MSc dissertation · Statistical genetics

Prioritising GPNMB as a Parkinson’s disease-relevant protein

Integrated two-sample Mendelian randomisation, LD-aware sensitivity analysis, Bayesian colocalisation and transcriptomic evidence to evaluate GPNMB as a Parkinson’s disease-relevant target. The SomaScan cis-instrument gave a strong association and colocalisation supported a shared signal, but heterogeneous Olink instruments meant the conclusion stops at prioritisation rather than causality.

View the code ↗

Context: MSc Bioinformatics dissertation, Atlantic Technological University, 2025–2026
Data: SomaScan and UKB-PPP Olink pQTLs; Parkinson’s disease GWAS (ieu-b-7); GTEx and Open Targets QTLs
Methods: Wald ratio, IVW (fixed and random effects), Cochran’s Q, LD clumping, coloc.abf, SuSiE
Stack: R, TwoSampleMR, coloc, susieR, OpenGWAS

The GPNMB locus on chromosome 7 (GRCh37): SomaScan protein QTL (165 variants) above, Parkinson’s disease GWAS (ieu-b-7) below.

1.32Odds ratio for PD from the SomaScan cis-instrument (95% CI 1.19–1.46)

0.94Posterior probability of a shared causal variant (PP.H4, 150 SNPs)

0.47Random-effects p for the strict Olink instrument set: the check that tempered the result

The question

Is GPNMB a plausible Parkinson’s disease protein?

GPNMB sits at a known Parkinson’s disease risk locus on 7p15.3. The dissertation asked two linked questions: are genetically predicted GPNMB protein levels associated with Parkinson’s disease risk, and are the protein and disease associations consistent with a single shared causal variant in the region, rather than two nearby signals in linkage disequilibrium?

Approach

Two proteomic platforms, three kinds of evidence

I analysed GPNMB protein QTLs from two platforms against Parkinson’s disease GWAS summary statistics, then tested how well the result held up under sensitivity analyses.

SomaScan MR. Harmonised the SomaScan GPNMB exposure with the PD outcome and estimated a single-instrument Wald ratio at rs5850.
UKB-PPP Olink MR. 4,338 harmonised regional records → 790 QC-filtered genome-wide significant records → European-reference LD clumping at strict, standard, sensitivity and liberal thresholds (5, 12, 12 and 25 instruments). Standard and sensitivity use the same r² threshold (0.01) and differ only in clumping window (10 Mb vs 1 Mb); at this cis locus both select the same 12 variants.
Heterogeneity. Fixed-effect IVW with Cochran’s Q; when Q showed strong heterogeneity, multiplicative random-effects IVW recalculated from the saved instrument sets.
Colocalisation. coloc.abf across 150 overlapping SNPs, a sweep of the shared-variant prior (p₁₂), and SuSiE to allow for more than one signal at the locus (SomaScan, 146 SNPs; UKB-PPP Olink, 335 SNPs with a 1000 Genomes EUR LD matrix).
Supporting layers. Cross-tissue eQTL evidence (GTEx, Open Targets) and brain RNA expression.

Results

A strong signal that weakens under scrutiny

The SomaScan instrument gave OR 1.32 (95% CI 1.19–1.46; P = 2.6×10⁻⁷). The LD-clumped Olink instrument sets pointed the same way under fixed effects (strict OR ≈ 1.08, p = 0.057; standard OR ≈ 1.09, p = 0.017), but Cochran’s Q was large (p ≈ 10⁻⁵) and the random-effects estimates were not significant (strict p = 0.473; standard p = 0.221).

Forest plot of GPNMB Mendelian randomisation estimates for Parkinson’s disease across SomaScan and Olink analyses — MR estimates across SomaScan and LD-clumped Olink instrument sets; heterogeneity-aware random-effects estimates are shown next to the fixed-effect results.

Under the main prior, colocalisation supported a shared causal variant (PP.H4 = 0.944). The prior sweep shows how much that depends on assumptions: at a stricter shared-variant prior (p₁₂ = 5×10⁻⁷ or lower) the evidence favours distinct variants instead. SuSiE did not resolve the multi-signal structure in the SomaScan data. In the UKB-PPP Olink data it separated the signals: the lead pQTL (rs75801644) does not share a causal variant with PD (PP.H3 = 0.999), but a secondary pQTL signal led by rs199347 does (PP.H4 = 0.99 under the default prior, 0.51 at p₁₂ = 10⁻⁷).

Line plot of colocalisation posterior probabilities PP.H4 and PP.H3 against the shared-variant prior p12 — Colocalisation prior sensitivity: PP.H4 is 0.944 at the default p₁₂ = 10⁻⁵ and falls to 0.144 at 10⁻⁷.

Interpretation

What I would and wouldn’t claim

The final interpretation prioritises GPNMB for further study rather than treating the evidence as causal closure.

What it supports

A strong single-instrument association between genetically predicted GPNMB and PD risk.
A shared regional signal under standard colocalisation priors.
GPNMB as a reasonable candidate for follow-up work.

What it doesn’t show

Causality: one SomaScan instrument cannot test heterogeneity or horizontal pleiotropy.
Robustness across platforms: Olink random-effects estimates were not significant.
A single clean causal variant: colocalisation depends on the prior and on multi-signal locus structure.

Reproducibility

Packaged so it can be audited

Analysis scripts, saved final instrument sets, derived results and figure-generation code in one supplementary package.
The random-effects IVW and Cochran’s Q results can be rerun from the saved clumped instruments without new OpenGWAS API calls.
R session information, source-file provenance and figure notes are included; Figures 2–5 regenerate from saved outputs with one script.