Genome-Wide Survey of MicroRNA - Transcription Factor Feed-Forward Regulatory Circuits in Human

Background

• The study extends transcriptional feed-forward loops to mixed motifs that combine transcriptional and post-transcriptional regulatory interactions.The proposed circuit has a master transcription factor regulating a microRNA, with both regulating a shared set of target genes.Source: Background · pp. 3–4
• MicroRNAs are approximately 22-nucleotide non-coding RNAs that negatively regulate gene expression, generally by binding target-gene 3′-UTRs through a short complementary seed sequence.The passage describes the seed as involving nucleotides 2–8 and notes that microRNA binding sites are often overrepresented in target 3′-UTRs.Source: Background · pp. 3–4
• The authors propose predicting transcriptional and post-transcriptional interactions by searching for overrepresented short sequences in promoters and 3′-UTRs, then applying evolutionary or functional constraints.The raw data could also be used to construct more complex network motifs and were made available as supplementary files.Source: Background · p. 4
• The genome-wide ab initio search identified 638 putative (merged) FFLs, including circuits with previously known interactions and predicted interactions for experimental validation.In some cases, the interplay of known regulatory interactions in a closed circuit had not previously been recognized.Source: Background · p. 4
• The study filtered circuits for Gene Ontology enrichment among joint targets, independent database evidence for regulatory interactions, and cancer relevance.The evidence sources included ECRbase, miRBase, PicTar, TargetScan, Oncomir, and the Cancer Gene Census.Source: Background · p. 4

Construction of a human transcriptional regulatory network

• At 0.1 FDR, the transcriptional network contained 2,031 oligos linked to 115 TFs and targeting 21,159 genes; 20,216 additional significant motifs lacked known TFBS associations.The miRNA-regulation subnetwork involved 110 TFs targeting 187 miRNAs.Source: Construction of a human… · p. 6

Construction of a human post-transcriptional regulatory network

• The post-transcriptional network retained 182 of 3,989 conserved 7-mer motifs, matching seeds in 140 of 193 mature miRNAs and targeting 17,266 protein-coding genes.The motifs were selected after multiple-testing correction and ranking at a 0.1 FDR threshold.Source: Construction of a human… · p. 6

Construction of the human mixed feed-forward loops catalogue

• The integration yielded 638 merged FFLs from 5,030 single-target circuits, involving 2,625 joint targets, 101 TFs, and 133 miRNAs.Each merged circuit had 1–38 joint targets, and 74% targeted up to 10 genes.Source: Construction of the human mixed… · p. 7
• Extending the construction to unidentified fixed motifs yielded 4,035 circuits with 1–5 joint targets, and 79% targeted one gene.Motif lengths numbered 170 at 6 nt, 128 at 7 nt, 440 at 8 nt, and 3,297 at 9 nt.Source: Construction of the human mixed… · p. 7

Circuits assessment I: functional analysis.

• Gene Ontology filtering retained 32 merged FFLs, corresponding to 380 single-target loops and involving 344 joint targets, 24 TFs, and 25 miRNAs.Enrichment was assessed using Fisher’s exact test with p < 10^-4.Source: Circuits assessment I: functional… · p. 8

Circuits assessment II: comparison with existing computational databases.

• The assessment validated 216 TF–target gene links and 98 TF–miRNA links; it also validated 64 TF–miRNA links with the tool discussed for Ref., and miRNA–target links for 607 circuits.The 607 circuits comprised 503 validated by miRBase, 343 by PicTar, and 560 by TargetScan.Source: Circuits assessment II: comparison… · p. 9

Circuits assessment III: looking for cancer related FFLs.

