, the flexibility of ncdDetect allows for different scoring schemes to be chosen for different types of candidate elements.

Figure 2

Download asset Open asset

Scoring schemes

A good scoring scheme must be able to discriminate well between events that constitute true cancer drivers and events that are neutral. The scoring scheme can, for example, evaluate the mutational burden and be defined by means of the number of mutations in a candidate region. This approach has been taken by existing non-coding cancer driver detection methods (Lochovsky et al., 2015). Here, the score for a given position is defined to be one if a mutation of any type occurs, and zero if no mutation occurs. Another approach is to evaluate the goodness of fit of the observed mutations to the null model, and define the scores as log-likelihoods, that is, minus the natural logarithm of the predicted position- and sample-specific mutation probabilities. This scoring scheme ensures that the more unlikely a mutational event, the higher the associated score. A third approach is to also evaluate the functional impact of mutations when defining the scores. However, for non-coding regions, we often lack the functional understanding to interpret and predict the functional impact. We therefore illustrate this approach using phyloP, a position-specific score of evolutionary conservation (Pollard et al., 2010), as a proxy score for functional impact (Material and methods: Scoring schemes).

The three proposed scoring schemes assign an observed score value of zero (number of mutations and conservation scores), or a value close to zero (log-likelihoods), to positions with no mutations, and a positive score to positions with mutations (Figure 2E). The assigned scores for mutated positions depend on the mutation type and the scoring scheme. Also, for each position, all possible score values and the associated predicted probabilities are integrated in the calculation of the background score distribution.

Driver detection

Although ncdDetect is designed for the analysis of non-coding elements, it can also be applied on protein-coding genes. We initially evaluate the performance of different versions of ncdDetect null models and different scoring schemes on protein-coding genes. We then use it to detect driver candidates among promoter elements, splice sites, 5' UTRs and 3' UTRs (Table 1, Material and methods: Candidate elements). While all the analyses presented here are pan cancer, individual cancer types can be analyzed separately.

Table 1

Element type	Number of elements	Median element length (bps)	Percentage of genome covered
protein-coding genes	20,153	1296	1.19
promoter elements	20,052	848	0.69
splice sites	18,682	30	0.03
3’ UTRs	19,346	1007	1.06
5’ UTRs	19,078	259	0.25

ncdDetect significance evaluation

With ncdDetect, significance evaluation of the observed mutations in a given genomic region of interest is performed (Figure 3). ncdDetect uses a two-step algorithm in which sample-specific calculations are followed by calculations across all samples in the dataset. The output is a p-value indicating if the region of interest is under recurrent positive selection across the sample set (Figure 3A).

Figure 3 with 1 supplement see all

Download asset Open asset

The test statistic used in the significance evaluation performed by ncdDetect is the observed score. This value is defined as the sum of sample- and position-specific observed scores across the specific element that is being tested. For a given sample and a given position, the observed score will depend on the chosen scoring scheme. For instance, the scoring scheme using the number of mutations will always give a score of 1 to a mutated position, and a score of 0 to an unmutated position. The scoring scheme using phyloP will give a score corresponding to the position-specific phyloP value to a mutated position, and a score of 0 to an unmutated position. Thus, the observed score for a specific element will depend on the chosen scoring scheme.

The observed score is significance evaluated in the background score distribution. Again, the shape of this distribution will depend on the chosen scoring scheme: All possible sample- and position-specific scores for the chosen scoring scheme are combined with the sample- and position-specific mutational probabilities to form the background score-distribution.

The algorithm to obtain the background score-distribution works as follows: for a specific sample i, the genomic region of interest is annotated with position-specific probabilities of mutation and scores (Figure 3B–C). As noted above, the observed score of sample i is defined as the sum of observed scores across all positions in the region (Figure 3D). In the current implementation, only the two highest scoring mutations are considered for each sample. The position-specific mutational probabilities and scores are aggregated using mathematical convolution; a method to efficiently calculate the exact probability of observing a given sample-level score by summing up the probabilities of all possible combinations of positional outcomes that could lead to it (Grinstead and Snell, 1997). The use of convolution is inspired by previously published protein-coding driver detection methods (Lawrence et al., 2013; Dees et al., 2012). These calculations lead to the sample-specific background score-distribution (Figure 3E). By repeating this process, background score-distributions are found for each individual sample (Figure 3F). These distributions are aggregated, again using convolution, to yield the overall background score-distribution across samples. The individual sample-specific observed scores are summed to give the overall observed score, which is significance evaluated in the overall background score-distribution. (Figure 3G) (more details are given in Appendix, section 3).

Protein-coding driver detection and model selection

To build a robust background null-model and evaluate the performance of ncdDetect, we apply it to protein-coding genes (Figure 4). While we lack a well-established true-positive driver set for the non-coding part of the genome, the COSMIC Cancer Gene Census (Cosmic, 2016) provides that for the protein-coding genes. As a performance measure, we therefore use the fraction of COSMIC genes recalled among the ncdDetect candidate sets.

Figure 4 with 2 supplements see all

Download asset Open asset

Figure 4—source data 1

P-values obtained on protein-coding genes for each of the five models considered.

The p-values are obtained using conservation scores. This data set relates to Figure 4A.

https://doi.org/10.7554/eLife.21778.012

Download elife-21778-fig4-data1-v3.txt

Figure 4—source data 2

COSMIC Gene Census recall data.

The fraction of recalled COSMIC genes in the top ncdDetect candidates, for the three different scoring schemes. This data set relates to Figure 4C.

https://doi.org/10.7554/eLife.21778.013

Download elife-21778-fig4-data2-v3.txt

Figure 4—source data 3

The background score distribution for the protein-coding gene TP53 obtained with conservation scores.

This data set relates to Figure 4D.

https://doi.org/10.7554/eLife.21778.014

Download elife-21778-fig4-data3-v3.txt

Figure 4—source data 4

The background score distribution for the protein-coding gene PIK3CA obtained with conservation scores.

This data set relates to Figure 4E.

https://doi.org/10.7554/eLife.21778.015

Download elife-21778-fig4-data4-v3.txt

Figure 4—source data 5

The background score distribution for the protein-coding gene SLFN11 obtained with conservation scores.

This data set relates to Figure 4F.

https://doi.org/10.7554/eLife.21778.016

Download elife-21778-fig4-data5-v3.txt

With ncdDetect, multiple hypothesis tests are performed. For example, protein-coding driver detection requires significance evaluation of 19,256 genes. In order to evaluate all these tests simultaneously, QQ-plots are used to assess the distribution of the p-values and the number of true hypotheses (Schweder and Spjotvoll, 1982). In these plots, the observed p-values are plotted against the expected (uniform) p-values of the null distribution. P-values, which follow the expected uniform distribution, will thus fall on the identity line, while smaller p-values will deviate from this line. Per construction, 90% of the expected values lie in the interval [1,10⁻¹], 99% lie in the interval [1,10⁻²], etc. (Figure 4A).

Model selection

The final model underlying ncdDetect is determined through a forward model-selection procedure. In each step, position- and sample-specific probabilities are predicted for the protein-coding genes, which are then evaluated with ncdDetect (Figure 4A, Figure 4—figure supplement 1). The basic model includes the genomic annotations sample id, replication timing and trinucleotides as these are all known to correlate with mutation rate. The resulting p-values appeared slightly inflated. To increase robustness of the predicted mutation-probabilities, we defined model 1a by adding the variable local mutation rate to the basic model. This addition resulted in less inflated p-values. However, the p-values were below the identity line in the QQ-plot for more than 99% of the analyzed genes, indicating that the predicted probabilities of mutation were too large. As we found the somatic mutation rate is elevated in intergenic regions compared to other functional elements, we defined model 1b by adding the variable genomic segment to the basic model (Figure 1—figure supplement 1). This had the desired effect of decreasing the final p-values, although leading to severe inflation. We defined model 1c that extended the basic model with the local mutation rate, genomic segment, and tissue-specific gene expression level. This lowered the p-values, although a small amount of inflation was still observed.

Since we do not know all the relevant genomic annotations that correlate with mutation rate for all of our samples, it is unavoidable that we observe a difference between the actual and predicted mutation rate (Appendix 1—figure 1). The effect of this difference will be accumulated along elements, and even small biases in the predicted versus observed mutation rate may become significant if elements are sufficiently long (Appendix 1—figure 2). The difference between the predicted and observed mutation rate will cause overdispersion of the mutation rate. In the final model, we thus correct for overdispersion by adjusting the sample- and position-specific probabilities of mutation with an overdispersion-based rate adjustment (Materials and methods: An overdispersion-based mutation rate adjustment, Appendix section 1). The resulting p-values follow the expected uniform distribution, with less extreme p-values for the top-ranked genes than for the previous models.

Recall of known protein-coding drivers

The p-values obtained with ncdDetect are corrected for multiple testing using a false discovery rate of 10% (Benjamini and Hochberg, 1995). The resulting ncdDetect candidate protein-coding drivers are compared to the COSMIC Gene Census list for each of the three proposed scoring schemes. We call 64 protein-coding genes significant using the conservation scores of which 15 (23%) are in COSMIC. In contrast, we call 109 protein-genes significant with log-likelihoods of which 19 (17%) are in COSMIC, and 52 with the number of mutations of which 14 (27%) are in COSMIC (Figure 4B, Supplementary files 1–3). The mean number of mutations per bp is on average eight times higher for the COSMIC genes detected by ncdDetect compared to the undetected COSMIC genes (Figure 4—figure supplement 2). The three proposed scoring schemes have similar recall graphs, although the use of conservation scores appears the most sensitive as it generally recalls the highest fraction of COSMIC genes (Figure 4C). For example, in the top-15 protein-coding genes called by ncdDetect with conservation scores, nine are COSMIC genes. This number is seven for the number of mutations, and seven for log-likelihoods (Figure 4C, Figure 5—figure supplement 1A). The use of log-likelihoods results in the highest number of elements called significant across most element types (Figure 5—figure supplement 2).

