Thread by @Joni_Coleman

Apologies to the morning folks in the education session (5am landing and an hour or so walking the streets of Florence required a bit of sleep...)

But I'm only chuffing here! Enjoying listening to @KKuchenbaecker explaining the vanishing boundaries of ancestry #WCPG2022

@KKuchenbaecker KK: Defining specific ancestry groups (e.g. through PCA) can be valuable, but comes with risks of abuse by bad actors. Definitions are not always straightforward and easy to get wrong - relatives have an influence on this (and this also --> misestimates of relatives) #WCPG2022

Roadmap for sustainable diverse gneomic studies: pubmed.ncbi.nlm.nih.gov/35145307/. Need to engage groups being studied as active partners and leads of the studies, and have strong focus on capacity building. Highlights efforts in Pakistan #WCPG2022

Next up is @egatkinson to talk about developing better methods to engage with ancestral diversity in GWAS. Starts with a point KK made earlier - EUR dominates GWAS and this isn't changing. Risks exacerbating existing health disparities #WCPG2022

@egatkinson Studying understudied populations is morally, clinically and scientifically justified. Particular among these are admixed groups, including African Americans and Latinos (1/3 of US population). Data often exists, but is not being used #WCPG2022

Why? Fear population stratification will bias GWAS results. Pop strat - coincidence of ubiquitous ancestral geographic differences in variant frequency with genographic differences in phenotype. Usually addressed with PCA - but crude, limited, tends to exclude admixed #WCPG2022

Ancestry essentially dealt with by avoiding it - discarding all but the bulk of collected samples. Enter TRACTOR - split individuals into their ancestral segments, and analyse all data by focussing on segments, not individuals #WCPG2022

How does TRACTOR work? Uses ubiquitous patchwork chromosomes of all organisms (segments of correlated DNA inherited from ancestors) --> segments of local ancestry #WCPG2022

Assess local ancestry (through chromosome painting), and then analyse by pieces of same local ancestry. Chromosome painting uses phased genotype data and compares variants on the basis of ancestry allele frequencies #WCPG2022

Local ancestry inference approach is very accurate - >97% correct calls from simulated data, and produces useful outputs (e.g. local ancestry specific weights for polygenic risk scoring). Resulting TRACTOR GWAS boosts power of GWAS compared to traditional methods #WCPG2022

Want to know more about TRACTOR works? Check out the GitHub: github.com/Atkinson-Lab/T… #WCPG2022

TRACTOR doesn't have major power loss when local ancestry irrelevant, and has good control of false positivity. Simulated effect sizes are recapitulated accurately by TRACTOR across multiple models. #WCPG2022

TRACTOR is also valuable for fine-mapping - reduces number of potentially causal variants, prioritises different variants to traditional analysis (because it incorporates information from disruptive LD blocks from different local ancestries) #WCPG2022

TRACTOR also improves phasing, effectively recovers some long range tracts of ancestry. Read the paper for more: nature.com/articles/s4158… #WCPG2022

Next, Ying Wang presents on incorporating diverse ancestry data in polygenic risk scores. Begins with a definition of polygenic risk scores - GWAS-weighted sum of risk alleles (ID'd from GWAS) in new genotype data independent of GWAS #WCPG2022

Polygenicity means different patterns of risk alleles across at-risk individuals are expected, so PRS summarises [covers up?] between-individual diversity #WCPG2022

Applications of PRS across differently ascertained cohorts, with different aims: pubmed.ncbi.nlm.nih.gov/32997097/ #WCPG2022

E.g. risk stratification and stratified screening in breast cancer - high risk (indexed by PRS) --> early screening, low risk --> later/no screening #WCPG2022

PRS analysis: do a GWAS. GWAS has errors e.g. winner's curse (overestimation of effect sizes due to preferentially selecting variants with true effect + positive noise). Numerous methods to deal with this, esp. machine learning methods but also simple LD-clumping #WCPG2022

Metrics used to assess PRS effectiveness - R2, AUC. Separation of PRS distribution in cases and controls. Stratification of sample into relevant strata and OR between strata. #WCPG2022

PRS accuracy is theoretically limited by (broad-sense) SNP-based heritability. Also practical limits of measurable heritability and GWAS, target sample sizes #WCPG2022

Transferability of PRS is limited across ancestries, esp. others --> AFR. Transferability variable across studies and phenotypes. Partly due to differences in allele frequency and LD (variable by phenotype due to different causal variants) #WCPG2022

Methods being developed to address this e.g. PRSCSx. But methods development not enough, need major efforts to prioritise diversity (H3Africa, AllofUS) etc. But major efforts are still EUR dominated. Need more from global south #WCPG2022

Examples from the Global Biobank Meta-analysis Initiative. 2.2M individuals. PRS accuracy heterogeneous across ancestries and biobanks. e.g. very different performance of EUR asthma PRS-->EAS in different EAS biobanks #WCPG2022

Leave-one-biobank out GWAS --> much more powerful PRS than biggest asthma GWAS #WCPG2022

Genetics are not enough to counter health disparities. Need to measure and model environments effectively. Systematic modelling of environments through identifying latent factors #WCPG2022

Check out the paper: medrxiv.org/content/10.110… #WCPG2022