• Cancer-related filtering identified 24 circuits containing at least two cancer-related genes, using oncomiRs and genes in the Cancer Gene Census to define cancer relevance.The oncomiR lists were reported in and.Source: Circuits assessment III: looking for… · pp. 9–10

Potential function of mixed feed-forward circuits

• Because sequence-based identification could not reveal whether master TF actions were excitatory or inhibitory, the study could not assign its circuits to coherent or incoherent types.Type II (coherent) circuits may reinforce regulation and remove mRNAs after transcription switches off; Type I circuits may buffer fluctuations or create time-limited protein expression.Source: Potential function of mixed… · pp. 10–11

Analysis of the mixed feed-forward circuits in terms of network motifs

• Real-network mixed FFL counts exceeded randomized counts (Z = 3.5, 8.3, and 8.9), but random networks also contained many circuits, limiting overrepresentation as a sole identification criterion.The randomizations reshuffled miRNA promoters and seeds, switched network edges while preserving node degrees, or replaced nodes without constraints; results agreed with.Source: Analysis of the mixed feed-forward… · pp. 11–12

Analysis of the Gene Ontology enrichment results.

• GO enrichment links the circuits to development, differentiation, proliferation, apoptosis, and migration, including cardiac, skeletal, neural, and hematopoietic lineages.Single-gene analysis annotated 108 targets to multicellular-organism development, 56 to cell differentiation, and 48 to proliferation and apoptosis; 57 annotated genes were associated with the ubiquitin cycle.Source: Analysis of the Gene Ontology… · pp. 12–13

Comparison of our results with the database [33] of ChIP-PET c-Myc targets.

• Against the 2,088 ChIP-PET c-Myc targets, 253 of 1,979 predicted targets overlapped (p = 1.1 × 10−6), versus 27 of 235 TRANSFAC targets (p = 0.21).With the (-500/+100) promoter region, 203 of 1,612 predicted targets overlapped the dataset (p = 8 × 10−5).Source: Comparison of our results with the… · p. 13, Dependence of our results on the… · p. 14

Dependence of our results on the choice of the promoter’s region.

• Using a (-500/+100) rather than (-900/+100) promoter region yielded 6,682 single-target FFLs versus 5,030, with 1,769 shared; discussed circuits and most assessment-surviving circuits persisted.Randomization and c-Myc comparisons were comparable, supporting the authors’ assessment of overall robustness.Source: Dependence of our results on the… · p. 14

Comparison with related works

• None of the 16 co-regulating TF–miRNA pairs reconstructed from appeared in this dataset; the authors highlight miRNA-promoter definition and binding-sequence search as key novelties.Their miRNA-targeting transcriptional subnetwork contains 110 TFs targeting 187 miRNAs.Source: Comparison with related works · pp. 15–16

Description of a few interesting circuits

• The c-Myc–miR-20a/miR-17-5p circuit includes 11 joint targets, including E2F1, and combining it with E2F2 reveals a more complex, jointly regulated motif.The AREB6/ZEB1–miR-375 circuit has 14 joint targets and is proposed to relate to embryonic development and pancreatic physiology, including insulin secretion.Source: Description of a few interesting… · pp. 16–18
• Other examples connect MEF-2–miR-133a with BRUNOL4 and cardiomyocyte hypertrophy, C-REL–miR-199a with ovarian cancer, and HSF2–let-7f with cell-cycle and cancer-related targets.The discussion also notes UST linked to miR-1 in a circuit without an identified master TF, connecting genes with roles in myogenesis and muscle differentiation.Source: Description of a few interesting… · pp. 18–19

Conclusions

• The prioritized FFL catalogue links mixed transcriptional/post-transcriptional circuits to organism development and differentiation and yields a putative TF regulatory network of human miRNA genes.The pipeline identifies loops in which a master TF regulates a miRNA and, together with it, joint protein-coding targets.Source: Conclusions · pp. 19–20
• The authors call this work a first step because technical and biological issues likely leave circuits unidentified, including the restrictive assumption that each gene has a well-defined promoter.Alternative splicing and transcription start sites can produce isoforms with different functions; the notion of a gene is also being redefined.Source: Conclusions · p. 20
• Agreement with some existing experimental data suggests the approach may help explain gene regulatory networks and the interplay of transcriptional and post-transcriptional layers.Source: Conclusions · p. 20