As the mutational process is stochastic, it varies which drivers are involved in cancer development, both within and between cancer types. COSMIC genes are identified from analyses across many individual cancer types and a large fraction are likely not drivers in the particular set of cancer samples analyzed here. Furthermore, there might exist true protein-coding cancer drivers not yet included in COSMIC. Out of the three proposed scoring schemes, the conservation scores appear to have the highest sensitivity. It is more conservative than the log-likelihoods, as it finds fewer significant protein-coding genes. Furthermore, it is compelling that the use of this scoring scheme incorporates a measure of functional mutational impact in the driver significance evaluation. In light of these considerations, we focus on the results obtained with conservation scores in the following, and include the remaining two scoring schemes for comparison.

To give an impression of how the calculated background score-distributions behave in practice, we highlight a few examples (Figure 4D–F). The top-two protein-coding genes called by ncdDetect are TP53 (spanning 1378 bps) and PIK3CA (spanning 3207 bps), which are both well-known cancer driver genes. An example of a protein-coding gene not called significant by ncdDetect is SLFN11 (spanning 2706 bps). The smoothness of the overall score-distribution is related to the length of the gene.

The performance of ncdDetect on protein-coding genes was benchmarked against a recent non-coding cancer driver detection method, ExInAtor (Lanzós et al., 2017) (Appendix, section 2). In general, ExInAtor predicts much fewer candidates than ncdDetect, and thus has a lower false-positive rate. However, ncdDetect performs better at ranking genes compared to ExInAtor (Appendix 1—figure 3).

Non-coding driver detection

Although the functional impact of non-coding mutations in cancer is not yet fully understood, it is widely believed that they may play an important role in cancer development (Diederichs et al., 2016). Here, we apply ncdDetect to gene-associated non-coding elements of various types (promoter elements, splice sites, 3’ UTRs and 5’ UTRs) to evaluate their cancer driver potential (Figure 5, Figure 5—figure supplement 1, Supplementary files 1–3, Material and methods: Candidate elements).

Recall and function of previously described non-coding drivers

Promoter mutations might dysregulate gene expression in cancer. In particular, such mutations might affect the expression levels of tumor suppressor genes or oncogenes (Diederichs et al., 2016). The average mutation rate in the analyzed promoter elements is 7.0 mutations per mega base (Mb) per sample (Figure 1—figure supplement 1). Of the investigated non-coding element types, the promoter elements have the most significant calls in the ncdDetect analyses. Approximately 1% of the evaluated promoter elements have a more significant p-value than expected under the null (Figure 5A). We find a total of 160 significant (q < 0.10) promoter elements. Within these, the observed mutation rate is 31.3 mutations per Mb per sample, which is a fourfold increase of the mutation rate among all promoter elements. Of the promoter elements, the TERT promoter is ranked most significant (q = 2.4 × 10⁻⁶⁹, ncdDetect). The promoter of the TERT gene is known to play an important role in telomerase expression, and cancers with TERT promoter mutations have been shown to exhibit an elevated expression of the TERT gene. This increased expression might ensure telomere maintenance, believed to enable cancer cells to divide (Heidenreich et al., 2014). Two other identified promoter elements are WDR74 (q = 4.1 × 10⁻⁴, ncdDetect) and PLEKHS1 (q = 4.3×10⁻⁵, ncdDetect). Mutations in the promoter region of WDR74 have been associated with increased gene expression and are thought to have functional relevance for tumorigenesis (Khurana et al., 2013). Mutations in the PLEKHS1 promoter are also previously found to be significant in non-coding driver screens (Weinhold et al., 2014; Melton et al., 2015). We note that out of the 863 whole genomes analyzed in Weinhold et al. (2014), 356 are sequenced by TCGA. These samples appear to be a subset of the 505 samples analyzed in the present work, and the data sets are thus not completely independent. In total, 29 of the 160 significant promoter elements called with ncdDetect are previously found to be significant in non-coding cancer driver studies (Weinhold et al., 2014) (Appendix 1—figure 4). As a benchmark of the performance of ncdDetect on regulatory non-coding regions, we compared our results on promoter elements to those obtained with another non-coding cancer driver detection method, LARVA (Lochovsky et al., 2015). The ncdDetect promoter candidates that are not detected by LARVA include the previously described WDR74, PLEKHS1 and promoters of COSMIC genes (Appendix, section 2).

Figure 5 with 3 supplements see all

Download asset Open asset

Figure 5—source data 1

P-values obtained on promoters and splice sites using conservation scores.

This data set relates to Figure 5A–B. P-values obtained using log-likelihoods and number of mutations as scores are in Supplementary files 2–3.

https://doi.org/10.7554/eLife.21778.022

Download elife-21778-fig5-data1-v3.txt

Another class of non-coding mutations are splice site mutations. They might disrupt the splicing code and have been linked to cancer development (Srebrow and Kornblihtt, 2006). The destruction of a splice site will typically introduce stop codons or frameshifts and ruin the function of the translated protein. The splice site mutation rate is 5.1 mutations per Mb per sample. Three splice sites are found significant in this analysis (Figure 5B). As many as 90% of the splice sites have zero observed mutations across the 505 samples. By construction, this means that the resulting p-values are 1, and the p-value distribution is thus not uniform. More samples would increase the detection power in these cases. Interestingly, in the top-ten ranking splice sites, we see a highly significant enrichment of splice sites associated to COSMIC genes (p=6.1 × 10⁻⁹, Fisher’s exact test). Within the three significant splice site elements, the mutation rate is 130.0 mutations per Mb per sample, corresponding to a 25-fold increase of the mutation rate among all splice site elements. Splicing mutations in TP53 are previously described in cancer studies (Lee et al., 2010; Varley et al., 2001). Here, we observe that 12 samples from six different cancer types are mutated in the 52 bps that make up the splice sites of TP53. These splicing mutations are highly significant (q = 1.2 × 10⁻²¹, ncdDetect), and might lead to inactivation of the tumor suppressor TP53 gene (Eicheler et al., 2002).

Finally, we investigate somatic mutations in the 3’ and 5’ UTRs (Figure 5—figure supplement 1B–C), which regulate mRNA stability and translation. UTR mutations might disrupt binding sites for miRNAs and RNA-binding proteins and thereby affect post-transcriptional regulation. They might also alter the structural conformations of the UTRs, which have previously been associated with cancer (Diederichs et al., 2016). The average number of mutations is 6.4 per Mb per sample for 5’ UTRs and 7.1 per Mb per sample for 3’ UTRs. We find a total of 16 significant 3’ UTRs and 86 significant 5’ UTRs (Figure 5—figure supplement 1B–C). Within the significant 5’ UTRs, the mutation rate is 36.3 mutations per Mb per sample, a sixfold increase compared to all 5’ UTRs. For the significant 3’ UTRs, the mutation rate is 20.6 mutations per Mb per sample, which is a threefold increase of the average 3’ UTR mutation rate. Two of the called 3’ UTRs (DRD5 and PCMTD1) have previously been detected in cancer driver studies (Weinhold et al., 2014). This is also the case for 12 of the 86 called 5’ UTRs, one of them SDHD (q = 2.3 × 10⁻³, ncdDetect) (Appendix 1—figure 4). A recent study identified the promoter region as well as the 5’ UTR of SDHD to be potential cancer drivers in melanoma. In particular, promoter mutations of SDHD were shown to be associated with reduced gene expression and poor survival prognosis (Weinhold et al., 2014). In the present data set, we observe six mutated melanoma samples in the 5’ UTR of SDHD, which covers 135 bps.

Case studies

The absence of a true-positive driver set for the non-coding part of the genome means that we must find alternative ways to validate the driver potential of the candidates found by ncdDetect. We thus seek to support the significance of the candidate elements and further characterize them with evidence from two additional data sources.

A first approach is to analyze the effects of mutations on gene expression. To be able to look up individual driver candidates, we gather expression values for each of the 505 considered whole genome samples (Fredriksson et al., 2014) and perform a Wilcoxon rank sum test for top-ranking candidates of each element type. To further support the findings, we obtain mutation calls and expression values from the larger set of TCGA exomes (Weinstein et al., 2013; Supplementary file 7) and likewise perform a rank sum test on these data (Supplementary file 5, Material and methods: Expression analysis). Reassuringly, we recover known differences in TERT gene expression levels between samples mutated and not mutated in the promoter region of the gene for the 505 whole genome samples (q = 1.4 × 10⁻³, Fisher’s method) (Vinagre et al., 2013). Similarly for the TCGA exome samples, splice-site mutations in TP53, which are known to drive cancer (Varley et al., 2001; Lee et al., 2010), correlated with differences in gene expression levels (q = 2.3 × 10⁻², Fisher’s method).

Another approach we take to validate the ncdDetect candidates is to look at correlation between mutation status and survival data for both the 505 whole genome samples and the TCGA exomes. For this purpose, we download clinical data from the TCGA data portal (TCGA Data Portal, 2016; Supplementary file 7). For a particular candidate driver, we test the significance of the difference in survival between mutated and non-mutated samples using a one-sided Log-rank test on the Kaplan-Meier estimated survival curves (Supplementary file 6, Material and methods: Survival analysis). This recovers some known prognostic markers, such as TP53 where splice site mutations correlate with a significant decrease in survival (q = 1.0 × 10⁻¹, Fisher’s method) (Yang et al., 2013). In the analysis of the 505 whole genome samples, we furthermore observe a significant decrease in survival associated with HLA-DRB1 promoter mutations (q = 2.0 × 10⁻², Fisher’s method). Although this finding is potentially interesting, further investigation of this candidate is beyond the scope of this paper, as genotyping in HLA regions is challenging due to the highly polymorphic nature of these genes (Ehrenberg et al., 2014).

In the following, we study a number of the top-ranking non-coding ncdDetect driver candidates in detail. For each of them, we further evaluate their driver potential by including results from expression analysis and survival analysis.

SMUG1 mutations and a uracil-DNA glycosylase deficiency mutational signature

We observe 19 mutations in the 997 bp-long SMUG1 promoter-region, which is approximately seven times more than expected under the null model (q = 1.1 × 10⁻⁶, ncdDetect) (Figure 6A). The mutations are distributed among 14 samples from three different cancer types (one breast cancer sample, two colorectal cancer samples and eleven melanoma samples). As 15 out of 16 of the melanoma mutations are of type C→T in a CC context (or its reverse complement), they are consistent with the mutational signature of ultraviolet (UV) light (Alexandrov and Stratton, 2014). They may be a result of a mutational mechanism (Sabarinathan et al., 2016); however, they may also affect SMUG1 function.

