Mutational hotspots in the TP53 gene and, possibly, other tumor suppressors evolve by positive selection

Background The mutation spectra of the TP53 gene and other tumor suppressors contain multiple hotspots, i.e., sites of non-random, frequent mutation in tumors and/or the germline. The origin of the hotspots remains unclear, the general view being that they represent highly mutable nucleotide contexts which likely reflect effects of different endogenous and exogenous factors shaping the mutation process in specific tissues. The origin of hotspots is of major importance because it has been suggested that mutable contexts could be used to infer mechanisms of mutagenesis contributing to tumorigenesis. Results Here we apply three independent tests, accounting for non-uniform base compositions in synonymous and non-synonymous sites, to test whether the hotspots emerge via selection or due to mutational bias. All three tests consistently indicate that the hotspots in the TP53 gene evolve, primarily, via positive selection. The results were robust to the elimination of the highly mutable CpG dinucleotides. By contrast, only one, the least conservative test reveals the signature of positive selection in BRCA1, BRCA2, and p16. Elucidation of the origin of the hotspots in these genes requires more data on somatic mutations in tumors. Conclusion The results of this analysis seem to indicate that positive selection for gain-of-function in tumor suppressor genes is an important aspect of tumorigenesis, blurring the distinction between tumor suppressors and oncogenes. Reviewers This article was reviewed by Sandor Pongor, Christopher Lee and Mikhail Blagosklonny.


Background
The p53 protein is called "the guardian of the genome" because this multifunctional transcription factor, which regulates cell cycle progression, repair and programmed cell death in mammals, targets for apoptosis those cells that accumulate unsustainable DNA damage [1][2][3][4]. In nearly 60% of human cancers, the TP53 gene carries muta-tions that are generally thought to abrogate the tumor suppressor function of p53 [1][2][3][4]. However, many independent studies have also revealed gain of new biochemical and biological functions as a result of TP53 mutations, suggesting that this gene additionally has properties of an oncogene [4][5][6][7][8][9][10]. Recent reports on mouse models of the Li-Fraumeni syndrome (LFS), a familial cancer predisposition syndrome caused by germline p53 mutations, revealed significant changes in the tumor spectra in mice carrying common p53 mutations, indicating that gain-of-function in p53 is important for tumorigenesis [11,12]. This notion has been supported and extended by bioinformatic analysis of the tumor-specific mutation spectra in the TP53 gene which show a highly significant excess of non-synonymous mutations over the neutral expectation, suggesting that p53 evolution in tumors is subject to positive selection [13] as a result of preferential fixation of missense mutations in p53 [14][15][16].
However, this apparent positive selection does not necessarily account for the strongly non-uniform distribution of mutations among the sites in the TP53 gene, i.e., the existence of hotspots. The presence of hotspots is compatible with either a mutational or a selectional scenario or a combination thereof [14,17]. A well-characterized case of apparent mutational origin of hotspots is the preponderance of G>T transitions in the TP53 gene in lung cancers which is usually perceived as a reflection of the mutagenesis specificity of polycyclic carcinogens [18,19]. However, this paradigm has been challenged on the grounds that the respective mutations localize predominantly in highly conserved, functionally important sites of p53 [20]. It is important to determine the relative contributions of selection and mutation specificity to the hotspot origin because it has been suggested that mutable nucleotide contexts could be used to infer mechanisms of mutagenesis and thus elucidate key mechanisms of tumor initiation and progression [18,19]. This type of analysis has proved informative for somatic hypermutation in immunoglobulin genes [21,22], some cancer-related genes in lymphomas [23], and germline mutations in human disease genes [24,25].
Here we apply three independent tests, accounting for non-uniform base compositions in synonymous and non-synonymous sites, to test whether selection makes a significant contribution to the origin of the hotspots. All three tests consistently indicate that the hotspots in the TP53 gene evolve, primarily, via positive selection. By contrast, only one, the least conservative test reveals the signature of positive selection in BRCA1, BRCA2, and p16. The results of this analysis seem to indicate that positive selection for gain-of-function in tumor suppressor genes is an important aspect of tumorigenesis, blurring the distinction between tumor suppressors and oncogenes.

