Thread by @Joni_Coleman

Lots of great posters and lunch later, I am now in the methods session for #WCPG2022 (it was a hard decision over depression...) First up is Wouter Peyrot talking about DDx-PRS

@caina89 and @hailianghuang introduce WP. DDX-PRS distinguishes between different psychiatric disorders. Early diagnosis is challenging - non specific symptoms often give way to more specific symptoms ten years on - would be valuable to be able to intervene early #WCPG2022

Genetics is well situated to act as a stable predictor of separating disorders early on. No specific methods to achieve this - enter DDX-PRS. Predicts probability of given disorder from individual genotypes. PRS --> liability scale #WCPG2022

--> covariance across PRS and liability to predict probability for each disorder. Can be tailored to local diagnostic environment #WCPG2022

Tuning needed? Generate PRS from discovery GWAS, want to use it in target dataset. Compared tuning vs no tuning. #WCPG2022

Is DDx-PRS better than other approaches? Compared across simulations.#WCPG2022

DDX-PRS well calibrated in simulation. Metric for clinical utility - 16% of individuals go from 25% --> >50% probability. #WCPG2022

DDX-PRS is not well calibrated in empirical data (better if use a tuning dataset) Issue of meta-analysis across population? Will compare results in iPSYCH alone #WCPG2022

Empirical results show robust clinical utility of 15%. With better PRS closer to h2 that could go up to 30-40%. Could be useful for targeted follow-up. Need to try to get robustness without tuning. #WCPG2022

Next up is Christian de Leeuw who will talk about decomposing genetic correlations. Obtain genetic correlations across the genome - just a summary of relationship. How heterogeneous is that relationship? #WCPG2022

Work builds on LAVA, which estimates local genetic correlation across multiple traits. See variability in local h2 - does this reflect measurement error or true heterogeneity #WCPG2022

Compare 13 phenotypes, 78 pairs. See good relationship of mean LAVA rg and LDSC rg - seem to estimate same underlying quantity #WCPG2022

Can measure variance of local genetic correlations, and look at difference of variance from 0. See significant heterogeneity for 75/78 pairs. #WCPG2022

Can also estimate intermediate level, looking per chromosome rather than genome or local. See significant explanation of heterogeneity in 24[?]/78 #WCPG2022

Next up in the VU session is Emil Uffelmann, who will talk about predicting an individual's disorder with PRS #WCPG2022

Same question as DDx-PRS - what is the probability this person you are seeing in the clinic has disorder X? Calibration is key - need agreement between proposed and actual diagnosis #WCPG2022

Multiple such methods exist - need a comparison paper in the future! EU's focus is on Bayesian PRS. Well calibrated for continuous traits, but not for binary traits due to ascertainment. Aim to recalibrate to fix this and predict disorder #WCPG2022

Can rescale PRS to make variance(PRS) equal to R2. Transform PRS to liability scale, and then onwards to ascertained observed scale. #WCPG2022

Simulated data like MDD and SCZ, ascertainment at 50% cases. PRS estimated with BPred. Raw PRS confirms not well calibrated, worse at low prev and high R2. Estimates also become more precisely miscalibrated #WCPG2022

Recalibrate - average converges to 1, so this is working. Does this help prediction? Yes (at least in more powerful cases, where recalibration is important) #WCPG2022

Want to expand to other methods and other diseases and disorders. #WCPG2022

Now we have Jette Steinbach, who will discuss improving the predictive value of family history in predicting disorder. #WCPG2022

Family history is an important risk factor for psychiatric disorders. Need more information than a simple yes/no - multiple members might indicate higher risk. Not simple to quantify (and trivial to think of multiple other relevant factors) #WCPG2022

Can build on the liability threshold model to incorporate other factors (as has previously been done for age, gender, birth year etc.) Can one include family history in the same way? #WCPG2022

Can use the age etc informed model to calculate thresholds for each family member. Can then use the resulting threshold plus the covariance matrix between family members to inform risk estimate for focal individual #WCPG2022

Built a model to do this in LTFH+ github.com/EmilMiP/LTFHPl… #WCPG2022

Applied in practice to Danish registry data. Compared prediction accuracy for case control status to various baseline models (such as gender, age, batch and PCs, PRS etc.) #WCPG2022

FH + other predictors outperforms other models, esp PRS + FH. See similar results with specific relative history, worst with sibling history (but iPSYCH is young) #WCPG2022

Next up is Clive Hoggard to talk about BRIDGE-PRS, a method for cross population PRS analysis #WCPG2022

Two strategies - refine variants using fine-mapping [see e.g. PolyPred, PRS-CSx] or aggregate info across putative loci, which assumes the existence of the causal variant but not that it necessarily exists in the data - this is what BRIDGE-PRS does #WCPG2022

Compare PRS-CSX with BRIDGE-PRS - CSX builds off fine mapping. Bridge uses ridge regression in base population to combat winner's curse. Then uses that as a prior for reestimate in the target population to update PRS. Applied within LD regions (but keep all SNPs) #WCPG2022

Calculate PRS across spectrum of prior parameters, and then create single PRS from all the parameters - avoids MCMC estimation #WCPG2022

Examined BRIDGE-PRS in simulated data [too many parameters to cover in a tweet, need to read the paper]. Outperforms PRS-CX when prop causal is high, and where causal variant is not in the data. True across ancestries #WCPG2022

Analyses in real data also outperform or match PRS-CSX for AFR, but not for EAS #WCPG2022