Gene-microRNA network module analysis for ovarian cancer

Transcription

1 Gene-microRNA network module analysis for ovarian cancer Shuqin Zhang School of Mathematical Sciences Fudan University Oct. 4, 2016

2 Outline Introduction Materials and Methods Results Conclusions

3 Introduction Networks are widely applied to model different types of complex systems. Many networks have the module structure property. Intuitively, a module is a cohesive group of nodes that connected more densely" to each other than to the nodes in other modules. Modules may correspond to some functional units or play similar roles.

4 Figure : Human liver cohort(hlc) gene co-expression network X.Yang,B.Zhang, et al. Genome Research, 20, ,2010

5 Introduction MicroRNAs (mirnas) are small ( 22 nucleotides) non-coding RNAs that have emerged as key gene regulators. Each mirna is potentially able to regulate around 100 or more mrna targets and over 30% of all human genes are supposed to be regulated by mirnas. Identification and validation of mirna targets may lead to new therapeutic methods. Most of the current methods mainly considered the down-regulatory effects from mirnas.

6 Introduction One mirna may regulate many mrnas, and be regulated by several mirnas, thus the intertwined relationship between mirnas and mrnas becomes very complex. The systems tools such as networks should be more appropriate for studying the relationships between mirnas and mrnas. A few papers have been published to study the complex interactions between genes and mirnas from the system point of view by using networks, especially the network modules. Published methods: SNMNMF, PIMiM36, Mirsynergy, etc..

7 Our aim: to find the gene-mirna modules and study the functions and properties. We called it the co-module of genes and mirnas: (1) both genes and mirnas are in the module, (2) all of gene-gene, mirna-mirna, and gene-mirna connections appear in the module.

8 Data sets Materials and Methods We downloaded the gene expression for ovarian cancer from The Cancer Genome Atlas (TCGA). The gene expression data is generated with UNC AgilentG4502A_07_03. There are a total of genes with annotations for the 562 samples. We downloaded the mirna expression data for ovarian cancer from TCGA, which is generated with UNC mirna_8x15kv2. There are a total of 590 mirnas for the 595 samples. We selected the first 3200 genes with the largest expression variance greater than 1. We chose both data sets for the common 556 samples. We downloaded the gene-mirna interaction data from mirtarbase ( There are a total of gene-mirna interactions.

9 Network Construction Gene (mirna, gene-mirna) coexpression network: Compute the Pearson correlation coefficient. Construct the adjacency matrix by hard thresholding. If the absolute value of the Pearson correlation coefficient between genes (mirnas, gene-mirnas) is greater than some given value, we assign an edge between them. Choose the threshold: Gene coexpression network and mirna coexpression network: try different thresholds and compute the linear regression coefficient between the log10 transformed degree frequency of degree d (log10 f(d)) and d (log10 d) to make the network has approximately scale free property. Gene-miRNA coexpression network: it depends on the AUCs for the known gene-mirnas being clustered in the same module. Given a fixed threhold, we use our proposed method in the following to get the score of one gene and one mirna in the same module. We choose the threshold that achieves the highest AUC.

10 Method: Co-module Indentification A gm the gene-mirna correlation adjacency matrix. C is an adjacency matrix with both A gm and the known gene-mirna relations.

11 Method: Co-module Indentification We let S g be the assignment of the N g genes into K modules for the network G g, S g (i, k) = { 1, if vertex i Vk, 0, otherwise, where i = 1, 2,, N g ; k = 1, 2,, K, V k denotes the k-th module. Similarly, we define the assignment of the N m mirnas into K modules for the network G m as S m.

12 Method: Within network module identification Taking network G g as an example. We define K S g (., k) T (2A g D g )S g (., k) Ψ g (S g ) = S g (., k) T. (1) S g (., k) k=1 The optimization problem is formulated as: max Ψ g (S g ), s.t. S g (i, k) {0, 1}, K S g (., k) = 1. (2) k=1 By letting S g (., k) = Sg(.,k) S g(.,k) 2, the problem is relaxed to: max Ψ g ( S g ) = Tr( S g T (2Ag D g ) S g ) s.t. SgT Sg = I K.

