DGPathinter: a novel model for identifying driver genes via knowledge-driven matrix factorization with prior knowledge from interactome and pathways

Somatic mutation datasets

For the somatic mutation data of cancers, we focus on three types of cancers from TCGA datasets, which include 507 tumors samples from breast invasive carcinoma (BRCA) (Cancer Genome Atlas Network, 2012), 83 tumor samples from glioblastoma multiforme (GBM) (Cancer Genome Atlas Research Network, 2008) and 401 tumor samples from thyroid carcinoma (THCA) (Cancer Genome Atlas Research Network, 2014). The somatic mutation data are downloaded from cBioPortal database (Gao et al., 2013). The somatic mutation data are formed as a binary matrix (sample × gene) X_n×p (Bashashati et al., 2012; Hofree et al., 2013; Kim, Sael & Yu, 2015), where n is the number of samples and p is the number of the tested genes. An entry of the matrix being 1 denotes a mutation occurs in the respective gene and tumor sample, when compared with the germline (Bashashati et al., 2012; Hofree et al., 2013; Kim, Sael & Yu, 2015). The network information used in this study is iRefIndex (Razick, Magklaras & Donaldson, 2008), a highly curated interaction network containing 12,129 nodes (genes) and 91,809 edges (interactions). For the pathway information, we follow previous studies (Park et al., 2015) and use the curated pathways from three databases, KEGG (Ogata et al., 1999), reactome (Joshi-Tope et al., 2005) and BioCarta (Nishimura, 2001), which are also downloaded from the previous study (Park et al., 2015).

Model of knowledge-driven matrix factorization

To efficiently identify potential driver genes from somatic mutation data, we use a knowledge-driven matrix factorization framework, which can successfully integrated information from pathways and interaction networks. A brief overview of DGFathinter is illustrated in Fig. 1. Since many matrix factorization-based methods have been used for detecting abnormal genes from heterogeneous tumor samples (Lee et al., 2010; Sill et al., 2011; Zhou et al., 2013, 2014; Xi & Li, 2016; Xi, Li & Wang, 2017), we introduce matrix factorization framework into DGPathinter. The matrix factorization-based methods factorize the data matrix as the multiplication of a low-rank sample matrix and a low-rank gene matrix (Lee et al., 2010; Sill et al., 2011; Zhou et al., 2013, 2014; Xi & Li, 2016; Xi, Li & Wang, 2017), where the entries of the sample matrix indicate the assignments of different samples to different subsets and the entries of the gene matrix indicate the scores of the abnormal genes in the related subsets of samples. In our previous study (Xi, Li & Wang, 2017), the matrix factorization framework has been shown to be an appropriate framework for the task of detecting driver genes from mutation data of heterogeneous cancers. Here we denote the matrix G_k×p = (g₁, …, g_k) as the gene matrix and the binary matrix S_n×k = (s₁, …, s_k) as the sample matrix, and use their multiplication SG^T to approximate the mutation matrix X. The entries of G represent the mutation scores of the tested genes related to the set of tumor samples indicated by the sample matrix S, which represent the assignments of the tested sample in different sample sets. Here the number k is the rank of reconstruction matrix of the mutation matrix X. Due to the constraint that the entries of matrix S are binary value, we use Boolean constraint on matrix S (Malioutov & Malyutov, 2012), i.e., S ∘ (S − J) = 0, where the operator ∘ indicates Hadamard product of two matrix, and the matrix J denotes an n×k matrix with all the entries being 1. The fitting problem of the multiplication of the two matrices S and G and the mutation matrix X can be formulate as X ≃ SG^T + ε, where the ε is the residual matrix between the matrix X and the multiplication SG^T, and matrix S is subject to the equality restriction S ∘ (S − J) = 0. By estimating the mutation scores of the tested genes in the matrix G from information of somatic mutation data, pathways and interaction network, we can identify the potential driver genes by ranking their mutation scores. The strategies of incorporating pathway and network information are present below.

