WeSME: uncovering mutual exclusivity of cancer drivers and beyond

Abstract

Motivation: Mutual exclusivity is a widely recognized property of many cancer drivers. Knowledge about these relationships can provide important insights into cancer drivers, cancer-driving pathways and cancer subtypes. It can also be used to predict new functional interactions between cancer driving genes and uncover novel cancer drivers. Currently, most of mutual exclusivity analyses are preformed focusing on a limited set of genes in part due to the computational cost required to rigorously compute P-values.

Results: To reduce the computing cost and perform less restricted mutual exclusivity analysis, we developed an efficient method to estimate P-values while controlling the mutation rates of individual patients and genes similar to the permutation test. A comprehensive mutual exclusivity analysis allowed us to uncover mutually exclusive pairs, some of which may have relatively low mutation rates. These pairs often included likely cancer drivers that have been missed in previous analyses. More importantly, our results demonstrated that mutual exclusivity can also provide information that goes beyond the interactions between cancer drivers and can, for example, elucidate different mutagenic processes in different cancer groups. In particular, including frequently mutated, long genes such as TTN in our analysis allowed us to observe interesting patterns of APOBEC activity in breast cancer and identify a set of related driver genes that are highly predictive of patient survival. In addition, we utilized our mutual exclusivity analysis in support of a previously proposed model where APOBEC activity is the underlying process that causes TP53 mutations in a subset of breast cancer cases.

Availability and Implementation: http://www.ncbi.nlm.nih.gov/CBBresearch/Przytycka/index.cgi#wesme

Contact: przytyck@ncbi.nlm.nih.gov

Supplementary information: Supplementary data are available at Bioinformatics online.

1 Introduction

Mutual exclusivity is a frequently observed property in two or more genes having mutations in cancer patients (Babur et al., 2015; Ciriello et al., 2012; Kim et al., 2015, 2016; Leiserson et al., 2013, 2015; Thomas et al., 2007; Vandin et al., 2012). Mutually exclusive gene pairs (i.e. gene pairs of which simultaneous mutations in the same patients are less frequent than it is expected by chance) are observed within many cancer types, while some are common across multiple different cancer types (Kim et al., 2015). The underlying cause of mutual exclusivity might vary depending on the context. In many cases, genes mutated in a mutually exclusive way have been found to be members of the same functional pathway. In such a case, the mutual exclusivity of their mutation pattern can be explained by the pathway centric view of cancer. Specifically, if two genes belong to the same cancer-driving pathway, then a mutation in just one of them might suffice to dysregulate the pathway and trigger cancer progression (Ciriello et al., 2012; Hofree et al., 2013; Kim et al., 2016; Kim and Przytycka, 2012; Vandin et al., 2012; Vogelstein and Kinzler, 2004). In contrast, in a combined analysis of multiple cancer types, mutually exclusive aberrations might represent tissue specific cancer drivers. In fact, the sets of mutually exclusive drivers obtained in Pan-Cancer analysis (Kandoth et al., 2013; Szczurek and Beerenwinkel, 2014) predominantly included the tissue type specific genes that, typically, do not share common pathways. Yet, even within the same tissue type, mutual exclusivity between genes from different pathways is not unusual (Kim et al., 2015). This type of exclusivity can be explained, for example, by synthetic lethality where simultaneous mutations in both genes are detrimental for cancer progression (Dwyer-Nield et al., 2010; Wang et al., 2014). Another (not necessarily unrelated) possibility is that the mutually exclusive drivers define different cancer subtypes within the same tissue type. Finally, mutual exclusivity can also be observed between less frequently mutated cancer drivers suggesting more subtle relationships between them and their contribution to cancer progression. For example, previous research found that even if the exclusivity between ARID1B and KRAS is not significant in any of individual cancer types separately, the pair was shown to be mutually exclusive when analyzing all cancer types together as statistical power was gained (Kim et al., 2015).

