Reprogramming of miRNA networks in cancer and leukemia

Abstract

We studied miRNA profiles in 4419 human samples (3312 neoplastic, 1107 nonmalignant), corresponding to 50 normal tissues and 51 cancer types. The complexity of our database enabled us to perform a detailed analysis of microRNA (miRNA) activities. We inferred genetic networks from miRNA expression in normal tissues and cancer. We also built, for the first time, specialized miRNA networks for solid tumors and leukemias. Nonmalignant tissues and cancer networks displayed a change in hubs, the most connected miRNAs. hsa-miR-103/106 were downgraded in cancer, whereas hsa-miR-30 became most prominent. Cancer networks appeared as built from disjointed subnetworks, as opposed to normal tissues. A comparison of these nets allowed us to identify key miRNA cliques in cancer. We also investigated miRNA copy number alterations in 744 cancer samples, at a resolution of 150 kb. Members of miRNA families should be similarly deleted or amplified, since they repress the same cellular targets and are thus expected to have similar impacts on oncogenesis. We correctly identified hsa-miR-17/92 family as amplified and the hsa-miR-143/145 cluster as deleted. Other miRNAs, such as hsa-miR-30 and hsa-miR-204, were found to be physically altered at the DNA copy number level as well. By combining differential expression, genetic networks, and DNA copy number alterations, we confirmed, or discovered, miRNAs with comprehensive roles in cancer. Finally, we experimentally validated the miRNA network with acute lymphocytic leukemia originated in Mir155 transgenic mice. Most of miRNAs deregulated in these transgenic mice were located close to hsa-miR-155 in the cancer network.

Figure 1.

The miRNA network in normal tissues (1107 samples, 50 tissues, 115 miRNAs). The network was inferred for all expressed and varying miRNAs, without preselecting for differential expression. Standard Banjo parameters were adopted with a q6 discretization policy. The consensus graph depicted here was obtained from the best 100 nets (after searching through 8.3 × 10⁹ networks). MCL graph-based clustering algorithm was applied to clusters extraction (miRNAs with highly related expression pattern have edges of the same color); thus, different clusters in the network are linked by different color edges. yEd graph editor (yFiles software) was employed for graphs visualization. See text for further discussion.

Figure 2.

Cellular pathways regulated by differentially expressed miRNAs in cancer. KEGG analysis by ClueGO (Bindea et al. 2009) of pathways (score = 3, P-value < 1 × 10⁻³) simultaneously targeted by both up-regulated and down-regulated miRNAs (listed in Table 1; Supplemental Tables IV, V). The KEGG pie-chart shows the functional effect of differentially expressed miRNAs on cellular pathways in cancer. The large majority of the affected pathways is related to cancer or signal transduction (i.e., Wnt, VEGF, TGF-beta, insulin, and phosphatidylinositol signaling, focal adhesion, and colorectal cancer). Target genes selection was performed with DIANA-miRpath, microT-V4.0 (Papadopoulos et al. 2009). The union of the target mRNAs with a score above 3 was used as an input to ClueGO. Right-sided hypergeometric test yielded the enrichment for GO terms. Benjamini-Hochberg correction for multiple testing controlled the P-values. GO term fusion was applied for redundancy reduction.

Figure 3.

The miRNA network in solid cancers (2532 samples, 31 cancer types, 120 miRNAs). The network was inferred for all expressed and varying miRNAs, without preselecting for differential expression. Standard Banjo parameters were adopted with a q6 discretization policy. The consensus graph depicted here was obtained from the best 100 nets (after searching through 8.4 × 10⁹ networks). MCL expression clusters are linked by different color edges. yEd graph editor (yFiles software) was employed for graph visualization. The miRNAs expressed differentially in the tumors are color-coded and in the graph miRNA neighbors are of the same color code: Closely clustered miRNAs are either pink (overexpressed) or green (down-regulated). The node labels, for which expression and physical alteration (CGH, see “miRNA copy number variations in cancer and leukemia” section) were concordant (i.e., overexpression and amplification), were emboldened and visually reinforced with a hexagonally shaped border.

Figure 4.

Comparison of miRNA networks in normal lung and adenocarcinoma. (A) Normal lung (71 samples). A single complete miRNA network is shown. (B) Lung adenocarcinoma (125 samples). In this graph, one major and eight minor subnetworks were detected. For example, hsa-miR-10a/b, hsa-miR-29a/b, hsa-miR-107, and hsa-miR-103 are in minor independent subnetworks disjoint from the main one.

Figure 5.

The KEGG functional analysis of eight disjointed minor miRNA networks in lung adenocarcinoma. The miRNA present in the unconnected cliques target genes are involved in many cancer-related terms, such as focal adhesion, small cell lung cancer, and calcium signaling. The detailed list of significant GO terms is shown in Supplemental Figure 7.

Figure 6.

The miRNA network in acute myeloid leukemia (589 samples, two subnetworks). Standard Banjo parameters were adopted with a q6 discretization policy. The consensus graph depicted here was obtained from the best 100 nets (after searching through 8.5 × 10⁹ networks). The miRNA network in AML has disjointed cliques. hsa-miR-155 and hsa-miR-181, two miRNAs with clinical relevance are in two separated subnetworks, as expected from their prognostic independence. hsa-miR-181 is associated with hsa-miR-146a in a detached yellow miniclique. mir-155 belongs to the main subnetwork, in the same red MCL clique of hsa-miR-223, hsa-miR-92a, hsa-miR-25, and hsa-miR-32. Finally, hsa-miR-29b has a key role in AML and acts as a hub in the AML net.

Figure 7.

The miRNA network in chronic lymphocytic leukemia (254 samples, three subnetworks). Standard Banjo parameters were adopted with a q6 discretization policy. The consensus graph depicted here was obtained from the best 100 nets (after searching through >1 × 10¹⁰ networks). The network graph shows a major net and two separated minicliques: hsa-miR-23a/b and the hsa-miR-15/16 cluster. hsa-miR-15 and hsa-miR-16, two miRNAs frequently deleted in CLL, have been shown to regulate apoptosis via BCL2. The key hsa-miR-29b, acting as a hub in AML, is only a branch in CLL. AML prognostic hsa-miR-181 is disjointed in AML but not in CLL, while the reverse happens in CLL for prognostic hsa-miR-15/16 genes.

Figure 8.

Deregulated miRNAs in leukemia from Mir155 transgenic mice are preferentially located close to hsa-miR-155 in the cancer network. We compared miRNA profiles of three leukemia samples from Mir155 transgenes to controls from wild-type mice. The deregulated miRNAs (Supplemental Table X) were mapped onto the cancer network and highlighted in yellow. Most of the other miRNAs are concentrated around hsa-miR-155 node (black). When a diagonal is drawn and the two sides compared the difference between yellow nodes is significant (Fisher's exact test, two-tail P-value < 0.009).