Figure 6 with 1 supplement see all

Download asset Open asset

Figure 6—source data 1

The defined uracil-DNA glycosylase deficiency signature statistic for each sample of the cancer types GBM, BLCA, CRC, BRCA, LUAD, SKCM and UCEC.

For each sample, the SMUG1 mutation status is indicated. This data set relates to Figure 6C.

https://doi.org/10.7554/eLife.21778.030

Download elife-21778-fig6-data1-v3.txt

Figure 6—source data 2

The defined uracil-DNA glycosylase deficiency signature statistic, as well as SMUG1 gene expression, UNG gene expression, and SMUG1xUNG gene expression for TCGA exome samples.

Expression values are RSEM values. This data set relates to Figure 6D as well as Figure 6—figure supplement 1.

https://doi.org/10.7554/eLife.21778.031

Download elife-21778-fig6-data2-v3.txt

SMUG1 is involved in base excision repair (BER). Together with UNG, it acts in BER as an uracil-DNA glycosylase, that is, an enzyme that removes uracil from DNA (Visnes et al., 2009). Uracil in DNA arises from spontaneous deamination of non-methylated cytosine, which causes the occurrence of U:G mismatches. If unrepaired, they give rise to G→A transition mutations. Mouse cell line experiments have shown additive effects of SMUG1 and UNG inactivation on G→A mutation rates (An et al., 2005). Furthermore, UNG and SMUG1 expression has recently been found to correlate negatively with genomic uracil levels in B cell lymphomas (Pettersen et al., 2015). We therefore hypothesize that SMUG1 mutations may affect the rate of G→A (and C→T) mutations. To investigate this further, we define the uracil-DNA glycosylase deficiency signature as the proportion of G→A (including C→T) mutations outside CpG sites (Figure 6B). For melanoma samples, we further deduct C→T mutations in a CC context, as potentially induced by UV light (Material and methods: Enrichment of G→A mutations in SMUG1 mutated samples).

There is a tendency for an increased value of the uracil-DNA glycosylase deficiency signature statistic in SMUG1 mutated melanoma samples (p=8.2 × 10⁻², one-sided Wilcoxon rank sum test), and a significantly increased value of this statistic for the one SMUG1 mutated uterus cancer sample (p=2.1 × 10⁻², one-sided Wilcoxon rank sum test), compared to samples not harbouring a SMUG1 mutation (Figure 6C). This is also the case when restricting the analysis to only include coding and splice site mutations (melanoma: p=5.3 × 10⁻², uterus cancer: p=2.1 × 10⁻², one-sided Wilcoxon rank sum tests). These findings indicate that SMUG1 mutations might perturb the uracil-DNA glycosylase function.

We further hypothesize that SMUG1 and UNG expression may correlate with the uracil-DNA glycosylase deficiency signature statistic. With the expression data available for the 505 analyzed TCGA samples, we are unable to detect a significant correlation between gene expression of SMUG1 or UNG and the signature statistic (SMUG1: p=9.4 × 10⁻¹, UNG: p=1.0 × 10⁻¹, Fisher’s method). To further investigate expression correlations, we look at the larger data set of TCGA exomes. From these data, the uracil-DNA glycosylase deficiency signature statistic is seen to be negatively correlated with SMUG1 gene expression (p=3.8 × 10⁻², Fisher’s method), and with UNG gene expression (p=5.1 × 10⁻⁴, Fisher’s method). As SMUG1 and UNG are thought to have complementary roles in BER (Pettersen et al., 2007), we also investigate the correlation between the signature statistic and the product of SMUG1 and UNG gene expression, which is negative and also significant (p=2.4 × 10⁻³, Fisher’s method) (Figure 6D, Figure 6—figure supplement 1).

Finally, we investigate whether survival correlates with SMUG1 mutation status. With the present data set, we are not able to detect such a pattern (p=8.2 × 10⁻¹, Fisher’s method).

The observed correlations combined with the existing literature (Pettersen et al., 2015; An et al., 2005) suggest that mutations that functionally impact SMUG1 and UNG may cause a mutational phenotype as captured by the defined deficiency signature. However, further validation must await availability of larger sets of cancer genomes.

Promoter and UTR candidates where mutations associate with decreased survival

The 5’ UTR of CD1A spans 533 bp. In the region we observe 11 mutations from 10 different samples, distributed across eight different cancer types. This corresponds to five times the amount of mutations expected under the null model (q = 1.1 × 10⁻², ncdDetect). For TCGA exome melanoma samples, we observe a highly significant decrease in survival associated with mutations in the region (p=1.2 × 10⁻⁹, Log-rank test) (Figure 7A), which is top-ranked among the non-coding regions tested (Supplementary file 6). CD1 proteins present antigens to T cells and are involved in eliciting adaptive immune responses. They are distantly related to HLA (MHC) proteins and similarly bind T cell receptors; however, they display glycoproteins and small molecules instead of peptides (Van Rhijn et al., 2015). Intriguingly, CD1A is generally lowly expressed in healthy tissue with high expression particularly in skin (GTEx Consortium, 2013), where it is found in the antigen-presenting Langerhans cells. CD1A has previously been implicated with cancer development, with expression and positive correlation to survival reported for some cancer types (Coventry and Heinzel, 2004). Although we cannot functionally interpret the observed TCGA melanoma mutations, the strong association with survival suggests potential clinical relevance, not-the-least given the success of immunotherapy in melanoma (Drake et al., 2014).

Figure 7

Download asset Open asset

Figure 7—source data 1

STK11 mutation status and STK11 gene expression (RSEM) for 469 LUAD TCGA exome samples.

This data set relates to Figure 7D.

https://doi.org/10.7554/eLife.21778.034

Download elife-21778-fig7-data1-v3.txt

A total of 22 mutations are observed in the 1976 bp-long promoter region of PRSS3. This is four times more mutations than expected under the null model, and they occur in 13 samples from seven different cancer types (q = 1.1 × 10⁻², ncdDetect). Previous studies have established the role of PRSS3 in the progression of pancreatic and ovarian cancer (Jiang et al., 2010; Ma et al., 2015). Not only the promoter region of this gene comes out significant in the driver detection screen; this is also the case for its 3’ UTR (q = 1.4 × 10⁻⁵, ncdDetect) as well as its protein-coding gene (q = 6.8 × 10⁻²⁰, ncdDetect). Based on the TCGA exome set, we observe a significantly decreased survival for head and neck cancer (HNSC) samples mutated in the promoter region of PRSS3 (p=1.8 × 10⁻³, Log-rank test) (Figure 7B) as well as in the PRSS3 coding gene (p=1.2 × 10⁻², Log-rank test). We also observe a tendency for decreased survival among melanoma samples with 3’ UTR mutations (p=8.3 × 10⁻², Log-rank test).

The 3’ UTR of SEC14L1 spans 3052 bps and contains 31 mutations from 27 samples distributed across 10 different cancer types. This is approximately four times the amount of expected mutations (q = 8.9 × 10⁻⁵, ncdDetect), and the majority (58%) of these are found in breast- and colorectal cancer. Although little is known about SEC14L1 in cancer, one study hypothesized that altered expression of the gene could contribute to breast tumorigenesis (Kalikin et al., 2001). Another study found SEC14L1 to be overexpressed in prostate cancer (Burdelski et al., 2015). For TCGA exome HNSC samples, we find a significant decrease in survival associated with mutations in the 3’ UTR region of the gene (p=5.8 × 10⁻⁵, Log-rank test).

STK11 splice sites mutations and their expression correlation

The combined splice site region of STK11 covers 36 bps and is mutated in four lung adenocarcinoma (LUAD) cancer samples, which is approximately 56 times more mutations than expected under the null model (Figure 7C). ncdDetect ranks the splice site region of STK11 second among all splice sites (q = 2.2 × 10⁻³, ncdDetect). STK11 is a known COSMIC tumour suppressor gene, which has been shown to be involved in lung and cervical cancers (Gill et al., 2011), and very recently, splice site mutations of the gene were described in relation to cancer (Mularoni et al., 2016; Wei et al., 2016). From the 505 whole genomes analyzed here, we are unable to associate the splice site mutations of STK11 with a changed level of gene expression (p=4.4 × 10⁻¹, Fisher’s method). Looking at the larger set of TCGA exomes, however, we detect a significantly lower expression level for LUAD samples mutated in the splice site region of STK11, compared to LUAD samples without such mutations (p=1.3 × 10⁻³, two-sided Wilcoxon rank sum test) (Figure 7D). We further observe a marginally significant decrease in survival associated with STK11 splice site mutations for LUAD TCGA exome samples (p=6.5 × 10⁻², Log-rank test) (Figure 7E).

Non-coding somatic mutations play part in tumour initiation and progression. With the advent of whole genome sequencing, the systematic screening of such mutations is possible. We have developed the method ncdDetect with the goal of detecting non-coding cancer driver elements and thereby gain an understanding of the underlying mechanisms of tumorigenesis. With ncdDetect, we model the heterogeneous neutral background mutation-rate, taking genomic annotations known to correlate with the mutation rate into account. We consider the mutational burden and functional impact to reveal signs of recurrent positive selection across cancer genomes.

The position- and sample-specific approach behind ncdDetect sets the stage for a number of distinct types of analyses. The analysis of one contiguous region is a straight-forward application of ncdDetect, as is the combined analysis of disjoint regions, potentially with vastly different background mutation-rates. The flexible setup conveniently ensures that no constraints are necessary when defining the size and location of a particular region of interest. Furthermore, the method can be used to evaluate more complex functional hypotheses than those presented here. For instance, different sets of regions in different samples can be jointly evaluated, and sample- or tissue-specific scoring schemes can be applied directly.