Mutual exclusivity can potentially provide important information about cancer, including its alternative progression pathways and subtypes. It can also reveal important relationships between genes and pathways. However, most of the analyses of mutual exclusivity so far have been focused on limited subsets of genes that were selected based on mutation rate cut-off (typically 3 or 5% depending on datasets) or restricted by other criteria. One of the factors limiting the applications of mutual exclusivity analysis beyond restricted gene sets was the cost of computing rigorous P-values for all gene pairs of interest. Statistically, pairwise mutual exclusivity can be estimated by a permutation test in which the original mutational profile is permuted by swapping random pairs of mutations iteratively while preserving the mutation rates of genes and patients. A possibly faster alternative approach is to use a hypergeometric test. However, the hypergeometric test treats all patients equally, ignoring patient specific mutation frequencies, which may limit its accuracy.

To provide a more in-depth understanding of mutual exclusivity in cancer data, we started by designing a new and fast heuristic for estimating statistical significance of mutual exclusivity relationships. Our method, Weighted Sampling based Mutual Exclusivity (WeSME), estimates P-values of mutual exclusivity while taking into account mutation frequencies of patients. It closely approximates the results of the permutation-based method and does so without using the costly permutations of a mutation matrix. Moreover, WeSME can compute P-values for a subset of genes independently from the rest of genome unlike the permutation method that requires whole genome permutations. By dynamically adjusting sampling depth, WeSME can provide high precision P-values without a significant increase in computational cost. Adopting the same sampling technique, we also developed a complementary test, Weighted Sampling based Co-Occurrence (WeSCO), to estimate the significance of co-occurrence in gene mutations.

With WeSME at hand, we began to explore the information that a comprehensive mutual exclusivity survey can provide. We included all genes with mutations in the analysis without pre-filtering. In particular, we did not exclude very long genes such as muscle protein Titin (TTN), gel forming Mucins, which often emerge as frequently mutated genes in cancer (even after correction with their lengths) but are generally assumed to be artifacts (Lawrence et al., 2013). Interestingly, we found that some of these genes have many mutual exclusivity partners. In particular, we identified TTN as one of the hubs in the mutual exclusivity network in breast cancer and endometrial carcinoma. The high number of mutual exclusivity partners of TTN prompted us to hypothesize that mutual exclusivity of its mutations could be a reflection of an underlying mutagenic process that occurs in a specific subgroup of patients.

To test this hypothesis, we reasoned that long and frequently mutated genes might harbor a recognizable signature of the mutagenic process acting on them. Specifically, it has been recently recognized that a subclass of APOBEC cytidine deaminases, which convert cytosine to uracil during RNA editing, are a source of mutagenicity in human tumors (Burns et al., 2013; Harris, 2015; Lawrence et al., 2013). Several other factors such as Pol ε mutation, patient ages at diagnosis, known mutagenic exposures, or defects in DNA maintenance are also associated with characteristic mutation signatures (Alexandrov et al., 2013). We found that, in the case of breast cancer, the mutation pattern of TTN is consistent with the APOBEC3B activity, while in endometrial carcinoma, it is consistent with the Pol ε mutation signature.

Including TTN gene in our mutual exclusivity analysis allowed us not only to observe interesting patterns of APOBEC activity in breast cancer but also to identify a set of driver genes that is highly predictive of patient survival. In addition, we showed that our mutual exclusivity analysis supports the model where APOBEC activity is the underlying process that causes TP53 mutations in highly mutated breast cancer cases. Finally, with our method, we have been able to uncover mutually exclusive cancer drivers and novel putative drivers despite their relatively low mutation rates in the breast cancer dataset.

Taken together, our results demonstrated the utility of our fast statistical test for mutual exclusivity and showed that the unrestricted analysis of mutual exclusivity enabled by the method can provide valuable biological insights.

2 Method: WeSME

Both hypergeometric and permutation based methods have been popular choices for estimating the significance of mutual exclusivity. The drawback of the permutation-based method is the computational cost, thus limiting gene sets to be considered and/or precision. On the other hand, the hypergeometric test does not take into account different mutation frequencies of patients and often underestimates the significance of exclusivity. WeSME aims to reduce the computational cost of estimating the significance of mutual exclusivity by utilizing weighted samplings instead of performing costly permutations of a mutation matrix, while accurately computing P-values by including patient mutation frequency information in sampling probability (see the discussion on the computational costs in Supplementary Section S-A).

