ClawBio WorkshopAgentic GWAS
From summary statistics to causal variants
Dr Manuel Corpas · University of Westminster · 2026
Where we are
In the previous session you annotated a single genome and found clinically relevant variants using VEP, ClinVar, gnomAD, and CPIC.
Now we scale up. GWAS looks across thousands of genomes to find variants associated with disease at the population level.
This session: ~10 min slides, then ~20 min hands-on in Google Colab. By the end you will have queried nine GWAS databases, computed a polygenic risk score, and fine-mapped a locus to identify causal variants.
Learning objectives
- Explain what a GWAS is and what summary statistics contain
- Query nine federated databases for variant associations in seconds
- Calculate polygenic risk scores from published PGS Catalog scores
- Apply SuSiE fine-mapping to identify credible sets of causal variants
- Recognise cross-ancestry gaps in GWAS representation
- Run all three analyses in Google Colab with zero infrastructure
Part 1
What is a GWAS?
GWAS in one slide
Test every variant in the genome for association with a trait across thousands of people.
The last number is the problem. Most GWAS findings may not transfer to non-European populations.
What are summary statistics?
Published GWAS release summary statistics: per-variant results without individual genotypes.
| Field | Meaning | Example |
|---|---|---|
rsid | Variant identifier | rs7903146 |
beta | Effect size (log-odds or per-allele) | 0.31 |
se | Standard error of beta | 0.02 |
p | P-value for association | 5.2 x 10-38 |
MAF | Minor allele frequency | 0.28 |
Key insight: Summary statistics are public, free, and enough to run PRS, meta-analysis, and fine-mapping. No HPC. No data access agreements. No individual-level data.
Part 2
Three skills, one pipeline
The GWAS pipeline
GWAS Lookup
Query 9 databases in parallel for a single variant: GWAS Catalog, Open Targets, UKB, FinnGen, Biobank Japan, GTEx, eQTL Catalogue.
PRS Calculator
Compute polygenic risk scores from 23andMe or AncestryDNA data using 3,000+ published scores from the PGS Catalog.
Fine-Mapping
Apply SuSiE to identify credible sets of likely causal variants from summary statistics. No individual-level data needed.
Skill 1: GWAS Lookup
Give it an rsID. It queries nine databases in parallel and returns a unified report.
| Database | What it returns | Ancestry |
|---|---|---|
| GWAS Catalog | Published trait associations | Mixed |
| Open Targets | Credible sets, L2G scores | Mixed |
| UKB-TOPMed PheWeb | PheWAS across 4,500 phenotypes | Multi-ancestry |
| FinnGen r12 | Finnish disease endpoints | Finnish |
| Biobank Japan | East Asian PheWAS | Japanese |
| GTEx v8 | eQTL tissue expression | Mostly European |
| EBI eQTL Catalogue | Multi-tissue eQTL associations | Mixed |
One command: python gwas_lookup.py --rsid rs7903146
Cross-ancestry: why it matters
The problem
- 86% of GWAS participants are of European ancestry
- Effect sizes and allele frequencies differ between populations
- PRS trained on Europeans perform poorly in African and South Asian populations
- Risk of widening health disparities
What ClawBio does
- Queries UKB, FinnGen, and Biobank Japan in one call
- Flags allele frequency differences across populations
- MVP (Million Veteran Program) is the most diverse GWAS cohort: 33% non-European
- Uganda Genome Resource: 6,407 samples, African ancestry GWAS
Skill 2: Polygenic Risk Scores
A PRS sums the effects of many variants into a single risk estimate.
Formula: PRS = ∑ (dosagei × effect_weighti) across all matched variants
ClawBio ships with 6 curated scores for instant demos:
| Trait | PGS ID | Variants |
|---|---|---|
| Type 2 diabetes | PGS000013 | 8 |
| Coronary artery disease | PGS000004 | 46 |
| Breast cancer | PGS000001 | 77 |
| Prostate cancer | PGS000057 | 147 |
| Atrial fibrillation | PGS000011 | 12 |
| BMI | PGS000039 | 97 |
Risk categories: Low (<20th) · Average (20-80th) · Elevated (80-95th) · High (>95th)
Skill 3: SuSiE Fine-Mapping
GWAS finds associated regions. Fine-mapping finds the causal variants within them.
Without fine-mapping
- A GWAS hit spans 10-200 correlated SNPs in LD
- Which one is causal? All look equally significant
- Manual triage: slow, subjective, error-prone
With SuSiE
- Credible sets: minimal set of SNPs capturing 95% of causal probability
- PIPs: posterior inclusion probability per variant
- Handles multiple causal signals per locus
- Works from summary statistics alone
Democratising GWAS
The infrastructure barrier is gone.
No HPC
Summary statistics are small. Everything runs in Google Colab on a free tier.
No data access agreements
Summary statistics are publicly released. No application, no waiting, no ethics board.
No bioinformatics team
ClawBio wraps the full pipeline. One command per analysis.
No cost
Google Colab is free. ClawBio is MIT-licensed. PGS Catalog is open. All databases are public APIs.
Implication for the Global South: A researcher in Lima, Kampala, or Dhaka can run the same GWAS analyses as one at the Broad Institute. Today.
Part 3
Hands-on Practical
Open Google Colab now
What you will do
| Step | Task | Time |
|---|---|---|
| 1 | Setup: install ClawBio in Colab (same as variant interpretation) | 2 min |
| 2 | GWAS Lookup: query rs7903146 (type 2 diabetes) across 9 databases | 5 min |
| 3 | Compare allele frequencies across UKB, FinnGen, and Biobank Japan | 3 min |
| 4 | PRS: compute polygenic risk scores for 6 traits using the Corpasome | 5 min |
| 5 | Fine-mapping: run SuSiE on a demo locus with 2 causal signals | 5 min |
Requirements: Same as before. A Google account and a web browser. Nothing to install.
Variants we will explore
| rsID | Gene / Locus | Trait | Why it matters |
|---|---|---|---|
| rs7903146 | TCF7L2 | Type 2 diabetes | Strongest common T2D signal. OR 1.4 per allele. |
| rs429358 | APOE | Alzheimer's | Already found in variant interpretation. Now see the GWAS context. |
| rs3798220 | LPA | Heart disease | Lipoprotein(a). Risk factor for coronary events. |
| rs1801282 | PPARG | Type 2 diabetes | Drug target for thiazolidinediones. |
Take-home messages
- GWAS summary statistics are free and public. You do not need individual-level data to do meaningful research.
- Three ClawBio skills cover the full GWAS workflow: lookup, PRS, and fine-mapping.
- Cross-ancestry analysis is not optional. Most GWAS are European-biased. Query multiple biobanks to check transferability.
- Fine-mapping narrows GWAS hits to causal variants. SuSiE credible sets are the state of the art.
- Infrastructure is no longer a barrier. Google Colab + ClawBio = publication-quality GWAS analysis for free.
Resources
| Resource | Link |
|---|---|
| ClawBio GitHub | github.com/ClawBio/ClawBio |
| Variant Interpretation Workshop | Previous workshop |
| GWAS Catalog | ebi.ac.uk/gwas |
| PGS Catalog | pgscatalog.org |
| Open Targets | genetics.opentargets.org |
| Corpasome (Zenodo) | doi:10.5281/zenodo.19297389 |
| Ensembl VEP | ensembl.org/vep |
| gnomAD | gnomad.broadinstitute.org |