Human genetics · Target discovery · Nextflow
Proteome-wide cis-MR and colocalisation scan for Parkinson’s disease
A Nextflow pipeline that tests every measured plasma protein as a potential cause of Parkinson’s disease. For each protein it uses variants near the protein’s own gene as instruments, estimates the causal effect with Mendelian randomisation, tests with colocalisation whether protein and disease share a causal variant, then ranks everything into evidence tiers. Run on both large plasma pQTL resources, 2,940 Olink assays from the UK Biobank Pharma Proteomics Project and 3,211 SomaScan measurements from Sun et al. 2018, it picks out known Parkinson’s risk loci without being told where to look.
Scope. Results come from two pQTL resources: UKB-PPP Olink (European discovery set, about 34,000 people, via Synapse) and SomaScan (Sun et al. 2018, 3,301 people, via the eQTL Catalogue). The scan prioritises loci and proteins for follow-up; it does not prove causality.

From one protein to thousands
My dissertation assessed a single protein, GPNMB. Human-genetics teams in drug discovery run the same logic across thousands of proteins at once. This project turns that analysis into a scalable target-prioritisation workflow: which plasma proteins have genetic evidence of a causal role in Parkinson’s disease, and how strong is it?
What runs for every protein
- cis window. Gene ± 500 kb on GRCh38, with INFO and MAF QC.
- Harmonisation. Alleles aligned on GRCh37 position between pQTL and GWAS, handling strand flips and palindromic SNPs.
- Instruments. p < 5×10−8, F ≥ 10, LD-clumped at r² < 0.001 on 1000 Genomes EUR.
- MR. Wald ratio or IVW (fixed and random effects), MR-Egger and weighted median.
- Colocalisation. Approximate Bayes factor coloc (p1 = p2 = 10−4, p12 = 10−5) with prior sensitivity.
- Across proteins. Benjamini–Hochberg FDR → evidence tiers → HTML report.
UKB-PPP summary statistics run to 1.6 TB, one archive per protein. The fetch script never downloads a whole archive: it reads the archive’s headers with HTTP byte-range requests, pulls only the chromosome carrying the gene and keeps the cis window, so the full proteome comes down as about 90 GB. For SomaScan, a preparation script splits the eQTL Catalogue file into one slice per protein and maps variants to GRCh37 through their rsIDs.
| Tier | Rule | Meaning |
|---|---|---|
| 1 | MR q < 0.05 and PP.H4 ≥ 0.8 | Protein and PD most likely share one causal variant |
| 2 | MR q < 0.05 and PP.H4 0.5–0.8 | Supportive, not conclusive |
| 3 | MR q < 0.05 and PP.H3 ≥ 0.5 | Different causal variants: MR probably reflects LD |
| 4 | MR q < 0.05, coloc inconclusive | Usually an underpowered outcome signal |
What the plasma proteome says about Parkinson’s disease
UKB-PPP Olink (2,940 assays, about 34,000 people)
2,016 proteins had a genome-wide significant cis-pQTL to test, 848 had none, and 90 sit on chromosome X, which the PD GWAS does not cover. Fifteen assays measure a protein encoded by two or three genes, and each gene’s cis window is tested separately. Eight proteins passed FDR < 0.05.
| Tier | Protein | Locus | OR per SD (95% CI) | Coloc |
|---|---|---|---|---|
| 1 | CTF1 | 16p11.2 | 8.07 (4.14–15.75) | PP.H4 0.95 |
| 1 | STX4 | 16p11.2 | 0.09 (0.04–0.21) | PP.H4 0.89 |
| 2 | LRRC37A2 | 17q21.31 | 0.83 (0.77–0.90) | PP.H4 0.58 |
| 3 | SNCA | SNCA | 0.10 (0.05–0.21) | PP.H3 1.00 |
| 3 | PRSS53, PRSS8 | 16p11.2 | 0.90, 0.46 | PP.H3 0.99, 1.00 |
| 3 | EFNA1 | GBA1 | 1.23 (1.11–1.36) | PP.H3 1.00 |
| 3 | HLA-DRA | HLA | 0.80 (0.72–0.90) | PP.H3 0.98 |
All eight sit at known Parkinson’s risk loci, and colocalisation is what separates them. For five, the protein and the disease signals come from different variants, so the MR association rides on linkage disequilibrium rather than on the protein. CTF1 and STX4 do colocalise, but they point to the 16p11.2 locus rather than to either gene. Each rests on one modest pQTL, the implied effects are implausibly large and opposite in direction within the same gene-dense region, and PP.H4 falls to 0.65 and 0.45 under a stricter prior. LRRC37A2 lies in the 17q21.31 inversion, where long-range LD makes gene-level attribution unreliable. The scan finds the disease loci without being told where to look, but no single protein comes out as a confident causal candidate. That is a realistic outcome for a proteome-wide scan, and the tiers and flags are there to make it visible.
Olink against SomaScan
| Protein | SomaScan | UKB-PPP Olink |
|---|---|---|
| FCGR2A | Tier 1 (OR 1.06, PP.H4 0.96) | No MR evidence (p = 0.07, PP.H4 0.11) |
| GPNMB | Tier 2–3 (OR 1.50–1.56) | No MR evidence (p = 0.53, PP.H3 1.00) |
| CTSB | Tier 2 (OR 0.82, PP.H4 0.73) | Same direction, below FDR (OR 0.87, p = 0.03, PP.H4 0.73) |
| PTGR1 | Tier 2 (OR 0.84, PP.H4 0.59) | Same direction, below FDR (OR 0.85, p = 8×10−4, PP.H4 0.49) |
| SNCA | Tier 3 | Tier 3 |
Most SomaScan hits do not carry over to Olink. For GPNMB this matches my dissertation: the Olink pQTL and the disease signal are distinct, while SomaScan gives a positive MR estimate. Disagreement between the platforms is common, because aptamers and antibodies can detect different forms of a protein and protein-altering variants can create binding artefacts on either. CTSB and PTGR1 agree in direction and size on both platforms, which makes them the most consistent candidates from the two scans.
SomaScan (Sun et al. 2018, 3,211 aptamers)

