Revealing differential modules in uveal melanoma by analyzing differential networks

  • Authors:
    • Li Han
    • Cui Chen
    • Chang‑Hui Liu
    • Min Zhang
    • Ling Liang
  • View Affiliations

  • Published online on: February 22, 2017
  • Pages: 2261-2266
Metrics: Total Views: 0 (Spandidos Publications: | PMC Statistics: )
Total PDF Downloads: 0 (Spandidos Publications: | PMC Statistics: )


The aim of the present study was to investigate differential modules (DMs) between uveal melanoma (UM) and normal conditions by examining differential networks. Based on a gene expression profile collected from the ArrayExpress database, the inference of DMs involved three steps: The first step was construction of a differential co‑expression network (DCN); second, the module algorithm was adapted to identify the DMs presented in DCN; finally, the statistical significance of DMs were assessed based on the null score distribution of DMs generated using randomized networks. A DCN with 309 nodes and 3,729 edges was obtained, and 30 seed genes from the DCN were examined. Subsequently, one DM, which had 179 nodes and 3,068 edges, was investigated. By utilizing randomized networks, the P‑value for DM was 0.034, therefore, the DM was statically significant between UM and baseline conditions. In conclusion, the present study successfully identified one DM in UM based on DCN and module algorithm, and this DM may be beneficial in revealing the pathological mechanism of UM and provide insight for future investigation of UM.


Uveal melanoma (UM) is the most frequent primary malignant tumor worldwide, which arises from neural crest-derived melanocytes of the uveal tract, and includes the iris, ciliary body and choroid of the eye (1). Options for treating UM consist of enucleation (eye removal), various forms of radiation therapy, laser hyperthermia and surgical resection (2,3). The 5-year local-tumor control rates in the majority of specialized treatment centers are >90%, however, despite successful treatment of the primary tumor, metastasis occurs via hematogenous spread in up to 50% of patients (4). The most common sites of involvement include the liver (93%), lung (24%) and bone (16%), with the overwhelming majority presenting initially in the liver (5). Due to large tumor size, involvement of the ciliary body and increased patient age, metastasis is life-threatening (6). Therefore, revealing the molecular characteristics may assist in understanding the pathological mechanism and provide potential target markers for the treatment of UM.

Previous studies have reported several molecular markers in revealing the pathology of UM (79). For example, activating mutations in the Gαq stimulatory subunit, guanine nucleotide binding protein (G protein), q polypeptide, appear to represent an initiating event (10), whereas inactivating mutations in breast cancer 1, early onset associated protein-1 show a threshold in tumor progression, beyond which lead to metastasis and associated mortality (11). However, mutated genes are not involved individually and use of a network strategy, which connects genes together, is a useful approach to solve the problem to a certain extent (12). If the network is too large, it may ignore a certain number of significant genes and interactions (13), however, evaluating sub-networks or modules of the complex network avoids this problem (14). In small modules, the functions of individual genes and gene-gene interactions can be detected and examined with more detail and precision (15).

Therefore, the present study aimed to reveal differential modules (DMs) in UM by analyzing differential networks. To achieve this, a differential co-expression network (DCN) for the differentially expressed genes (DEGs) of UM were constructed, based on weighted gene co-expression network analysis (WGCNA) and one-sided t-tests. A module algorithm was then implemented to identify the DMs and the statistical significance of the DMs were evaluated. These DMs may be beneficial to disease progression and provide insight for future investigations of UM.

Materials and methods

Inference of DMs

Using the gene expression profile, the inference of DMs involved three steps: The first step involved construction of the DCN. Two genes were connected in a DCN if they exhibited correlated expression profiles across conditions, and their expression levels were significantly different between the UM and baseline conditions (normal control). Subsequently, the module algorithm was adapted to identify DMs present in multiple DCNs. Finally, the statistical significance of DMs was assessed based on the null score distribution of DMs, generated using randomized networks.

Gene expression profile

A gene expression profile (accession no. E-GEOD-44295) for the UM samples and normal controls was obtained from the online public ArrayExpress database ( The E-GEOD-44295 profile comprised 66 samples, of which 63 were UM samples and three were normal samples, and these were presented on an A-MEXP-1172-Illumina HumanRef-8 v3.0 Expression BeadChip platform (Illumina, Inc., San Diego, CA, USA). By converting the data of the microarray profile at the probe-level into gene symbols, a total of 18,631 genes were obtained for further examination.

Construction of the DCN

Prior to constructing the DCN for UM, a binary co-expression network was built, and a weight was assigned to each edge based on differential gene expression between UM and baseline condition.

