Skip to content

Instantly share code, notes, and snippets.

@sinarueeger

sinarueeger/eshg-2019.md

Last active April 8, 2022 07:05
Show Gist options
  • Save sinarueeger/0d04308ebfcb14fb197231c37a405dd4 to your computer and use it in GitHub Desktop.
Save sinarueeger/0d04308ebfcb14fb197231c37a405dd4 to your computer and use it in GitHub Desktop.
Download ZIP
ESHG Göteborg 2019 https://2019.eshg.org/
Raw
eshg-2019.md

Notes from ESHG Göteborg 2019

Content

Saturday

PL1.1 Ulrike Peters: Genetic epidemiology of colorectal cancer - from discovery to prevention

Sunday

C09: Cancer genetics

C09.1: Noah Zaitlen: Germline genetic variation drives the somatic landscape of tumors

Article

1. Method

  • n=100K with sequenced cancer genome
  • low X sequencing
  • SBT (seeing beyond the target): examine off-target reads in (software stitch for low X): imputation of low X reads.
  • correlation between genetic ancestry and germline genotypes is high

2. Profile cohort

  • long tail of somatic mutations
  • smoking signature

3. Impact of genetic ancestry on somatic mutation

  • smoking signature differs between ancestries
  • more with PCA genetics ancestry

4. Impact of germline risk on the somatic landscape

PRS for smoking + total mutational burden (?)

Models for germline

  • germine gt > exposure > distal somatic mutations (heritable exposure)
  • germline gt alters the somatic mutation > leads to local and distal somatic mutations

5. Conclusion

C11: Statistical and population genetics

C11.1: Zoltán Kutalik: Maximum likelihood method quantifies the overall contribution of gene-environment interaction to complex traits: an application to obesity traits

Article

C11.2: Ninon Mounier: Leveraging correlated risks to increase power in Genome-Wide Association Studies

Article

C11.3: Hieab H. Adams: One and a half million genome-wide association studies of brain morphometry: a proof-of-concept study

About: Genetics of brain structure

By: ENIGMA, cohorts for heart and aging research in genomic epidemiology

Article

1. How do you measure brain structure?

  • what is the total size of the brain (GWAS done): crude measure - similar to only studying genome with karyogram
  • subcortical structure - similar to FISH
  • cortical brain region - similar to array CGH
  • shape of subcortical structures, results in 54K variables: shape heritable - similar to WES
  • voxel-based morphometry (similar to WGS), 1.5 mio variables

2. Aim: voxel-wise GWAS

  • n=18K, middle age and elderly
  • European ancestry
  • used for proof of concept

3. Challenges

  1. amount of computation time
  2. data transfer
  3. multiple testing correction (trillion of tests)

4. Method

5. QC

  • for genotypes: MAF > 0.05, R2 > 0.3
  • for brain images: removing low density voxel

6. Results

  • genome: large number of SNPs
  • brain: 62K unique voxels, 55 brain regions

C11.4: Maarja Lepamets: Genome-wide copy number variant association study reveals several novel disease-associated loci

About: CNV studies

  • genotyping arrays for CNVs are not ideal, can only capture CNV values between 0 and 4
  • CNV detection done with pennCNV

1. Validation

  • WGS from 983 Estonians
  • half (0.4) of the CNVs are missing (FN)
  • 0.008 probability of FP

2. PennCNV + scores

Previous method by Aurelien Mace

3. Additionally GEX and MET

  • CNV quality evaluation with gene expression (GE) and methylation data (MET).
  • mix of three score: WGS score, MET score and GE Score

4. CNV GWAS

  • harmonise CNVs (various lengths and breakpoints)
  • only del, only dup, only minor effects

Phenotypes

  • T2D
  • IBD: previous score did not work
  • CD: new hits

C11.5: Ross Byrne: Population structure and demographic change

  • n = 1626 dutch samples (from an ALS study)
  • ~ 300K SNPs overlap
  • geolocation

1. Background

  • pop structure = existince of subgroups in an super populations

2. Method

use of chromosome painting to avoid removing SNPs in LD

3. Results

  • finds local population structure (PC1 and PC2) - not working with normal PCA on SNPs
  • regional ancestry with German (north influence), French, Belgium (south influence), Danish
  • lower diversity of the north (split south-north in 600 AD)
  • looking at the length of haplotypes: short haplotypes = older. then doing PCA on these binned clusters

4. Conclusion

C11.6: Marie Verbanck: The landscape of pervasive horizontal pleiotropy in human genetic variation is driven by extreme polygenicity of human traits and diseases

Articles

1. Pleiotropy

Definition: single genetic element having multiple distinct phenotypic effects

Two types

  • vertical pleiotropy: leveraged in causal inference
  • biological (single-locus) horizontal pleiotropy: not well measured today

2. How to create a pleiotropic score?

  • using a collection of summary statistics
  • define for each phenotypes a vector of z scores, normalise them to be N(0,1)
  • whitening procedure to remove dependence between traits

Two types of pleiotropic score

  • total magnitude of the effect the variant has (Pm, m for magnitude)
  • number of traits pleiotropy score (Pn)

3. How to take LD between SNPs into account?

SNV1 in LD with SNV2. SNV1 has an effect of T1, SNV2 has an effect on T2.