2.1 General description of WeSME method

WeSME first computes the mutation frequencies of patients based on the observed distribution (Fig. 1A and Algorithm 1 in Supplementary Section S-E). Let G = {g₁, g₂…g_m} and S = {s₁, s₂ … s_n} be the set of all genes and cancer patients, respectively. Let mf(s) be the observed mutation frequency of a patient s. Next we obtain the null distribution of the mutation pattern of a gene by performing samplings of patients based on the weights (mf(s₁), mf(s₂) … mf(s_n)) (Algorithm 2). Specifically, for a gene g in G with m(g) mutated patients, we randomly sample m(g) patients without replacement while making sure that the probability of a patient s being sampled is proportional to the observed mutation frequency mf(s) (i.e. sampling weighted by mf(s)). Let ($R_g^1,R_g^2,R_g^3\dots R_g^{\text{W}}$) denote the sampling results when W weighted samplings were performed for gene g. $R_{\text{g}}^l$ corresponds to the random subset of patients for gene g in the lth weighted sampling and the size of $R_{\text{g}}^l$ must be m(g) for all l. Note that the samplings depend only on m(g) and W, and therefore two genes g₁ and g₂ with m(g₁)=m(g₂) can reuse the same samplings. We use this fact to further reduce the sampling time (see Section 2.2 and Supplementary Section S-A).

Fig. 1.

Illustration of WeSME method

To estimate the P-value of mutual exclusivity of two genes g_i and g_j in G, we compare the number of patients exclusively covered by the two genes and the numbers of such patients from the null distribution (Algorithm 3). Formally, for two subsets of patients S_i and S_j, let us define ExCover(S_i, S_j) to be |XOR(S_i, S_j)|. Suppose that S(g_i) is the set of patients who have mutations in gene g_i. Then ExCover(S(g_i), S(g_j)₎ represents the exclusive cover size by either gene g_i or g_j. As in the permutation-based method, the empirical P-value of mutual exclusivity of g_i and g_j is given as

$$\left(g_{\text{i}},g_{\text{j}}\right)=\left|\{l=1\dots W\right|ExCover\left({\text{R}}_{\mathrm{i}}{}^l,{\text{ R}}_{\text{j}}{}^l\right)>ExCover\left(S\left(g_{\text{i}}\right),S\left(g_{\text{j}}\right)\right)\left\}\right|/W$$

where ($R_i^1,R_i^2,R_i^3\dots R_i^{\text{W}}$) and ($R_j^1,R_j^2,R_j^3\dots R_j^{\text{W}}$) are the random samplings of g_i and g_j, respectively. We note that the precision of P-values depends on the number of weighted sampling pairs W. In Section 2.2, we discuss how we increase W to obtain a better precision without increasing the computational time significantly.

One noteworthy advantage of using WeSME over the permutation method is that except Algorithm 1, the analysis can be carried out for a subset of genes independently from the rest of genome whereas in the permutation method, the whole genome permutations need to be performed to obtain random profiles.

2.2 WeSME provides a fast and accurate approximation to the permutation based method

In this section, we discuss some optimization techniques to further reduce the computational costs of WeSME and present the results applied to TCGA BRCA dataset.

Compact sampling: given the observed mutation frequency vector (mf(s₁), mf(s₂), … mf(s_n)) of patients, the weighted sampling of a gene g depends only on the number of mutated patients m(g) and the number of samplings W. Utilizing this fact, we ran Algorithm 2 for each different m(g) instead of running for each gene. More specifically, let K be {k | m(g) = k for some g}. We performed W weighted samplings for each different k in K (denoted as WS(k)). As discussed in Supplementary Section S-A, we found that it reduces the sampling time and storage requirement significantly.

