Systematic analysis of somatic mutations driving cancer: uncovering functional protein regions in disease development

Significantly mutated protein regions are identified based on mutation pattern

In this work we collected cancer associated non-synonymous mutations from the COSMIC database. We developed a novel method called iSiMPRe that is able to identify cancer genes at the level of protein regions based on these mutations. The underlying assumption of iSiMPRe is that cancer mutations affecting protein coding regions are not distributed evenly, rather they are accumulated in specific regions (Significantly Mutated Protein Regions - SiMPRes) that play an active role in tumorigenesis. Regions that harbor a significantly enriched amount of somatic mutations compared to neutral local mutations can highlight not only cancer genes but also specific functional regions within them that actively contribute to the development of cancer. Since the mutation frequency is calculated for each gene separately from the observed number of mutations, different background mutation rates apply for each gene, that can take into account that the background mutation rate is different for individual genes [10]. In our analysis, we only considered exonic mutations that have a local effect, therefore only missense mutations and in-frame insertions and deletions are considered. The presented method is the first that is able to consider in-frame insertions and deletions as well as missense point mutations in a unified statistical framework. This sets it apart from two conceptually similar methods, eDriver [17] and OncoDriveClust [16], which do not consider in-frame insertions and deletions. The input of iSiMPRe is a set of cancer-related missense mutations and in-frame insertions and deletions. The background mutation rate is calculated simply from these using only a few empirical parameters. This is in contrast to OncoDriveClust which estimates the background mutation rate from a set of silent substitutions that has to be supplied as a separate set of input data. iSiMPRe is described in detail in Additional file 1 (iSiMPRe protocol) and in short in Fig. 1. In order to enable potential users to apply the method to updated versions of COSMIC datasets or other sources of cancer mutation data, the source code of iSiMPRe is available for download. In our experience, the identified significantly mutated regions change very little with updates of COSMIC datasets.

Known cancer genes harbor significantly mutated regions

We have tested the overlap between the identified SiMPRes and the cancer-related genes in the SCGD (see Methods), and separately on annotated tumor suppressors and oncogenes. iSiMPRe can identify at least one SiMPRe in 74 % of the genes in SCGD (Fig. 2). The performance of the method is nearly perfect for oncogenes with at least one identified region (53 out of 54, accounting for 98 %) with only one exception, DNMT1. In this case, the available mutational data is insufficient to yield high significance. Unlike its relative, DNMT3A, where mutations are highly enriched at a C-terminal region near the catalytic site, studies indicate that mutations of DNMT1 appear to be a rare event [21].

It is typically assumed that oncogenes harbor mainly missense mutations, while tumor suppressor genes are mainly affected by inactivating mutations, most often truncating nonsense or frameshift mutations [22]. As iSiMPRe considers only mutations that are traditionally linked to oncogenes and not to tumor suppressors, the recovery rate calculated on the latter group of genes is expected to be lower compared to oncogenes. In accordance with the hypothesis, a more moderate performance was observed in this case. Nevertheless, the majority of tumor suppressor genes (50 out of 71, accounting for 70 %) were still identified as containing at least one region with a significant accumulation of local somatic mutations. This is quite a striking result given that the contribution of missense mutations to tumor suppressor gene alterations is considered to be secondary to inactivating and truncating mutations [22].

iSiMPRe outperforms already existing methods with the same scope

Additionally to iSiMPRe, two recently developed methods OncoDriveClust [16] and eDriver [17] were also tested. These methods - similarly to our approach - aim to identify regions where mutations are clustered in the amino acid sequence. The major difference between eDriver and iSiMPRe is that eDriver relies on regions with predefined boundaries based on domain annotations and disorder predictions, while iSiMPRe identifies regions that maximize mutational significance without relying on predefined regions. While the background model of eDriver is similar to the one used in our model, OncoDriveClust uses different input data that includes frameshift and nonsense mutations together with missense point mutations and compares it to the distribution of silence mutations.