13 Method: Co-module Indentification For the co-module, we expect that the genes and mirnas with dense connections are clustered into one module. We maximize to make it. S T g (.,k)cs m(.,k) S g(.,k) 2 S m(.,k) 2 Putting all terms together, our objective becomes: Ψ(S g, S m ) = Ψ g (S g )+Ψ m (S m )+λ K k=1 where λ controls the within and between network connections. The optimization problem is formulated as: max Ψ(S g, S m), s.t. S g(i, k) {0, 1}, S m(j, k) {0, 1}, S T g (., k)cs m (., k) S g T (., k) 2 S m (., k) 2, K S g(, k) = 1 k=1 K S m(, k) = 1 k=1

14 Method: Co-module Indentification We define L g = 2A g D g, L m = 2A m D m, ( ) ( ) 0 C L w = diag(l g, L m ), L b = C T, 0 S Sg =, and S m L = L w + λl b. The above optimization problem can be relaxed to: max Ψ( S) = Tr( S T L S), s.t. S T S = 2IK. We take S as a data set composed of N g + N m nodes and do k means clustering to get the assignment label for each node.

15 Algorithm Input: Adjacency matrix A g, A m, C, and K, which is the number of modules. 1. Compute the matrices L g, L m ; 2. Construct the matrix L; 3. Compute the K eigenvectors v 1, v 2,, v K corresponding to the K largest eigenvalues of matrix L; 4. Construct a new matrix T R (Ng+Nm) K, with columns v 1, v 2,, v K ; 5. Cluster the points constructed from each row of matrix T with k-means clustering into K clusters; Output: Index of nodes in each module.

16 Results: Gene-microRNA Network Module Analysis for Ovarian Cancer Cutoff for building gene/mirna coexpression network: 0.6. Cutoff for building gene-mirna coexpression network: 0.3. λ = 1. Figure : AUCs for diffferent cutoffs. With our method, we finally got 46 modules.

17 MiRNA module enrichment analysis The mirna cluster data are downloaded from the mirbase website ( se.org/), with the inter-mirna distance cutoff 10kb. This criterion resulted in 153 clusters containing from 2 to 46 mirnas. There are a total of 14 modules enriched by clusters, and 8 clusters enriched by modules with the overlap size between clusters and modules being at least 3. Example, 5 of 12 mirnas in module 35 belong to a cluster with size 6 in Chr13. The total distance of this cluster is about 700bp.

18 Table : MiRNA module enrichment results. No. p-value MiRNAs Loci E-18 mir-411, mir-299, mir-758, mir-329-1, mir-543, mir-495, Chr mir-654, mir-376b, mir-376a-1, mir-381, mir-487b, mir-539, mir-487a, mir-382, mir-154, mir-377, mir-409, mir-369, mir-376c,mir-889,mir E-14 mir-379, mir-411, mir-299, mir-758, mir-329-1, mir-543, Chr mir-376c, mir-654, mir-376b, mir-376a-1, mir-381, mir-487a, mir-382, mir-154, mir-377, mir-409, mir-369, mir-495, mir-487b, mir-539, mir E-15 mir-411, mir-758, mir-329-1, mir-543, mir-495, Chr mir-376b, mir-376a-1, mir-487b mir-539, mir-889, mir-382 mir-154, mir-409, mir-369, mir-654,mir-487a, mir E-02 mir-379, mir-299, mir-376c, mir-376a-1, mir-381, mir-377 Chr E-03 mir-379, mir-299, mir-376c, mir-376a-1, mir-381, mir-377 Chr E-05 mir-493, mir-337, mir-433, mir-127, mir-432, mir-136 Chr E-03 mir-379, mir-299, mir-376c, mir-376a-1, mir-381, mir-377 Chr E-02 mir-379, mir-299, mir-376c, mir-376a-1, mir-381, mir-377 Chr E-06 mir-17, mir-18a, mir-19a, mir-20a, mir-19b-1 Chr

