Marie Sadler

A huge thanks to all the co-authors Alex, Caterina, Chiara, Diogo, Russ and to @zkutalik.bsky.social for supervising this work. Also many thanks to the #Fulbright program for funding my research stay at Stanford where this work started.

#DBC @unil.bsky.social @sib.swiss #Unisanté

10/10 To conclude, we show that EHRs present great opportunities to study longitudinal phenotypes such as drug response. Although we found low heritabilities for PGx traits, there are more genetic predictors to be identified by combining resources and conducting GWAS in more diverse ancestries.

9/10 #PRS of the underlying biomarker level can predict drug efficacy, and high baseline cholesterol PRS were associated with increased absolute, albeit lower relative LDL reduction following statin treatment. However, PRS explained less than 2% of the variance.

8/10 More generally, a GWAS on cholesterol progression—defined as the difference between two cholesterol meas. in statin-free individuals —identified loci associated with cholesterol levels when adjusting for baseline meas. A GWAS on progression without baseline adjustment did not detect any signal

7/10 For example, in the analysis of LDL-C response to statin, the genetic association with the rs7412 SNP (APOE gene) is inflated when adjusting for baseline levels, since this SNP is also strongly associated with LDL-C baseline levels.

6/10 To avoid spurious associations with genetic variants that are also associated with baseline biomarker genetics, we propose a theoretical framework to model drug response and longitudinal phenotypes in an unbiased manner which implies not adjusting biomarker differences for baseline levels.

5/10 The contribution of rare variants to drug efficacy was found to be modest compared to common variants. We identified GIMAP5, a gene associated with autoimmune diseases, to impact absolute HbA1c response to metformin, however, the association did not replicate in the All of Us data.

4/10 In a systematic comparison with the literature, we find that GWAS derived from EHRs have comparable statistical power to GWAS conducted on PGx consortia and clinical trial data.

3/10 A major challenge with EHR data is that medication start, baseline and post-treatment measures are not reported as such. We tested multiple strategies and found prescription regularity to be important as well as the time between prescription start and the first post-treatment measure.

2/10 We then conducted #pharmacogenetic GWAS and rare variant burden tests. Discovery GWAS were conducted in the @ukbiobank.bsky.social and replication analyses in the #AllOfUSResearch. We don’t discover any new pharmacogenetic signals, but also don’t miss any major signal reported previously.

1/10 We identified drug regimens from medication prescriptions and extracted biomarker levels before and after drug initiation for ten cardiometabolic drug-biomarker pairs. For example, following statin treatment, LDL-C levels levels drop by ~40% which is below the levels of statin-free controls.

We study the genetics of drug efficacy using longitudinal biomarker and medication prescription data from #EHRs. 🧬💊

Have a look at the 🧵 on i) modelling drug response and biomarker progression, ii) using #PRS to predict efficacy and iii) the challenges with #EHRs.

www.nature.com/articles/s41...

Tissue-specific co-expression networks to identify key genes 🧬 Have a look at Patrick's thread that explains how Downstreamer can link common genetic variation to rare diseases. Thank you for including Zoltán and me on this project!