Binary co-expression network

To construct the binary co-expression network, DEGs between the UM and baseline conditions were identified according to the Significance Analysis of Microarrays (SAM) (16). SAM assigns a score to each gene on the basis of alterations in gene expression relative to the standard deviation of repeated measurements. The relative difference d(i) and expected relative difference dE(i) in gene expression were defined. For the majority of genes, d(i)≅dE(i), however, certain genes were represented by points displaced from the d(i)=dE(i) line by a distance above the threshold (Δ). As Δ decreased, the number of genes considered significant by SAM increased. In the present study, genes which met the criterion of Δ=3.701 were selected as DEGs.

The binary co-expression network for the DEGs of UM was constructed using WGCNA, which describes the correlation structure between gene expression profiles, image data, genetic marker data, proteomics data and other high-dimensional data (17). The first step was to define a gene co-expression similarity (Sij) for each pair of genes (i and j). The second step involved transforming the similarity matrix into an adjacency matrix with the assistance of an adjacency function. The adjacency was defined by increasing the Sij to a soft threshold power β≥1. This allowed the adjacency to take on values in succession between 0 and 1 to preserve the continuous nature of the co-expression information. The continuous measure to assess the strength of a gene connection was calculated as aij=Sijβ and the interaction between two genes was proportional to their Sij in the following formula: log (aij)=βxlog(Sij).

DCN construction

In order to remove indirect correlations resulting from a third gene in the binary co-expression network, first order partial Pearson's correlation coefficient was used (18) and only edges with correlations above the pre-defined threshold (δ=0.4) were selected for the DCN. Subsequently, weights were assigned to edges in the binary co-expression network based on the P-value of differential gene expression between the UM and baseline conditions, which was calculated using a one-sided t-test (19). The weight (Wi,j) of an edge (i,j) in the differential network was defined as follows:


Where pi and pj represent P-values for genes i and j, respectively. V is the node set of the co-expression network, and cor(i,j) is the absolute value of Pearson's correlation between genes i and j based on their expression profiles.

Identification of DMs

A module algorithm was used to identify the DMs present in the DCN (20), which comprised three steps: Seed prioritization, module identification by seed expansion and refinement of candidate modules.

Seed prioritization

For each network Gk=(V, Ek) (k=1) with an adjacency matrix Ak=(aijk)nxn, a function, g(i), was proposed to assess the importance of vertex i in the corresponding network. Gk represented a network, V and Ek stood for the variables in the network. V→R represented a function that builds a relationship between V and R.




In which Nk (i) represents the set of neighbors of i in Gk; A′ijk represents the degree of the normalized weighted adjacency matrix; D denotes a diagonal matrix with element Dij=∑jAijk. A′g is the information propagation on network via the edges of networks, which indicates that the importance of a node depends on the number of neighbors, strength of connection and importance of neighbors. For each gene, the g(i)=z-score was obtained, and the genes were ranked g=(g(1), g(2), …, g(M)) based on their z-score, with the top 10% of genes selected as the seed genes.

Module identification by seed expansion

Module identification was accomplished with the assistance of seed genes. With each seed, the module identification step iteratively included genes whose addition led to the maximum decrease in the graph entropy-based objective function until there is no further decrease in the objective function. Regarding each seed gene (v) as a module C=(v), for each vertex u in its neighborhood in all networks, N(v)=∪iNi(v)[uϵN(v)] was determined, in which Ni(ν) was the neighbor set in Gi as the candidate for C. The new module, C′=C∪() and the entropy decrease between C and C' was defined as follows:


The graph entropy for C across all networks and normalized for the size of C was:


Where Hk(C)=∑i∈cH(Cj) represents the sum of all vertices in C and network k. Ci (1≤i≤τ) was a candidate DM. H(C') was calculated in a similar manner. ΔH(C', C)>0 indicated that the addition of vertex u improved the connectivity of the former module C. The vertex u, whose addition maximized ΔH, was added to C. The expansion step terminated when no additional vertex could further reduce the entropy of the evolving module.

Refinement of candidate modules

During the refinement step, candidate DMs with sizes >5 genes were removed. To merge overlapping candidate DMs into DMs, the Jaccard index (21) was used, which is the ratio of intersection over union for two sets. A Jaccard index of 0.5 was used in the present study.