  • therefore regress out the effect of LD scores from the pleiotropy score
  • additionally take polygenicity into account

4. Data

  • UKBB
  • only heritable phenotypes (p=372)

5. Results

Enrichment analysis:

  • more pleiotropy in UTR, Promotor, Transcription and regulation and enhancer
  • less in quiescent/ heterochromatin
  • more pleiotropy in eqtls
  • GWPS (GWAS with Pm and Pn as outcome)

6. Conclusions

Also: Marie Verbanck is moving to a new position in Paris.

W08: Investigating genotype-phenotype data using the GWAS Catalog

  • manual curation and ML methods
  • the basis of the catalog are GWAS publications
  • surprised to see that as criterium, only GWAS from array data (?)

S07: Polygenic risk scores coming of age

S07.1: Krista Fischer: Polygenic risk scores in genetic epidemiology

Articles

1. Introduction

  • history of GRS, emerged around 2012
  • weighted sum of SNP effects
  • what is the optimal selection of SNPs: not only significant

2. Basics

How to build a PGRS?

  1. Use a large GWAS (how to choose? large sample size, ancestry?)
  2. How to handle LD? (either LD pruned/clumped SNPs OR use method that accounts for LD structure)

Winner's curse: relationship between estimated effect size and p-value (smaller p-value the more effect size is overestimated)

Solution: use extra weight to account for this winners curse: doubly weighted score.

They looked at the incident T2D: analysis of 638 incident cases in overweight (but not obese).

3. If different large-scale GWAS meta-analyses are chosen, do they produce similar PGRS-s?

Läll et al 2019, BMC Cancer: different companies would report different risks.

4. metaGRS: combine different PRGRS to one score

  • combine m PGRSs scores
  • works similarly like doubly-weighted approach

5. Other approaches

  • LDpred: would it work better?
  • Lasso regression? may work slightly better

6. What else do we need to know?

  • Disease-free baseline survival
  • Effect of non-genetic risk factors
  • Effect of PGRS?

Implemented in Estonian biobank: genetic risk prediction and environment prediction (e.g. CVD GRS prediction, and depending on BMI).

7. Caution

8. Summary

  • no best methodology
  • PGRS is not unique (depends on GWAS used)
  • PGRS is ancestry-specific
  • moving towards personalised risk prediction

S07.2: Samuli Ripatti: Polygenic risks and their impact on behavior

  • Presents two parts
  • Finriski calculator
  • kardioKompassi
  • Framingham risk score vs 13 SNPs (2010) and 6.4 m SNPs
  • what has changed? better input with larger GWAS, better algorithms and better tests sets (UKBB)

Article

1. FinnGen

  • 500K = 10% of the population
  • combination of existing cohorts + new biobanking efforts
  • past 50 years phenotypic data of registry

2. LDPred is used

  • Summary stats + LD panel

3. Examples

  • T2D: 17K cases
  • Prostate Cancer
  • Prostate Cancer Death

4. Looking at high impact variants

  • e.g. effect of certain mutations, frameshift mutation

5. 2nd part: reporting back, does it help?

  • Does it matter to communicate back the risk based on a few weak effects SNPs? Nope, not really
  • But genetic information based on a few strong effect SNPs (BRCA) does.
  • What about GRS in Finland? GeneRISK

6. GeneRISK

7. Follow-up

  • 40% were at high risk previously
  • follow-up after 1.5 years
  • 15% stopped smoking
  • 16% started losing weight
  • females change a lot

Done by Elisabeth Widen.

8. Conclusion

Good PRGS + communicative tools = change in behaviour

S07.3: Rosalind Eeles: Polygenic risk scores in prostate cancer

Article

1. Introduction

  • prostate cancer prevalent and underdiagnosed in Africa
  • prostate cancer is 50% more common in identical twins than non-identical ones
  • what contributes to risk: rare (small effect size) and common variants (large effect size)

2. Data

3. PGRS

  • 27 of the loci were associated with aggressive and early-onset prostate cancer

Monday

S10: From genome-wide association study to mechanisms: fine-mapping

S10.1: Christian Benner: From association to causal variant(s): statistical methods for finemapping

