Proteome-Wide Drug-Target Mendelian Randomization - Usage Guide

Overview

Cis-pQTL Mendelian randomization is the genetic-instrument analogue of a placebo-controlled drug trial: a genetic variant near the gene encoding a target protein perturbs that protein's plasma level from conception, and the variant's effect on a downstream phenotype estimates the causal effect of pharmacological inhibition of that target. The PCSK9-LDL-cholesterol-CAD story is the canonical positive example: cis-MR at the PCSK9 locus correctly predicted the direction and magnitude of clinical effect of evolocumab and alirocumab years before the FOURIER and ODYSSEY trials. The PCSK9 -> type 2 diabetes signal (Schmidt 2017 Lancet Diabetes Endocrinol) was the canonical on-target adverse-effect discovery.

This skill operationalises the Schmidt 2020 cis-MR framework (Nat Commun 11:3255) for plasma-proteome drug-target validation. It handles cis-pQTL instrument selection from UKB-PPP (Olink antibody), deCODE / Fenland / INTERVAL / ARIC (SomaScan aptamer), and FinnGen-PPP; cis-IVW with correlated instruments; colocalization triangulation; phenome-wide cis-MR for on-target adverse-effect scanning; cross-platform replication; and PAV (protein-altering-variant) sensitivity. The agent enforces the triangulation rule that a single significant cis-MR is necessary but never sufficient for a drug-target nomination.

Prerequisites

install.packages(c('remotes', 'MendelianRandomization', 'coloc', 'susieR', 'dplyr'))
remotes::install_github('MRCIEU/TwoSampleMR')
remotes::install_github('MRCIEU/ieugwasr')
remotes::install_github('MRCIEU/genetics.binaRies')   # bundles plink (1.90) binary
remotes::install_github('rondolab/MR-PRESSO')

# Ensembl VEP for PAV annotation
conda install -c bioconda ensembl-vep
vep_install -a cf -s homo_sapiens -y GRCh38 -c $HOME/.vep

# 1000 Genomes EUR plink reference (for local clumping and LD matrix)
# https://mrcieu.github.io/ieugwasr/articles/local_ld.html

pQTL summary statistics: - UKB-PPP Olink (Sun 2023): ukb-ppp.gwas.eu portal (download by protein UniProt ID) - deCODE SomaScan (Ferkingstad 2021): https://www.decode.com/summarydata/ - Fenland SomaScan (Pietzner 2021): EBI GWAS catalog - FinnGen-PPP Olink: FinnGen DF12 release portal

Quick Start

Tell the AI agent what to do: - "Run a cis-MR of PCSK9 on coronary artery disease using UKB-PPP cis-pQTLs" - "Test IL6R inhibition on rheumatoid arthritis with cis-pQTL instruments from deCODE plus coloc triangulation" - "Run a phenome-wide cis-MR of PCSK9 across OpenGWAS to find on-target adverse effects" - "Replicate this cis-MR target across UKB-PPP Olink and deCODE SomaScan and flag PAV-confounded instruments" - "Test a target with two independent cis-signals using coloc.susie per credible set" - "Apply the Burgess 2016 sample-overlap correction to a UKB-PPP exposure with a UKB phenotype outcome"

Example Prompts

Single Drug Target, Single Outcome

"I have UKB-PPP cis-pQTLs for PCSK9 within +/-500 kb of the gene and CARDIoGRAMplusC4D CAD summary stats. Run cis-IVW with weak-IV filtering, harmonise with action=2, triangulate with coloc.abf at p12=5e-6, and report PP.H4 plus PAV-excluded sensitivity. Annotate every cis-pQTL with VEP first."

"Test IL6R protein -> rheumatoid arthritis using cis-pQTLs from deCODE SomaScan plus colocalization. Flag any cis-pQTL coloc'd with neighbouring genes' eQTLs in GTEx whole blood."

Phenome-Wide On-Target Adverse-Effect Scan

"Hold the PCSK9 cis-pQTL instrument set fixed and run cis-MR against all OpenGWAS outcomes with sample size >= 50,000 in European-ancestry cohorts. Bonferroni-correct over outcomes and report a phewas-style forest plot of significant hits."

"Screen IL23R cis-pQTL effects across all FinnGen DF12 disease endpoints for on-target adverse-effect discovery prior to a clinical-trial protocol."

Target with Multiple Independent Cis-Signals (Allelic Heterogeneity)

"ANGPTL3 has two independent cis-pQTLs in low LD. Run coloc.susie on the cis-window with in-sample LD, report per-credible-set PP.H4 against triglycerides GWAS, and run a Wald ratio per credible set."

Cross-Platform Replication

"Run the same cis-MR of target X on disease Y independently in UKB-PPP (Olink) and deCODE (SomaScan). Report direction agreement, magnitude ratio, and flag the protein as platform-discordant if direction disagrees."

Sample-Overlap Correction

"Both my exposure GWAS (UKB-PPP cis-pQTL for protein X) and outcome GWAS (UKB HES-derived phenotype) are from UK Biobank. Apply MR-RAPS with the one-sample-equivalent treatment, or switch the outcome to FinnGen for an independent-cohort test."