Figure 2a shows the number of genes identified by the three methods together with the overlap between them. The overlap between iSiMPRe and eDriver is high with most eDriver genes and regions being identified by iSiMPRe as well. However, the overlap between genes identified by iSiMPRe and OncoDriveClust was significantly lower. In terms of performance a similar trend was followed for the three methods (Fig. 2b–d). Oncogenes were recognized best, while tumor suppressors are more challenging, especially for OncoDriveClust. In all three datasets, iSiMPRe achieves the best performance, followed by eDriver. The lower performance of concurrent methods can be traced back to various factors. In the case of eDriver, averaging mutation numbers in larger predefined regions can sometimes mask significant regions, which can have a negative effect on correctly identifying cancer driving genes. Trivially, no significantly mutated regions could appear in proteins composed of a single domain or one disordered region alone. The performance of OncoDriveClust was significantly lower, with only one tumor suppressor and only less than third of oncogenes identified. This lower performance of OncoDriveClust suggests that its expected mutation distribution statistical model should be revisited and that the larger scale effect of frameshift/nonsense mutations cannot be compared to the local effect of missense mutations and short indels.

In-frame indels are important for finding cancer genes

The used version of the COSMIC database contains 404,903 missense mutations and 9977 in-frame insertions and deletions. While the number of short insertions and deletions is only less than 2.5 % of that of all local somatic mutations, they still make significant contribution to the identification of significantly mutated regions and contribute to the increased performance of iSiMPRe. These types of genetic alterations are dominant in about 20 % of the found regions, meaning indels in these cases contribute more to the significance of the region than missense mutations. In accordance, one-fifth of all regions are undetectable without taking indels into account. Omitting indels from the calculations would result in a decreased performance. Nevertheless, iSiMPRe would still outperform the other two methods (Fig. 2b–d). The incorporation of indel type mutations is especially important for the identification of tumor suppressors. Omitting indels does not have an impact on the recognition of oncogenes, but it does affect the results on both SCGD and tumor suppressors.

Some of the significantly mutated indel regions occur within genes with a well-established connection with cancer development, such as KIT, BRAF or PTEN. The proteins encoded by these genes harbor multiple SiMPRes, many of which are dominated by missense mutations, making the genes themselves identifiable using missense mutations alone. However, in certain cancer genes the primary mechanism of genetic alteration is the accumulation of short insertions or deletions. Discarding indels in the region identification process would make it impossible to reliably identify such known cancer genes. Notable examples include the interleukin-7 receptor subunit alpha (IL7R), which is known to be involved in the occurrence and development of various forms of acute leukemias and solid tumors [23]. Other examples include TSC2, where mutations have been linked to the development of hamartomas in multiple organs [24]; and ATAD5, the somatic mutations of which were identified in endometrial tumors [25].

Other types of genetic alterations are also often associated with local somatic mutations

Known cancer genes are defined based on various dominant type of genetic alterations and do not necessarily contain an accumulation of local somatic mutations. In order to gain insights into how the dominant type of genetic alteration for a given gene influences the chance of iSiMPRe to find a significantly mutated region at the protein level, the performance was further analyzed on specific datasets. For this analysis the KEGG database was used that provides information about a broad set of genetic alterations associated with specific genes. Figure 3 shows the ratio of genes for each type of dominant genetic alteration where SiMPRes were identified in the protein counterparts. Ten categories were analyzed in which somatic mutations represent only one of the categories. The results indicate that chromosomal translocations and rearrangements represent a roughly independent modulation of the genome and genes primarily subject to such changes are largely devoid of significantly mutated regions. Also, relatively modest recovery rate was observed for genes altered dominantly via germline mutations. In general, more than 80 % of the proteins dominantly altered by other mechanisms also contain at least one significant region identified by iSiMPRe in seven out of ten categories.

Significantly mutated regions can be specific to one cancer type or can be associated with a broad set of cancer types

Based on the supplied annotations in the respective datasets, genes were linked to one or more of the 27 standardized cancer types (see Methods and Table 2). Taking advantage of this annotation, the performance of iSiMPRe could be evaluated for individual cancer types. Figure 4 indicates that the overlap between cancer genes and significantly mutated regions found by iSiMPRe is highly uneven across various cancer types. In certain cancer types, such as head and neck carcinomas, iSiMPRe is able to identify all known cancer genes, exhibiting a perfect performance. In other cases the recovery rate is lower (such as the 40 % in the case of prostate and bone cancer). However, iSiMPRe is able to recover 74 % of cancer genes on average with recovering at least one third of genes even in the poorest case. Although perfect performance can be seen in some cancer types with low numbers of COSMIC mutations (such as stomach cancer, head and neck carcinoma, basal cell carcinoma and cervical cancer); in general there is no obvious relationship between the recovery rates and cancer types, sample number or the number of mutations.