Dynamic depth of precision: Suppose that we run Algorithm 3 for two genes g_i, g_j. Let k_i=m(g_i) and k_j= m(g_j) and let ($R_{\text{i}}^1,R_{\text{i}}^2,R_{\text{i}}^3\dots R_{\mathrm{i}}^{\text{W}}$) and ($R_{\text{j}}^1,R_{\text{j}}^2,R_{\text{j}}^3\dots R_{\mathrm{j}}^{\text{W}}$) denote the weighted random samplings for k_iand k_j respectively. Then there are at least W(W - 1) different ways to create the pairs (discarding $R_{\text{i}}^l$ and $R_{\text{j}}^l$ when k_i=k_j in which case $R_{\text{i}}^l$ and $R_{\text{j}}^l$ are identical) and compute ExCover. For added randomness and optimizing the running time, we started by randomly selecting a small number of pairs (W₀) and compute ExCover for them initially. For example, we start with W₀ =1000 pairs at first and if empirical P-value < 0.1 (there are less than 100 random pairs whose ExCover is bigger than the original ExCover), we increased the number of sampling pairs by 10 times to obtain a better precision. The maximum precision we used in this analysis is 10 ^− 6 (i.e. the maximum number of pairs =10⁶ ≤ W(W - 1))

Analysis of TCGA BRCA somatic mutations: we ran WeSME with the somatic mutation profile of 665 TCGA BRCA patients. We considered all genes with at least one mutation without pre-filtering and computed the P-values of all gene pairs. Figure 2A shows the comparison of the P-value distributions obtained by WesME, permutation test and hypergeometric test for the gene pairs with their P-values ≤ 0.25. The P-values of the permutation-based method were computed using 1000 randomly permuted mutation profiles (permuted while preserving the mutation frequencies of patients and the mutation rates of genes). We found that the results obtained with WeSME approximate the ones with the permutation test very well (relative error rate ≅ 0.17, Pearson corr. coef. ≅ 0.98) while the hypergeometric test, in general, underestimates the mutual exclusivity (relative error rate ≅ 0.97, Pearson corr. coef. ≅ 0.89). We noted that WeSME P-values tend to be slightly more significant than the P-values based on the permutation test especially for the gene pairs with significant P-values (see more discussion in Supplementary Section S-B). Still WeSME approximates the permutation test much better than the hypergeometric test for the significant gene pairs (for pairs with P-value ≤ 0.05, relative error rate—0.30 versus 3.3, Pearson corr. coef.—0.89 versus 0.76). This trend of over-estimation was less prominent when we considered all gene pairs in HumanNet (http://www.functionalnet.org/humannet/) (Fig. 2A inset, the whole spectrum of P-values are shown).

Fig. 2.

(A) Comparison of the P-values computed based on permutation, hyergometric and WeSME tests. Only the data for the pairs with their P-values ≤ 0.25 are shown. The P-values for the gene pairs in HumanNet are shown in the inset. (B) The plot shows the WeSME P-values (P ≤ 0.25) and their corresponding FDRs. FDRs are computed separately depending on the mutation rates of genes

False discovery rate (FDR): in order to control FDR, we obtained the null distribution by generating 100 permuted mutational profiles and computing WeSME P-values for all gene pairs with the permuted mutations (Fig. 1B and Algorithm 4). Observing that the null distribution of P-values showed different patterns depending on the mutation rates of involved genes, we divided the genes into two groups based on their mutation rates (highly or rarely mutated), which in turn divided all gene pairs into three different bins. The null distribution for each bin was created separately using the P-values from the permuted mutation profiles. We then computed the FDR of the P-value for the gene pairs in each bin by comparing them with the null distribution for the corresponding bin. Figure 2B shows the different FDRs obtained using three bins—(R, R), (R, H) and (H, H) for BRCA where R (Rarely mutated) includes all genes with the mutation rate of 2% or lower, and the remaining genes are in H (Highly mutated). There were no pairs in bin (R, R) with P-value ≤ 0.25. In Figure 2B, we can observe that the pairs in bin (R, H) have higher FDRs (shown in purple) than the pairs in bin (H, H) (shown in cyan) for the same P-values.

2.3 Mutual exclusivity network in BRCA and UCEC

Figure 3A shows the network with significant mutually exclusive (ME) edges created with BRCA somatic mutation profiles. We included the gene pairs with P value ≤ 0.01 for highly mutated gene pairs in (H, H) and applied a stricter P-value threshold of 0.001 for pairs in (R, H). The corresponding FDR ≤ 0.125. We also computed the ME network with TCGA endometrial carcinoma (UCEC) somatic mutation data with a stricter FDR cut-off of 0.0025 (Fig. 3B). The more stringent FDR cutoff was used because the analysis of UCEC patients resulted in a larger number of ME pairs, presumably due to its partition into subtypes with drastically different mutational profiles. Additional list of ME pairs with FDR ≤ 0.01 is included in Supplementary Section S-F. We also performed the analysis with TCGA Acute Myeloid Leukemia dataset and the resulting ME network is shown in Supplementary Section S-D.