Article

1. Introduction

  • PON1 gene with PON1 enzymatic activity in finrisk
  • Genomic regions can harbour many genetic variants
  • Summary statistics fine-mapping
  • Uses shotgun stochastic search (Hans et al. 2007)

2. Building blocks

  • y = x beta + eps (x genotype, beta effect size, y phenotype)
  • vary x as the genotypes (singletons, pairs, triplets, ...)
  • using bayes factor that uses summary level data
  • speeding up search with shotgun search (similar to MCMC)

3. Examples

  • LIPC association with HDL cholesterol (Surakka et al. 2015): pinpointing to 3 other variants than when doing conditional analysis.
  • PON1 gene with PON1 enzymatic activity in finrisk: difficult to pinpoint variants

4. Intermediate summary

  • Uses only summary stats
  • scales to large genomic regions

5. What if no LD reference panel?

  • Ex: APOE association with LDL cholesterol: 3 variants found with GWAS manual fine mapping + FINEMAP
  • if 1KG-FIN was used to calc LD, then other variants are found
  • simulation with NFBC and FINRISK: NFBC used for summary statistics + LD and FINRISK only for LD
  • Relationship between GWAS size and LD panel size: misspecified LD can cause FP and FN

6. fine-mapped heritability

  • hypothesis: large part of heritability could be explained by the fine-mapped variants
  • motivation: PON1. 5 lead variants explain 70%, bolt explains a bit less.
  • setting: BOLT vs FINEMAP
  • when small effects: FINEMAP underestimates
  • when large effects: FINEMAP has smaller MSE
  • Utility of FINEMAP heritability?

S10.2: Andrew P. Morris: Leveraging genome-wide association studies in diverse populations to fine-map complex human trait loci

Article

1. Introduction

  • LD: enables efficient efficient discovery of complex traits through GWAS array design + imputation
  • LD: hinders fine-mapping, because similar/exchangeable association
  • Extent of LD is weakest in African ancestry populations: hence optimal for fine-mapping
  • Trans-ethnic fine-mapping includes summary statistics from diverse populations

2. Are causal variants shared across populations?

Genomewise trans-ancestry meta-analysis provides insight into the genetic architecture of type2 diabetes susceptibility gives evidence to a model that the underlying causal variants are shared across global populations and arose prior to human pop migration out of Africa. Argues against "synthetic association" hypothesis.

Smiley face

^[From [Trans-Ethnic Fine-Mapping of Rare Causal Variants](https://link.springer.com/chapter/10.1007/978-1-4939-2824-8_18) (Wang & Teo 2011)]

3. Trans-ethnic meta-analysis

  • Benefits in modelling heterogeneity that is correlated with differences between populations

4. MANTRA

  • Article: Morris 2011
  • Similar to a Bayesian partition model
  • Uses summary statistics
  • Takes account of the expected similarity in allelic effects between the most closely related populations, while allowing for heterogeneity between more diverse ethnic groups

5. How does it perform in T2D and multi-ethnic cohorts?

  • n=22K cases + 40Kcontrols
  • from EUR, SAS, EAS, Hispanic, African American ancestry groups
  • first step: on basis of pair-wise mean allele frequency differences across the five loci: three clusters (Asian vs American vs European)

6. MR-MEGA: trans-ethnic meta regression

  • Article: Mägi et al. 2017
  • projects studies onto axes of genetic variation
  • includes principle components as covariates

7. Methodological challenges

S10.3: Sara L. Pulit: Large scale integration of genetic and -omics data to find susceptibility genes for obesity and fat distribution

Article

1. Introduction

  • Difference between obesity (measured with BMI) vs. fat distribution (measured with WHR adjusted for BMI)
  • WHR adj: pear vs apple shaped (decreased vs increased risk for metabolic syndrome/T2D)
  • UKB LMM BOLT + meta-analysis with GIANT: > 300 signal
  • WHR is genetically sex dimorphic: genetics effect tend to be stronger in women than men
  • next: check enrichment of GWAS signals in human tissue: FUMA
  • WHR not associated with Brain Tissue
  • obesity (BMI) not associated with brain tissue
  • next: fine-mapping

2. Fine-mapping

  • separate between GIANT vs UKBB
  • 1st problem: difference in number of SNPs: 2.5 M vs 27 M analyse separately, using GIANT as replication
  • 2nd problem: GIANT compromised between sub-cohorts, LD not accessible vs. UKB homogeneous and LD accessible

3. Follow-up

4. UKBB results

  • finds enrichment of likely causal EZH2 binding sides

5. Takeaways

  • fine-mapping and enrichment testing reveals sex-specific effects (potential role for EZH2)
  • what phenotype results from perturbing EZH2?