Tissue/organ	Cancer types	Number of local somatic mutations in COSMIC	Number of samples	Average number of mutations per sample
Bladder	Bladder cancer	6 265	3 027	2.07
Blood	Acute myeloid leukemia	11 650	5 857	1.99
	Chronic myeloid leukemia	1 603	1 041	1.54
	Lymphoma	24 572	6 944	3.54
Bone	Bone cancer	2 814	706	3.99
Brain	Glioblastoma	9 361	2 909	3.22
	Neuroblastoma	4 932	619	7.97
	Glioma	12 793	2 947	4.34
	Medulloblastoma	4 701	728	6.46
Breast	Breast cancer	31 544	4 404	7.16
Cervix	Cervical cancer	136	132	1.03
Colorectal	Colorectal cancer	37 727	25 806	1.46
Esophagus	Esophageal cancer	3 659	433	8.45
Head and neck	Thyroid cancer	19 975	13 908	1.44
	Head and neck carcinoma	75	67	1.12
Kidney	Renal cell carcinoma	27 897	1 750	15.94
Liver	Hepatocellular carcinoma	19 100	2 091	9.13
Lung	Small cell lung cancer	976	208	4.69
	Non-small cell lung cancer	19 991	11 236	1.78
Ovary	Ovarian cancer	19 286	2 936	6.57
Pancreas	Pancreatic cancer	33 776	5 609	6.02
Prostate	Prostate cancer	18 813	967	19.46
Skin	Melanoma	12 374	7 923	1.56
	Squamous cell carcinoma	15 726	3 459	4.55
	Basal cell carcinoma	292	251	1.16
Stomach	Stomach cancer	7 795	1 430	5.45
Uterus	Endometrial cancer	5 722	1 627	3.52
Total		353 555	109 015	3.24

At the level of individual genes, cancer genes can be specific to one or a very limited number of (usually related) forms of cancer or can be involved in multiple cancer types. In a number of cases, multiple regions showed a significant enrichment of local mutations within a single gene. In these cases two basic scenarios could emerge. In the first scenario, mutations from various tissue samples are typically distributed in roughly the same way and their accumulation along the sequence outline the same functional regions. This general trend is demonstrated in Fig. 5a in the case of DNA methyltransferase 3A (DNMT3A) and for phosphatidylinositol 3-kinase regulatory subunit alpha (PIK3-R1). In the case of DNMT3A, the dominant part of mutations come from various hematopoietic cancers, such as acute myeloid leukemia (AML), chronic myeloid leukemia (CML) or lymphoma. These mutations cluster in three distinct regions: one region falls into the middle of the ADD (ATRX-DNMT3-DNMT3L) domain, responsible for the interaction with the polycomb repressive complex 2 (PRC2), and it roughly covers the PHD-type zinc finger; the second and the third regions both fall into the catalytic (DNA cytosine methyltransferase) domain. In all three regions, the source of these mutations is fairly homogeneous with respect to the originating cancer types, even though the dominant mutation types differ for the three regions (deletions + insertions, deletions + missense and missense only, respectively). Similarly, the mutations in PI3K-R1 outline two distinct regions, covering the two terminal parts of the inter-SH2 coiled-coil region, which is responsible for the interaction with the catalytic subunit. Similarly to the DNMT3A case, the regions differ slightly in their dominant mutation types; however, in all regions the majority of the mutations come from the same, endometrial and breast cancer, samples.