Not all the significant non-coding elements can be regarded as true cancer drivers. ncdDetect might falsely identify driver elements (‘false positives’) for both technical and biological reasons. The false positives stemming from predicting too low a mutation rate in certain regions can be reduced by adding relevant genomic annotations as explanatory variables to the null model. In general, failing to include an explanatory variable that explains variation in the mutation rate will cause too little variation in the predicted mutational probabilities. We handle such overdispersion of the mutation rate by adjusting each sample- and position-specific probability of mutation with an overdispersion-based correction factor. This improves the model fit, although some inflation appears to remain for long elements. We acknowledge that the false-positive rate among long genes is not properly controlled and likely higher than the applied FDR threshold. We therefore continue to strive to improve our model of the site-specific mutational process, which will also improve power. The observed mutation rate also varies for technical reasons. For instance, the power to call mutations and the rate with which mutations are missed, will vary with genomic complexity, including repeats, pseudogenes, etc. The predicted mutational probabilities can thus be further improved by including genomic annotations that correlate with the rate of either false-negative or false-positive mutation calls as explanatory variables.

Likewise, ncdDetect might miss true driver elements (‘false negatives’). Especially, with the size of the current whole cancer-genome data sets, we lack statistical power to detect infrequently mutated driver elements, or driver elements that may operate within an individual cancer type. This issue will be remedied as larger sets of sequenced whole genomes become available. In the near future, more than 2500 whole genomes will be available from the Pan-Cancer Analysis of Whole Genomes (PCAWG) project. However, it is becoming evident that some instances of detected potential driver regions may be explained by local mutational mechanisms rather than recurrent selection (Sabarinathan et al., 2016). This emphasizes the importance of critical scrutinization and eventually independent validation of driver candidates emerging from ncdDetect.

With ncdDetect, we screen for non-coding cancer drivers and highlight cases of special interest. To gain further evidence for the identified candidates, we correlate the presence of non-coding mutations with gene expression as well as patient survival: We find that mutations in the promoter and in the coding region of a gene in the Base Excision Repair pathway, SMUG1, correlate with an increase of C→T mutations. We hypothesize that SMUG1 mutations might perturb uracil-DNA glycosylase function and cause a specific mutational phenotype. Although our study is limited to correlative observations between the expected mutational signature for uracil-DNA glycosylase deficiency and mutational presence as well as gene expression, perturbation experiments in cell lines support the hypothesis (An et al., 2005). We also identify non-coding regulatory regions that associate with patient survival, including the potential clinically important 5’ UTR of CD1A, the promoter and 3’UTR of PRSS3, and the 3’UTR of SEC14L1. Finally, we identify lung cancer mutations in the splice sites of STK11 as potential driver events. By extending the analysis to the larger TCGA data set, we show that these correlate significantly with expression. The patients also show a strong tendency for poorer survival.

In this work, we have addressed the challenges associated with distinguishing driver and passenger non-coding mutations. We evaluated three different scoring schemes and found that a conservation-based scheme performed better than mutation counts and log-likelihoods in our setting. For selected candidate cases, we found a significant effect on expression levels and a significant decrease in survival for mutated samples. The combined analyses of mutational impact on expression and survival across cancer types allowed us aggregate evidence and gain power. The screen identified candidates of potential clinical relevance. However, sample sizes remain small and further studies in large independent cohorts are necessary to establish their potential as prognostic biomarkers or therapeutic targets.

As we continue to gain larger cancer genomics data sets for driver screens, accurate modelling of the mutational heterogeneity will become increasingly important. This will help control the false-positive rate as the power of the data increases. As our understanding of the general differences in mutational mechanisms between cancer types improves further, this knowledge should be incorporated in ncdDetect.

1. Book

   1. Agresti A

   (2013)

   Categorical Data Analysis

   John Wiley & Sons.

   • Google Scholar
2. 1. Alexandrov LB
   2. Nik-Zainal S
   3. Wedge DC
   4. Aparicio SA
   5. Behjati S
   6. Biankin AV
   7. Bignell GR
   8. Bolli N
   9. Borg A
   10. Børresen-Dale AL
   11. Boyault S
   12. Burkhardt B
   13. Butler AP
   14. Caldas C
   15. Davies HR
   16. Desmedt C
   17. Eils R
   18. Eyfjörd JE
   19. Foekens JA
   20. Greaves M
   21. Hosoda F
   22. Hutter B
   23. Ilicic T
   24. Imbeaud S
   25. Imielinski M
   26. Imielinsk M
   27. Jäger N
   28. Jones DT
   29. Jones D
   30. Knappskog S
   31. Kool M
   32. Lakhani SR
   33. López-Otín C
   34. Martin S
   35. Munshi NC
   36. Nakamura H
   37. Northcott PA
   38. Pajic M
   39. Papaemmanuil E
   40. Paradiso A
   41. Pearson JV
   42. Puente XS
   43. Raine K
   44. Ramakrishna M
   45. Richardson AL
   46. Richter J
   47. Rosenstiel P
   48. Schlesner M
   49. Schumacher TN
   50. Span PN
   51. Teague JW
   52. Totoki Y
   53. Tutt AN
   54. Valdés-Mas R
   55. van Buuren MM
   56. van 't Veer L
   57. Vincent-Salomon A
   58. Waddell N
   59. Yates LR
   60. Zucman-Rossi J
   61. Futreal PA
   62. McDermott U
   63. Lichter P
   64. Meyerson M
   65. Grimmond SM
   66. Siebert R
   67. Campo E
   68. Shibata T
   69. Pfister SM
   70. Campbell PJ
   71. Stratton MR
   72. Serena N-Z
   73. Samuel AJ
   74. Sam B
   75. Australian Pancreatic Cancer Genome Initiative
   76. ICGC Breast Cancer Consortium
   77. ICGC MMML-Seq Consortium
   78. ICGC PedBrain

   (2013) Signatures of mutational processes in human Cancer

   Nature 500:415–421.

   https://doi.org/10.1038/nature12477

   • PubMed
   • Google Scholar
3. 1. Alexandrov LB
   2. Stratton MR

   (2014) Mutational signatures: the patterns of somatic mutations hidden in Cancer genomes

   Current Opinion in Genetics & Development 24:52–60.

   https://doi.org/10.1016/j.gde.2013.11.014

   • Google Scholar
4. 1. An Q
   2. Robins P
   3. Lindahl T
   4. Barnes DE

   (2005) C --> T mutagenesis and gamma-radiation sensitivity due to deficiency in the Smug1 and ung DNA glycosylases

   The EMBO Journal 24:2205–2213.

   https://doi.org/10.1038/sj.emboj.7600689

   • PubMed
   • Google Scholar
5. Website

   1. Bates D
   2. Maechler M

   (2015) MatrixModels: modelling with sparse and dense matrices

   http://CRAN.R-project.org/package=MatrixModels
6. 1. Begg CB
   2. Gray R

   (1984) Calculation of Polychotomous Logistic regression parameters using individualized regressions

   Biometrika 71:11–18.

   https://doi.org/10.2307/2336391

   • Google Scholar
7. 1. Benjamini Y
   2. Hochberg Y

   (1995) Controlling the false discovery rate: a practical and powerful approach to multiple testing

   Journal of the Royal Statistical Society. Series B Statistical Methodology 57:289–300.

   https://doi.org/10.2307/2346101

   • Google Scholar
8. 1. Berens EB
   2. Sharif GM
   3. Schmidt MO
   4. Yan G
   5. Shuptrine CW
   6. Weiner LM
   7. Glasgow E
   8. Riegel AT
   9. Wellstein A

   (2017) Keratin-associated protein 5-5 controls cytoskeletal function and Cancer cell vascular invasion

   Oncogene 36:593–605.

   https://doi.org/10.1038/onc.2016.234

   • PubMed
   • Google Scholar
9. Preprint

   1. Bertl J
   2. Guo Q
   3. Rasmussen MJ
   4. Besenbacher S
   5. Nielsen MM
   6. Hornshoej H
   7. Pedersen JS
   8. Hobolth A

   (2017) A site specific model and analysis of the neutral somatic mutation rate in Whole-Genome Cancer data

   bioRxiv.

   http://biorxiv.org/content/early/2017/03/31/122879

   • Google Scholar
10. 1. Burdelski C
    2. Barreau Y
    3. Simon R
    4. Hube-Magg C
    5. Minner S
    6. Koop C
    7. Graefen M
    8. Heinzer H
    9. Sauter G
    10. Wittmer C
    11. Steurer S
    12. Adam M
    13. Huland H
    14. Schlomm T
    15. Tsourlakis MC
    16. Quaas A
    17. Ysé B
    18. Ronald S
    19. Claudia H-M
    20. Sarah M
    21. Christina K

    (2015) Saccharomyces cerevisiae-like 1 overexpression is frequent in prostate Cancer and has markedly different effects in Ets-related gene fusion-positive and fusion-negative cancers

    Human Pathology 46:514–523.

    https://doi.org/10.1016/j.humpath.2014.06.006

    • PubMed
    • Google Scholar
11. 1. Chen CL
    2. Rappailles A
    3. Duquenne L
    4. Huvet M
    5. Guilbaud G
    6. Farinelli L
    7. Audit B
    8. d'Aubenton-Carafa Y
    9. Arneodo A
    10. Hyrien O
    11. Thermes C

    (2010) Impact of replication timing on non-CpG and CpG substitution rates in mammalian genomes

    Genome Research 20:447–457.

    https://doi.org/10.1101/gr.098947.109

    • PubMed
    • Google Scholar
12. 1. Cornish-Bowden A

    (1985) Nomenclature for incompletely specified bases in nucleic acid sequences: recommendations 1984

    Nucleic Acids Research 13:3021–3030.

    https://doi.org/10.1093/nar/13.9.3021

    • PubMed
    • Google Scholar
13. Website

    1. Cosmic

    (2016) COSMIC: Cancer Gene Census

    Accessed May 2016.

    http://cancer.sanger.ac.uk/census/
14. 1. Coventry B
    2. Heinzel S

    (2004) CD1a in human cancers: a new role for an old molecule

    Trends in Immunology 25:242–248.