Results and discussion
Many independent studies have revealed gain of new biochemical and biological functions as a result of TP53 mutations, suggesting that this gene additionally has properties of an oncogene [4][5][6][7][8][9][10][11][12]. This notion has been supported by bioinformatic analysis of the tumor-specific mutation spectra in the TP53 gene which show a highly significant excess of missense mutations over the neutral expectation, suggesting that p53 evolution in tumors is subject to positive selection [14,15]. In Table 1, we present an update of this comparison based on the latest somatic mutation data; the substantial excess of non-synonymous substitutions suggests positive selection [13] acting on p53 in all tumor types for which sufficient information was available.
We developed a simple statistical test (hereinafter NSMC test, after Non-Synonymous Monte Carlo) that specifically addressed the dilemma of mutational vs. selectional origin of the hotspots. This test included comparison of samples of synonymous and non-synonymous sites selected such that both the number of sites and the number of mutations in the samples was the same (Fig.  1). Only positions in which mutations were found were analyzed. Sites in which both synonymous and non-synonymous substitutions were observed (e.g., third positions in two-codon series) were analyzed independently for the two types of substitutions. The NSMC test was designed to account for differences in the nucleotide compositions and the frequencies of substitutions in synonymous and non-synonymous sites (non-synonymous sites were sampled to mimic the nucleotide composition in synonymous sites). With this normalization, the comparison of the number of hotspots, i.e., sites with at least two substitutions, between the samples of synonymous and non-synonymous (designated NSH and NNH, respectively) sites gives a measure of the skewness of the distribution of mutations (Fig. 1). Monte Carlo simulations, repeated 100,000 times, were used to assess the statistical significance of differences between the distributions of hotspots in the synonymous and non-synonymous sites. Two alternative statistical hypotheses were tested: H0mutational bias (no difference between the distributions of hotspots in the synonymous and non-synonymous sites) and H1 -selectional bias (the distributions of hotspots in the synonymous and non-synonymous sites are different). The fraction of simulated sets in which NNH > NSH is the probability P(H1) of the rejection of H0. Large values of P(H1) (≥ 0.95) indicate that the hypothesis H0 is rejected and there is a significant excess of hotspots in non-synonymous sites. Using the NSMC test, we detected a statistically significant excess of hotspots in non-synonymous sites in 50% of the tumors for which extensive mutational data was available (Table 1). When the data from all tumor types were pooled, the excess of hotspots in non-synonymous sites was highly significant: the null hypothesis, i.e., that the distributions of mutations in the synonymous and nonsynonymous sites were identical, was rejected with P < 0.001 (Table 2). Since we accounted for differences in nucleotide compositions, mutational biases are not expected to differ between synonymous and non-synonymous positions. Thus, the greater skew of the mutation distribution in non-synonymous positions should be viewed as evidence of, primarily, selectional origin of the hotspots.
This conclusion was further supported by analysis of mutation spectra after removal of CpG dinucleotides, the most prominent mutational hotspots in the human genome [24,25]. Under this test, many hotspots in CpG sites overlapping arginine, glycine and valine codons were removed but the selection hypothesis was nevertheless supported for several tissues and for the combined spectrum (Additional file 1). Furthermore, the results of the NSMC test performed before or after removal of the CpG sites did not depend on the threshold used for hotspot identification (Additional file 2).
We also applied the NSMC test to compare the distributions of hotspots in nonsense and synonymous sites. An excess of hotspots in nonsense sites would be indicative of positive selection for loss of p53 function. A significant excess of hotspots in nonsense sites was detected only in colorectal cancers as opposed to 8 of the 16 analyzed tumor types in which hotspots non-randomly associated with non-synonymous sites were identified (Table 1). The difference between the excess of hotspots in non-synonymous sites and the excess of hotspots in nonsense sites was statistically significant (P = 0.015 by the Fisher's exact test). This observation is compatible with the notion that non-synonymous hotspots in p53 evolve under positive selection for gain of function.
The procedure used for random sampling of mutations at non-synonymous sites in the NSMC test Figure 1 The procedure used for random sampling of mutations at non-synonymous sites in the NSMC test.
Step 1 includes the selection of a sample of non-synonymous sites such that the number of sites and their base composition were the same as in the entire set of synonymous substitutions in the given gene; sampling was performed without replacement. The comparison of the number of hotspots, i.e., sites with at least two substitutions, between the samples of synonymous and nonsynonymous sites gives a measure of the skewness of the distribution of mutations (Step 2). Synonymous sites and sampled non-synonymous sites are shown by inverse shading, synonymous hotspots are shown in yellow, and simulated non-synonymous hotspots are shown in green. In the last column, the P-value for the NSMC test (numerator) and the χ 2 value for the NSCS test (denominator) are shown along with the NSMC test P-value for the hotspots in nonsense sites. The statistically significant values (rejection of H0) are shown in bold type. We further tested the hypothesis of independence between the mutation class (hotspot vs. non-hotspot) and site class (non-synonymous vs. synonymous). The data for all analyzed spectra were represented as 2 × 2 contingency tables which were analyzed using the χ 2 test (hereinafter NSCS test, after Non-Synonymous Chi-Square).
Using the NSCS test, we observed a significant excess of hotspots in non-synonymous sites compared to the expectation under the independence hypothesis. Thus, two independent statistical tests show that, in the spectra of somatic mutations in the TP53 gene from most tumors, the hotspots are highly non-randomly associated with non-synonymous sites. In a direct analogy to the classical Ka/Ks signature of positive selection [13], this preferential occurrence of hotspots in non-synonymous positions indicates that the hotspots result, mostly, from positive selection for new functions of the p53 protein.
Both the NSMC and the NSCS tests produced opposite results when applied to the available mutational spectra of three other tumor suppressor genes, BRCA1, BRCA2, and p16 INK4a (Table 3) [26]. The hypothesis that hotspots are randomly distributed among synonymous and nonsynonymous sites could not be rejected for these genes. This observation suggests that p53 might be unique among tumor suppressors in that its somatic evolution in many tumors involves intense positive selection for gain of function. Alternatively, however, it cannot be ruled out that the available mutation data for the other tumor suppressors is insufficient to detect statistically significant association of hotspots with non-synonymous sites.
We also developed a third statistical test (hereinafter NSB test, after Non-Synonymous Binomial) to identify nonsynonymous substitution hotspots (analyzed, for this purpose, at the level of codons), i.e., those with a statistically significant excess of non-synonymous substitutions over the random expectation. The expected numbers of non-synonymous and synonymous substitutions were calculated using a Monte-Carlo simulation procedure, which was repeated 1,000 times for each codon. Each step involved random shuffling of transitions and transversions among the three positions of a codon. The statistical significance of the observed excess of the number of the detected non-synonymous substitution hotspots over the random expectation was assessed using the binomial test and the Bonferroni correction for multiple tests.
The NSB test revealed from 1 (p16) to 59 (TP53) hotspots non-randomly associated with non-synonymous sites in each of the tumor suppressors (Table 4). Thus, it appears that positive selection might affect not only somatic evolution of p53 but also that of other tumor suppressors albeit, seemingly, to a lesser extent. The failure of the NSMC and NSCS tests to detect the signature of positive selection in genes other than p53 could be due to the fact that these tests require a large number of synonymous substitutions which is currently available only for p53. Alternatively, however, it cannot be ruled out that synonymous substitutions are underrepresented in the databases for BRCA1, BRCA2, and p16. Such an artifact would affect the NSB test, potentially resulting in false-positives, but not the NSMC or the NSCS tests. Expanded compendia of somatic mutations for these genes and thorough database curation are critical for a reliable assessment of the contribution of positive selection to their evolution in tumors. Even the largest available database of somatic mutations, that for p53, is not large enough for some statistical experiments. For example, we were unable to apply our tests to G>T substitutions in lung tumors [18][19][20] because only a few unique synonymous G>T substitutions associated with lung tumors were found in the p53 database, whereas the non-synonymous G>T mutations are the most frequent type of substitutions in lung tumors [14,[18][19][20]. Furthermore, more data on somatic mutations is required to explore the effect of nucleotide context other than that of CpG that was examined here.
It should be emphasized that, although we detected a highly statistically significant association of non-synonymous sites with hospots, in particular, for p53, the results of the present analysis do not allow us to assign any individual mutation to the gain-of-function or loss-of-function category. Nevertheless, these results can be used for devising experimental studies of gain-of-function by tumor suppressors mutated in specific sites (hotspots) and/or specific tumor types with particularly strong evidence of positive selection.