Less frequently, a different scenario can also be observed, where a single gene is associated with different types of cancer depending on the location of the mutations. The most interesting examples involve certain tyrosine kinase receptors. Figure 5b shows the domain structure and the significantly mutated regions encountered in c-KIT and FGFR3. Both are single-pass transmembrane proteins: in their extracellular region they contain multiple Ig-like domains, while their intracellular regions harbor a tyrosine kinase domain. The latter element is one of the most frequently mutated domains in various cancers (Table 3) and it also contains significantly mutated regions. However, other regions in these two proteins also exhibit an increased number of genetic variations that can involve mostly missense mutations, deletions, insertions or mixed types of variations. Importantly, these specific regions are associated with different cancer types. In the case of KIT, the majority of known somatic mutations were sequenced in AML, melanoma and gastrointestinal stromal tumor (GIST) samples. The two regions found in the last Ig-like domain of the extracellular receptor part is linked to AML and GIST, respectively. The intracellular regions are either GIST, AML or melanoma specific, or a combination of two of the three cancer types, but interestingly KIT does not seem to harbor a ubiquitously mutated region. Another uncommon cancer-type specific partition of mutations can be seen in the case of FGFR3, where both extracellular regions are linked exclusively to bladder cancer while the intracellular regions are present in various forms of skin tumors as well.

Significantly mutated protein regions correspond to a diverse set of functional modules

The identified SiMPRes enable us to take a deeper look into which functional modules of proteins are involved in cancer development. For this, we collected biological annotations of the identified regions/genes. Information was collected from UniProt database [26] which is a rich source of information about active sites of enzymes and additional functional sites including nucleotide phosphate-, DNA- and other binding regions. Another annotation category corresponds to ‘regions of interest’. This category used by UniProt can refer to basically any experimentally characterized protein region that can be of interest concerning protein function (such as interaction sites, different regions of multifunctional enzymes or regions crucial for biological processes/sub-cellular localizations). UniProt annotations were complemented with annotations from the Pfam database [27] collecting evolutionarily conserved protein family information with associated functions and from the ELM database which provides information about linear motifs operating largely outside of globular domains [28]. The occurrence of SiMPRes in various types of annotations were tested for over and under-representation using a basic statistical model (see Methods).

In most cases the identified SiMPRes could be associated with some structural or functional annotations. The mean overlaps and standard errors of randomized regions is shown in Table 3 together with the overlaps of real SiMPRes. The data in Table 3 shows that local genetic alterations are significantly concentrated on the structured parts of proteins. This is in accordance with earlier findings [29]. Within structured domains, there is a clear tendency for enzymes, especially for kinases, confirming earlier observations [30, 31]. The most commonly mutated domain is the Tyrosine kinase domain observed in 20 different proteins, including such well-known cancer driver genes as ALK, EGFR or BRAF (see Additional file 3). Other domains mutated in multiple cases include the Serine/threonine-protein kinase domain (e.g. in Activin receptor type-1 or in Serine/threonine-protein kinase STK11) and the SH2 domains (e.g. in PIK3-R1 and STAT3). In theory, the easiest way to disrupt enzyme function would be the modulation of the active site via the introduction of mutations. However, data show that the number of active sites overlapping with SiMPRes is much lower than the number of targeted enzyme/kinase domains. Instead, iSiMPRe showed a highly significant overlap with other types of functional regions, including nucleotide phosphate-, DNA- and other binding regions, as well as ‘regions of interest’. Protein binding regions were suggested to be the primary hot-spots for tumor suppressor proteins [15]. The high accumulation of mutations was not restricted to structured protein modules. The non-globular parts of proteins – especially motifs, motif switches and post-translational modification sites – tended to coincide with a high accumulation of mutations as well [32, 33].