Of 3,211 SomaScan aptamers, 767 (713 genes) had a genome-wide significant cis-pQTL to test; the rest had no instrument at p < 5×10−8 in 3,301 people. Nine passed FDR < 0.05.
| Tier | Protein | OR per SD (95% CI) | Coloc PP.H4 |
|---|---|---|---|
| 1 | FCGR2A | 1.06 (1.03–1.09) | 0.96 |
| 2 | GPNMB (two aptamers) | 1.50–1.53 | 0.54–0.56 |
| 2 | CTSB | 0.82 (0.74–0.91) | 0.73 |
| 2 | PTGR1 | 0.84 (0.76–0.92) | 0.59 |
| 3 | SNCA (two aptamers) | 0.40–0.42 | 0.00 |
| 3 | GPNMB (third aptamer) | 1.56 (1.31–1.85) | 0.38 |
| 3 | BAG3 | 1.62 (1.30–2.03) | 0.01 |
Without being told where to look, the scan lands on genes at known Parkinson’s risk loci: FCGR2A, CTSB, GPNMB, SNCA and BAG3. Two GPNMB aptamers reproduce my dissertation’s single-instrument estimate in a sample 3.3 times larger. SNCA and BAG3 show why the colocalisation tier matters. Their MR signals are among the strongest in the scan, but the protein and disease associations come from different variants (PP.H3 ≈ 1), so the MR result most likely reflects linkage disequilibrium rather than the protein.
Checked against R, then against known answers
The statistics are implemented in Python and checked against the reference R packages. The test suite reproduces my dissertation’s R results to at least eight significant figures.
| Check | R (dissertation) | This pipeline |
|---|---|---|
| coloc.abf, SomaScan GPNMB vs PD (150 SNPs) | PP.H4 = 0.943955 | PP.H4 = 0.943955 |
| coloc.abf, UKB-PPP Olink GPNMB vs PD (4,018 SNPs) | PP.H3 = 0.999804 | PP.H3 = 0.999804 |
| IVW fixed + random effects, Cochran’s Q (strict, standard and liberal clumping) | e.g. OR 1.079, RE p = 0.473 | identical |
| Wald ratio, lead variant rs75801644 | β = −0.06717, SE 0.04832 | identical |
The test profile combines the real GPNMB locus with four proteins simulated from a genotype panel, so LD, clumping and colocalisation behave realistically. Each lands where it should.
| Protein | Scenario | Result |
|---|---|---|
| SYN_SHARED | One causal pQTL that also drives PD | Tier 1 (OR 1.32, PP.H4 0.99) |
| SYN_MULTI | Three causal pQTLs, all mediated by the protein | Tier 1, 3 instruments (OR 1.42, PP.H4 1.00) |
| SYN_DISTINCT | PD signal from a nearby variant in LD | Tier 3 (MR p = 0.014, PP.H3 1.00) |
| SYN_NULL | pQTL with no effect on PD | No MR evidence (PP.H1 0.99) |
| GPNMB (real) | UKB-PPP Olink vs ieu-b-7 | No MR evidence at the lead variant; flagged for multi-signal follow-up |
Why GPNMB is flagged, not discarded
The strongest Olink pQTL for GPNMB is a low-frequency variant (rs75801644) with no PD association, so lead-variant MR is null and single-variant colocalisation strongly prefers distinct signals (PP.H3 = 0.9998). But the pQTL has several independent signals and PD’s peak lies over them. SuSiE-based colocalisation of the same Olink data finds a secondary pQTL signal overlapping the PD signal (PP.H4 = 0.93–0.99 under the default prior), though its lead variant depends on the LD reference panel and the evidence weakens under a strict prior. A naive filter would throw GPNMB away; the pipeline flags loci like this for multi-signal follow-up instead.

What the scan can’t tell you
- Single-variant colocalisation assumes one causal variant per trait. Multi-signal loci need SuSiE-based colocalisation; the eQTL Catalogue’s in-sample SuSiE fine-mapping of Sun et al. 2018 is the planned route.
- In gene-dense regions such as 16p11.2, 17q21.31 and HLA, several proteins can colocalise with the same disease signal; colocalisation then points to the locus, not to a gene.
- Many estimates rest on a single instrument (with more than one: 1,364 of 2,016 Olink proteins, 255 of 767 SomaScan), and SomaScan aptamers can bind more than their target protein.
- cis-MR assumes instruments act on PD only through the protein. Protein-altering variants can change Olink antibody binding without changing protein level.
- Results are genetic prioritisation evidence, not proof of causality or druggability.