Case studies
Projects
Eleven projects, each written up as a case study: the question, the approach, the results and their limits. Code for all of them is on GitHub.

MSc dissertation · Statistical genetics
GPNMB and Parkinson’s disease: Mendelian randomisation and colocalisation
Two-sample MR, LD-aware sensitivity analysis and Bayesian colocalisation to test whether genetically predicted GPNMB protein levels are linked to Parkinson’s disease risk. The evidence supports prioritisation, not causality.
Read case study →
Human genetics · Target discovery · Nextflow
Proteome-wide cis-MR and colocalisation scan for Parkinson’s disease
Scaled the dissertation from one protein to a Nextflow pipeline that tests every plasma protein with cis-pQTL MR and colocalisation, FDR control and evidence tiers. Run on 2,940 UKB-PPP Olink and 3,211 SomaScan proteins, it finds known PD risk loci without being told where to look and separates colocalising hits from LD-driven ones.
Read case study →
Single-nucleus transcriptomics · Parkinson’s disease
GPNMB in Parkinson’s disease: snRNA-seq re-analysis
Re-analysed 83,484 substantia nigra nuclei from 29 donors to localise GPNMB and test PD–control differences with donor-level pseudobulk models. GPNMB was highest in microglia and about two-fold higher in PD microglia (P = 0.009), a nominal effect that did not survive transcriptome-wide FDR.
Read case study →
Biostatistics · Survival analysis
GPNMB and glioma survival: separating a prognostic signal from confounding
Pre-specified Cox analysis of 624 TCGA diffuse gliomas. GPNMB’s hazard ratio fell from 2.00 per SD unadjusted to 1.15 (95% CI 0.96–1.38) after adjusting for grade and IDH/1p19q status, with an optimism-corrected C-index gain of 0.0009.
Read case study →
Biostatistics · Longitudinal and survival analysis
Parkinson’s disease progression in PPMI: trajectories and clinical milestones
Mixed-effects and Cox models of 1,789 PPMI participants. Medication-adjusted motor progression slowed from 2.0 to 1.2 MDS-UPDRS III points a year, MoCA held for five years before declining, and age was the strongest predictor of milestones.
Read case study →
Microbial genomics · WGS
M. tuberculosis WGS variant analysis (Galaxy)
End-to-end Galaxy workflow from SRA reads to annotated resistance variants, run on 6 clinical isolates. It called rpoB and katG mutations, confirmed in IGV, and classified 4 isolates as genotypically MDR-TB.
Read case study →
Clinical genomics · NGS · Pipelines
Germline variant calling benchmarked against GIAB HG002
Nextflow pipeline from reads to germline SNVs and indels with GATK and DeepVariant, scored against the Genome in a Bottle truth set. On chr20, DeepVariant reaches SNV F1 0.997 and indel F1 0.995; stratified results show where each caller loses accuracy.
Read case study →
Transcriptomics · Pipelines
RNA-seq pipeline (Nextflow + Docker)
Samplesheet-driven Nextflow DSL2 pipeline from raw FASTQ to differential expression in one command, with every tool in a versioned container. Validated on real human data: 8 of 8 known dexamethasone-response genes recovered.
Read case study →
Regulatory genomics · Machine learning
RegulonML: regulatory variant effects in unseen promoters and enhancers
Predicted the effects of 41,724 single-base variants in 21 regulatory elements from DNA sequence alone, scoring binding-site gain and loss for 1,019 JASPAR motifs and testing only on loci the model had never seen.
Read case study →
Health data engineering · GenAI evaluation
From FHIR to OMOP to RAG: answering clinical questions from notes, and measuring it honestly
Synthetic FHIR records built into a tested OMOP CDM with dbt, notes with exact ground truth, and a retrieval-augmented LLM scored once on 1,200 held-out questions. Accuracy 0.925, with errors traced to retrieval, against a SQL baseline that is exact on structured data.
Read case study →
Data modelling · Ontologies · AI curation
A Parkinson’s biomarker data model, and an AI mapping agent held to a frozen gold standard
A LinkML model with ontology-bound value sets, 16 validation rules and FAIR provenance; a blind AI agent maps a 68-variable study dictionary, a curator reviews every proposal in SSSOM. Precision 0.906, recall 0.967, 67 of 68 proposals accepted.
Read case study →
Health economics · Cost-effectiveness
Pembrolizumab in PD-L1 ≥ 50% NSCLC: a reproducible partitioned survival model
A NICE TA531-style model rebuilt in R from KEYNOTE-024 five-year data: reconstructed patient data checked against published results, DSU-compliant extrapolation, PSA and a price-threshold analysis. At list price the ICER is £52,670 per QALY.
Read case study →Get in touch
I’m looking for bioinformatics roles in statistical genetics, transcriptomics and NGS analysis, especially where clinical or cell and gene therapy experience helps. Based in London, open to hybrid and remote.