19 Gene module enrichment analysis We did enrichment analysis for Gene Ontology biological process (GO-BP) terms and KEGG pathways with DAVID. By taking the cutoff of the Benjamini p-values as 0.05, 15 modules are enriched by GO-BP terms and 7 modules are enriched by KEGG pathways significantly.

20 Enriched KEGG pathways for the modules Table : Enriched KEGG No. Enriched Pathways p-value 41 Small cell lung cancer 1.30E-03 pathways for the modules. Chronic myeloid leukemia 3.10E-02 Pathways in cancer 2.40E-02 Colorectal cancer 1.90E-02 No. Enriched Pathways p-value Cell cycle 3.40E p53 signaling pathway 3.50E-03 Thyroid cancer 1.30E-01 Small cell lung cancer 2.70E-03 Bladder cancer 1.60E-01 Cell cycle 4.00E-03 Endometrial cancer 1.80E-01 Pathways in cancer 2.10E-02 Non-small cell lung cancer 1.60E-01 Non-small cell lung cancer 8.20E-02 Acute myeloid leukemia 1.60E-01 Glioma 8.00E-02 Glioma 1.60E-01 Melanoma 7.70E-02 p53 signaling pathway 1.50E-01 Pancreatic cancer 6.90E-02 Melanoma 1.50E-01 Chronic myeloid leukemia 6.40E-02 Pancreatic cancer 1.40E-01 Prostate cancer 6.80E-02 Prostate cancer 1.60E-01

21 Gene-miRNA modules are strongly associated with cancers We checked the cancer related mirnas from the website: ecu.edu. There are 295 different mirnas related to cancer, of which 122 are in our identified modules. 57 of the 295 mirnas are related to ovarian cancer, of which 29 are in our identified modules, which achieves a p-value Table : Number of cancer associated mirnas for the modules enriched by clusters. Module No No. of mirnas No. of c-mirnas p-value E E E E E E-06 Module No No. of mirnas No. of c-mirnas p-value 2.17E E E E E E-03

22 In module 1, all the mirnas are associated with ovarian cancer. The genes in this module take part in the process of transcription, gene expression etc.. In module 41, 5 mirnas are associated with ovarian cancer. By checking the GO-BP terms, we found that the most enriched term is sexual reproduction, which has a p-value 7.50E-06. This module also enriches the GO-term: gamete generation, male gamete generation, and spermatogenesis significantly.

23 (a) Figure : Module 37: 156 genes are regulated by the 12 mirnas. There are a total of 47 known regulations. (b)

24 Comparison with Mirsynergy Mirsynergy operates in two steps: It detects the mirna modules based on gene-mirna relationship. It expands each mirna module by greedily including (excluding) mrnas into (from) the mirna module to maximize the synergy score, which is a function of mirna-mrna and gene-gene interactions. Comparision with Mirsynergy: Table : Module enrichment performance of Mirsynergy and our method. Method N module Ng Nm N en module N en cluster N GO N KEGG Mirsynergy Our method

25 Conclusions We proposed an optimization model to study the gene and mirna co-modules in networks. We applied this model to an ovarian cancer data set. 14 modules are enriched by the mirna clusters with overlap size being at least 3, 15 modules are enriched by GO-BP terms, and 7 modules are enriched by KEGG pathways significantly. In the identified modules, 122 mirnas are cancer associated and 29 mirnas are related to ovarian cancer, which has a p-value Compared to the existing method Mirsynergy, our method can find more modules with the number of genes and mirnas having a good balance. The models and algorithms can be extended to analyze more complex networks.

26 Thank you!