    https://doi.org/10.1016/j.it.2004.03.002

    • PubMed
    • Google Scholar
15. 1. Dees ND
    2. Zhang Q
    3. Kandoth C
    4. Wendl MC
    5. Schierding W
    6. Koboldt DC
    7. Mooney TB
    8. Callaway MB
    9. Dooling D
    10. Mardis ER
    11. Wilson RK
    12. Ding L

    (2012) MuSiC: identifying mutational significance in Cancer genomes

    Genome Research 22:1589–1598.

    https://doi.org/10.1101/gr.134635.111

    • PubMed
    • Google Scholar
16. 1. Denisova E
    2. Heidenreich B
    3. Nagore E
    4. Rachakonda PS
    5. Hosen I
    6. Akrap I
    7. Traves V
    8. García-Casado Z
    9. López-Guerrero JA
    10. Requena C
    11. Sanmartin O
    12. Serra-Guillén C
    13. Llombart B
    14. Guillén C
    15. Ferrando J
    16. Gimeno E
    17. Nordheim A
    18. Hemminki K
    19. Kumar R

    (2015) Frequent DPH3 promoter mutations in skin cancers

    Oncotarget 6:35922–35930.

    https://doi.org/10.18632/oncotarget.5771

    • PubMed
    • Google Scholar
17. 1. Diederichs S
    2. Bartsch L
    3. Berkmann JC
    4. Fröse K
    5. Heitmann J
    6. Hoppe C
    7. Iggena D
    8. Jazmati D
    9. Karschnia P
    10. Linsenmeier M
    11. Maulhardt T
    12. Möhrmann L
    13. Morstein J
    14. Paffenholz SV
    15. Röpenack P
    16. Rückert T
    17. Sandig L
    18. Schell M
    19. Steinmann A
    20. Voss G
    21. Wasmuth J
    22. Weinberger ME
    23. Wullenkord R

    (2016) The dark matter of the Cancer genome: aberrations in regulatory elements, untranslated regions, splice sites, non-coding RNA and synonymous mutations

    EMBO Molecular Medicine 8:442–457.

    https://doi.org/10.15252/emmm.201506055

    • PubMed
    • Google Scholar
18. 1. Ding J
    2. McConechy MK
    3. Horlings HM
    4. Ha G
    5. Chun Chan F
    6. Funnell T
    7. Mullaly SC
    8. Reimand J
    9. Bashashati A
    10. Bader GD
    11. Huntsman D
    12. Aparicio S
    13. Condon A
    14. Shah SP

    (2015) Systematic analysis of somatic mutations impacting gene expression in 12 tumour types

    Nature Communications 6:8554.

    https://doi.org/10.1038/ncomms9554

    • PubMed
    • Google Scholar
19. 1. Drake CG
    2. Lipson EJ
    3. Brahmer JR

    (2014) Breathing new life into immunotherapy: review of melanoma, lung and kidney Cancer

    Nature Reviews Clinical Oncology 11:24–37.

    https://doi.org/10.1038/nrclinonc.2013.208

    • PubMed
    • Google Scholar
20. Website

    1. Eddelbuettel D

    (2016) Rcpp: seamless R and C++ integration

    Accessed June 2016.

    http://www.rcpp.org
21. 1. Ehrenberg PK
    2. Geretz A
    3. Baldwin KM
    4. Apps R
    5. Polonis VR
    6. Robb ML
    7. Kim JH
    8. Michael NL
    9. Thomas R

    (2014) High-throughput multiplex HLA genotyping by next-generation sequencing using multi-locus individual tagging

    BMC Genomics 15:864.

    https://doi.org/10.1186/1471-2164-15-864

    • PubMed
    • Google Scholar
22. 1. Eicheler W
    2. Zips D
    3. Dörfler A
    4. Grénman R
    5. Baumann M

    (2002) Splicing mutations in TP53 in human squamous cell carcinoma lines influence immunohistochemical detection

    Journal of Histochemistry & Cytochemistry 50:197–204.

    https://doi.org/10.1177/002215540205000207

    • PubMed
    • Google Scholar
23. 1. ENCODE Project Consortium

    (2012) An integrated encyclopedia of DNA elements in the human genome

    Nature 489:57–74.

    https://doi.org/10.1038/nature11247

    • PubMed
    • Google Scholar
24. 1. Forbes SA
    2. Beare D
    3. Gunasekaran P
    4. Leung K
    5. Bindal N
    6. Boutselakis H
    7. Ding M
    8. Bamford S
    9. Cole C
    10. Ward S
    11. Kok CY
    12. Jia M
    13. De T
    14. Teague JW
    15. Stratton MR
    16. McDermott U
    17. Campbell PJ

    (2015) COSMIC: exploring the world's knowledge of somatic mutations in human cancer

    Nucleic Acids Research 43:D805–D811.

    https://doi.org/10.1093/nar/gku1075

    • PubMed
    • Google Scholar
25. 1. Fredriksson NJ
    2. Ny L
    3. Nilsson JA
    4. Larsson E

    (2014) Systematic analysis of noncoding somatic mutations and gene expression alterations across 14 tumor types

    Nature Genetics 46:1258–1263.

    https://doi.org/10.1038/ng.3141

    • PubMed
    • Google Scholar
26. Website

    1. GENCODE

    (2016) GENCODE - Release history

    Accessed September 2016.

    http://www.gencodegenes.org/releases/
27. 1. Gerstein MB
    2. Kundaje A
    3. Hariharan M
    4. Landt SG
    5. Yan KK
    6. Cheng C
    7. Mu XJ
    8. Khurana E
    9. Rozowsky J
    10. Alexander R
    11. Min R
    12. Alves P
    13. Abyzov A
    14. Addleman N
    15. Bhardwaj N
    16. Boyle AP
    17. Cayting P
    18. Charos A
    19. Chen DZ
    20. Cheng Y
    21. Clarke D
    22. Eastman C
    23. Euskirchen G
    24. Frietze S
    25. Fu Y
    26. Gertz J
    27. Grubert F
    28. Harmanci A
    29. Jain P
    30. Kasowski M
    31. Lacroute P
    32. Leng J
    33. Lian J
    34. Monahan H
    35. O'Geen H
    36. Ouyang Z
    37. Partridge EC
    38. Patacsil D
    39. Pauli F
    40. Raha D
    41. Ramirez L
    42. Reddy TE
    43. Reed B
    44. Shi M
    45. Slifer T
    46. Wang J
    47. Wu L
    48. Yang X
    49. Yip KY
    50. Zilberman-Schapira G
    51. Batzoglou S
    52. Sidow A
    53. Farnham PJ
    54. Myers RM
    55. Weissman SM
    56. Snyder M

    (2012) Architecture of the human regulatory network derived from ENCODE data

    Nature 489:91–100.

    https://doi.org/10.1038/nature11245

    • PubMed
    • Google Scholar
28. 1. Gill RK
    2. Yang SH
    3. Meerzaman D
    4. Mechanic LE
    5. Bowman ED
    6. Jeon HS
    7. Roy Chowdhuri S
    8. Shakoori A
    9. Dracheva T
    10. Hong KM
    11. Fukuoka J
    12. Zhang JH
    13. Harris CC
    14. Jen J

    (2011) Frequent homozygous deletion of the LKB1/STK11 gene in non-small cell lung Cancer

    Oncogene 30:3784–3791.

    https://doi.org/10.1038/onc.2011.98

    • PubMed
    • Google Scholar
29. 1. Gonzalez-Perez A
    2. Lopez-Bigas N

    (2012) Functional impact Bias reveals Cancer drivers

    Nucleic Acids Research 40:e169.

    https://doi.org/10.1093/nar/gks743

    • PubMed
    • Google Scholar
30. Book

    1. Grinstead CM
    2. Snell JL

    (1997)

    Introduction to Probability

    American Mathematical Soc.

    • Google Scholar
31. 1. GTEx Consortium

    (2013) The Genotype-Tissue expression (GTEx) project

    Nature Genetics 45:580–585.

    https://doi.org/10.1038/ng.2653

    • PubMed
    • Google Scholar
32. 1. Heidenreich B
    2. Rachakonda PS
    3. Hemminki K
    4. Kumar R
    5. Barbara H
    6. Sivaramakrishna Rachakonda P
    7. Kari H
    8. Rajiv K

    (2014) TERT promoter mutations in Cancer development

    Current Opinion in Genetics & Development 24:30–37.

    https://doi.org/10.1016/j.gde.2013.11.005

    • PubMed
    • Google Scholar
33. 1. Hockla A
    2. Miller E
    3. Salameh MA
    4. Copland JA
    5. Radisky DC
    6. Radisky ES

    (2012) PRSS3/mesotrypsin is a therapeutic target for metastatic prostate Cancer

    Molecular Cancer Research 10:1555–1566.

    https://doi.org/10.1158/1541-7786.MCR-12-0314

    • PubMed
    • Google Scholar
34. 1. Horn S
    2. Figl A
    3. Rachakonda PS
    4. Fischer C
    5. Sucker A
    6. Gast A
    7. Kadel S
    8. Moll I
    9. Nagore E
    10. Hemminki K
    11. Schadendorf D
    12. Kumar R

    (2013) TERT promoter mutations in familial and sporadic melanoma

    Science 339:959–961.

    https://doi.org/10.1126/science.1230062

    • PubMed
    • Google Scholar
35. 1. Huang FW
    2. Hodis E
    3. Xu MJ
    4. Kryukov GV
    5. Chin L
    6. Garraway LA

    (2013) Highly recurrent TERT promoter mutations in human melanoma

    Science 339:957–959.