To make the driver gene prioritization procedure in our model driven by prior knowledge from pathways, we introduce a non-negative matrix V_m×k as the pathway score matrix. The row number of the matrix m is total number of pathways used in our model. The column vectors in the matrix V = (v₁, …, v_k) represent the scores of the pathways, and a higher score of a pathway indicates a larger potential that the pathway is dysregulated in the related set of tumor samples. To incorporate pathway information into gene scores for different sets of samples, we project the gene scores onto their related pathways and maximize the covariance between the projection scores and pathway scores as $R_C=-{\sum }_{j=1}^k\text{Cov}(F{g}_j,v_j)=-\text{Tr}\left\{G^{T}FV\right\}$ (Chen & Zhang, 2016), where the matrix V is subject to the inequality restriction V ≥ 0. Here the matrix F_m×p represents the relationships of the tested genes and their related pathways. The entry F_ij equaling 1 denotes that the jth gene belongs to the ith pathway. In addition, to avoid an overfitting problem, we also use Frobenius norm-based regularization on the pathway scores V as $R_V={\Vert V\Vert }_F^2$ (Pan et al., 2008). Furthermore, to integrate interaction network information into our model, we utilize Laplacian regularization to encourage the smoothness between the scores of the interacted genes (Xie, Wang & Tao, 2011). The regularization term is formulated as R_L = Tr{G^TLG}, where the matrix L = D − A is the Laplacian matrix of interaction network, the matrix A is the adjacency matrix, and the matrix D is its corresponding degree matrix.

Consequently, we estimate the mutation scores of the tested genes in gene matrix G, pathway score matrix V and sample indicator matrix S by optimizing an optimization function of a knowledge-driven model with integrated data fitting and regularization terms formulated as: (1) $$\begin{array}{l}\underset{S,G,V}{\text{min}}\frac{1}{2}\Vert X-S{G}^{\text{T}}{\Vert }_F^2-{\text{λ}}_C\text{Tr}\left\{G^{\text{T}}{F}^{\text{T}}V\right\}+\frac{1}{2}{\text{λ}}_L\text{Tr}\left\{G^{\text{T}}LG\right\}+\frac{1}{2}{\text{λ}}_V\Vert V{\Vert }_F^2\hfill \\ \text{\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}s}.\text{t}.S\circ (S-J)=0,V\ge 0\hfill \end{array}$$ where λ_C, λ_L and λ_V are used to balance the data fitting, the coherence gene scores and pathway scores according to their relations, the smoothness between scores of interacted genes and the regularization term of pathway scores. The tuning parameters λ_C, λ_V and λ_L are empirically set to 0.01, 0.01 and 0.1, respectively. We have also investigated the results of DGPathinter when the three parameters changes. For the three parameters, we can see that the detection results of the top 100 genes show little variance when the tuning parameters are changed (Figs. S1–S3), demonstrating the robustness of our model with respect to these parameters. In Fig. 1, we illustrate an overview of DGPathinter through schematic diagram.

Optimization of knowledge-driven matrix factorization

Due to the equivalence between the matrix multiplication SG^T and the summation of multiple rank-one layers ${\sum }_{i=1}^k{s}_i{g}_i^T$, we incorporate a layer-by-layer procedure to solve the optimization problem iteratively (Lee et al., 2010; Sill et al., 2011; Xi & Li, 2016). Note that the first layer is the best rank-one estimation of the data matrix. We estimate the first layer by minimizing the following objective function (2) $$\begin{array}{l}\underset{s_1\mathrm{,}{g}_1\mathrm{,}{v}_1}{{\displaystyle \text{min}}}\frac{1}{2}{\Vert X-s_1{g}_1^{\text{T}}\Vert }_F^2-{\text{λ}}_C\text{Tr}\left\{g_1^{\text{T}}{F}^{\text{T}}{v}_1\right\}+\frac{1}{2}{\text{λ}}_L\text{Tr}\left\{g_1^{\text{T}}L{g}_1\right\}+\frac{1}{2}{\text{λ}}_V{\Vert v_1\Vert }_F^2\\ \text{\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}s}.\text{t}.s_1\circ (s_1-1_{\text{n}})=0,v_1\ge 0,\end{array}$$ where s₁, g₁ and v₁ are the first column vectors of matrices S, G and V respectively, and 1_n×1 indicate a vector with all coefficients being 1. The $v_1^T{v}_1$ is the inner product of vector v₁, which is equivalent to squared Frobenius norm of the vector.