Drug Repurposing / Target Nomination

"Cross-reference cis-MR estimates with Open Targets L2G scores and the Open Targets Drug platform to nominate druggable proteins for an autoimmune indication. Require cis-MR P < 1.7e-5, coloc PP.H4 >= 0.7, and cross-platform replication."

What the Agent Will Do

Identify the target gene's coordinates (hg38), define the +/-500 kb cis-window
Download or load cis-pQTL summary stats for the target protein from UKB-PPP (Olink) and ideally also deCODE / Fenland (SomaScan)
Filter to genome-wide-significant cis-SNPs (P < 5e-8) inside the cis-window
Compute per-SNP F-statistic from the EXPOSURE GWAS (not outcome); drop F < 10
Annotate every retained cis-SNP with Ensembl VEP; tag PAVs (missense/nonsense/splice/protein_altering)
LD-clump within the window at r2 < 0.1 using local plink + 1KG (ancestry-matched)
Harmonise with the outcome GWAS via harmonise_data(action = 2); document palindromic drops
Run cis-MR panel: Wald ratio (sentinel) + cis-IVW (random) + Egger (if >=10 SNPs) + weighted median + MR-PRESSO outlier test
Run cis-IVW with correl = TRUE if instruments are in moderate LD; supply the in-window LD matrix from plink
Run coloc.abf (or coloc.susie if multi-causal evidence) on the full cis-window with p12 = 5e-6; report PP.H4 + sensitivity
For each cis-pQTL, run coloc against non-target genes' eQTL/pQTL in the same window (drop if PP.H4 >= 0.5 with neighbour)
Run PAV-excluded sensitivity: re-run cis-MR after dropping all PAV cis-SNPs; report both
If cross-platform replication: repeat steps 2-11 on the alternate platform; report agreement
If sample overlap (UKB-on-UKB): apply Burgess 2016 correction or switch outcome to FinnGen
For pheWAS: loop step 7-8 over the OpenGWAS outcome catalogue; Bonferroni-correct
Produce STROBE-MR-compliant report listing: cis-MR estimate + 95% CI, PP.H4, PAV-excluded estimate, platform agreement, neighbour-gene coloc results, sample-overlap statement, claim-strength ladder

Tips

The exclusion-restriction in cis-MR is relaxed (the protein product directly mediates) but NOT eliminated; neighbour-gene mediation is the dominant residual concern
Always VEP-annotate cis-pQTLs and run a PAV-excluded sensitivity panel; PAVs can produce apparent platform-specific pQTLs without changing actual protein abundance
Olink (antibody PEA) and SomaScan (aptamer SOMAmer) show only modest cross-platform correlation, with roughly half of cis-pQTL signals not shared across platforms; replicate every clinically-actionable claim on both
Cross-platform direction disagreement is a strong flag for measurement artifact; do not advance such targets to nomination
Sample-overlap: treating UKB-PPP exposure + UKB outcome as two-sample is a common methodological error; bias is one-sample-equivalent, toward observational
Cis-MR clumping uses r2 < 0.1 within window (not r2 < 0.001 polygenic-MR convention); the window itself is the LD-pruning mechanism
Egger needs >=10 cis-pQTLs to be powered; most cis-windows have fewer, so cis-MR-Egger is usually exploratory only
Phenome-wide cis-MR is the canonical use case for discovering on-target adverse effects (PCSK9 -> T2D is the textbook example)
For correlated cis-pQTLs (r2 0.1-0.7), build the input with MendelianRandomization::mr_input(..., correlation = ld_matrix) then call mr_ivw(mr_obj, model = 'default', correl = TRUE); supply the matrix from local plink against an ancestry-matched reference
coloc PP.H4 >= 0.7 is the Open Targets clinical bar; PP.H4 >= 0.5 alone is exploratory; require PP.H3 to be substantially less than PP.H4 for confidence
The Steiger filter has a known caveat under unmeasured confounding (Lutz 2022); cross-validate direction with bidirectional cis-MR
L2G score (Mountjoy 2021) plus cis-MR plus coloc plus existing-drug evidence is the modern Open Targets target-prioritization stack
Trans-pQTLs (outside +/-500 kb) cannot serve as primary instruments; their inclusion violates exclusion-restriction by definition
Some proteins (e.g. complement factors, immunoglobulins) have many PAVs in cis; the PAV-excluded sensitivity may discard most instruments and the cis-MR claim cannot be made
The cis-window default is +/-500 kb (Schmidt 2020); some pipelines use 1 Mb. Pre-specify and document the window in methods

causal-genomics/mendelian-randomization - Parent polygenic-MR framework; cis-MR is the drug-target specialization causal-genomics/colocalization-analysis - Required PP.H4 triangulation for any cis-MR drug-target claim causal-genomics/fine-mapping - Credible-set construction prior to coloc.susie at the cis-locus causal-genomics/pleiotropy-detection - MR-PRESSO / Egger diagnostics adapted to cis-window causal-genomics/transcriptome-wide-association - eQTL-based parallel evidence for the same target causal-genomics/mediation-analysis - Step from cis-MR to downstream mediator pathway population-genetics/association-testing - Source GWAS pipelines for pQTL discovery population-genetics/linkage-disequilibrium - LD-matrix construction for cis-IVW correl=TRUE variant-calling/variant-annotation - VEP PAV annotation for sensitivity analysis clinical-databases/clinvar-lookup - Pathogenic-variant context for nominated targets