Conclusion
The previous computational analysis of the tumor-specific mutation spectra of the TP53 gene has suggested that positive selection made a substantial contribution to the evolution of this gene during tumorigenesis [14,15]. Here, we show that positive selection, as opposed to mutational biases, is, largely, responsible for the formation of hotspots in the TP53 gene (of course, this does not rule out the existence of true mutational hotspots and their substantial role in tumorigenesis; we only show that such hotspots are in the minority). This finding is compatible with the previous observations that hotspots are located primarily in highly conserved, functionally important regions of p53 [14,17]. Together, the results of computational analyses of the mutational spectra strongly support the crucial role of gain-of-function in the tumorigenic evolution of p53, which agrees with the results of several experimental studies [4][5][6][7][8][9][10], in particular, the recent work on mouse models of the Li-Fraumeni syndrome [11,12]. It has been shown that hotspots are tumor-specific and, furthermore, include both residues that are directly involved in DNA-binding and residues that are important for maintaining the conformation of the p53 protein [14]. Thus, along with loss of function mutations, tumorigenesis might involve positive selection for a diverse set of novel activities of p53. The present analysis also yielded preliminary evidence of the role of positive selection in the evolution of BRCA1, BRCA2, and p16, suggesting that mutational gain-of-function in tumor suppressors might be a widespread and important aspect of tumorigenesis and blurring the boundary between tumor suppressors and oncogenes.