We then apply an alternatively strategy to estimate the three vectors s₁, g₁ and v₁ in Eq. (2). When the other two vectors v₁ and s₁ are fixed, the minimization problem for the mutation score vector g₁ can be reformulated as below: (3) $$\begin{array}{l}{\min }_{{\text{g}}_1}\frac{1}{2}\Vert s_1{\Vert }_2^2{g}_1^{\text{T}}{g}_1-(X^{\text{T}}{s}_1)g_1-{\text{λ}}_C{(F^{\text{T}}{v}_1)}^{\text{T}}{g}_1+\frac{1}{2}{\text{λ}}_L{g}_1^{\text{T}}L{g}_1.\hfill \end{array}$$

Through Karush–Kuhn–Tucker (KKT) conditions, the mutation score vector g₁ can be estimated as (4) $$\begin{array}{l}{g}_1\leftarrow {\left({\Vert s_1\Vert }_2^2{I}_p+{\text{λ}}_{L}L\right)}^{-1}\left(X^{\text{T}}{s}_1+{\text{λ}}_C{F}^{\text{T}}{v}_1\right)\mathrm{,}\hfill \end{array}$$ where I_p is a p×p identity matrix.

Likewise, the optimization function to solve the pathway score vector v₁ in optimization problem in Eq. (2) is formulated as (5) $$\begin{array}{l}\underset{v_1}{{\displaystyle \text{min}}}\frac{1}{2}{\text{λ}}_V{v}_1^{\text{T}}{v}_1-{\text{λ}}_C(F{g}_1{)}^{\text{T}}{v}_1\\ \text{\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}s}.\text{t}.v_1\ge \mathrm{0,}\end{array}$$ which is a non-negative quadratic programming problem. The estimation of the vector v₁ in Eq. (5) can be calculated as: (6) $$\begin{array}{l}{v}_1\leftarrow {\{({\text{λ}}_C/{\text{λ}}_V)F{g}_1\}}^{+}\mathrm{,}\hfill \end{array}$$ where {·}⁺ is an operator which replace the negative coefficients of the input vector with zeros.

For sample indicator vector s₁, the optimization function of Eq. (2) is formulated as a Boolean constraint problem (7) $$\begin{array}{c}\underset{g_1}{{\displaystyle \text{min}}}\frac{1}{2}{\Vert g_1\Vert }_2^2{s}_1^{\text{T}}{s}_1-(X{g}_1)s_1\\ \text{\hspace{1em}s}.\text{t}.s_1\circ (s_1-1_n)=0.\end{array}$$

Through KKT conditions, the problem in Eq. (7) can be solved as: (8) $$\begin{array}{l}{s}_1\leftarrow I_{[0,+\infty )}(X{g}_1-\frac{1}{2}{\text{‖}{g}_1\Vert }_2^2)\hfill \end{array}$$ where I_Φ (z) is indicator function, of which the coefficients of the output vector are assigned to 1 when the corresponding coefficients of input vector z belongs to the set Φ, and 0 otherwise. Consequently, the minimizing function is optimized by alternatively estimating the three vectors g₁, v₁ and s₁ in Eqs. (4), (6) and (8) until convergence (Pseudo-code in Table 1).

After convergence, the first rank-one layer $s_1{g}_1^T$ from the mutation matrix X, along with the related pathway score vector v₁, are obtained. Since the cancer data may display heterogeneity, it is not sufficient to utilize only one layer to fit the mutation data matrix. Subsequently, we apply the one layer estimation strategy aforementioned on the remaining samples to obtain the next layer. When the mutation matrix is factorized iteratively until no sample remains, we can obtain the rank number k automatically (Lee et al., 2010; Sill et al., 2011; Xi & Li, 2016). The multiple layers estimation yielded by our model can effectively incorporate information from the mutation matrix, the interaction network and pathways.

Experimental design and evaluation

For the driver gene prioritization of our approach, we select the maximum entries of each row of gene matrix G as the score of the tested genes to be potential driver genes, which represent the intensities of the mutation of tested genes among different sets of tumor samples. We then prioritize the investigated genes according to their mutation scores and select the top ranked genes as potential driver genes. Due to the lack of gold standard for driver genes that are generally accepted, we further evaluate the detected genes via a list of known benchmarking cancer genes from a highly curated database, the Network of Cancer Genes (NCG4.0) (An et al., 2014). The NCG4.0 gene list contains both experimentally supported cancer genes from the Cancer Gene Census (CGC) (Futreal et al., 2004) and statistical inferred candidate genes from previous studies (An et al., 2014). The cancer specific genes from the two benchmarking gene lists are used to assess the prioritizing results of the investigated methods.

