🏆 AI UK Agent Gold Prize Winner
ClawBio
The open standard for biological AI skills.
A community-driven library where domain experts across 3 continents
encode, share, and validate agent-discoverable genomics skills.
Defining how biological skills are shared, verified, and trusted.
724
GitHub Stars
46
Skills
3
Continents
1,162
Tests
Manuel Corpas · clawbio.ai · Life & Health Track · Haiju Yingcai 2026
The Paradigm Shift
Genomics has entered the agentic era.
📄
Before: code production
Researchers write code, configure tools, run pipelines, interpret results. The bottleneck is building the analysis. Only bioinformaticians can participate.
🤖
Now: agents execute
AI agents discover skills, plan multi-step analyses, execute them, and adapt when things fail. The researcher describes intent in natural language.
⚠️
New bottleneck: validation
Agents produce results faster than humans can verify them. Silent degradation to plausible-looking errors is the defining risk of this era.
The bottleneck has shifted from production to judgement. The field needs infrastructure for trust.
The Community
Domain experts encode knowledge. Agents discover and run it.
🌐
3 continents contributing
Wet-lab biologists with limited coding experience contribute functional, tested skills using LLMs as the programming intermediary. Domain experts encode knowledge without becoming software engineers.
🔍
Agent-discoverable via SKILL.md
Each skill has a natural-language specification (SKILL.md) that lets any AI agent discover the right skill, understand its requirements, and invoke it without human configuration.
🛠
The semantic layer
Skills encapsulate the domain expert's decisions: parameter rationale, expected failure modes, test data for verification. They complement Galaxy, Nextflow, and nf-core, not replace them.
A qualitatively new model: domain experts encode their knowledge into agent-executable modules without needing to become software engineers. ClawBio is the largest open-source bio skill community.
The Core Problem
Skills can be shared faster than they can be audited.
💥
Silent degradation
A skill given an empty input file silently returned "all normal" results, including recommended dosages for 51 drugs. Discovered by an external auditor, not automated testing.
🧠
Hallucination risk
Agents can fabricate gene-disease associations, cite non-existent references, or generate variant annotations that conflate unrelated loci. Confident but wrong.
📈
Convergent failure
CellAtria (AstraZeneca), AutoBA, Bio-Copilot, and ClawBio all independently discovered the same pattern: plausible-looking errors that pass basic checks.
Agentic genomics lowers the barrier to producing analyses. It does not lower the barrier to evaluating them. Patient data demands a structured response.
The Solution
Tiered validation: from research to clinical grade.
Tier 1
Research-grade
- → Unit tests, adversarial inputs
- → Empty files, malformed data, edge cases
- → Community review encouraged
- → False positives tolerable if caught during interpretation
Tier 2
Benchmarked
- → Validated on public reference datasets (GIAB, scRNA-seq benchmarks)
- → Published metrics with confidence intervals
- → Reproducible across environments
- → Independent benchmarking required
Tier 3 ✓
Clinical-grade 🏆
- → External multi-site validation
- → FDA/EMA regulatory alignment
- → Signed reproducibility bundles
- → BioCompute Object audit trails
- → CLIA/CAP compliance ready
- → Verification badge issued
Platforms must enforce tier boundaries: research-grade skills cannot be used in clinical contexts without explicit override. The badge is the trust signal.
The Business
Open community. Verified trust layer.
Open (free forever)
✓ Skill library: open source, MIT licence
✓ SKILL.md specification: community-governed
✓ Agent discovery and execution
✓ Research-grade validation tools
✓ Community contributions and review
✓ SKILL.md specification: community-governed
✓ Agent discovery and execution
✓ Research-grade validation tools
✓ Community contributions and review
Verified (revenue)
🏆 Clinical-grade verification badge
🏆 Multi-site concordance testing
🏆 Signed reproducibility bundles
🏆 CLIA/CAP compliance documentation
🏆 Regulatory audit trail generation
🏆 Enterprise SLA and support
🏆 Multi-site concordance testing
🏆 Signed reproducibility bundles
🏆 CLIA/CAP compliance documentation
🏆 Regulatory audit trail generation
🏆 Enterprise SLA and support
The model: Red Hat for biology. The community builds and shares freely. The business provides the verification, compliance, and trust layer that clinical labs and pharma companies require before deploying agent-driven genomics on patient data.
Traction
The community is already building. Organically.
724
GitHub Stars
+100/month, zero marketing
719
Event Registrations
single virtual event, Apr 22
4
Hackathons
Imperial, Westminster, Madrid, virtual
Gold
AI UK Agent Prize
top European hackathon
✓ Contributors from Europe, Africa, Latin America
✓ Cell Genomics Perspective paper (under review)
✓ Wet-lab biologists contributing skills via LLMs
✓ LLM-UKB accepted at PLOS Computational Biology
✓ External adversarial audit by independent developer
✓ Co-authors: UK, Uganda, Peru
Team & Ask
Setting the standard for trusted biological AI.
Manuel Corpas, PhD
→ 20+ years in bioinformatics and genomics
→ Former EMBL-EBI team lead
→ Created BioJS (180K+ stars as OpenClaw)
→ First published personal genome interpretation
→ AI UK Agent Gold Prize (2026)
→ University of Westminster
→ Former EMBL-EBI team lead
→ Created BioJS (180K+ stars as OpenClaw)
→ First published personal genome interpretation
→ AI UK Agent Gold Prize (2026)
→ University of Westminster
Co-authors:
Segun Fatumo (MRC/UVRI Uganda, QMUL)
Heinner Guio (UTEC, Lima, Peru)
Segun Fatumo (MRC/UVRI Uganda, QMUL)
Heinner Guio (UTEC, Lima, Peru)
Shanghai Landing
Seeking clinical partners to pilot the verification badge on patient-facing genomic skills. Shanghai's biotech ecosystem and data sovereignty alignment make it ideal for the Asia-Pacific clinical-grade rollout.
The Opportunity
ClawBio has the potential to define the global standard for how biological AI skills are shared, verified, and deployed. The community builds the commons. The verification layer builds the business.