The search for cancer driver genes continues

Cancer emerges due to genetic and epigenetic changes. With the dramatic expansion of catalogued mutations detected in diverse tumor samples and the advent of exome sequencing screens, we have an overwhelming amount of data on our hands. Although this – in theory – should enable easier insights into tumorigenesis, it is still not a simple task to adequately distinguish between relevant, cancer-associated mutations and background genetic alterations without clinical significance. Therefore, one of the major focuses of cancer research is still the identification of genes that are responsible for cancer formation, tumor progression and metastasis. However, the definition of these cancer genes is far from trivial. Usually a prerequisite of a cancer driver gene is that its genetic alterations should have a positive contribution to the development of cancer. In practice, the decision of whether a gene fulfills these criteria or not can depend on the cancer type and the number of samples already analyzed, as well as the type of genetic alterations analyzed. It was also suggested that significantly larger sample size is needed to obtain a complete set of cancer driver genes [34]. The subjectivity of cancer drivers is transferred to the publicly available databases in which many such genes are collected. These datasets use different levels of evidence for the inclusion of a gene or gene product and they can also contain erroneously identified cancer genes, for example based on mouse models or paralog sequences. Since currently there is no consensus on the list of cancer genes, we used four different sources to collect and combine manually assembled and curated sets of cancer genes (see Methods and the figure in Additional file 4).

The lack of consensus regarding cancer genes is evident from the limited overlap between the four included datasets. The fact that only around 27 % of OMIM genes have support in other databases can be attributed to the fact that OMIM is primarily aimed at collecting germline affected genes that only partially overlap with somatically altered genes. Unexpectedly, even more limited overlap was observed in the case of COSMIC census database. For this dataset over half of the listed genes were absent from all other databases, as opposed to the 20 % and 4 % of the genes listed in KEGG and Driver genes, respectively. It could mean that the COSMIC census list includes novel cancer genes that are not present in other datasets as of yet, or it could also indicate lack of supportive evidence and possible biases of the inclusion system.

In addition to manually curated databases, various computational methods have also been developed to identify cancer driver genes [35, 36]. Most commonly used approaches seek to identify cancer driver genes either in the context of pathways and protein interaction networks [37] or by detecting signs of positive selection at the level of genes. Such methods can be based on the increased number of observed mutations compared to the background mutation rate [10]; a high rate of non-silent mutations compared to silent mutations [2]; or on the bias towards the accumulation of mutations with high functional impact [38]. However, the accumulation of mutations can highlight not only genes but also specific functional regions at the protein level that are involved in disease development. Recent methods, such as OncoDriveClust [16] or e-Driver [17] identify putative cancer driver genes based on this concept, similarly to the iSiMPre method presented in this work. Clustering of mutations can also be observed in three-dimensional protein structures that often correspond to perturbed protein-protein interaction sites [18, 19]. Given the complexity and heterogeneity of the molecular basis of cancer, the combination of different signals of positive selection can more reliably indicate mutational drivers [39]. Nevertheless, in our experience, the iSiMPre method is able to identify the majority of cancer driver genes based on the clustering of mutations and outperforms methods with similar scope. The increased performance of iSiMPre can be attributed to several factors, including the cleaning of mutational data (e.g. eliminating likely neutral polymorphisms and mutations occurring within tandem repeats that are more likely to accommodate neutral mutations as well as sequencing errors). Additional factor is the incorporation of all genetic variations with positional information, which include short in-frame insertions and deletions while excluding frameshift and non-sense mutations. iSiMPre is based on an unbiased approach that does not rely on previous knowledge of structure or domain, which could be especially important to detect cancer driver mutations located in intrinsically disordered proteins for example (manuscript in preparation).

Based on the presented analysis, the local accumulation of somatic mutations detected by iSiMPRe can also be used to pinpoint novel genes not yet (fully) represented in the available databases. We found 7 genes that contained high significance mutated regions that were absent from SCGD. In these cases, the high significance of the enrichment of somatic mutations lends a very strong support that these proteins indeed correspond to true cancer drivers. These include the lymphoma gene IL7R, the K⁺ channel KCNJ5, a hyperaldosteronism-linked gene that is also known to be mutated in adrenal cancers [40], CD79B known to be involved in B-cell lymphomas [41], the breast cancer gene ESR1 [42], the cytokine receptor IL6ST known to be involved in a range of cancers, the known cancer gene RAC1 regulating cell motility and RHOA also involved in various cancers, eg. gastric carcinomas [43]. For these 7 genes that are unique to the COSMIC census dataset, iSiMPRe confirms their cancer driver statuses.

iSiMPRe also identified 211 low significance regions that reside in genes not present in any of the cancer gene datasets (see table in Additional file 2). Although some of these genes have recently been linked to tumorigenesis (eg. FRK was shown to be involved in hepatocellular adenomas, which matches our annotations [50]), most genes have no direct indication of being cancer genes. Some of these candidate genes might be linked causatively to tumorigenesis, but they can only be expected to exhibit a weak phenotype (small growth advantage versus wild-type cells) in agreement with their lower mutation hit rate. However, they can also correspond to genes with a locally increased passenger mutation rate, without true biological significance. In these cases more data is needed to be able to discriminate low significance regions that might correspond to false positives from those regions that are genuinely involved in various cancers even though just with a weak phenotype.