miRNA transcriptional units

• Pre-miRNAs were classified as genic when located in annotated exons, introns, or flanking untranslated regions of KNOWN protein-coding genes.Human and mouse gene annotations came from Ensembl release 46, assemblies hg18 and mmu8, and miRGen supplied overlaps and hairpin locations.Source: miRNA transcriptional units · p. 20
• The analysis used human and mouse pre-miRNA coordinates from miRBase release 9.2 and retained hairpins only when a mouse orthologue was identified using the human-to-mouse orthology table [7].The orthology table was compiled by and provided as Supplementary Table S15.Source: miRNA transcriptional units · pp. 20–21
• Pre-miRNA-to-gene overlaps and precise hairpin locations were obtained from mirGen, while only protein-coding genes labelled “KNOWN” were selected in human and mouse.Coordinates and annotations were drawn from miRBase and Ensembl, respectively.Source: miRNA transcriptional units · p. 20
• Transcriptional units (TUs) grouped nearby miRNA hairpins presumed to be co-transcribed as one poly-miRNA precursor.cDNA and EST expression data supported co-transcription of miRNAs in the same TU.Source: miRNA transcriptional units · p. 21
• The study treated each TU as a single miRNA gene and assigned all its miRNAs the promoter corresponding to the TU’s transcriptional start site.This promoter-sharing approach also applied when treating isolated miRNAs and TUs together in the precursor set.Source: miRNA transcriptional units · p. 21
• 130 human miRNA precursors and 130 mouse orthologues corresponded to 193 mature miRNAs, whose sequences and seeds formed the target-search input motifs.Of the human precursors, 68 were non-genic and 62 were within a KNOWN gene; 53 genic precursors shared the host gene’s orientation and 9 had the opposite orientation.Source: miRNA transcriptional units · p. 21

Definition of promoter regions

• The protein-coding and miRNA promoter datasets were extracted from Ensembl version 46.The promoter regions were prepared as distinct datasets for the two gene classes.Source: Definition of promoter regions · p. 21
• Protein-coding promoters used the longest known transcript’s −900 to +100 TSS region, with repeat and coding-exon masking, yielding 21,316 human and 21,814 mouse sequences.Genes and transcripts were restricted to those labeled “KNOWN”; the TSS was position +1.Source: Definition of promoter regions · pp. 21–22
• For miRNAs, promoters were assigned by transcription unit: the most upstream member’s promoter represented the TU, with host-gene promoters used for same-orientation genic miRNAs and −900 to +100 regions for non-genic or opposite-orientation miRNAs.For an isolated miRNA, the TU rule was trivial; the sequences were repeat- and exon-masked.Source: Definition of promoter regions · p. 22
• 21,446 human and 21,944 mouse regulatory sequences remained after merging protein-coding and miRNA promoters.These were the final collection sizes.Source: Definition of promoter regions · p. 22

Definition of 3’-UTR regions

• Complete 3′-UTR sequences came from Ensembl version 46 for protein-coding genes labeled “KNOWN”.Source: Definition of 3’-UTR regions · p. 22
• For each gene, the analysis retained only its longest transcript among those labeled “KNOWN”.Source: Definition of 3’-UTR regions · p. 22
• The resulting dataset contained 17,486 human and 15,921 mouse genes with defined 3′-UTRs.Source: Definition of 3’-UTR regions · p. 22
• The selected 3′-UTR sequences were Repeat-Masked using Ensembl’s default masking parameters.Source: Definition of 3’-UTR regions · p. 22
• Human and mouse 3′-UTRs averaged ∼1157 nts and ∼982 nts, respectively.Their lengths varied, unlike the promoter sequences.Source: Definition of 3’-UTR regions · p. 22

Oligos analysis

• Oligo-analysis details were described in the Supporting Text, and the promoter and 3′-UTR sequences were available on the Supporting website.The software implementation was available upon request from the authors.Source: Oligos analysis · p. 23

TF-miR pairs and their joint target genes

• Crossing predicted TF and miRNA target lists yielded feed-forward circuits in which a transcription factor regulates a miRNA that coregulates target genes.When a mature miRNA had multiple genomic loci, all possible promoters were considered.Source: TF-miR pairs and their joint target… · p. 23

Assessment of miRNA targets using existing databases

• TargetScan, PicTar, and miRBase supplied in silico miRNA target predictions, based essentially on sequence conservation across species.TargetScan predictions came via miRGen Release 3 and TargetScan Release 4.2; PicTar targets came from UCSC hg17, and miRBase targets from version 4.Source: Assessment of miRNA targets using… · p. 23
• Target IDs from the three resources were mapped to Ensembl Gene IDs, which served as the main namespace across differing genomic annotations.Gene Symbol and RefSeq IDs were mapped to Ensembl IDs.Source: Assessment of miRNA targets using… · p. 23

Comparison with ECRbase

• ECRbase TFBS predictions were downloaded for CoreECR regions at least 355 nt long with 77% identity.The data came from the tfbs_coreEcrs.hg18mm8.v94.txt file.Source: Comparison with ECRbase · p. 23
• Predicted TFBSs were mapped by genomic coordinates onto the study’s promoter regions for protein-coding and miRNA genes.The mapping used TFBS predictions stored in the relevant databases.Source: Comparison with ECRbase · pp. 23–24
• An ECRbase TF was assigned to a gene only when the complete ECRbase sequence was present in the study’s promoter sequence, avoiding mismatches from differing masking or misannotations.Source: Comparison with ECRbase · p. 24

Gene Ontology analysis

• GO annotations included genes directly annotated to a term or to any of its descendants in the GO graph.The study used GO annotation DAGs and Ensembl gene product annotations, version 46.Source: Gene Ontology analysis · p. 24
• An exact Fisher’s test estimated the probability of observing at least as many term-annotated genes in a randomly selected set of equal size.Source: Gene Ontology analysis · p. 24
• Only Gene Ontology enrichment P-values below 10−4 were reported to account for multiple testing.Source: Gene Ontology analysis · p. 24

Identification of cancer related genes

• OncomiRs were identified using sources and, while the Cancer Gene Census supplied the working list of genes causally implicated in cancer by mutation.Source: Identification of cancer related genes · p. 24
• A circuit counted as cancer-related if it included at least one oncomiR or one gene in the Cancer Gene Census catalogue.Source: Identification of cancer related genes · p. 24
• Cancer Gene Census annotations included chromosomal location, tumour types, oncogenic mutation classes, and other genetic properties; detailed circuit lists appear in Supplementary files S9 and S10.Source: Identification of cancer related genes · p. 24

Supporting website

• Supplementary Files 1–5 provide miRNA lists and transcriptional and post-transcriptional regulatory-network data.They cover pre-miRNAs and mature miRNAs, promoter-region regulatory data, and 3′-UTR microRNA-seed associations.Source: Supporting website · p. 35
• Supplementary Files 6–8 contain raw circuit data, including circuits with unknown transcription factors and final datasets with complete annotations.The complete circuit datasets and their annotations are also reported in Supplementary File S8.Source: Supporting website · p. 35, Table 1, Table 2
• Supplementary Files 9 and 10 list cancer-relevant circuits associated with known and unknown transcription factors, respectively.Source: Supporting website · p. 35
• Supplementary File 11 provides raw circuit data using an alternative promoter region, defined as −500/+100 around the TSS.Source: Supporting website · p. 35

Oligos analysis

• Human and mouse promoter and 3′-UTR oligos were tested after CG-content stratification, using 6–9-nt promoter words and 7-nt UTR words.The UTR sequences were merged only when on the same strand. Genes with P < 0.01 entered each oligo set without multiple-testing correction at this stage.Source: Oligos analysis · pp. 36–37