Materials and methods
The TP53 mutation data were from the IARC database [27]. Sixteen tumor types were analyzed (the IARC "short topology" keyword = "BLADDER, BRAIN, BREAST, LIVER, LUNG, PANCREAS, OVARY, PROSTATE, COLON, COLORECTAL, ESOPHAGUS, HEMATOPOIETIC, LAR-YNX, MOUTH, SKIN, STOMACH"). The p16, BRCA1 and the BRCA2 mutation data were from the databases of mutation spectra in the p16 gene [28] and the BRCA1 and BRCA2 genes [29]. The statistical tests were implemented as ad hoc programs written in C++ or Perl.

Reviewer's report 1 Sandor Pongor, International Centre for Genetic Engineering and Biotechnology, Padriciano 99, I-34012, Trieste, Italy
Detecting positive selection at the DNA sequence level is of substantial interest in view of the role it may play in pathogenic events. Glazko and coworkers show, using three statistical tests developed for the purpose, that the mutational hotspots of the TP53 gene evolve by positive selection. In view of the general interest of the topic I sought the advice of Dr. Lawrence Banks who is a biologist working on p53 mutations. We both felt that the calculations are thoroughly planned and the results support the main message of the paper.
The presentation of the paper could however be improved with special respect to the wide audience of Biology Direct. The most important criticism is that p53 mutations can lead both to inactivation and/or to GOF type changes, and these two groups may need to be analyzed separately. Currently, the reader may not see clearly if positive selection was found only in the case of GOF-type mutations or also in the case of inactivating mutations.

Author response:
We agree that it should be stated with full clarity that both gain-of-function and loss-of-function mutations in tumor suppressors, in particular, p53, are important. Therefore such changes have been made in several places in the manuscript; in particular, see the last paragraph in the Results and Discussion section.
Minor points: -In the Background section it might be useful to add 1) a brief description of mutation types found in p53 as well as their biological roles; and 2) a paragraph describing the mathematical approaches to detecting positive selection. These sections may help the reader in understanding what has been done and what is being accomplished in this work.  1. Both in the abstract and introduction, the authors emphasize the importance of taking into consideration the effect of "nucleotide context" on mutational bias, as a motivation for this study. However, as I understand it, this study takes into consideration nucleotide composition (i.e. frequency of single nucleotides), not nucleotide context (e.g. frequency of nucleotide triplets, to consider the effect of one adjacent nucleotide on either side of the nucleotide under study). Since nucleotide context can have large effects on mutation rate (e.g. CpG effects), this is an important issue. For the very reasons that the authors articulated in their Introduction, many readers will expect direct tests of whether nucleotide context affects the authors' results.
The difficulty, of course, is that it is harder to match nucleotide context (e.g. triplet frequencies, 64 different numbers) than nucleotide composition (just 4 numbers). The NSMC procedure would probably not be able to construct samples with matching triplet frequencies, without some modifications. One possible solution would be to include ALL sites (including unmutated sites, instead of just sites where mutations were observed) in the analysis. First, generate a random sample of synonymous sites (a specific number of sites, with a specific triplet profile, and a specific number of observed mutations).
Now generate a random sample of non-synonymous sites of the same size, with the same triplet profile. Finally, generate equal-sized random samples of mutations from each set of sites, and analyze the number of "hotspots" as in the NSMC method. Including non-mutated sites in this sampling process should make it possible to match the triplet profiles between the syn vs. non-syn samples, and I don't see a reason why non-mutated sites should be excluded.
If such analysis is practical I think it could greatly strengthen the paper, by directly addressing the question of nucleotide context. At any rate, the existing analysis in the manuscript should be clearly described as testing "nucleotide composition" not "nucleotide context", and the difference between these should be emphasized. The authors should point out that even if composition is controlled for, nucleotide context could have large effects on mutation rate, so the current results should be interpreted with some caution.
Author response: The CpG effects have been accounted for in the NSMC test; to emphasize this, we mention this control in the revised abstract. However, the currently available data on somatic mutations is insufficient to examine other, subtler effects of the nucleotide context. As for including non-mutated sites, we were concerned that this approach could lead to uncontrollable increase in the error rate due to the different and unknown intrinsic mutation rates of different sites.
2. The NSMC analysis, while conceptually simple, needs to be described in more detail, in the Methods section. Currently, there is only an outline of NSMC, presented in the Results section, which leaves out many details (e.g. sampling with replacement or without replacement? I assume the latter), such that one could not replicate the calculation with any confidence that equivalent results would be obtained from the same input data.