Title: Seeking Advice on an AI Precision Medicine & Drug Discovery Project

[justin-mbca]

Hello ClawBio community,

My name is Justin Zhang. I have a background in bioinformatics, oncology data engineering, and healthcare interoperability, with experience in clinico-genomic data integration, FHIR, OMOP, and precision oncology.

I’m organizing a collaborative project for LLM Zoomcamp https://github.com/DataTalksClub/llm-zoomcamp and would appreciate feedback from the ClawBio community.

Our vision is to build an AI Precision Medicine & Drug Discovery Platform composed of several independent but connected RAG/Agent projects, including:

• Variant Interpretation Assistant
• Biomarker Assistant
• Biomedical Literature Assistant
• Target Discovery Assistant
• Drug Repurposing Assistant
• Clinical Trial Matching Assistant
• FHIR / mCODE Oncology Assistant

The idea is that each project can function independently while together supporting a workflow such as:

Genomic Variants → Biomarkers → Literature Evidence → Target Discovery → Drug Discovery → Clinical Trials → Healthcare Interoperability

I’d love to hear your thoughts on:

1. Which of these project areas would align best with the ClawBio ecosystem?
2. Are there existing ClawBio skills or datasets we should learn from?
3. Which project(s) might be most valuable as future ClawBio contributions?
4. Do you see any gaps in the current ecosystem where oncology, FHIR, or clinical informatics skills would be useful?

Thank you for any advice. I’m excited to learn from the community and explore ways our work could eventually contribute back to ClawBio.

[manuelcorpas]

Thanks for your question I think there’s a lot of overlap The point of ClawBio is to automate capabilities that can be discovered by agents. There’s a lot in what you want to achieve. It would be good to have a benchmark that allows and shows ClawBio capability as trustworthy to deliver on specific functionality. Perhaps we should take this in a call Best Manuel

…

On Sunday, 7 June 2026, Justin Zhang ***@***.***> wrote: Hello ClawBio community, My name is Justin Zhang. I have a background in bioinformatics, oncology data engineering, and healthcare interoperability, with experience in clinico-genomic data integration, FHIR, OMOP, and precision oncology. I’m organizing a collaborative project for LLM Zoomcamp and would appreciate feedback from the ClawBio community. Our vision is to build an AI Precision Medicine & Drug Discovery Platform composed of several independent but connected RAG/Agent projects, including: • Variant Interpretation Assistant • Biomarker Assistant • Biomedical Literature Assistant • Target Discovery Assistant • Drug Repurposing Assistant • Clinical Trial Matching Assistant • FHIR / mCODE Oncology Assistant The idea is that each project can function independently while together supporting a workflow such as: Genomic Variants → Biomarkers → Literature Evidence → Target Discovery → Drug Discovery → Clinical Trials → Healthcare Interoperability I’d love to hear your thoughts on: 1. Which of these project areas would align best with the ClawBio ecosystem? 2. Are there existing ClawBio skills or datasets we should learn from? 3. Which project(s) might be most valuable as future ClawBio contributions? 4. Do you see any gaps in the current ecosystem where oncology, FHIR, or clinical informatics skills would be useful? Thank you for any advice. I’m excited to learn from the community and explore ways our work could eventually contribute back to ClawBio. — Reply to this email directly, view it on GitHub <#287?email_source=notifications&email_token=AAC5SR5ALR6DMRVH3L5577D46VUTDA5CNFSNUABBM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63RPGEYDEMJTGUYDNJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAC5SRZR5JAW5WS6F5C5CNL46VUTDAVCNFSM6AAAAACZ55ZKCSVHI2DSMVQWIX3LMV43GRDJONRXK43TNFXW4OZRGAZDCMZVGA3A> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- -- Dr Manuel Corpas Mob: +44 7939 807 507 Blog: http://manuelcorpas.com/ ORCID: 0000-0002-4417-1018 <http://orcid.org/0000-0002-4417-1018> Publons: 1168880 <https://publons.com/author/1168880> Twitter: @manuelcorpas <https://twitter.com/manuelcorpas> LinkedIn: manuelcorpas <https://www.linkedin.com/in/manuelcorpas> Skype: manu_corpas [image: https://congresogenomica.com/] <https://congresogenomica.com/>

[justin-mbca]

Hi Manuel, Thank you for your quick response and confirmation! I would love to schedule a call with you later this week. Do you have any availabilities on Thursday or Friday? Best, Justin

[vladimir-mitev]

For this architecture, I would separate three contracts before choosing tools:

1. Public variant identity and evidence.
2. Literature retrieval and publication provenance.
3. Patient-specific, FHIR, or clinical decision context.

The first two can be composed with public identifiers. The third should remain a separately governed layer and must not leak patient context back into public retrieval tools.

A useful benchmark for the Variant Interpretation Assistant would be:

• Submit a supported public variant and require a closed outcome such as resolved, ambiguous, not found, invalid, unsupported, or unavailable.
• If ambiguous, preserve the candidate list and forbid a silent downstream choice.
• If resolved, pass the returned canonical identifier byte-for-byte into the literature component.
• Follow one returned PMID into publication details and require the source link and match basis to survive the chain.
• Reject patient, phenotype, family, segregation, and private case fields at the public-tool boundary.

That benchmark remains useful across implementations because it tests composition and refusal behavior rather than whether a generated explanation sounds plausible.

I represent Helena Bioinformatics. Folklore Clinical Variant Interpretation MCP is one public, read-only implementation of the first two stages: public variant evidence, automated ACMG/AMP decision support, variant-linked literature, publication details, and semantic biomedical literature search. It does not accept patient or case data, and its results require qualified professional review rather than serving as diagnosis or treatment advice: https://folklore.helena.bio/integrations

If this project is still active, I can contribute a small output-to-input conformance fixture against the public interface while keeping the FHIR/mCODE layer explicitly separate.