C16: Personalized and predictive medicine

C16.4: Nina J. Mars: High polygenic risk contributes to an early disease onset in common cardiometabolic diseases and cancers

C21.5: Hanieh Yaghootkar: Genome-wide association study of MRI liver iron content in 9,800 individuals yields new insights into its link with hepatic and extrahepatic diseases

S16: Methods for genetic epidemiology

S16.1: Hilary Finucane: Leveraging polygenic signals for insight into disease biology

Articles

1. Which gene is it?

Ways to prioritize genes at a locus:

  1. fine-map coding variant?
  2. use eQTLs (e.g. EWAS, eCaviar)
  3. local functional genomics
  4. closest gene
  5. similarity/network closeness
  6. Leveraging genome-wide enrichment: what about patterns that are constant across phenotypes (e.g. similar GEXPR enrichment patterns in BMI and schizophrenia)

2. Polygenic priority score (PoPS)

  • n Gene by p Gene features
  • vector with (likely disease gene) and (less likely disease gene)
  • predict "how likely a disease gene" is from a new gene data features

MAGMA model

  • LOCO approach
  • pvalue

Input:

  • Summary stats
  • LD panel
  • Gene feature matrix

Output:

  • gene scores and p-values

3. Validation

  • With gold standard genes: not lots of gold standard
  • "Benchmarker" for polygenic signal: Fine et al 2019 AJHG

4. High-confidence genes

  • how to quantify confidence?
  • can we estimate P(causal|prioritised & in locus)? see Stacey et al 2019 nucleic acids reserach
  • for 30-70% of loci, closest gene = causal gene
  • vs. P(causal | prioritised & closest)

5. Example with UKB CVD self-reported

Top features

  • Mouse and GO terms
  • ACE gene

Asthma

  • FCER1G gene

S16.2: George Davey-Smith: Genetic instruments in mendelian randomization studies

S16.3: Manuel Rivas: Large-scale inference of human genetic data

Material

https://biobankengine.stanford.edu/

Tuesday

S18: Our genetic history and its phenotypic consequences

S18.3: Alicia Martin: Consequences of population genetic differences in genetic risk prediction across diverse human populations

Articles

1. Motivation

  • Genetics has a huge diversity problem 70% European
  • How do biases in genetics impact the generalisability of knowledge with respect to different populations?

2. What differs across populations?

  • Causal variant effects are mostly shared across populations
  • AF, LD, environment and other factors differ?

3. PRS

4. Example GIANT vs. UKBB

5. Example FINRISK

6. UKBB

In UKBB

  • Europeans are best predicted,
  • American 1.6 less
  • SAN
  • African 4.5 x less

7. Why is that?

GWAS is best powered to discover common variants in the population we are studying

  • LD differences across populations
  • Biology and causal variants mostly shared, but estimations differ
  • Environment, selection, and other complicated differences (e.g. heritability differences)

UKBB and Biobank Japan (BBJ)

  • Ancestry matches or not
  • If discovery is UKBB, then the prediction in BBJ is less good. But when BBJ is uses, then UKBB prediction is almost equally good (differences in BBJ)

East Asian schizophrenia

EUR 3x larger sample size than EAS (13K cases)

Addressing genomics gaps in Africa

MAMA: multi ancestry meta-analysis

  • GWASs from three populations
  • approach: consider cross population to calibrate effect sizes in each pop
  • related methods: LD score and MTAG

Summary

  • PRS increase disparities due to Eurocentric GWAS biases
  • more diverse GWAS studies and new methods
  • communicate culturally sensitive topics responsibly and wildly

C25: Bioinformatics and multiomics

C25.1: Tom G. Richardson: A transcriptome-wide Mendelian randomization study to uncover tissue-dependent regulatory mechanisms across the human phenome

Article

C25.2: Sanni E. Ruotsalainen, Inst for Molecular Med Finland, Univ of Helsinki, Helsinki, Finland

Multivariate GWAS of inflammatory markers reveals novel disease associations

  • each variant with multiple phenotypes are tested

  • apply multivariate GWAS + find out which traits and SNPs were driving the associations

  • multivariate GWAS: MetaCCA (Cichonska et al. 2016)

  • CCA does not output effect size and se, hence no FINEMAP

  • solved by decomposing the multivariate GWAS results again on a single variant level.

  • MetaPhat (Lin et al. 2019): tool to detect and decompose multivariate associations from univariate GWAS stats.

C25.6: Tzung-Chien Hsieh: GestaltMatcher: Identifying the second patient of its kind in the phenotype space

Article

Notes

  • How to identify another patient with a rare phenotype?
  • This is important to increase power for analysis
  • Using facial representation
  • matching patient to other people in a database
  • Data used: PEDIA
  • extract facial features, deep convolutional network

Posters

Nordin et al. 2019 https://www.biorxiv.org/content/10.1101/660241v1

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment