This document proposes a framework for identifying key regulators of complex traits from genome-wide association studies (GWAS). It utilizes cell type-specific regulatory networks constructed from epigenomic maps of 127 tissues to assign GWAS variants to regulatory elements without directly mapping variants to genes. Applying this approach to GWAS of various complex traits revealed candidate regulators for traits like BMI, cholesterol and type 2 diabetes. Experimental validation in model systems provided further support for some candidates, including TBX15 which reduces lipid accumulation in human adipocytes and IRX3 whose adipocyte-specific knockdown increases fat mass. However, identifying target genes and completing cell type-specific regulatory networks remains challenging.
A framework for identifying key regulators of complex traits
1. A framework for identifying key
regulators of complex traits
Gerald Quon
Manolis
Kellis
Melina
Claussnitzer
Soheil
Feizi
Michal
Grzadkowski
Daniel
Marbach
2. Identifying functional mechanisms of
GWAS variants is challenging
chr16 position
SNP-log10(p)
0
70
Recombinationrate
(cM/Mb)
0
100
• > 90% of GWAS variants do not tag a coding variant (Welter et al.,
2014)
• Mechanism of action (target gene, or disrupted regulatory
element) is typically unknown
3. Segrè et al. (2010) PLoS Genet 6(8): Rossin EJ et al. (2011) PLoS Genet 7(1): e1001273.
Most network and gene set approaches
explicitly map SNPs to genes
Gene set enrichment analysis
(MAGENTA)
Network enrichment analysis
DAPPLE
4. Segrè et al. (2010) PLoS Genet 6(8): Rossin EJ et al. (2011) PLoS Genet 7(1): e1001273.
Most network and gene set approaches
explicitly map SNPs to genes
MAGENTA
DAPPLE
How do we assign variants to genes?
Do physical interaction networks make sense?
5. Maurano et al., 2012
Few regulatory motifs are directly
disrupted
GWAS variants
Cardiovasc.Cancer
Cancer Cardiovasc.
TranscriptionFactors
6. Finding key regulators of complex traits
• Goal: identify the key regulators driving complex traits (obesity and
cholesterol)
• Previous work attempt to identify regulators whose binding is
directly disrupted
• We relax the constraint that regulator motifs have to be directly
disrupted
• We also expand the analysis to be cell type specific
• Our networks do not depend on mapping variants to target genes
TFs
Regulatory elements
(blue = GWAS target)
GWAS variants
8. Combinations of chromatin marks are associated with regulatory elements
• H3K4me3
• H3K9ac
• DNase
•H3K36me3
•H3K79me2
•H4K20me1
•H3K4me1
•H3K27ac
•DNase
•H3K9me3
•H3K27me3
•DNAmethyl
Enhancers Promoters Transcribed Repressed
FTO intron
SNPr2
127Roadmapcelltypes
ChromHMM (Ernst et al,. 2010)
9. 659 motifs (Kheradpour et al., 2014)
(JASPAR, TRANSFAC, ENCODE)
CTCF (H1 ESC)
0 0.2 0.4
0
1
Precision
Recall
0.6
USF1 (H1 ESC)
0 0.1
0
1
Precision
Recall
0.2
Cell type specific regulatory network construction
Sum all motif instances
in a given regulatory
element
Generate shuffled
motifs
Estimate
background expected
# motif hits
Estimate # motif instances above background
(threshold at 0.5)
10. #SNPs
# tagged enhancers (Liver)
0 30
0
100
GWAS SNPs can tag more than one
regulatory element
• ~50% of 197 total cholesterol variants tagging liver
regulatory elements, tag >1 element
Regulatory elements SNPs
LD block
Lead GWAS hit
GWAS variants
Regulatory element
targets
11. TFs
Infer regulators of
GWAS target
elements
M-step:
Refine target
elements
of variants
E-step:
GWAS variants (input)
Regulatory element
targets
Pruned GWAS
variants Apply to 47 traits:
CARDIoGRAM: LDL, HDL,
total cholesterol,
triglycerides, CAD
GLGC: LDL, HDL, total
cholesterol, triglycerides
WTCCC2: Multiple sclerosis
IBDG: Crohn’s, Ulcerative
colitis
MAGIC: Glycemic traits
DIAGRAM: T2D
ICBP: Systolic and diastolic
blood pressure
GIANT: BMI, weight, height
Network (input)
13. PWM-regulatory element incidence matrix
111targetregulatoryelements
Recurring PWMs
• More recurring PWMs compared
to previous incidence matrix for
disrupted binding elements (but
also by design.)
Total cholesterol
15. TBX15 over-expression yields decrease in lipid
accumulation in isolated human adipocyte cells
• TBX15 still reduces lipid accumulation even after constitutive
upregulation of a key adipocyte differentiation factor (PPARG)
• Now looking into mouse models for TBX15 knockdown as well
to look for closer connection to BMI
Control TBX15++ Control TBX15++
(++PPARG)
Gesta et al., 2011
16. IRX3 KD yields an increase in fat
accumulation
• Lipid accumulation in fat cells is a clear cellular phenotype
related to BMI
WT IRX3 KD
Perigonadal (visceral) fat
17. Whole body KD
Hypothalamus-
specific KD
Adipocyte-
Specific KD
Fat Mass Ratio (% of control)
0 9050
IRX3 KD effect is adipocyte-specific
• Lipid accumulation is only inhibited when knocked down in
adipocytes
Normal diet
High fat diet
IRX3 KD, Normal diet
IRX3 KD, High fat diet
19. Summary
• Using only enhancer and promoter maps, we can
recover known regulators and have prioritized
new ones for followup
• Cell type specific regulatory networks are still far
from complete
– ~25% of enhancers have no predicted binding
– PWM library is incomplete (~3,000+ regulators)
– Linking regulatory elements to genes is still a huge
challenge
20. MIT Computational Biology Group
Wouter
Meuleman
Jason ErnstSoheil FeiziGerald Quon
Daniel
Marbach
Bob
Altshuler
Anshul
Kundaje
Matt
Eaton
Abhishek
Sarkar
Pouya
KheradpourMariana
Mendoza
Jessica
Wu
Manasi
Vartak
David
Hendrix
Mukul
Bansal
Matt
Rasmussen
Stefan
Washietl
Andreas
Pfenning
Hayden
Metsky
Luis
Barrera
Manolis
Kellis
ENCODE Project Consortium
Roadmap Epigenome Mapping Consortium
Editor's Notes
To disentangle the molecular phenotype predictions on genotype level, we performed qPCR-based gene expression analysis of potential target genes using primary human adipose-derived mesenchymal cells from genotyped patient samples. A longstanding debate has been what exact target gene(s) might underlie the obesity association at FTO. Knock-out of genes flanking the intronic association region, including Fto (Fischer et al., 2009) and Rpgrip1l (Stratigopoulos et al., 2014), show body weight-related phenotypes in mice. Further, studies have pointed to long-range regulation of other flanking genes that could mediate the obesity signal of FTO intron 1 and 2, such as IRX3 (Ragvin et al., 2010)(Smemo et al., 2014) and RBL2 (Jowett et al., 2010).
Diverse tissues and cells, from adult, fetal, and progenitor cells
Ok, so attached already a knockdown of the TBX15 factor. I would include in your talk the Oil-Red-O staining experiment. Oil-Red-O stains lipids in fat cells. You identified TBX15 as a master regulator of BMI. You see that TBX15 indeed inhibits lipid accumulation in fat cells which is a clear cellular phenotype related to body mass index.
those are human fat cells/adipocytes isolated from subcutaneous adipose tissue from healthy subjects. I can send you a slide of the isolation process from whole tissue if you want (always impresses ;-).
Collaborating with jocelyn diabetes center to look at mouse models of conditional knockdowns
Primary human adopocytes cells -- siRNA
A primary driver of increased BMI is an increase in fat cell size (Arner Nature 2008), which involves a reduced lipid removal rate in existing adipocytes due to reduced lipolysis and fatty acid oxidation (Arner Nature 2011).
To demonstrate the causal effect of this cell autonomous, adipose lineage-specific regulation of body mass and fat mass on the organismal level, we analyzed a mouse model expressing a dominant negative form of Irx3 specifically in fat (Rosa26EnR-Irx3;ap2-Cre). The loss of Irx3 activity in fat reduced body size compared to Rosa26EnR control mice (Fig. 5a). In contrast to control mice, Irx3-EnR;ap2-Cre mice did not gain body mass on high-fat diet (Fig. 5b-c). This protection against diet-induced obesity was most pronounced in the fat-specific dominant negative Irx3 mouse model when compared to hypothalamic dominant negative Irx3 mice and whole body Irx3 knockout mice (Figure 5 d). This effect could be ascribed to lower mass of adipose tissues (Fig. 5d). Histological analysis confirmed diminished lipid accumulation in individual adipocytes in all measured fat pads resulting in decreased adipocyte size, with no apparent reduction in adipocyte number in Irx3-EnR;ap2-Cre compared to control mice (Fig. 5e). Further, Irx3-EnR;ap2-Cre mice were protected against high-fat diet induced fatty liver. White adipocytes store lipids in the form of a unilocular lipid droplet. Indeed, the expression of several genes involved in lipid accumulation was decreased in the transgenic mice compared to wild-type controls (Fig. 5f). White adipocytes store lipids in the form of a unilocular lipid droplet. The flux of lipids through lipid droplets largely depends on perilipins, an ancient family of droplet-coating proteins14. By qPCR analysis of perigonadal WAT, we observed an x-fold down-regulation of perilipin 2 (Plin2), while there was no effect on Plin1 mRNA (Fig. 5f). Of note, PLIN2 promotes lipid droplet formation for early adipocyte growth 15 and, in contrast to PLIN1, prevents access of adipose triglyceride lipase (ATGL) to the lipid droplet 16. Consistent with decreased lipid accumulation in the fat-specific Irx3 transgenic mice, the mice also showed a marked reduction in leptin mRNA (Fig. 5f).
In line with human adipose cells from risk allele carriers, expression of the adipogenic nuclear receptor Pparγ2 was increased in adipose tissue of Irx3-EnR;ap2-Cre mice, indicating that Irx3 does not promote generation of new adipocytes but instead promotes obesity via increased lipid accumulation in existing adipocytes. An adipocyte autonomous effect of IRX3 and IRX5 on lipid storage was further supported by mouse embryonic fibroblasts (Mefs) obtained from Irx3/5 double knock-out mice, which showed reduced lipid staining after adipogenic differentiation (Fig. 5g). Gene expression analysis of these cell cultures showed consistent effects on genes related to mitochnodrial function and lipid accumulation (Fig. 5h).
Han and CC to edit this paragraph
Han to generate Figure and to include quantification of percentage body fat for all models
Dominant negative allele === null allele
Promoter driven cre construct
In line with human adipose cells from risk allele carriers, expression of the adipogenic nuclear receptor Pparγ2 was increased in adipose tissue of Irx3-EnR;ap2-Cre mice, indicating that Irx3 does not promote generation of new adipocytes but instead promotes obesity via increased lipid accumulation in existing adipocytes. An adipocyte autonomous effect of IRX3 and IRX5 on lipid storage was further supported by mouse embryonic fibroblasts (Mefs) obtained from Irx3/5 double knock-out mice, which showed reduced lipid staining after adipogenic differentiation (Fig. 5g). Gene expression analysis of these cell cultures showed consistent effects on genes related to mitochnodrial function and lipid accumulation (Fig. 5h).
Han and CC to edit this paragraph
Han to generate Figure and to include quantification of percentage body fat for all models
Dominant negative allele === null allele
Promoter driven cre construct