Fig. 1

Pipeline for inferring methylation regulation networks. a Top: Schematic of the MICMIC pipeline that uses information theoretic approaches to distinguish direct regulation from indirect correlation, where variables X and Y are connected only via variable A, then CMI(X,Y|A) will be close to zero, suggesting that there is no direct connection between X and Y. Bottom: A PC algorithm is used to infer the regulatory network from the observed data matrix, eliminating the indirect edges by CMI testing. b MICMIC is designed to identify the regulatory relationship between the methylation level of a CpG probe and the expression level of its potential gene target. For every target gene tested, we included all nearby genes and CpGs ± 300 kb from the transcription start site (TSS) of the test gene and merged the related expression and methylation matrix together. Then MICMIC applies the CMI-based PC algorithm to infer the regulatory network. CpG probes that passed the test were named direct regulatory elements (DREs). The DRE and its gene targets were denoted as DRE-target pairs. c A representative example of the MICMIC output for the CDCA5 gene, where ten DREs (nine shown here) were identified to be associated with CDCA5 expression in gastric cancer (TCGA STAD), four of which were at least 240 kb away from the TSS of the target gene. One of these DREs, cg02933228 (blue oval), was experimentally verified. In the lollipop diagram, green represents significant CpG probes, i.e. DREs. The Pearson correlation coefficient (PCC) for each DRE-target pair was represented by a vertical line (red for negative PCC and green for positive PCC). d Simulation test to justify the MICMIC p value cut-off. The number of actual DREs identified (blue) vs the number of DREs identified by chance (red) at various p value cut-offs