Conservation of overrepresentation

• An oligo was deemed conserved overrepresented when human and mouse sets shared significantly more orthologous genes than expected, using one-to-one orthologs and an FDR threshold of 0.1.The study used Benjamini-Yekutieli correction for multiple testing.Source: Conservation of overrepresentation · pp. 37–38

Search of putative TF binding sequences in the promoter motifs database

• Putative TF binding sequences were assigned using complementary TRANSFAC Match searches and exact IUPAC-overlap comparisons with Xie et al. consensus sequences.Match searches used vertebrate-specific high-quality matrices, predefined cutoffs, human sites, and no mismatches; compatible matrices could yield multiple candidate TFs.Source: Search of putative TF binding… · pp. 38–39

Search of putative miRNA binding sequences in the 3’-UTR motifs database

• MiRNA binding sites were defined by perfect Watson-Crick matches to 7-nt seeds beginning at nucleotide 1, 2, or 3 of the mature miRNA’s 5′ end.This approach produced putative target-gene lists for 193 mature miRNAs by merging genes associated with their binding-site words.Source: Search of putative miRNA binding… · p. 39

Randomizations for network motifs analysis

• Three protocols randomized the regulatory data in different ways: miRNA promoter and seed reshuffling, edge switching, and complete node replacement.The authors evaluated the mean and standard deviation of FFL counts across randomizations.Source: Randomizations for network motifs… · pp. 39–40
• Promoter and seed reshuffling used 50 randomizations, swapping nucleotides NlogN times while preserving promoter repeat-masking and CG content and excluding known miRNA seeds.The same pipeline parameters were used to construct putative FFLs from the randomized sequences.Source: Randomizations for network motifs… · pp. 39–40
• Edge switching used 1000 randomizations with 106 exchanges per cycle, preserving each TF or miRNA’s out-degree and each target gene’s in-degree while changing connections.Target genes were swapped only when neither was already regulated by the other selected TF or miRNA.Source: Randomizations for network motifs… · p. 40
• Complete node replacement used 1000 randomizations with 106 exchanges per cycle and no swap filter, completely modifying the original network topology.The mean and standard deviation of FFL counts were evaluated for this protocol as well.Source: Randomizations for network motifs… · p. 40
• The figure compares real-network single-target mixed FFL counts with normalized randomized-count histograms for promoter windows of (-900/+100) and (-500/+100) nucleotides.The histograms distinguish promoter-and-seed reshuffling, edge switching, and complete node replacement; the caption does not report numerical outcomes.Source: Randomizations for network motifs… · p. 41

• Other listed circuits pair AP-1–miR-142-3p with DDIT3, ATF-1 or ATF6–miR-199a* with MTCP1, NF-Y–miR-223 with APC and ATF1, and OCTAMER or PAX-4–miR-125b with IRF4.Source: FFL id TF miRNA JTs · p. 46
• HIF-1–hsa-miR-199a* is listed with MTCP1, while HNF-3–let-7a/let-7f is listed with CCND2; other entries include HSF2–let-7a/let-7f or miR-199a* with MYCN, and IRF–miR-125b with BCL2.Source: FFL id TF miRNA JTs · p. 46
• The entries list IY–miR-296 with RPL22 and BCL2, and MYC with miR-17-5p, miR-19a, and miR-20a; each of the latter entries lists MYC as a joint target.Source: FFL id TF miRNA JTs · p. 46
• SOX-5–miR-29a is listed with EXT1 and COL1A1, while SRY–miR-221 is listed with CCND2 and SRY–miR-412 with BRAF and ATIC.Source: FFL id TF miRNA JTs · p. 46
• The sequence-entry list associates motifs including CAGACAATG, GGACTGCAA, and GGATCTTAA with miR-125b/IRF4, miR-200c/MTCP1, and let-7a/CCND2, respectively.Source: FFL id TF miRNA JTs · p. 46