Fig. 3.

ME networks for (A) BRCA (FDR ≤ 0.125) and (B) UCEC (FDR ≤ 0.0025). The node size of each gene represents the mutation rate of the gene for the given cancer type. The widths of edges correspond to their WeSME P-values (thicker edges for more significant pairs) whereas darker colors of the edges correspond to lower FDRs. The networks are created using Cytoscape (Shannon et al., 2003)

The BRCA network (Fig. 3A) contains 34 mutually exclusive gene pairs. There were 13 pairs where one of the genes in the pair was mutated in 3% or less patients including two pairs where the mutation rate was below 2%. These genes would have been filtered out if 3% of mutation rate threshold was applied. Interestingly, these pairs include five pairs of well-known drivers: (ARID1A, TP53), (MTOR, PIK3CA), (AKT1, PIK3CA), (AKT1, TP53), (CTCF, TP53); four pairs where one member is a known cancer driver and its partner is a likely (and previously implicated) cancer driver: (TP53, MED23), (PIK3CA, PEG3)(Su et al., 1999), (CBFB, TP53) (Banerji et al., 2012) (USP36, PIK3CA) (Sun et al., 2015); and three pairs where one member is a long gene—Titin (TTN) or Cytoplasmic Dynein 2 Heavy Chain (DYNC2H1)—while its partner is a known cancer driver (TTN, ARID1A), (AKT1, TTN), (DYNC2H1, PIK3CA). Previous work suggested that TTN and DYNC2H1 are unlikely to be cancer drivers (Lawrence et al., 2014). Yet, TTN has rather a large number of partners in the mutual exclusivity network, suggesting that patients carrying mutations in this gene may have some distinguishing properties.

As for endometrial carcinoma, the previous analysis on TCGA UCEC dataset identified 12 genes (PTEN, TP53, PIK3CA, PIK3R1, ARID1A, ARID5B, KRAS, CTCF, CTNNB1, FBXW7, PPP2R1A, RPL22) as recurrently mutated in different subtypes (Cancer Genome Atlas Research et al., 2013). All but ARID5B and RPL22 are included in the ME network obtained with FDR cut-off of 0.0025. RPL22 and TP53 pair had FDR = 0.00256 while the most significant ME partner of ARID5B was CTNNB1, for which FDR = 0.02. Though overlooked in the TCGA publication, TRRAP (transformation/transcription domain-associated protein) is likely to be an important driver. The gene encodes a protein that plays a role in transcription and DNA repair. With our very stringent FDR cut-off, TRRAP is mutually exclusive with CTNNB1 and PIK3CA. Finally, as in the case of BRCA, the TTN gene has four neighbors, all of which are prominent cancer drivers.

In our subsequent analysis in Section 3, we show that although TTN may not be a cancer driver itself, it is likely to be a passenger of a mutagenic process that is specific to a subgroup of patients in these two cancer types.

3 Results: beyond mutual exclusivity of drivers—discovering driver signatures from mutual exclusivity partners of passengers

3.1 APOBEC and POL ε mutational signatures of TTN

Recent studies identified several genes that are highly mutated but still unlikely to be cancer drivers (Lawrence et al., 2013). These genes are often extremely long genes, including the muscle protein Titin (TTN), gel forming mucins (MUC) and ryanodine receptors (RYR1). Those genes were mostly excluded from candidate cancer driver lists in subsequent analyses. However, some of the apparent false-positive findings may not be simply caused by their lengths since the methods for identifying significantly mutated genes typically account for gene length information.

TTN is one of the genes identified as the ones that are not likely a cancer driver despite its high mutation rate but the prominent position of TTN in both BRCA and UCEC mutual exclusivity network strongly suggests that the mutagenesis of TTN may provide valuable information about patient subgroups. That is, even if TTN is not a cancer driver, the gene may be a passenger that is mutated in a specific group of patients and be a ‘witness’ of a mutagenic process occurring in the subgroup.