Evaluation of the statistical significance of DMs

The statistical significance of DMs was computed based on the null score distribution of DMs generated using randomized networks. Each randomized network was composed of edges captured from interactions in the binary co-expression network, and the number of edges in the randomized network was the same as that in the DCN. Each network was completely randomized 100 times by degree-preserved edge shuffling. To construct the null distribution for DM scores, module identification was performed on the randomized networks. Based on the null distribution, the empirical P-value of a DM was defined as the probability of the module having the observed score or less by chance. P-values were corrected for multiple testing utilizing the method of Benjamini-Hochberg (22), and DMs with P<0.05 were considered significant.


Construction of DCN

By accessing SAM, 443 DEGs were obtained between UM and baseline conditions with the threshold of Δ=3.701. Subsequently, on the basis of the DEGs, a binary co-expression network comprising 97,903 interactions was constructed through WGCNA. To remove indirect correlation and improve the network confidence, interactions that met the δ≥0.4 criteria were selected to form the DCN (Fig. 1). There were a total of 309 nodes and 3,729 interactions. In addition, a weight value was assigned to each interaction or edge, based on a one-sided t-test. As a consequence, for genes i and j, and their edges e(i,j), a three-dimensional matrix for the weight of the edges, e(i,j), Wi,j, in the network, Gk, were determined. Therefore, the DCN was termed a three-dimensional network, which may be more feasible and stable, compared with the binary network, and the subsequent analyses were dependent on it.

Identification of DMs

For the purpose of extracting DMs from the DCN, the present study used a module algorithm, which comprised seed prioritization, module identification by seed expansion and refinement of candidate modules. A total of 30 seed genes were detected from the DCN, as shown in Fig. 1, and the z-scores for these are listed in Table I. The top five seed genes with the highest z-scores were KIAA1913 (z-score=5,030.874), RAE1 (z-score=4,654.952), EDIL3 (z-score=4,288.795), FBN2 (z-score=4,280.710) and KCNE1L (z-score=3,907.409). With the 30 seed genes as a starting point, modules were identification was performed based on the entropy decrease ΔH(C, C) between C and C', and candidate modules were obtained. Following elimination of candidate modules with sizes <5, the modules between which the Jaccard index was ≥0.5 were merged, and one DM was identified for UM. In this DM, there were 179 nodes and 3,068 edges (Fig. 2A). To further validate the connectivity of two genes in DM, the present study focused on the weight distributions of edges (Fig. 2B). The majority of edges were distributed in the section of 0.9–1, which suggested that the genes were closely correlated to each other and that the DM possessed good topological properties, meaning that the constructed DM correlated to UM closely.

Table I.

Seed genes in the differential co-expression network.

Table I.

Seed genes in the differential co-expression network.

  1 KIAA19135,030.874
  6 C11orf703,893.342
13 C12orf243,753.696
22 C6orf1663,542.197
29 C18orf103,451.743
Statistical significance of the DM

To evaluate the statistical significance of DM, the present study constructed a randomized network of the 3,729 edges captured from 97,903 interactions of the binary co-expression network at random to identify modules. This type of randomized network was constructed 100 times, and a total of 4,298 modules were obtained. The empirical P-value of DM was defined as the probability of the module having the observed score or a lower score by chance, and this was adjusted using the Benjamini-Hochberg test. The P-value for DM was determined as 0.034, which indicated that this DM was a statically significant DM across the UM condition and baseline condition.


A network-based approach is capable of extracting informative and significant genes dependent on biomolecular networks, including protein-protein interaction networks, co-expression networks and gene regulatory networks, compared with individual genes (23,24). In addition, it offers a quantifiable description of the molecular networks, which characterize the complex interactions and intricate associations governing cellular functions among tissues and disease-associated genes, to explain the molecular processes during the development and progression of disease (25). Traditionally, if one interaction between a gene pair shows high correlation in one condition, the interaction is selected as an edge in the network (26). By contrast, if one gene in the interaction is differently expressed but the other one is not, it may not be considered a significant interaction in the entire dataset. These challenges are overcome to a certain extent by constructing a DCN.

Therefore, in the present study, a DCN for UM was constructed based on DEGs using WGCNA, which comprised 309 nodes and 3,729 edges. Based on a one-sided t-test, a total of 30 seed genes were examined from the DCN. Subsequently, the identification of DMs was performed on the DCN and its seed genes according to the module algorithm. Following evaluation using randomized networks and adjusting the P-value of DM using the Benjamini-Hochberg test, one DM was obtained with P=0.034 across the UM and baseline conditions, and was considered a statistically significant DM between the UM and baseline conditions. The significant DM possessed 179 nodes and 3,068 edges. In addition, the majority of edges were distributed in section 0.9–1, which suggested that the genes were closely correlated with each other and the DM possessed good topological properties.