    https://doi.org/10.1126/science.1229259

    • PubMed
    • Google Scholar
36. 1. Hudson TJ
    2. Anderson W
    3. Artez A
    4. Barker AD
    5. Bell C
    6. Bernabé RR
    7. Bhan MK
    8. Calvo F
    9. Eerola I
    10. Gerhard DS
    11. Guttmacher A
    12. Guyer M
    13. Hemsley FM
    14. Jennings JL
    15. Kerr D
    16. Klatt P
    17. Kolar P
    18. Kusada J
    19. Lane DP
    20. Laplace F
    21. Youyong L
    22. Nettekoven G
    23. Ozenberger B
    24. Peterson J
    25. Rao TS
    26. Remacle J
    27. Schafer AJ
    28. Shibata T
    29. Stratton MR
    30. Vockley JG
    31. Watanabe K
    32. Yang H
    33. Yuen MM
    34. Knoppers BM
    35. Bobrow M
    36. Cambon-Thomsen A
    37. Dressler LG
    38. Dyke SO
    39. Joly Y
    40. Kato K
    41. Kennedy KL
    42. Nicolás P
    43. Parker MJ
    44. Rial-Sebbag E
    45. Romeo-Casabona CM
    46. Shaw KM
    47. Wallace S
    48. Wiesner GL
    49. Zeps N
    50. Lichter P
    51. Biankin AV
    52. Chabannon C
    53. Chin L
    54. Clément B
    55. de Alava E
    56. Degos F
    57. Ferguson ML
    58. Geary P
    59. Hayes DN
    60. Hudson TJ
    61. Johns AL
    62. Kasprzyk A
    63. Nakagawa H
    64. Penny R
    65. Piris MA
    66. Sarin R
    67. Scarpa A
    68. Shibata T
    69. van de Vijver M
    70. Futreal PA
    71. Aburatani H
    72. Bayés M
    73. Botwell DD
    74. Campbell PJ
    75. Estivill X
    76. Gerhard DS
    77. Grimmond SM
    78. Gut I
    79. Hirst M
    80. López-Otín C
    81. Majumder P
    82. Marra M
    83. McPherson JD
    84. Nakagawa H
    85. Ning Z
    86. Puente XS
    87. Ruan Y
    88. Shibata T
    89. Stratton MR
    90. Stunnenberg HG
    91. Swerdlow H
    92. Velculescu VE
    93. Wilson RK
    94. Xue HH
    95. Yang L
    96. Spellman PT
    97. Bader GD
    98. Boutros PC
    99. Campbell PJ
    100. Flicek P
    101. Getz G
    102. Guigó R
    103. Guo G
    104. Haussler D
    105. Heath S
    106. Hubbard TJ
    107. Jiang T
    108. Jones SM
    109. Li Q
    110. López-Bigas N
    111. Luo R
    112. Muthuswamy L
    113. Ouellette BF
    114. Pearson JV
    115. Puente XS
    116. Quesada V
    117. Raphael BJ
    118. Sander C
    119. Shibata T
    120. Speed TP
    121. Stein LD
    122. Stuart JM
    123. Teague JW
    124. Totoki Y
    125. Tsunoda T
    126. Valencia A
    127. Wheeler DA
    128. Wu H
    129. Zhao S
    130. Zhou G
    131. Stein LD
    132. Guigó R
    133. Hubbard TJ
    134. Joly Y
    135. Jones SM
    136. Kasprzyk A
    137. Lathrop M
    138. López-Bigas N
    139. Ouellette BF
    140. Spellman PT
    141. Teague JW
    142. Thomas G
    143. Valencia A
    144. Yoshida T
    145. Kennedy KL
    146. Axton M
    147. Dyke SO
    148. Futreal PA
    149. Gerhard DS
    150. Gunter C
    151. Guyer M
    152. Hudson TJ
    153. McPherson JD
    154. Miller LJ
    155. Ozenberger B
    156. Shaw KM
    157. Kasprzyk A
    158. Stein LD
    159. Zhang J
    160. Haider SA
    161. Wang J
    162. Yung CK
    163. Cros A
    164. Cross A
    165. Liang Y
    166. Gnaneshan S
    167. Guberman J
    168. Hsu J
    169. Bobrow M
    170. Chalmers DR
    171. Hasel KW
    172. Joly Y
    173. Kaan TS
    174. Kennedy KL
    175. Knoppers BM
    176. Lowrance WW
    177. Masui T
    178. Nicolás P
    179. Rial-Sebbag E
    180. Rodriguez LL
    181. Vergely C
    182. Yoshida T
    183. Grimmond SM
    184. Biankin AV
    185. Bowtell DD
    186. Cloonan N
    187. deFazio A
    188. Eshleman JR
    189. Etemadmoghadam D
    190. Gardiner BB
    191. Gardiner BA
    192. Kench JG
    193. Scarpa A
    194. Sutherland RL
    195. Tempero MA
    196. Waddell NJ
    197. Wilson PJ
    198. McPherson JD
    199. Gallinger S
    200. Tsao MS
    201. Shaw PA
    202. Petersen GM
    203. Mukhopadhyay D
    204. Chin L
    205. DePinho RA
    206. Thayer S
    207. Muthuswamy L
    208. Shazand K
    209. Beck T
    210. Sam M
    211. Timms L
    212. Ballin V
    213. Lu Y
    214. Ji J
    215. Zhang X
    216. Chen F
    217. Hu X
    218. Zhou G
    219. Yang Q
    220. Tian G
    221. Zhang L
    222. Xing X
    223. Li X
    224. Zhu Z
    225. Yu Y
    226. Yu J
    227. Yang H
    228. Lathrop M
    229. Tost J
    230. Brennan P
    231. Holcatova I
    232. Zaridze D
    233. Brazma A
    234. Egevard L
    235. Prokhortchouk E
    236. Banks RE
    237. Uhlén M
    238. Cambon-Thomsen A
    239. Viksna J
    240. Ponten F
    241. Skryabin K
    242. Stratton MR
    243. Futreal PA
    244. Birney E
    245. Borg A
    246. Børresen-Dale AL
    247. Caldas C
    248. Foekens JA
    249. Martin S
    250. Reis-Filho JS
    251. Richardson AL
    252. Sotiriou C
    253. Stunnenberg HG
    254. Thoms G
    255. van de Vijver M
    256. van't Veer L
    257. Calvo F
    258. Birnbaum D
    259. Blanche H
    260. Boucher P
    261. Boyault S
    262. Chabannon C
    263. Gut I
    264. Masson-Jacquemier JD
    265. Lathrop M
    266. Pauporté I
    267. Pivot X
    268. Vincent-Salomon A
    269. Tabone E
    270. Theillet C
    271. Thomas G
    272. Tost J
    273. Treilleux I
    274. Calvo F
    275. Bioulac-Sage P
    276. Clément B
    277. Decaens T
    278. Degos F
    279. Franco D
    280. Gut I
    281. Gut M
    282. Heath S
    283. Lathrop M
    284. Samuel D
    285. Thomas G
    286. Zucman-Rossi J
    287. Lichter P
    288. Eils R
    289. Brors B
    290. Korbel JO
    291. Korshunov A
    292. Landgraf P
    293. Lehrach H
    294. Pfister S
    295. Radlwimmer B
    296. Reifenberger G
    297. Taylor MD
    298. von Kalle C
    299. Majumder PP
    300. Sarin R
    301. Rao TS
    302. Bhan MK
    303. Scarpa A
    304. Pederzoli P
    305. Lawlor RA
    306. Delledonne M
    307. Bardelli A
    308. Biankin AV
    309. Grimmond SM
    310. Gress T
    311. Klimstra D
    312. Zamboni G
    313. Shibata T
    314. Nakamura Y
    315. Nakagawa H
    316. Kusada J
    317. Tsunoda T
    318. Miyano S
    319. Aburatani H
    320. Kato K
    321. Fujimoto A
    322. Yoshida T
    323. Campo E
    324. López-Otín C
    325. Estivill X
    326. Guigó R
    327. de Sanjosé S
    328. Piris MA
    329. Montserrat E
    330. González-Díaz M
    331. Puente XS
    332. Jares P
    333. Valencia A
    334. Himmelbauer H
    335. Himmelbaue H
    336. Quesada V
    337. Bea S
    338. Stratton MR
    339. Futreal PA
    340. Campbell PJ
    341. Vincent-Salomon A
    342. Richardson AL
    343. Reis-Filho JS
    344. van de Vijver M
    345. Thomas G
    346. Masson-Jacquemier JD
    347. Aparicio S
    348. Borg A
    349. Børresen-Dale AL
    350. Caldas C
    351. Foekens JA
    352. Stunnenberg HG
    353. van't Veer L
    354. Easton DF
    355. Spellman PT
    356. Martin S
    357. Barker AD
    358. Chin L
    359. Collins FS
    360. Compton CC
    361. Ferguson ML
    362. Gerhard DS
    363. Getz G
    364. Gunter C
    365. Guttmacher A
    366. Guyer M
    367. Hayes DN
    368. Lander ES
    369. Ozenberger B
    370. Penny R
    371. Peterson J
    372. Sander C
    373. Shaw KM
    374. Speed TP
    375. Spellman PT
    376. Vockley JG
    377. Wheeler DA
    378. Wilson RK
    379. Hudson TJ
    380. Chin L
    381. Knoppers BM
    382. Lander ES
    383. Lichter P
    384. Stein LD
    385. Stratton MR
    386. Anderson W
    387. Barker AD
    388. Bell C
    389. Bobrow M
    390. Burke W
    391. Collins FS
    392. Compton CC
    393. DePinho RA
    394. Easton DF
    395. Futreal PA
    396. Gerhard DS
    397. Green AR
    398. Guyer M
    399. Hamilton SR
    400. Hubbard TJ
    401. Kallioniemi OP
    402. Kennedy KL
    403. Ley TJ
    404. Liu ET
    405. Lu Y
    406. Majumder P
    407. Marra M
    408. Ozenberger B
    409. Peterson J
    410. Schafer AJ
    411. Spellman PT
    412. Stunnenberg HG
    413. Wainwright BJ
    414. Wilson RK
    415. Yang H
    416. International Cancer Genome Consortium

    (2010) International network of Cancer genome projects

    Nature 464:993–998.

    https://doi.org/10.1038/nature08987

    • PubMed
    • Google Scholar
37. 1. Jiang G
    2. Cao F
    3. Ren G
    4. Gao D
    5. Bhakta V
    6. Zhang Y
    7. Cao H
    8. Dong Z
    9. Zang W
    10. Zhang S
    11. Wong HH
    12. Hiley C
    13. Crnogorac-Jurcevic T
    14. Lemoine NR
    15. Wang Y

    (2010) PRSS3 promotes tumour growth and metastasis of human pancreatic Cancer

    Gut 59:1535–1544.