By using these benchmarking genes as ground truth genes in the evaluation studies, we firstly compute the precisions and recalls under different rank thresholds and draw precision–recall curves of the competing methods, where a curve closer to the top and right indicates a better performance (Wu, Hajirasouliha & Raphael, 2014; Yang et al., 2017). The precision is calculated as the fraction of selected genes that are also benchmarking genes, and the recall is computed as the fraction of benchmarking genes that are selected by the rank threshold. Next, we calculate the average rank of known genes in the prioritization results to comprehensively assess the prioritization performance, which is a traditional metric for evaluating the performance of retrieval (Ma & Zhang, 1998; Gargi & Kasturi, 1999; Müller et al., 2001). Furthermore, we select the top 100 genes from the results of the competing methods, and compare the proportions of known driver genes detected by the competing methods. Fisher’s exact test is also applied on the results, which can evaluate whether the selected genes are significantly enriched for known driver genes by p values of the test. In addition, for the highly scored pathways given by our approach, we also investigate whether these pathways are correlated with cancer progressions.

Driver gene identification

To evaluate the identification performance of DGPathinter, we compare our model with three existing methods, HotNet2 (Leiserson et al., 2014), NBS (Hofree et al., 2013) and MUFFINN (Cho et al., 2016), on datasets of three types of cancers: BRCA, GBM and THCA. In the performance evaluation, HotNet2, NBS and MUFFINN are set to their default parameters. For MUFFINN, there are two versions (MUFFINN-DNmax and MUFFINN-DNsum) based on different strategies (Cho et al., 2016), and we use both versions in the comparison study. The interactome information of all the investigated methods is the iRefIndex gene interaction network from Razick, Magklaras & Donaldson (2008). The pathway information used in DGPathinter is from KEGG, Reactome and BioCarta (Ogata et al., 1999; Joshi-Tope et al., 2005; Nishimura, 2001). In DGPathinter, the genes are ranked by their mutation scores from the matrix G. In the identification result of HotNet2, a higher delta score of a gene indicates a larger potential of being driver genes (details in Supplemental Information). For NBS, the genes are sorted according to their scores in the NBS profiles. In MUFFINN, the prediction scores of genes for MUFFINN-DNmax and MUFFINN-DNsum are used to prioritize the genes. To give a comprehensive view of the identification performance, we analyze all the investigated genes by precision–recall curves and average ranks of known driver genes over the prioritization results. For the top ranked genes, we further compare the fractions of known benchmarking genes among the results of these methods, and their related p values of Fisher’s exact test. Venn diagrams of the top ranked genes among the competing methods are also analyzed.

Performance comparison

The overall performance of DGPathinter, HotNet2, NBS, MUFFINN-DNmax and MUFFINN-DNsum are illustrated as precision–recall curves in Fig. 2. When we use the known benchmarking cancer genes in NCG4.0 as a gold-standard, the precision–recall curves of DGPathinter are located over the other curves clearly for all the three types of cancers, indicating that DGPathinter yields the best identification performance among the four results on the datasets of the three types of cancers. Taking BRCA result as an example, the precisions of DGPathinter, HotNet2, NBS, MUFFINN-DNmax and MUFFINN-DNsum are 37.7%, 4.1%, 16.4%, 4.4% and 4.3% respectively when the recalls of the results are fixed at 5.0%. In the GBM results, the precisions of GBM-specific NCG4.0 genes are 37.6% for HotNet2, 45.8% for NBS, 0.8% for MUFFINN-DNmax and 3.6% for MUFFINN-DNsum when the recalls are 5.0%. In comparison, DGPathinter achieves a precision of 100.0% in the same situation. For the known experimental validated driver genes curated by CGC, we also draw the precision–recall curves of the four investigated results for the CGC gene list. In consistency with the NCG4.0 results, a similar phenomenon can also be observed that the identification performance of our approach outperforms the results of the other competing methods (Fig. S4). For example, DGPathinter gives a precision of 77.5% on the BRCA data and 100.0% on the GBM data when recalls are at 10.0%, which are also higher than those of the other competing methods in the same situation. To assess whether and to which extent the difference between the performance of our method and previous approaches is statistically significant, we apply the non-parametric Friedman test on the areas under the curve (AUCs) of precision–recall curves among the three investigated cancers (Table S1). The AUCs for known NCG4.0 and CGC genes yields p values of 0.03 and 0.02, respectively, indicating that the difference between the performance of the investigated methods is statistically significant.