From a systems biology point of view, diseases are caused by perturbations to the gene network (27). These perturbations are dynamic as disease progresses. The identification of small modules may provide assistance in investigating the dynamic perturbations with more detail and precision. Therefore, the key focus of the present study was to investigate DMs between UM and normal conditions utilizing differential networks. A key innovation of the method used in the present study is the ability to identify unique and shared modules from multiple differential gene networks, each of which represents a different perturbed condition (28). In addition, sets of genes, which are differentially expressed under the UM state, but do not exhibit correlated expression patterns are not identified as a module (29). Therefore, the DM offers more detailed evidence of the pathological mechanism underlying the progression of UM.

In conclusion, the present study successfully identified a DM in UM, based on the DCN and module algorithm. This DM may be useful in revealing the pathological mechanism of UM and provide insight for future investigations of UM.



Harbour JW, Onken MD, Roberson ED, Duan S, Cao L, Worley LA, Council ML, Matatall KA, Helms C and Bowcock AM: Frequent mutation of BAP1 in metastasizing uveal melanomas. Science. 330:1410–1413. 2010. View Article : Google Scholar : PubMed/NCBI


Harbour JW: Clinical overview of uveal melanoma: introduction to tumors of the eyeOcular Oncology. Albert DM and Polans A: Marcel Dekker; New York, NY: pp. 1–18. 2003, View Article : Google Scholar


Singh AD, Turell ME and Topham AK: Uveal melanoma: Trends in incidence, treatment, and survival. Ophthalmology. 118:1881–1885. 2011. View Article : Google Scholar : PubMed/NCBI


Landreville S, Agapova OA and Harbour JW: Emerging insights into the molecular pathogenesis of uveal melanoma. 4:629–636. 2008.


Diener-West M, Reynolds SM, Agugliaro DJ, Caldwell R, Cumming K, Earle JD, Hawkins BS, Hayman JA, Jaiyesimi I, Jampol LM, et al: Development of metastatic disease after enrollment in the COMS trials for treatment of choroidal melanoma: Collaborative ocular melanoma study group report No. 26. Arch Ophthalmol. 123:1639–1643. 2005. View Article : Google Scholar : PubMed/NCBI


Abdel-Rahman MH, Pilarski R, Cebulla CM, Massengill JB, Christopher BN, Boru G, Hovland P and Davidorf FH: Germline BAP1 mutation predisposes to uveal melanoma, lung adenocarcinoma, meningioma, and other cancers. J Med Genet. 48:856–859. 2011. View Article : Google Scholar : PubMed/NCBI


Ambrosini G, Musi E, Ho AL, de Stanchina E and Schwartz GK: Inhibition of mutant GNAQ signaling in uveal melanoma induces AMPK-dependent autophagic cell death. Mol Cancer Ther. 12:768–776. 2013. View Article : Google Scholar : PubMed/NCBI


Cruz C, Teule A, Caminal JM, Blanco I and Piulats JM: Uveal melanoma and BRCA1/BRCA2 genes: A relationship that needs further investigation. J Clin Oncol. 29:e827–e829. 2011. View Article : Google Scholar : PubMed/NCBI


Singh N, Bergman L, Seregard S and Singh AD: Uveal melanoma: epidemiologic aspectsClinical Opthalmic Oncology: Uveal Tumors. Damato B and Singh AD: 2nd edition. Springer; Berlin Heidelberg, New York, NY: pp. 75–87. 2014, View Article : Google Scholar


Field MG and Harbour JW: GNAQ/11 mutations in uveal melanoma: Is YAP the key to targeted therapy? Cancer Cell. 25:714–715. 2014. View Article : Google Scholar : PubMed/NCBI


Harbour JW: The genetics of uveal melanoma: An emerging framework for targeted therapy. Pigment Cell Melanoma Res. 25:171–181. 2012. View Article : Google Scholar : PubMed/NCBI


Zhang L, Li S, Hao C, Hong G, Zou J, Zhang Y, Li P and Guo Z: Extracting a few functionally reproducible biomarkers to build robust subnetwork-based classifiers for the diagnosis of cancer. Gene. 526:232–238. 2013. View Article : Google Scholar : PubMed/NCBI


Nibbe RK, Chowdhury SA, Koyutürk M, Ewing R and Chance MR: Protein-protein interaction networks and subnetworks in the biology of disease. Wiley Interdiscip Rev Syst Biol Med. 3:357–367. 2011. View Article : Google Scholar : PubMed/NCBI


