ClawBio WorkshopAgentic Genomics
Leverage the power of AI in genomics
Dr Manuel Corpas · University of Westminster · 2026
Part 1
What is ClawBio?
ClawBio in one sentence
An open-source toolkit of AI agent skills for genomic analysis.
You give it genetic data. It runs the right analysis. You get a structured report.
The problem ClawBio solves
Without ClawBio
- Download VEP, configure databases, debug dependencies
- Submit variants to web APIs one by one
- Parse raw JSON output by hand
- Cross-reference ClinVar, gnomAD, CPIC separately
- AI chatbots invent gene-drug associations
Hours to days per genome
With ClawBio
- One command runs the full pipeline
- Built-in demo data for every skill
- Structured reports in markdown and JSON
- Grounded in CPIC, ClinVar, gnomAD
- All processing is local: your data never leaves
Minutes per analysis
How a skill works
SKILL.md
The contract. Defines inputs, outputs, domain decisions, safety rules. The AI reads this; it never overrides it.
Python script
The implementation. Runs the pipeline, writes structured output. All processing is local.
Demo data
Synthetic test data with every skill. Run a demo instantly without real patient data.
Report
Markdown + JSON. Human-readable and machine-parseable. Includes disclaimers and provenance.
Part 2
Variant Interpretation
From a single genome to clinical findings
The annotation pipeline
| Database | What it tells you |
|---|---|
| Ensembl VEP | What gene is affected? Missense, synonymous, frameshift? |
| ClinVar | Has this variant been seen in patients? Pathogenic or benign? |
| gnomAD | How common is it across 76,000+ genomes? |
| CPIC | Does it affect drug metabolism? Which drugs? |
ACMG classification
The American College of Medical Genetics uses five categories:
| Category | What it means | Clinical action |
|---|---|---|
| Pathogenic | Causes or strongly contributes to disease | Report. Genetic counselling. |
| Likely pathogenic | Strong evidence, not conclusive | Report with caveat. |
| VUS | Not enough evidence | Not acted upon. May be reclassified. |
| Likely benign | Probably harmless | Usually not reported. |
| Benign | Definitively harmless | Not reported. |
More than half of all known variants are VUS. "We don't know yet" is often the most honest answer.
The Corpasome: a real open genome
In 2013, Manuel Corpas published his genome under CC0 public domain. Real findings you will discover:
| Variant | Gene | Clinical meaning |
|---|---|---|
| rs6025 | F5 (Factor V Leiden) | 3-8x increased blood clotting risk |
| rs1800562 | HFE (C282Y) | Carrier for haemochromatosis |
| rs113993960 | CFTR (deltaF508) | Carrier for cystic fibrosis |
| rs9923231 | VKORC1 | Warfarin: AVOID standard dose |
| rs429358 | APOE | Elevated Alzheimer's risk (e3/e4) |
Pharmacogenomics: drugs meet DNA
| Gene | Drugs affected | Risk |
|---|---|---|
| CYP2D6 | Codeine, tamoxifen, antidepressants | Poor metabolisers get zero pain relief from codeine |
| CYP2C19 | Clopidogrel (blood thinner) | Drug not activated; stroke risk unchanged |
| CYP2C9 + VKORC1 | Warfarin | Wrong dose causes bleeding or clots |
| DPYD | 5-fluorouracil (chemo) | Standard dose can be fatal |
The Corpasome needs a significantly lower warfarin dose than standard. Without genetic testing, a doctor might prescribe a dose that causes a haemorrhage.
Part 3
Agentic GWAS
From one genome to population-level analysis
GWAS in one slide
Test every variant for association with a trait across thousands of people.
Summary statistics are public, free, and sufficient for PRS, meta-analysis, and fine-mapping. No HPC needed.
Three GWAS skills, one pipeline
GWAS Lookup
Query 9 databases in parallel: GWAS Catalog, Open Targets, UKB, FinnGen, Biobank Japan, GTEx, eQTL Catalogue.
PRS Calculator
Compute polygenic risk scores from 23andMe/AncestryDNA data using 3,000+ PGS Catalog scores.
Fine-Mapping
SuSiE credible sets to identify causal variants from summary statistics. No individual data needed.
