Bio Causal Genomics Proteome Mr Drug Target

Version Compatibility

Reference examples tested with: TwoSampleMR 0.5.11+, MendelianRandomization 0.10+, MR-PRESSO 1.0+, coloc 5.2.3+, susieR 0.12.35+, ieugwasr 1.0+, plink2 2.00a5+, R 4.4+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters
• CLI: plink2 --version; VEP vep --help

If code throws OAuth or rate-limit errors from OpenGWAS, or a missing dataset$N from coloc, introspect the installed API and adapt the example rather than retrying. UKB-PPP, deCODE, and Fenland summary statistics changed file layouts between 2023 and 2025; verify column headers before passing into format_data().

Proteome-Wide Drug-Target Mendelian Randomization

"Does genetically lowering plasma protein X cause a change in disease Y, mimicking a drug?" -> Use cis-pQTLs in the gene window for protein X as instruments under the Schmidt 2020 framework (Nat Commun 11:3255), restrict the exclusion-restriction violation to the geometric neighbourhood of the encoding gene, triangulate with colocalization (3-tier PP.H4 ladder below) and cross-platform replication (Olink vs SomaScan), and flag protein-altering-variant (PAV) confounding. A single significant cis-MR estimate is necessary but not sufficient for a drug-target claim; the operational bar is MR + coloc + cross-platform agreement + PAV-excluded sensitivity.

PP.H4 Three-Tier Threshold Ladder

Operational rule: drug-target nomination requires PP.H4 >= 0.8 minimum; industry-grade clinical claim requires PP.H4 >= 0.95 plus the full triangulation panel.

• R (canonical): TwoSampleMR::mr() orchestrates the cis-IVW + Egger + median + Wald-ratio panel
• R (correlated cis-pQTLs in a window): MendelianRandomization::mr_input(..., correlation = ld_matrix) then mr_ivw(mr_obj, model='default', correl = TRUE)
• R (triangulation): coloc::coloc.abf() or coloc::coloc.susie() on the same cis-window
• pheWAS: ieugwasr::associations() against the OpenGWAS catalogue, looped over outcomes
• VEP CLI: annotate every cis-pQTL with vep --species homo_sapiens --canonical --check_existing for PAV flagging

Data Source Taxonomy

Methodology evolves; check the UKB-PPP portal (ukb-ppp.gwas.eu), deCODE Genetics summary-stat releases, and the eQTL Catalogue / GTEx Portal for the current data version. UKB-PPP was substantially re-released in 2024 (extended ancestry meta-analyses); pin a download date in the methods.

Platform and Cohort Versioning (2024-2026)

• UKB-PPP Phase 2 (2024) -- cross-ancestry meta-analyses; ~54k EUR plus multi-ancestry expansion; file layouts differ from Phase 1 (per-protein parquet vs flat TSV); the 14,287 primary pQTL count from Phase 1 shifts in Phase 2 results. Verify the freeze used.
• FinnGen-PPP -- Olink Explore platform; DF12 (2024) release is current; Finnish-specific allele frequencies require ancestry-aware downstream analysis.
• AoU pQTL (2024) -- All-of-Us proteomics; pre-symptomatic-cohort design strength for reverse-causation control and longitudinal follow-up.
• SomaScan v4 vs v5 -- v5 expands to ~11k SOMAmers (vs ~5k in v4); binding consistency for shared SOMAmers documented in deCODE / SomaLogic technical notes but not guaranteed; pin platform version.
• Olink Explore HT -- ~5,400 proteins (vs ~3,072 in Explore Expansion, ~1,536 in Explore 1536); UKB-PPP Phase 1 used Explore 3072 -- do not assume Explore HT coverage when reading Phase 1 papers.
• Pin specific platform version + release date in the methods section of every cis-MR drug-target manuscript.

Cis-MR Methodological Taxonomy

Verify against Burgess 2023 Wellcome Open Res "Guidelines for performing Mendelian randomization" (v3+) and the Open Targets Genetics drug-target pipeline (Mountjoy 2021) before pinning a primary method for a publication.

Decision Tree by Scenario

Per-Method Failure Modes

Olink vs SomaScan platform discordance

Trigger: A cis-pQTL effect size or even direction differs between UKB-PPP (Olink antibody) and deCODE/Fenland (SomaScan aptamer) for the same protein.

Mechanism: Olink uses paired antibody proximity-extension assay (PEA) binding distinct epitopes; SomaScan uses single-aptamer SOMAmer binding a folded epitope. A missense SNP that alters one epitope produces an apparent pQTL on that platform only (a pseudo-PAVQTL / aptamer-affinity QTL, AAVQTL). Splice/isoform differences also produce platform-specific signal. A direct Olink-vs-SomaScan cross-platform comparison reported only modest cross-platform correlation, with roughly half of cis-pQTL signals not shared across platforms (72% of Olink vs 43% of SomaScan assays had a cis-pQTL) (Eldjarn G et al 2023 Nature 622:348).

Symptom: Cis-MR significant on one platform, null on the other; or significant on both but opposite direction.

Fix: Replicate every cis-MR claim on at least one alternate platform; annotate cis-pQTLs with VEP and flag missense / nonsense / splice variants in the gene; consult Olink and SomaScan documentation for the protein's antibody / aptamer binding region; perform a PAV-excluded sensitivity analysis and report both estimates. If platforms disagree irreconcilably, report the protein as platform-discordant and do NOT advance for clinical claim.

Olink panel coverage pre-check: Olink Explore 3072 covers ~3,000 proteins; Explore Expansion ~3,000; Explore HT ~5,400. Always confirm the target is on the panel BEFORE running cis-MR: grep -i <GENE> olink_panel_proteins.tsv (panel manifest from olink.com). Sun 2023 Nature supplementary Table S1 lists every UKB-PPP-covered protein explicitly; absence from Table S1 means the target was not measured in Phase 1 (regardless of biology) and the analysis must use SomaScan.

LD-based pleiotropy in the cis-window

Trigger: A cis-pQTL for target gene X is also a strong eQTL or pQTL for a neighbouring gene within +/-500 kb.

Mechanism: The instrument's effect on outcome may be mediated by the neighbour gene's protein, not by target X. Cis-window proximity does NOT guarantee specificity.

Symptom: Colocalization with the non-target gene's pQTL or eQTL returns PP.H4 >= 0.7; cis-MR using only "clean" cis-pQTLs (those not coloc'd with neighbours) gives a substantially different estimate.

Fix: For every cis-pQTL, run colocalization against all eQTL/pQTL signals within +/-500 kb in the relevant tissue; drop cis-pQTLs that coloc (PP.H4 >= 0.5) with any non-target gene; coloc.susie is preferred when multiple credible sets exist in the window. Reports must list which cis-pQTLs were retained and the rationale.

Reverse causation from disease state on plasma protein

Trigger: The outcome trait elevates the protein as a downstream consequence (e.g. CRP elevated in coronary disease patients; TNF in autoimmune disease).

Mechanism: Plasma proteomes measured in observational cohorts (including UKB-PPP) include people who have or will develop the outcome; the observed protein-disease association may be downstream not upstream.

Symptom: Observational protein-disease association is large; cis-MR estimate is much smaller, null, or in the opposite direction.

Fix: Apply Steiger filtering on each cis-pQTL (steiger_filtering()); restrict to pre-symptomatic samples where possible (pediatric or early-adult cohorts); replicate in longitudinal cohorts measuring protein years before disease onset. Note: Steiger has its own caveat under unmeasured confounding (Lutz SM et al 2022 Genet Epidemiol 46:139); cross-validate via bidirectional cis-MR.

library(TwoSampleMR)
dat <- harmonise_data(exposure_pQTL, outcome_GWAS)
dat <- steiger_filtering(dat)
dat_forward <- dat[dat$steiger_dir, ]   # drop reverse-direction SNPs
dir_test <- directionality_test(dat_forward)

PAV (protein-altering-variant) confound

Trigger: A cis-pQTL is itself a missense, nonsense, frameshift, splice-site, or stop-gain variant in the target gene.

Mechanism: The variant changes the protein sequence, which can change antibody affinity (Olink) or SOMAmer affinity (SomaScan) without changing the actual protein abundance in plasma. The pQTL appears strong but reflects measurement artifact.

Symptom: The strongest cis-pQTL is a coding variant; cis-MR effect magnitude shrinks substantially when PAVs are excluded; the pQTL is platform-specific.

Fix: Annotate ALL cis-pQTLs with Ensembl VEP (--check_existing --canonical). Tabulate every cis-pQTL's most-severe consequence. Run two cis-MR analyses: (a) all cis-pQTLs, (b) PAV-excluded. Report both; require concordance for a publishable claim (Sun 2023 supplementary). For aptamer panels, also annotate the SOMAmer binding-region overlap if available.

PAV-excluded concordance rule: Concordance between all-cis and PAV-excluded estimates requires (i) effect direction preserved, (ii) |effect| within 2x of the all-cis estimate, AND (iii) p-value still nominally significant (P < 0.05) after PAV exclusion. If ANY of the three criteria fails, report both estimates and downgrade the claim from "drug-target" to "suggestive cis association requiring orthogonal confirmation."

Trans-pQTL pleiotropy if used as instrument

Trigger: Including trans-pQTLs (outside +/-500 kb of the gene) in the instrument set to gain power.

Mechanism: A trans-pQTL acts through some other gene's protein that then regulates target X; using it as an instrument violates the exclusion-restriction by definition (horizontal pleiotropy).

Symptom: Effect estimate shifts when trans-pQTLs are added; Egger intercept significant.

Fix: Restrict instrument set to cis-window only (Schmidt 2020). Trans-pQTLs may be confirmatory ("does the regulator gene also predict outcome?") but never primary instruments for a drug-target claim.

Sample overlap when both ends are UKB

Trigger: Using UKB-PPP for the exposure (Olink protein) AND a UKB phenotype (HES, cancer registry, ICD-10) for the outcome.

Mechanism: The same individuals contribute to both summary statistics; weak-instrument bias is now one-sample-equivalent and points TOWARD the confounded observational estimate, not the null.

Symptom: UKB-on-UKB cis-MR effect is substantially larger than the same protein-disease pair tested in an independent cohort.

Fix: Use an independent outcome cohort whenever possible (FinnGen, Biobank Japan, MVP). If UKB-on-UKB is necessary, apply MR-RAPS with one-sample-equivalent treatment OR the Burgess 2016 (Genet Epidemiol 40:597) sample-overlap correction. Document the overlap fraction. MRlap (Mounier 2023 Genet Epidemiol 47:314) is the recommended modern tool for joint sample-overlap and winner's-curse correction; see causal-genomics/mendelian-randomization for the canonical implementation.

Triangulation Requirement (Operational Postdoc Rule)

A drug-target causal claim that survives peer review and informs pharmacology requires MULTIPLE concordant streams, not a single significant cis-MR p-value:

1. Cis-MR estimate with P < 0.05 / N_proteins (Bonferroni proteome-wide ~ 1.7e-5 for 2923 Olink proteins) OR P < 0.05 / N_outcomes (Bonferroni pheWAS-wide) -- depending on the testing regime
2. Colocalization PP.H4 >= 0.8 between cis-pQTL and outcome GWAS at the gene locus -- standard publication tier; >= 0.95 for industry-grade claim (cross-reference causal-genomics/colocalization-analysis)
3. Cross-platform replication -- significant on both Olink (UKB-PPP) AND SomaScan (deCODE or Fenland); direction agreement is mandatory, magnitude within 2x is acceptable
4. Cross-cohort replication (ideal but not strictly required) -- UKB-PPP -> deCODE -> FinnGen-PPP step-up
5. PAV-excluded sensitivity -- the cis-MR survives when missense / nonsense / splice / aptamer-binding-region SNPs are dropped
6. No neighbour-gene coloc -- no cis-pQTL in the instrument set coloc-shares a causal variant with any non-target gene's eQTL or pQTL within +/-500 kb
7. Open Targets L2G concordance -- Open Targets Platform locus-to-gene score for the target gene >= 0.5 at the disease GWAS lead (cross-reference causal-genomics/effector-gene-prioritization)

Operational claim ladder: cis-MR significant alone = exploratory; +coloc PP.H4 >= 0.7 = consistent with shared-causal; +cross-platform = consistent across detection chemistries; +PAV-excluded + neighbour-clear = publication-grade target nomination; +cohort replication + Open Targets L2G >= 0.5 = clinical-pharmacology-grade. Clinical-pharmacology claims require ALL 6 original criteria PLUS L2G concordance.

Phenome-Wide Drug-Target MR

Goal: For a single drug target (single protein), test causal effect across hundreds of outcomes to discover on-target adverse effects.

Approach: Hold cis-pQTL instrument set fixed; loop outcome over OpenGWAS catalogue or FinnGen DF12; multi-test correct over outcomes.

library(TwoSampleMR); library(ieugwasr)

target_pqtl <- read.table('pcsk9_cis_pqtls.tsv', header = TRUE)
exposure_dat <- format_data(target_pqtl, type = 'exposure',
    snp_col = 'SNP', beta_col = 'BETA', se_col = 'SE',
    effect_allele_col = 'A1', other_allele_col = 'A2', eaf_col = 'EAF', pval_col = 'P')

curated_endpoints <- read.table('finngen_DF12_endpoints.tsv', header = TRUE)  # ~3000 curated endpoints from finngen.fi
outcomes <- available_outcomes()
outcomes_filt <- subset(outcomes, id %in% curated_endpoints$id & sample_size >= 50000 & population == 'European')

results <- lapply(outcomes_filt$id, function(out_id) {
    outcome_dat <- extract_outcome_data(snps = exposure_dat$SNP, outcomes = out_id)
    if (nrow(outcome_dat) < 2) return(NULL)
    dat <- harmonise_data(exposure_dat, outcome_dat, action = 2)
    mr(dat, method_list = 'mr_ivw')
})

results_df <- do.call(rbind, Filter(Negate(is.null), results))
n_tests <- nrow(curated_endpoints)
results_df$p_bonf <- pmin(results_df$pval * n_tests, 1)
top_hits <- subset(results_df, pval < 0.05 / n_tests)   # 0.05 / 3000 = 1.7e-5

Document the curated endpoint list (FinnGen DF12, Open Targets curated trait map, or a manuscript-specific phecode hierarchy) in methods. The outcomes_filt$id[1:200] pattern is a debug shortcut, not a defensible pheWAS protocol. The PCSK9 -> T2D signal (Schmidt 2017 Lancet Diabetes Endocrinol 5:97) was discovered exactly via curated-endpoint pheWAS: cis-MR of LDL-lowering instruments revealed on-target T2D risk before clinical trials confirmed it. Drug-target pheWAS is the canonical use case.

Cis-MR Standard Workflow

Goal: Produce a defensible cis-MR estimate with triangulation, given a target gene and an outcome.

Approach: Extract cis-pQTLs in +/-500 kb of the gene -> compute F per instrument from exposure -> clump within window at r2 < 0.1 -> harmonise with outcome -> run TwoSampleMR -> run coloc.abf on the same window -> PAV-annotate via VEP -> report panel.

library(TwoSampleMR); library(coloc); library(ieugwasr)

cis_window_kb <- 500  # +/- 500 kb per Schmidt 2020 standard cis-window
gene_chr <- 1; gene_start <- 55039548; gene_end <- 55064852  # PCSK9 hg38

pqtl <- read.table('ukbppp_pcsk9.tsv', header = TRUE)
pqtl_cis <- subset(pqtl, CHR == gene_chr &
    POS > (gene_start - cis_window_kb * 1000) &
    POS < (gene_end + cis_window_kb * 1000) &
    P < 5e-8)

pqtl_cis$f_stat <- (pqtl_cis$BETA / pqtl_cis$SE)^2  # F from exposure (Burgess 2011)
pqtl_cis <- subset(pqtl_cis, f_stat >= 10)  # Staiger-Stock 1997 weak-IV floor

exposure_dat <- format_data(pqtl_cis, type = 'exposure',
    snp_col = 'SNP', beta_col = 'BETA', se_col = 'SE',
    effect_allele_col = 'A1', other_allele_col = 'A2', eaf_col = 'EAF', pval_col = 'P')

clumped <- ld_clump(
    dplyr::tibble(rsid = exposure_dat$SNP, pval = exposure_dat$pval.exposure),
    clump_r2 = 0.1, clump_kb = cis_window_kb,  # cis-MR clumping per Schmidt 2020
    plink_bin = genetics.binaRies::get_plink_binary(),
    bfile = '1kg_EUR/EUR'
)
exposure_dat <- subset(exposure_dat, SNP %in% clumped$rsid)

outcome_dat <- read_outcome_data('cad_gwas.tsv', snps = exposure_dat$SNP, sep = '\t',
    snp_col = 'SNP', beta_col = 'BETA', se_col = 'SE',
    effect_allele_col = 'A1', other_allele_col = 'A2', eaf_col = 'EAF', pval_col = 'P')

dat <- harmonise_data(exposure_dat, outcome_dat, action = 2)

primary <- mr(dat, method_list = c('mr_wald_ratio', 'mr_ivw',
                                    'mr_egger_regression', 'mr_weighted_median'))

# Triangulation step 1: colocalization on the same window
gwas_window <- read.table('cad_gwas_pcsk9_window.tsv', header = TRUE)
pqtl_window <- read.table('ukbppp_pcsk9_full_window.tsv', header = TRUE)

coloc_res <- coloc.abf(
    dataset1 = list(beta = pqtl_window$BETA, varbeta = pqtl_window$SE^2,
                    snp = pqtl_window$SNP, type = 'quant', N = 54219, sdY = 1),
    dataset2 = list(beta = gwas_window$BETA, varbeta = gwas_window$SE^2,
                    snp = gwas_window$SNP, type = 'cc', N = 122733, s = 0.34),
    p1 = 1e-4, p2 = 1e-4, p12 = 5e-6  # conservative p12 for drug-target
)

cat('PP.H4 =', coloc_res$summary['PP.H4.abf'], '\n')

Cis-IVW with Correlated Instruments

Goal: When several cis-pQTLs in the window are correlated (r2 between 0.1 and 0.7), use the generalized IVW that takes the LD matrix as an explicit parameter.

Approach: Compute or load the LD matrix in the cis-window from an ancestry-matched plink reference; build the input with MendelianRandomization::mr_input(..., correlation = ld_matrix) then call mr_ivw(mr_obj, model = 'default', correl = TRUE).

library(MendelianRandomization); library(ieugwasr)

ld <- ld_matrix(exposure_dat$SNP, bfile = '1kg_EUR/EUR', plink_bin = genetics.binaRies::get_plink_binary())

mr_obj <- mr_input(bx = dat$beta.exposure, bxse = dat$se.exposure,
                   by = dat$beta.outcome, byse = dat$se.outcome,
                   corr = ld)

result_correl <- mr_ivw(mr_obj, model = 'default', correl = TRUE)

Numerical caveat: when any pair of cis-pQTLs has r2 ~ 1 (e.g. perfect proxies), the LD matrix is rank-deficient and the SE explodes. Pre-prune at r2 < 0.95.

API caveat for MendelianRandomization::mr_ivw: Correlation between cis-pQTLs is supplied at MRInput construction via the correlation-matrix argument, whose formal name is correlation=; corr= also works only as an abbreviation resolved by R partial-matching, so prefer the explicit correlation=: mr_input(bx, bxse, by, byse, correlation = ld_matrix). mr_ivw() then reads the correlation slot directly; passing correl = TRUE as an argument is redundant when MRInput already has a non-NA correlation matrix. A common bug is supplying both, which produces inconsistent behaviour across versions; prefer the MRInput-slot approach and treat correl = TRUE as a legacy flag.