To confirm the hypothesis that TTN is a witness of a particular mutational process, we analyzed the mutational spectrum of this gene for both cancer types (BRCA and UCEC), and compared it to the signatures of mutational processes in human cancer (Alexandrov et al., 2013). We found that the mutational signature of TTN in BRCA dataset was consistent with the APOBEC signature. It showed a significant bias towards C > T and C > G mutations—67 and 44 mutations respectively (38% and 25% of 175 mutations in total and 50 and 33% of 132 C cite mutations). In addition, most of these mutations are in the context where C is preceded by A or T (71%).

In UCEC dataset, on the other hand, TTN has a different mutation signature, which is consistent with Pol ε mutations. TTN showed a strong preference for C > A and C > T mutations, having 231 and 347 mutations respectively, which represents 30 and 46% of 761 mutations in total and 40 and 59% of C cite mutations. Moreover, 123 of C > A mutations (53%) are TCT > TAT and 116 mutations of C > T mutations (33%) are TCG > TTG, strongly demonstrating that TTN mutations in UCEC is associated with Pol ε mutation signature.

3.2 Mutual exclusivity patterns of TTN and TP53 indicate a group of genes predictive of patient survival and suggest a role of APOBEC3 activity in TP53 mutations

ME network and survival analysis: An interesting property of mutational landscape of breast cancer is that frequently mutated genes such as TP53, PIK3CA, TTN and GATA3 in breast cancer do not have a strong discriminative power of patient survival (only TP53 mutations are weakly discriminative, P-value = 0.08). We therefore asked the question whether the mutual exclusivity network could suggest a set of genes with a better discriminative power.

Previous studies have suggested that APOBEC3B activity correlates with unfavorable prognosis in breast cancer. For example, it has been shown that APOBEC3B gene expression correlates with the Genomic Grade Index (GGI) score and that high APOBEC3B gene expression was associated with recurrence after treatment (Cescon et al., 2015). According to a separate study, APOBEC3B was associated with poor survival in ER +  breast cancer patients (Periyasamy et al., 2015). Building on our observation that TTN mutations are largely due to APOBEC3B activity (as discussed in the previous section), we hypothesized that putative cancer drivers exclusive with TTN mutations would be characterized by better prognosis. This was indeed confirmed in our survival analysis. We found that four genes that are exclusive with TTN (MAPK3K1, NCOA3, AKT1, ARID1A) were a good predictor of survival jointly and also predict disease free survival after treatment very well (Fig. 4B). Interestingly, TTN and TP53 share the discriminative gene set as their direct ME neighbors in the BRCA ME network (Fig. 3A). On the other hand, adding other putative drivers exclusive with TP53 but not with TTN (FOX1A, MED23, MAP2K4) (CTCF was not included because it shows a slight tendency of exclusivity with TTN, P-value < 0.15) reduced the discriminatory power of the gene set.

Fig. 4.

Mutual exclusivity of TTN and TP53 defines a group of related genes that are predictive of survival. (A) Genes mutually exclusive with TP53 (purple) and TTN (blue) with FDR cutoff of 0.125 (except of FOXA1 for which P-value is 0.00365 and FDR = 0.3) with their context of functional interactions based on (Tamborero et al., 2013). The interactions between FOXA1 and GATA3 are based on (Theodorou et al., 2013), and the relationships between EP300 and SWI/SNF and between CTCF and SWI/SNF are from (Dallas et al., 1998; Ogiwara et al., 2011) and (Euskirchen et al., 2011), respectively. (B) Cancer driver genes AKT1, NCOA3, ARID1A and MAP3K1 that mutually exclusive with both TTN and TP53 provide a better survival prognosis. (C) Comparison of mutation frequencies for patients with TTN and TP53 mutations. The numbers in the Venn diagram represent the number of patients and the average mutation frequencies of the corresponding groups are shown above the diagrams

Role of APOBEC3B activity in TP53 mutations: The similarity of exclusivity patterns of TTN and TP53 is consistent with the recently discovered relationship between APOBEC enzymatic activity and TP53 mutations (Burns et al., 2013). In their research, Burns et al. hypothesized that APOBEC3B-catalyzed deamination provides a chronic source of DNA damage that could cause TP53 mutations.