We also evaluate average rank of the known cancer genes predicted by the investigated methods. For the NCG4.0 gene list, DGPathinter yields average rank of 67.5 for known breast cancer specific genes, which is smaller than the those of 829.7 for HotNet2, 91.8 for NBS, 1731.4 for MUFFINN-DNmax, 1271.9 for MUFFINN-DNsum (Table 2) and 6064.5 for random selection. This result demonstrates that the benchmarking cancer genes in the result of our approaches are ranked much closer to the top in average, when compared with those of other three results. When we examine the CGC experimentally validated genes, our approach also yields the smallest average rank among the competing methods (Table 2). The average ranks for breast cancer specific CGC gene list are 67.5, 829.7, 91.8, 1731.4, 1271.9 and 6064.5 for DGPathinter, HotNet2, NBS, MUFFINN-DNmax, MUFFINN-DNsum and random selection, respectively. We also apply the non-parametric Friedman test on the average ranks of known NCG4.0 and CGC genes in the detection results of the competing methods, yielding p values of 0.02 and 0.02 respectively. The aforementioned investigations suggest that DGPathinter shows a promising capability for prioritizing known cancer genes than those of the other competing approaches.

Evaluation of top ranked genes

For the top ranked genes, the top 100 genes of the five prioritization results are selected for further evaluation. In BRCA result, the top 100 genes for HotNet2, NBS, MUFFINN-DNmax and MUFFINN-DNsum include 3, 15, 5 and 4 genes in NCG4.0 list respectively. In contrast to the top 100 genes for DGPathinter, there are 30 genes matched in the NCG4.0 benchmarking genes (Fig. 3). The p value of Fisher’s exact tests on the result of the DGPathinter on BRCA is 1.13–24, indicating that the selected NCG4.0 genes are significantly enriched for NCG4.0 genes. For the GBM results, DGPathinter, HotNet2, NBS, MUFFINN-DNmax and MUFFINN-DNsum identify 13, 12, 8, 1 and 1 NCG4.0 genes, with related p values of 3.63e−14, 1.01e−12, 1.91e−07, 4.68e−01 and 4.68e−01 respectively. Compared with the p values of the other identification results, DGPathinter yields the smallest p values among the competing results (Fig. 3). These results demonstrate that DGPathinter performs the best among these methods in detecting NCG4.0 benchmarking cancer genes. Furthermore, we investigate the numbers of selected top 100 genes that are also CGC experimentally validated driver genes. For BRCA data, there are 10, 1, 6 CGC driver genes detected by DGPathinter, HotNet2 and NBS, with p values of 6.88e−15, 2.01e−01 and 6.95e−08 respectively. For GBM datasets, there are nine CGC genes captured by DGPathinter, of which the Fisher’s exact test p value is 5.57e−15 and is smaller than those of other investigated methods (eight CGC genes for HotNet2 with p value of 6.57e−13, six CGC genes for NBS with p value of 4.43−e09 and one CGC gene for both MUFFINN-DNmax and MUFFINN-DNsum with p values of 1.39e−01). For THCA-specific driver genes detected by the competing methods, the NCG4.0 genes completely overlap the CGC genes, and it can be observed that DGPathinter achieves the best performance among the competing methods. When we apply the Friedman test on the proportions of known NCG4.0 and CGC genes in the top 100 genes detected by DGPathinter on the investigated cancers, we obtain p values of 0.05 and 0.02 respectively. We further change the number of top-ranked genes considered to see how the changes affect the results of the statistical test. For the proportions of known NCG4.0 and CGC genes in the top 200 genes of the results of the competing methods, the p values of Friedman test are 0.04 and 0.02 respectively; for the top 300 genes, the p values are 0.03 for known NCG4.0 genes and 0.02 for known CGC genes. These results demonstrate that there is a statistically significant difference between the performance of the competing methods.