    https://doi.org/10.1136/gut.2009.200105

    • PubMed
    • Google Scholar
38. 1. Kalikin LM
    2. Bugeaud EM
    3. Palmbos PL
    4. Lyons RH
    5. Petty EM

    (2001) Genomic characterization of human SEC14L1 splice variants within a 17q25 candidate tumor suppressor gene region and identification of an unrelated embedded expressed sequence tag

    Mammalian Genome 12:925–929.

    https://doi.org/10.1007/s00335-001-2073-3

    • PubMed
    • Google Scholar
39. 1. Kaplan EL
    2. Meier P

    (1958) Nonparametric estimation from incomplete observations

    Journal of the American Statistical Association 53:457.

    https://doi.org/10.1080/01621459.1958.10501452

    • Google Scholar
40. 1. Kent WJ
    2. Sugnet CW
    3. Furey TS
    4. Roskin KM
    5. Pringle TH
    6. Zahler AM
    7. Haussler D

    (2002) The human genome browser at UCSC

    Genome Research 12:996–1006.

    https://doi.org/10.1101/gr.229102

    • PubMed
    • Google Scholar
41. 1. Khurana E
    2. Fu Y
    3. Chakravarty D
    4. Demichelis F
    5. Rubin MA
    6. Gerstein M

    (2016) Role of non-coding sequence variants in Cancer

    Nature Reviews Genetics 17:93–108.

    https://doi.org/10.1038/nrg.2015.17

    • PubMed
    • Google Scholar
42. 1. Khurana E
    2. Fu Y
    3. Colonna V
    4. Mu XJ
    5. Kang HM
    6. Lappalainen T
    7. Sboner A
    8. Lochovsky L
    9. Chen J
    10. Harmanci A
    11. Das J
    12. Abyzov A
    13. Balasubramanian S
    14. Beal K
    15. Chakravarty D
    16. Challis D
    17. Chen Y
    18. Clarke D
    19. Clarke L
    20. Cunningham F
    21. Evani US
    22. Flicek P
    23. Fragoza R
    24. Garrison E
    25. Gibbs R
    26. Gümüs ZH
    27. Herrero J
    28. Kitabayashi N
    29. Kong Y
    30. Lage K
    31. Liluashvili V
    32. Lipkin SM
    33. MacArthur DG
    34. Marth G
    35. Muzny D
    36. Pers TH
    37. Ritchie GR
    38. Rosenfeld JA
    39. Sisu C
    40. Wei X
    41. Wilson M
    42. Xue Y
    43. Yu F
    44. Dermitzakis ET
    45. Yu H
    46. Rubin MA
    47. Tyler-Smith C
    48. Gerstein M
    49. 1000 Genomes Project Consortium

    (2013) Integrative annotation of variants from 1092 humans: application to Cancer genomics

    Science 342:1235587.

    https://doi.org/10.1126/science.1235587

    • PubMed
    • Google Scholar
43. 1. Lanzós A
    2. Carlevaro-Fita J
    3. Mularoni L
    4. Reverter F
    5. Palumbo E
    6. Guigó R
    7. Johnson R

    (2017) Discovery of Cancer driver long noncoding RNAs across 1112 tumour genomes: new candidates and distinguishing features

    Scientific Reports 7:41544.

    https://doi.org/10.1038/srep41544

    • PubMed
    • Google Scholar
44. 1. Lawrence MS
    2. Stojanov P
    3. Polak P
    4. Kryukov GV
    5. Cibulskis K
    6. Sivachenko A
    7. Carter SL
    8. Stewart C
    9. Mermel CH
    10. Roberts SA
    11. Kiezun A
    12. Hammerman PS
    13. McKenna A
    14. Drier Y
    15. Zou L
    16. Ramos AH
    17. Pugh TJ
    18. Stransky N
    19. Helman E
    20. Kim J
    21. Sougnez C
    22. Ambrogio L
    23. Nickerson E
    24. Shefler E
    25. Cortés ML
    26. Auclair D
    27. Saksena G
    28. Voet D
    29. Noble M
    30. DiCara D
    31. Lin P
    32. Lichtenstein L
    33. Heiman DI
    34. Fennell T
    35. Imielinski M
    36. Hernandez B
    37. Hodis E
    38. Baca S
    39. Dulak AM
    40. Lohr J
    41. Landau DA
    42. Wu CJ
    43. Melendez-Zajgla J
    44. Hidalgo-Miranda A
    45. Koren A
    46. McCarroll SA
    47. Mora J
    48. Lee RS
    49. Crompton B
    50. Onofrio R
    51. Parkin M
    52. Winckler W
    53. Ardlie K
    54. Gabriel SB
    55. Roberts CW
    56. Biegel JA
    57. Stegmaier K
    58. Bass AJ
    59. Garraway LA
    60. Meyerson M
    61. Golub TR
    62. Gordenin DA
    63. Sunyaev S
    64. Lander ES
    65. Getz G

    (2013) Mutational heterogeneity in Cancer and the search for new cancer-associated genes

    Nature 499:214–218.

    https://doi.org/10.1038/nature12213

    • PubMed
    • Google Scholar
45. 1. Lee EB
    2. Jin G
    3. Lee SY
    4. Park JY
    5. Kim MJ
    6. Choi JE
    7. Jeon HS
    8. Cha SI
    9. Cho S
    10. Kim CH
    11. Park TI
    12. Jung TH
    13. Son JW
    14. Park JY

    (2010) TP53 mutations in korean patients with non-small cell lung Cancer

    Journal of Korean Medical Science 25:698–705.

    https://doi.org/10.3346/jkms.2010.25.5.698

    • PubMed
    • Google Scholar
46. 1. Lochovsky L
    2. Zhang J
    3. Fu Y
    4. Khurana E
    5. Gerstein M

    (2015) LARVA: an integrative framework for large-scale analysis of recurrent variants in noncoding annotations

    Nucleic Acids Research 43:8123–8134.

    https://doi.org/10.1093/nar/gkv803

    • PubMed
    • Google Scholar
47. 1. Ma R
    2. Ye X
    3. Cheng H
    4. Ma Y
    5. Cui H
    6. Chang X

    (2015) PRSS3 expression is associated with tumor progression and poor prognosis in epithelial ovarian Cancer

    Gynecologic Oncology 137:546–552.

    https://doi.org/10.1016/j.ygyno.2015.02.022

    • PubMed
    • Google Scholar
48. 1. Marsit CJ
    2. Okpukpara C
    3. Danaee H
    4. Kelsey KT

    (2005) Epigenetic silencing of the PRSS3 putative tumor suppressor gene in non-small cell lung Cancer

    Molecular Carcinogenesis 44:146–150.

    https://doi.org/10.1002/mc.20125

    • PubMed
    • Google Scholar
49. 1. Marx V

    (2014) Cancer genomes: discerning drivers from passengers

    Nature Methods 11:375–379.

    https://doi.org/10.1038/nmeth.2891

    • PubMed
    • Google Scholar
50. 1. Melton C
    2. Reuter JA
    3. Spacek DV
    4. Snyder M

    (2015) Recurrent somatic mutations in regulatory regions of human Cancer genomes

    Nature Genetics 47:710–716.

    https://doi.org/10.1038/ng.3332

    • PubMed
    • Google Scholar
51. 1. Mularoni L
    2. Sabarinathan R
    3. Deu-Pons J
    4. Gonzalez-Perez A
    5. López-Bigas N

    (2016) OncodriveFML: a general framework to identify coding and non-coding regions with Cancer driver mutations

    Genome Biology 17:128.

    https://doi.org/10.1186/s13059-016-0994-0

    • PubMed
    • Google Scholar
52. 1. Pettersen HS
    2. Galashevskaya A
    3. Doseth B
    4. Sousa MM
    5. Sarno A
    6. Visnes T
    7. Aas PA
    8. Liabakk NB
    9. Slupphaug G
    10. Sætrom P
    11. Kavli B
    12. Krokan HE
    13. Anastasia G
    14. Berit D
    15. Antonio S
    16. Torkild V

    (2015) AID expression in B-cell lymphomas causes accumulation of genomic uracil and a distinct AID mutational signature

    DNA Repair 25:60–71.

    https://doi.org/10.1016/j.dnarep.2014.11.006

    • PubMed
    • Google Scholar
53. 1. Pettersen HS
    2. Sundheim O
    3. Gilljam KM
    4. Slupphaug G
    5. Krokan HE
    6. Kavli B

    (2007) Uracil-DNA glycosylases SMUG1 and UNG2 coordinate the initial steps of base excision repair by distinct mechanisms

    Nucleic Acids Research 35:3879–3892.

    https://doi.org/10.1093/nar/gkm372

    • PubMed
    • Google Scholar
54. 1. Polak P
    2. Karlić R
    3. Koren A
    4. Thurman R
    5. Sandstrom R
    6. Lawrence MS
    7. Reynolds A
    8. Rynes E
    9. Vlahoviček K
    10. Stamatoyannopoulos JA
    11. Sunyaev SR

    (2015) Cell-of-origin chromatin organization shapes the mutational landscape of Cancer

    Nature 518:360–364.

    https://doi.org/10.1038/nature14221

    • PubMed
    • Google Scholar
55. 1. Pollard KS
    2. Hubisz MJ
    3. Rosenbloom KR
    4. Siepel A

    (2010) Detection of nonneutral substitution rates on mammalian phylogenies

    Genome Research 20:110–121.

    https://doi.org/10.1101/gr.097857.109

    • PubMed
    • Google Scholar
56. 1. Pon JR
    2. Marra MA

    (2015) Driver and passenger mutations in Cancer

    Annual Review of Pathology: Mechanisms of Disease 10:25–50.

    https://doi.org/10.1146/annurev-pathol-012414-040312

    • Google Scholar
57. Software

    1. R Development Core Team

    (2008) R: A Language and Environment for Statistical Computing

    R Foundation for Statistical Computing, Vienna, Austria.

    http://www.R-project.org
58. 1. Reimand J
    2. Bader GD

    (2013) Systematic analysis of somatic mutations in phosphorylation signaling predicts novel Cancer drivers

    Molecular Systems Biology 9:637.

    https://doi.org/10.1038/msb.2012.68

    • PubMed
    • Google Scholar
59. 1. Reva B
    2. Antipin Y
    3. Sander C

    (2011) Predicting the functional impact of protein mutations: application to Cancer genomics

    Nucleic Acids Research 39:e118.

    https://doi.org/10.1093/nar/gkr407

    • PubMed
    • Google Scholar
60. 1. Sabarinathan R
    2. Mularoni L
    3. Deu-Pons J
    4. Gonzalez-Perez A
    5. López-Bigas N

    (2016) Nucleotide excision repair is impaired by binding of transcription factors to DNA

    Nature 532:264–267.

    https://doi.org/10.1038/nature17661

    • PubMed
    • Google Scholar
61. 1. Samur MK

    (2014) RTCGAToolbox: a new tool for exporting TCGA Firehose data

    PLoS One 9:e106397.

    https://doi.org/10.1371/journal.pone.0106397

    • PubMed
    • Google Scholar
62. Book

    1. Schafer JL

    (1997)

    Analysis of Incomplete Multivariate Data

    CRC Press.

    • Google Scholar
63. 1. Schweder T
    2. Spjotvoll E

    (1982) Plots of P -values to evaluate many tests simultaneously

    Biometrika 69:493.

    https://doi.org/10.1093/biomet/69.3.493

    • Google Scholar
64. 1. Srebrow A
    2. Kornblihtt AR

    (2006) The connection between splicing and Cancer

    Journal of Cell Science 119:2635–2641.

    https://doi.org/10.1242/jcs.03053

    • PubMed
    • Google Scholar
65. 1. Stratton MR
    2. Campbell PJ
    3. Futreal PA
    4. Andrew Futreal P

    (2009) The cancer genome

    Nature 458:719–724.

    https://doi.org/10.1038/nature07943

    • PubMed
    • Google Scholar
66. 1. Tamborero D
    2. Gonzalez-Perez A
    3. Lopez-Bigas N

    (2013) OncodriveCLUST: exploiting the positional clustering of somatic mutations to identify Cancer genes

    Bioinformatics 29:2238–2244.

    https://doi.org/10.1093/bioinformatics/btt395

    • PubMed
    • Google Scholar
67. Website

    1. TCGA Data Portal

    (2016) The cancer genome atlas (TCGA) data portal

    Accessed July 2016.

    https://ocg.cancer.gov/resources/cancer-genome-atlas-tcga-data-portal
68. Software

    1. Tools to Make Developing R Packages Easier

    (2016) Comprehensive R Archive Network (CRAN), version 1.12.0

    R Package Devtools Version.

    http://CRAN.R-project.org/package=devtools
69. 1. Touzet H
    2. Varré JS

    (2007) Efficient and accurate P-value computation for position weight matrices

    Algorithms for Molecular Biology 2:15.

    https://doi.org/10.1186/1748-7188-2-15

    • PubMed
    • Google Scholar
70. Website

    1. UCSC Genome Browser

    (2016) UCSC Genome Bioinformatics

    Accessed May 2016.

    http://hgdownload.cse.ucsc.edu/downloads.html
71. 1. Van Rhijn I
    2. Godfrey DI
    3. Rossjohn J
    4. Moody DB
    5. Branch Moody D

    (2015) Lipid and small-molecule display by CD1 and MR1

    Nature Reviews Immunology 15:643–654.

    https://doi.org/10.1038/nri3889

    • PubMed
    • Google Scholar
72. 1. Varley JM
    2. Attwooll C
    3. White G
    4. McGown G
    5. Thorncroft M
    6. Kelsey AM
    7. Greaves M
    8. Boyle J
    9. Birch JM

    (2001) Characterization of germline TP53 splicing mutations and their genetic and functional analysis

    Oncogene 20:2647–2654.

    https://doi.org/10.1038/sj.onc.1204369

    • PubMed
    • Google Scholar
73. 1. Vinagre J
    2. Almeida A
    3. Pópulo H
    4. Batista R
    5. Lyra J
    6. Pinto V
    7. Coelho R
    8. Celestino R
    9. Prazeres H
    10. Lima L
    11. Melo M
    12. da Rocha AG
    13. Preto A
    14. Castro P
    15. Castro L
    16. Pardal F
    17. Lopes JM
    18. Santos LL
    19. Reis RM
    20. Cameselle-Teijeiro J
    21. Sobrinho-Simões M
    22. Lima J
    23. Máximo V
    24. Soares P

    (2013) Frequency of TERT promoter mutations in human cancers

    Nature Communications 4:2185.

    https://doi.org/10.1038/ncomms3185

    • PubMed
    • Google Scholar
74. 1. Visnes T
    2. Doseth B
    3. Pettersen HS
    4. Hagen L
    5. Sousa MM
    6. Akbari M
    7. Otterlei M
    8. Kavli B
    9. Slupphaug G
    10. Krokan HE

    (2009) Uracil in DNA and its processing by different DNA glycosylases

    Philosophical Transactions of the Royal Society B: Biological Sciences 364:563–568.

    https://doi.org/10.1098/rstb.2008.0186

    • PubMed
    • Google Scholar
75. 1. Wei S
    2. LiVolsi VA
    3. Brose MS
    4. Montone KT
    5. Morrissette JJ
    6. Baloch ZW

    (2016) STK11 mutation identified in thyroid carcinoma

    Endocrine Pathology 27:65–69.

    https://doi.org/10.1007/s12022-015-9411-6

    • PubMed
    • Google Scholar
76. 1. Weinhold N
    2. Jacobsen A
    3. Schultz N
    4. Sander C
    5. Lee W

    (2014) Genome-wide analysis of noncoding regulatory mutations in Cancer

    Nature Genetics 46:1160–1165.

    https://doi.org/10.1038/ng.3101

    • PubMed
    • Google Scholar
77. 1. Weinstein JN
    2. Collisson EA
    3. Mills GB
    4. Shaw KR
    5. Ozenberger BA
    6. Ellrott K
    7. Shmulevich I
    8. Sander C
    9. Stuart JM
    10. Cancer Genome Atlas Research Network

    (2013) The Cancer genome Atlas Pan-Cancer analysis project

    Nature Genetics 45:1113–1120.

    https://doi.org/10.1038/ng.2764

    • PubMed
    • Google Scholar
78. 1. Yang P
    2. Du CW
    3. Kwan M
    4. Liang SX
    5. Zhang GJ

    (2013) The impact of p53 in predicting clinical outcome of breast Cancer patients with visceral metastasis

    Scientific Reports 3:2246.

    https://doi.org/10.1038/srep02246

    • PubMed
    • Google Scholar
79. 1. Zhou X
    2. Li X
    3. Cheng Y
    4. Wu W
    5. Xie Z
    6. Xi Q
    7. Han J
    8. Wu G
    9. Fang J
    10. Feng Y

    (2014) BCLAF1 and its splicing regulator SRSF10 regulate the tumorigenic potential of Colon cancer cells

    Nature Communications 5:4581.

    https://doi.org/10.1038/ncomms5581

    • PubMed
    • Google Scholar

Author details

1. Malene Juul

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   MJ, Data curation, Software, Formal analysis, Visualization, Writing—original draft, Writing—review and editing, Developed, implemented and applied the ncdDetect algorithm and performed downstream analyses of results, Wrote the manuscript

   For correspondence

   malene.juul.rasmussen@clin.au.dk

   Competing interests

   The authors declare that no competing interests exist.

   "This ORCID iD identifies the author of this article:" 0000-0001-9722-0461

2. Johanna Bertl

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   JB, Data curation, Formal analysis, Writing—review and editing, Developed and implemented the statistical null model and provided the predicted mutational probabilities used in ncdDetect, Assisted with preparing the manuscript

   Competing interests

   The authors declare that no competing interests exist.

3. Qianyun Guo

   Bioinformatics Research Centre (BiRC), Aarhus University, Aarhus, Denmark

   Contribution

   QG, Data curation, Formal analysis, Developed and implemented the statistical null model and provided the predicted mutational probabilities used in ncdDetect

   Competing interests

   The authors declare that no competing interests exist.

4. Morten Muhlig Nielsen

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   MMN, Data curation, Formal analysis, Provided genomic elements and contributed expression analysis, Helped visualise perform downstream analysis of the candidate elements, Assisted with preparing the manuscript

   Competing interests

   The authors declare that no competing interests exist.

5. Michał Świtnicki

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   MŚ, Data curation, Formal analysis, Visualization, Provided correlation analysis of survival data and mutational status for candidate elements

   Competing interests

   The authors declare that no competing interests exist.

6. Henrik Hornshøj

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   HH, Data curation, Formal analysis, Visualization, Methodology, Participated in visualising and performing downstream analysis of the candidate elements, Further helped to curate the analysed tumor samples

   Competing interests

   The authors declare that no competing interests exist.

7. Tobias Madsen

   Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark

   Contribution

   TM, Software, Formal analysis, Methodology, Implemented the overdispersion-based mutation rate correction factor, and provided feedback and insights into dynamic programming and the use of the R-package Rcpp

   Competing interests

   The authors declare that no competing interests exist.

8. Asger Hobolth

   Bioinformatics Research Centre (BiRC), Aarhus University, Aarhus, Denmark

   Contribution

   AH, Conceptualization, Supervision, Writing—review and editing, Conceived the research project and supervised the development of ncdDetect, Revised final version of the manuscript

   Competing interests

   The authors declare that no competing interests exist.

9. Jakob Skou Pedersen

   1. Department of Molecular Medicine (MOMA), Aarhus University Hospital, Aarhus, Denmark
   2. Bioinformatics Research Centre (BiRC), Aarhus University, Aarhus, Denmark

   Contribution

   JSP, Conceptualization, Supervision, Funding acquisition, Visualization, Writing—original draft, Writing—review and editing, Conceived the research project and supervised the development of ncdDetect, Wrote the manuscript

   For correspondence

   jakob.skou@clin.au.dk

   Competing interests

   The authors declare that no competing interests exist.

   "This ORCID iD identifies the author of this article:" 0000-0002-7236-4001

• 3,159

  views

• 606

  downloads

• 37

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.