To support the model proposed by Burns et al., we need to test if it is possible that the mutations of both genes in BRCA depend on APOBEC activity but are independent of each other after controlling for their dependency on APOBEC. That is, under this model, we expect TTN and TP53 to co-occur but the significance of co-occurrence to disappear after correcting with mutation frequencies of patients. To test if this is the case, we developed WeSCO by adopting the same sampling technique as WeSME for estimating the significance of co-occurrence of mutations corrected with patient mutation frequency. Indeed, we did not find a significant co-occurrence (Fig. 4C, P-value = 0.29) of TTN and TP53 when applying WeSCO test while the co-occurrence was highly significant without the correction ( P-value = 0.0002, hypergeometric test).

Although it is currently not known how APOBEC activity leads to TP53 mutations, we conjecture that immune response might be the precursor step leading, via APOBEC activity, to TP53 mutation for some cancer patients. It has been demonstrated that the AID/APOBEC family plays important roles in adaptive and innate immunity (Wang et al., 2012) and APOBEC up-regulation is correlated with HPV-positive status, which implicates an HPV-mediated mechanism of APOBEC3B up-regulation for these cancer types (Henderson et al., 2014). In particular, the authors provided evidence that APOBEC activity is responsible for the generation of helical domain hot spot mutations in PIK3CA across multiple cancers. The key role of APOBEC family in natural and adaptive immunity strongly suggests that APOBEC up-regulation may be also related to immune response in non-HPV mediated cases and our results support a similar mechanism for TP53 mutations in some breast cancer cases (but not all cases as discussed below).

Underlying causes of TP53 mutations may be heterogeneous: We observed that the subgroup of patients with TP53 mutations but no TTN mutations had significantly lower mutation frequencies compared to the subgroup of patients with TTN mutations (Fig. 4C;P-value = 0.025). This suggests that the TP53 only group also contains patients whose TP53 mutations are not due to APOBEC activity as otherwise we would not expect to see a difference in mutation frequencies between the TP53 only and the TTN patient groups. The difference of mutation frequencies in the group suggests that underlying causes of TP53 mutations in breast cancer are heterogeneous.

Interestingly, TP53 and TTN mutations are mutually exclusive in the analysis of endometrial carcinoma data where TTN mutations show Pol ε mutation signature (Fig. 3B). This observation suggests that the mutations in TP53 are generally not caused by the mutagenic process related to Pol ε mutations.

In summary, our results are consistent with the previously proposed model in which APOBEC3B can be a source of mutations in TP53 for highly mutated cancer cases, however reject the model in which this is true across all breast cancer patients with TP53 mutations.

4 Conclusions

We performed a whole genome scale analysis of mutual exclusivity of genes with mutations in breast cancer and endometrial carcinoma. For this purpose, we first developed WeSME—a new, fast method to estimate statistical significance of mutual exclusivity. Similar to the permutation based method, WeSME controls mutation frequencies of patients and mutation rates of genes in the null model, but can be performed quickly with a nearly arbitrary level of precision. The P-values estimated by WeSME closely matched the P-values obtained with the permutation-based method. In addition, our proposed method to control FDR takes into account the differences of P-value distributions from gene pairs with different mutation rates.

The development of WeSME allowed us to perform a large-scale mutual exclusivity analysis without restricting it to highly mutated genes or genes known a priori to be cancer drivers. The approach led us to uncover mutually exclusive cancer drivers, some of which have relatively low mutation rates. In addition, our mutual exclusivity analysis allowed us to pinpoint differences between mutagenic processes in different cancer groups. In particular, the inclusion of TTN (a very long and frequently mutated gene) not only allowed us to observe the patterns of APOBEC activity in breast cancer and Pol ε mutation signature in endometrial carcinoma but also help identify a set of driver genes that is highly predictive of patient survival. TTN was considered to be an artifact and typically filtered out in cancer analyses.

Taken together, we developed a tool enabling a fast genome-wide analysis of mutual exclusivity of mutations in cancer and used it to demonstrate that the genome-wide mutual exclusivity analysis can provide information beyond relationships between cancer drivers.

Funding

This work was supported in part by the Intramural Research Program of the National Institutes of Health, National Library of Medicine.

Conflict of Interest: none declared.