Furthermore, we compare the top 100 genes among the five prioritization results and draw Venn diagram of their results for three types of cancers respectively (Fig. 4). For BRCA dataset, 14 genes identified by DGPathinter are also detected by at least of one of the other results. For example, GATA3 gene is identified by DGPathinter, HotNet2 and NBS. As reported in a previous study (Usary et al., 2004), variants in GATA3 gene may have contribution to tumorigenesis in ESR1-positive breast cancers. Another study also shows that GATA3 mutations have the potential to be associated with aberrant nuclear localization, reduced transactivation and cell invasiveness in breast cancers (Gaynor et al., 2013). For GBM dataset, there are totally 16 genes shared in the detection results of DGPathinter, HotNet2 and NBS, including CGC curated GBM specific driver genes TP53, PTEN, PIK3R1, EGFR and PIK3CA (Futreal et al., 2004), and NCG4.0 inferred GBM specific driver candidates ERBB2 and RB1 (An et al., 2014). For THCA dataset, HRAS is co-detected by DGPathinter, HotNet2 and NBS. Although HRAS is not a THCA specific driver gene curated by neither CGC nor NCG4.0, it is reported as driver gene for infrequent sarcomas and some other rare tumor types by CGC (Futreal et al., 2004).

Moreover, some genes detected by DGPathinter are also captured by MUFFINN-DNmax or MUFFINN-DNsum. Taking GBM results as an example, MDM4 gene is shared by the results of DGPathinter, MUFFINN-DNmax and MUFFINN-DNsum, which is reported to be the driver gene in bladder cancer, glioblastoma and retinoblastoma by CGC (Futreal et al., 2004). The KRAS gene is identified by DGPathinter, HotNet2 and MUFFINN-DNsum, which is also a driver gene in several types of cancers reported by CGC (Futreal et al., 2004). The FLT4 gene is included in the results of DGPathinter, NBS and MUFFINN-DNsum, and is curated as a driver gene of soft tissue sarcoma by CGC (Futreal et al., 2004). In addition to the genes shared by DGPathinter and other methods, there are also some genes unique to the results of DGPathinter. For example, TSH gene is a breast cancer driver gene curated by CGC, which is only detected by DGPathinter on the BRCA dataset. A number of CGC curated GBM specific driver genes, including AKT1, CDH1, MAP2K4, NCOR1 and TBX3 (Futreal et al., 2004), are also unique to the results of DGPathinter on the GBM dataset. For the results of THCA dataset, the CGC-curated THCA specific driver gene TERT is only identified by DGPathinter but not the other competing methods (Futreal et al., 2004). The full lists of the top 100 genes detected by DGPathinter on BRCA, GBM and TCHA, along with the methods that co-detect them, are demonstrated in Tables S2–S4, respectively.

Pathway analysis

In addition to driver gene identification, DGPathinter can also provide highly scored pathways during the driver gene detection processing. We further analyze the top 30 scored pathways in the results of DGPathinter, and find some well-known cancer related pathways such as P53 pathway, PTEN pathway, P38MAPK events pathway, ATM pathway (Table 3). In the results of the BRCA dataset, the top one pathway is the GATA3 pathway curated by the BIOCARTA database, which is reported to be highly associated with breast cancer. For example, the GATA3 pathway is reported to play an important role in reducing E-cadherin in breast cancer tissues (Tu et al., 2017). Meanwhile, the top ranked pathway in the GBM results is the RB pathway curated by BIOCARTA, which is also found in the BRCA results. Reported by previous studies (Chow et al., 2011; Sherr & McCormick, 2002), mutated RB1 pathway is one of the obligate events in the pathogenesis of glioblastomas. Especially, thyroid cancer pathway is found in the results of DGPathinter on THCA dataset, and glioma pathway is in the results on GBM dataset. Some other cancer-related pathways are also included in the lists of top ranked pathways, such as the GAB1 signalosome pathway, the signaling to RAS, the MTOR signaling pathway, the non-small cell lung cancer pathway, the melanoma pathway, the pancreatic cancer pathway, the prostate cancer pathway, the bladder cancer pathway and the endometrial cancer pathway.