Our results indicate that iSiMPRe is able not only to identify known and novel cancer genes, but the assigned confidence levels of the identified regions correlate well with our current knowledge of cancer driver genes. Furtermore, iSiMPRe can also effectively target the main issue of interpreting genomic sequencing data: the discrimination of driver and passenger mutations. Overall, the regions marked by iSiMPRe only contain about one third of the mutations listed in COSMIC, which is a drastic reduction of data. As there is a significant association between identified regions and functional protein regions, iSiMPRe may be useful to help in classifying mutations as drivers or passengers; simply based on their location inside or outside of a SiMPRe.

The fundamental differences between passenger and driver mutations can influence the set of preferential amino acid substitutions. In recent analyses, COSMIC mutations in the case of EGFR [51] and also for all genes [52] showed an uneven distribution of missense substitutions among cancer driver mutations defined as reoccurring mutation compared to likely passenger mutations recorded only once in COSMIC. Partitioning driver and passenger mutations based on iSiMPRe showed good agreement with preferential distributions observed earlier (see Additional file 2), with the seven and 16 most frequent substitutions appearing in both lists for driver and passenger mutations respectively. These results also hint at the partitioning power of iSiMPRe between driver and passenger missense mutations.

The knowledge of significantly mutated regions involved in cancer can guide treatment and drug development choices

One of the main advantages of the method proposed in this work is that it is able to identify not only cancer driver genes but also specific regions that are involved in the disease. The importance of this more detailed view is apparent in cases where a single protein has multiple regions that are specifically mutated in different cancer types. An example for this behavior is exhibited in the case of cytokine receptors (Fig. 5b). In these cases various forms of cancer target different protein regions, which implies a tissue-specific biological selection for particular mutations. Unfortunately, little is known about the mechanistic differences of the same receptors in different tissues that would explain the selection for non-overlapping mutational hotspots. One possible explanation could be the homo- or heterodimerization of the same receptor with different partners in different tissues. Facultative receptor tyrosine kinase heterodimerization is a common phenomenon and already documented for several related proteins [53–55].

The current collections of genes that have been causatively linked to tumor formation are heterogeneous concerning their dominant genetic alterations by which their structure/function may be modulated. An important question is how well genes containing SiMPRes (calculated from somatic missense mutations and short in-frame indels) represent all cancer genes. A surprising outcome of our analyses was that significantly mutated regions can be found in the majority of cancer genes, regardless of their dominant genetic alterations (Figs. 3 and 4). The two main exceptions are chromosomal translocations and rearrangements, which represent roughly independent modulations of the genome (Fig. 3). In many other cases, however, somatic mutations can have similar effects to other types of genetic alterations. (Note that the detailed analysis of non-coding intergenic, promoter or intronic mutations was not done, yet the data published by other groups hints that the targets of these alterations might also be somewhat separate from those targeted by somatic mutations [56–58].)

The observed interchangeability of various genetic alterations and the accumulation of somatic mutations is relatively straight-forward in the case where the dominant alterations diminish or abolish protein function (eg. deletion/loss, downregulation/underexpression and promoter hypermethylation). An example is provided by the SMAD4 gene (Fig. 6, top row). In this case the clustering of somatic mutations affects the C-terminal MH2 domain that is essential for both homo- and hetero-oligomerization. Normally SMAD4 transmits signals in the TGF-beta pathway, which is a negative regulator of epithelial growth. The deletion of the SMAD4 gene or the abolishment of the protein function via somatic mutations leads to cancer through the breaking of a negative regulatory pathway.

In the case of genetic alterations that enhance protein function, the interchangeability with missense mutations is slightly less evident. Overexpression/amplification typically can happen in receptors containing intracellular kinase domains that are able to modulate various pathways (such as ERK, JNK, Akt, etc.) via their downstream targets. Both amplification and their accumulated local somatic mutations force these proteins to be constitutively active independently from the binding of their extracellular ligands. For example in the case of EGFR (Fig. 6, middle row) the active form of the receptor is brought about by homo- or hetero-dimerization, which is normally achieved by the binding of EGF/TGFα (or other ligands). This dimerization can occur without ligand binding when the protein is overexpressed; however, the same effect can also be brought about by the accumulation of local somatic mutations in the N-terminal part of the intracellular kinase domain.

As opposed to the previously discussed genetic alterations, in general the effects of translocations and chromosomal rearrangements cannot be understood through the function of a single gene/protein. The resulting fusion protein can carry functional parts from both original proteins and can fulfill novel biological roles. In most cases, the effect of translocations cannot be effectively mimicked by the accumulation of local mutations. However, in certain cases, genes undergoing translocations as their primary genetic alteration still show a significant accumulation of somatic mutations. A prime example for this behavior is presented by RUNX1 (Fig. 6, bottom row). This gene encodes a transcription factor that consists of the runt domain that is responsible for the interaction with the DNA and a long, presumably disordered transactivation region that binds to partner proteins. A common genetic alteration of RUNX1 is the fusion of the runt domain coding segment (first 5 exons) of the gene to the almost complete ETO gene. The resulting chimeric protein termed AML-ETO retains the ability to bind RUNX1 target genes while exerting the transcription repression function of ETO, silencing RUNX1 genes contributing to the development of various hematopoietic cancers, most notably AML. Apart from translocation, the RUNX1 gene also harbors significant accumulation of somatic mutations in the runt domain. The majority of these mutations were derived from AML samples indicating the possibility of these mutations mimicking the effect of the translocation. The mutation hotspots in the runt domain coincide with three regions that are directly involved in DNA binding mostly affecting arginine/lysine residues that are necessary for the charge complementarity. This diminishes the DNA binding capability of RUNX1 and inhibits the activation of its target genes, achieving the gene silencing effect of the gene fusion.

COSMIC somatic mutations

Cancer related somatic mutations were collected from version 73 of the COSMIC database (http://cancer.sanger.ac.uk/cosmic/) [8]. The complete dataset contained 20,337 protein sequences with at least one annotated mutation in the 73 version of the COSMIC database. All missense point mutations together with in-frame insertions and deletions were collected and checked against the protein sequences given in COSMIC. Mutations that were sequenced in samples having over 100 mutations were discarded as hypermutated samples, as these likely include a large fraction of passenger mutations. Furthermore, point mutations coinciding with positions of common single nucleotide polymorphisms were discarded using SNP data collected from the UCSC Genome Browser [64]. In order to filter out clonal variants, samples having at least five mutations and sharing at least 50 % of the local variations (missense mutations and insertions and deletions) were filtered and only mutations coming from the most mutation-rich sample were considered to avoid the inclusion of duplicated entries. Altogether, this resulted in the inclusion of 404,913 missense mutations and 9977 in-frame insertions and deletions. Correctly defined mutations were linked to one of the cancer types based on their given sites of occurrence and histology according to the rules given in Additional file 2.

In the COSMIC database mutational data can be mapped to multiple isoforms. The isoform with the largest number of mutations often does not necessarily correspond to the canonic or the main isoform used in other databases. In this work, the isoforms of the same gene were clustered together and the isoform with the the highest number of mutations was selected. For functional and structural annotations and crosslinks to other databases, COSMIC gene ID’s were also mapped to UniProt protein accession codes from the 2016/01 version of UniProt. The entries in each database and the identified regions were mapped to this main isoform, where possible.

Databases of known cancer related genes

In our study we surveyed four source databases that contained known cancer genes based on expert curation.

Identification of significantly mutated protein regions

A novel method was developed to identify significantly mutated protein regions. We refer this method as iSiMPRe (identification of significantly mutated protein regions) and the identified regions as ‘SiMPRe’s. iSiMPRe seeks to find regions that harbor a significantly enriched amount of somatic mutations compared to neutral local mutations which are assumed to be distributed evenly throughout the sequence. To discard a large portion of possible sequencing artefacts, iSiMPRe first filters out mutations that fall into genomic regions with low sequence complexity measured by TRF [68]. Next, the method follows a hierarchical, stepwise algorithm at the determination of possible significant regions, described in detail in Additional file 1 and in short in Fig. 1. Each identified region has a p-value assigned that characterizes the significance of the given region based on all three types of mutations. Based on this total p-value, regions are classified as high significance (if p < 10⁻²⁰), medium significance (10⁻²⁰ < p < 10⁻⁰⁵) or low significance (10⁻⁰⁵ < p < 10⁻⁰²). Within that, however, the dominant mutation type (missense, insertion or deletion) is established for a region based on the relative contribution of each mutation type to the total significance of the region. Here the p-values of the found regions were re-calculated with only considering one class of the three possible mutations types. The mutation type yielding the lowest p-value on its own was considered to be the dominant one. Furthermore, the resulting regions were assigned to one or more of the 27 cancer types based on the histology and site annotations of the mutations present in the region. These cancer types are shown in Table 2.

Comparison of iSiMPRe with oncodriveclust and edriver

The performance of our method was compared to two methods with similar scope, OncoDriveClust [16] and e-Driver [17]. For both methods the programs were installed and run locally with the default parameters. For the sake of direct comparison, all three methods were tested on the same set of mutations assembled from COSMIC.

Association of functional regions with SiMPRes

For each protein in the COSMIC dataset, functional and structural annotations were collected from the corresponding UniProt entry from the UniProt version 2016/01. Additionally, Pfam families, and known instances of linear motif sites collected in the ELM database were also mapped to the studied protein sequences. Altogether 18 categories were assigned to analyze the annotation of the significantly mutated regions (SiMPRe) including domains, motifs, repeats, kinases and enzymes (from Pfam [27]), DNA binding regions, regions of interest, transmembrane regions, nucleotide phosphate binding regions, active sites, binding sites and post-translational modification sites (from UniProt [26]), motifs, switches and phosphorylation sites (from ELM [28] and phosphoELM [69]). The possible over- and under-representation of these functional protein regions in SiMPRes was also determined and tested for statistical significance. For this, a random baseline was established. Each of the 534 found SiMPRe was moved to a randomly selected protein in a way that the length and disorder content of the original and the new, random protein had to be the same within 10 % to avoid structural bias. Next, both real and randomly selected SiMPRes were assessed regarding their overlap with various functional protein units. The random region selection procedure was repeated 1000 times and the average and the standard deviation of the overlap between randomized regions and functional protein regions was calculated. The over-representation of SiMPRe in functional regions compared to randomized regions is given in standard error (StdErr) units in Table 3. As values calculated on random regions closely follow Gaussian distributions, the threshold for p < 0.01 statistical significance in StdErr units is 2.326, therefore all color-coded over-representations are statistically significant.

Assessment of protein disorder

In the process of generating random regions, the disorder content of proteins were assessed using IUPred [70, 71], ANCHOR [72, 73] and Pfam [27]. The output of IUPred was smoothed in a window of 31 residues. Positions with a smoothed IUPred score over 0.5 or with an ANCHOR score over 0.5 were considered disordered, except for positions included in specific Pfam families which were annotated as domains.

Substitution rates

Substitution rates were calculated from missense mutations from the COSMIC database. Each type of missense mutation was considered only once for each position. First, all types of occurring missense mutations were collected separately for all positions. The number of substitutions changing the original amino acid of type i to type j is denoted by n _i-j and substitution percentages are calculated as follows: R _i-j   = (n _i-j /N)*100, where N is the total number of missense mutations. These substitution percentages were calculated separately for mutations inside and outside of SiMPRes and are shown in Additional file 2.

Reviewer’s report 1: Sándor Pongor, International Centre for Genetic Engineering and Biotechnology (ICGEB), Italy

Reviewer’s report 2: Michael Gromiha, Indian Institute of Technology Madras, India

Reviewer’s report 3: Zoltán Gáspári, Pázmány University, Budapest