Cross-ancestry: the equity gap in GWAS
The problem
- 86% of GWAS participants are European
- PRS lose up to 80% accuracy in non-European 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 is the most diverse GWAS cohort: 33% non-European
Part 4
30x Whole-Genome Sequencing
The same genome at full resolution
SNP array vs whole-genome sequencing
| SNP Array (~600K) | 30x WGS | |
|---|---|---|
| SNPs | ~600,000 | 3,716,648 |
| Indels | 0 | 912,009 |
| Structural variants | 0 | 8,925 |
| Copy number variants | 0 | 1,387 |
| Coverage | Pre-selected positions | Every base |
SNP arrays answer pre-defined questions. WGS lets you ask questions you did not know to ask.
Structural variants: the hidden layer
| Type | What happens | Count |
|---|---|---|
| DEL | DNA segment missing | 5,854 |
| BND | Translocation or complex rearrangement | 1,413 |
| DUP | Segment copied (alters gene dosage) | 778 |
| INV | Segment flipped in orientation | 673 |
| INS | New DNA inserted (mobile elements) | 207 |
Structural variants cause an estimated 20% of rare genetic disease but are invisible to SNP arrays.
QC metrics: is this genome trustworthy?
| Metric | Expected | This genome | What it tells you |
|---|---|---|---|
| Ti/Tv ratio | ~2.0 | 2.03 | Below 2.0 suggests errors or contamination |
| Het/Hom ratio | 1.5-1.7 | 1.63 | Outside range may indicate contamination |
| Total SNPs | 3.5-4.5M | 3,716,648 | Too few = low coverage; too many = errors |
Always check QC before interpreting. These three numbers tell you immediately whether the data is trustworthy.
Part 5
Equity in Genomics
Whose genomes are studied? Whose are left out?
The numbers
- 86% of GWAS participants are of European ancestry
- Databases have 30x more BRCA variant data for Europeans than other populations
- Polygenic risk scores lose up to 80% accuracy in non-European populations
- A variant classified as "benign" may simply be unstudied in the relevant population
AI amplifies this. Models trained on biased data produce biased results. Every analysis should acknowledge which populations the reference data represents.
Democratising genomics
The infrastructure barrier is gone.
No HPC
Everything runs in Google Colab on a free tier.
No data agreements
Summary statistics are publicly released. No application, no waiting.
No bioinformatics team
ClawBio wraps the full pipeline. One command per analysis.
No cost
Google Colab is free. ClawBio is MIT-licensed. All databases are public.
A researcher in Lima, Kampala, or Dhaka can run the same analyses as one at the Broad Institute. Today.
Part 6
Hands-on Practical
Open Google Colab now
Workshop schedule
| Session | What you'll do | Time |
|---|---|---|
| Setup | Install ClawBio in Colab (one click) | 2 min |
| Session 1: Variant Interpretation | Annotate 21 clinical variants from a real genome. VEP, ClinVar, gnomAD, CPIC. | 25 min |
| Session 2: GWAS | Query rs7903146 across 9 databases. Compute PRS for 6 traits. Fine-map a locus with SuSiE. | 25 min |
| Session 3: 30x WGS | Explore structural variants, QC metrics, and WGS vs SNP chip comparison. | 25 min |
| Q&A | Discussion and questions | 15 min |
Requirements: A Google account and a web browser. Nothing to install. No API keys. No payment.
Take-home messages
- AI makes the mechanical steps faster, but domain expertise in molecular biology, genetics, and clinical context remains essential.
- Pharmacogenomics is saving lives today. Drug-gene interactions like warfarin/CYP2C9/VKORC1 are implemented in hospitals now.
- "We don't know yet" is a valid answer. Over half of all variants are VUS. Honest uncertainty is a strength, not a weakness.
- Cross-ancestry analysis is not optional. Always check whether findings transfer across populations.
- Infrastructure is no longer a barrier. Google Colab + ClawBio = publication-quality analysis, for free, anywhere.
Resources
| Resource | Link |
|---|---|
| ClawBio GitHub | github.com/ClawBio/ClawBio |
| Documentation | docs.clawbio.ai |
| Variant Interpretation tutorial | docs.clawbio.ai/tutorials/variant-interpretation-workshop |
| GWAS tutorial | docs.clawbio.ai/tutorials/gwas-workshop |
| 30x WGS tutorial | docs.clawbio.ai/tutorials/30x-wgs-workshop |
| Corpasome (Zenodo) | doi:10.5281/zenodo.19297389 |
github.com/ClawBio/ClawBio