Wu Y, Jing R, Jiang L, Jiang Y, Kuang Q, Ye L, Yang L, Li Y and Li M: Combination use of protein-protein interaction network topological features improves the predictive scores of deleterious non-synonymous single-nucleotide polymorphisms. Amino Acids. 46:2025–2035. 2014. View Article : Google Scholar : PubMed/NCBI


Emilsson V, Thorleifsson G, Zhang B, Leonardson AS, Zink F, Zhu J, Carlson S, Helgason A, Walters GB, Gunnarsdottir S, et al: Genetics of gene expression and its effect on disease. Nature. 452:423–428. 2008. View Article : Google Scholar : PubMed/NCBI


Li J and Tibshirani R: Finding consistent patterns: A nonparametric approach for identifying differential expression in RNA-Seq data. Stat Methods Med Res. 22:519–536. 2013. View Article : Google Scholar : PubMed/NCBI


Zhang B and Horvath S: A general framework for weighted gene co-expression network analysis. Stat Appl Genet Mol Biol. 4:Aug 12–2005.(Epub ahead of print). PubMed/NCBI


Watson-Haigh NS, Kadarmideen HN and Reverter A: PCIT: An R package for weighted gene co-expression networks based on partial correlation and information theory approaches. Bioinformatics. 26:411–413. 2010. View Article : Google Scholar : PubMed/NCBI


Cohen J, Cohen P, West SG and Aiken LS: Applied multiple regression/correlation analysis for the behavioral sciences. Routledge; 2013


Ma X, Gao L and Tan K: Modeling disease progression using dynamics of pathway connectivity. Bioinformatics. 30:2343–2350. 2014. View Article : Google Scholar : PubMed/NCBI


Bouchard M, Jousselme AL and Doré PE: A proof for the positive definiteness of the Jaccard index matrix. Int J Approximate Reasoning. 54:615–626. 2013. View Article : Google Scholar


Benjamini Y and Hochberg Y: Controlling the false discovery rate: A practical and powerful approach to multiple testing. J Royal Statistical Society. Series B (Methodological). 57:289–300. 1995.


Liu ZP, Wang Y, Zhang XS and Chen L: Network-based analysis of complex diseases. IET Systems Biology Institution of Engineering and Technology. 22–33. 2012. View Article : Google Scholar


Chen L, Wang R-S and Zhang X-S: Reconstruction of Gene Regulatory NetworksBiomolecular Networks. John Wiley & Sons, Inc.; pp. 47–87. 2009, View Article : Google Scholar


Sun SY, Liu ZP, Zeng T, Wang Y and Chen L: Spatio-temporal analysis of type 2 diabetes mellitus based on differential expression networks. Sci Rep. 3:22682013. View Article : Google Scholar : PubMed/NCBI


Chen L, Wang RS and Zhang XS: Transcription regulation: Networks and modelsBiomolecular Networks. John Wiley & Sons, Inc.; pp. 23–45. 2009, View Article : Google Scholar


Valentini G, Paccanaro A, Vierci HC, Romero AE and Re M: Network integration boosts disease gene prioritization. Network Biology SIG 2013 ISMB 2013. 2013.


Ma X, Gao L, Karamanlidis G, Gao P, Lee CF, Garcia-Menendez L, Tian R and Tan K: Revealing pathway dynamics in heart diseases by analyzing multiple differential networks. PLoS Comput Biol. 11:e10043322015. View Article : Google Scholar : PubMed/NCBI


Segal E, Shapira M, Regev A, Pe'er D, Botstein D, Koller D and Friedman N: Module networks: Identifying regulatory modules and their condition-specific regulators from gene expression data. Nat Genet. 34:166–176. 2003. View Article : Google Scholar : PubMed/NCBI

Related Articles

Journal Cover

April 2017
Volume 15 Issue 4

Print ISSN: 1791-2997
Online ISSN:1791-3004

Sign up for eToc alerts

Recommend to Library

Copy and paste a formatted citation
Han, L., Chen, C., Liu, C., Zhang, M., & Liang, L. (2017). Revealing differential modules in uveal melanoma by analyzing differential networks. Molecular Medicine Reports, 15, 2261-2266.
Han, L., Chen, C., Liu, C., Zhang, M., Liang, L."Revealing differential modules in uveal melanoma by analyzing differential networks". Molecular Medicine Reports 15.4 (2017): 2261-2266.
Han, L., Chen, C., Liu, C., Zhang, M., Liang, L."Revealing differential modules in uveal melanoma by analyzing differential networks". Molecular Medicine Reports 15, no. 4 (2017): 2261-2266.