DuckDB powering cancer genomics research
- Hartwig Medical Foundation is a non-profit whose mission is to improve cancer care for patients today and in the future.
- It specializes in whole genome sequencing to help doctors make the best treatment decisions.
- A lab in Amsterdam receives patient samples from over 40 hospitals in the Netherlands, extracts the DNA, and runs it through sequencing machines.
- The raw sequencer output is fed into an open-source analysis pipeline called OncoAnalyser, used across the world.
- The pipeline runs various genetic analyses and produces an actionable report that clinicians use for treatment.
More detail
The intro slide shows the end-to-end diagnostic flow: patient sample intake, DNA extraction, sequencing machines, and the OncoAnalyser pipeline that emits a clinical report (with genome analysis DAGs and circos-style plots).
- Beyond diagnostics, patients can consent to have their data stored in Hartwig's research database.
- The database combines genetic and clinical data for almost 8,000 patients, making it one of the larger cancer genome databases in the world.
- The data is available for free to researchers and clinicians for cancer research.
- Access has been granted to over 300 research groups, producing almost 200 scientific publications.
- Hartwig is continually seeking new ways to give access to the data and maximize its impact.
More detail
The catalog UI (catalog.hartwigmedicalfoundation.nl) presents a sunburst chart of 7,883 total patients as of June 21, 2026, sliceable by primary tumor location and treatment type, with filters and Request Access / Export Data actions.
- GenOnco is a tool that runs fully in the browser, locally on the researcher's machine, powered by DuckDB WASM.
- Its goal is easier, faster access to the data for qualifying researchers, complementing the existing data request process.
- It is meant for cases where researchers do not need the raw DNA but want to explore the structured data.
- It provides a familiar SQL querying interface.
More detail
The GenOnco branding slide reveals it is a WASM (WebAssembly) application, so all query execution happens client-side in the browser rather than on a central server.
- GenOnco supports drawing diagrams using the Mosaic framework (idl.uw.edu/mosaic).
- Diagrams are specified declaratively in YAML, which Mosaic uses to render the data returned by a query.
- Mosaic hooks natively into DuckDB, enabling advanced and interactive visualizations.
- Example: a gene-by-tumor-type mutation heatmap built from a SQL query over the driver catalog.
More detail
The workbench shows a split view: a SQL query (WITH gene_totals AS ... over dc.driver_catalog) on top, its YAML Mosaic spec on the left, and the rendered mutation-type heatmap on the right.
- Hartwig believes its data has more impact when combined with similar data from other countries.
- Because the data is sensitive, each external source requires a separate legal vetting process, making a central middleman infeasible.
- Instead, DuckDB WASM pushes federation down to the researcher's own machine, i.e. the browser.
- Each data-source institution can set up its own domain, which can hold multiple data sources, each made of multiple tables.
- Data can be statically hosted, but an active server component is still needed for authorization.
More detail
A world map highlights whole-genome-sequencing-in-clinical-care partners (routine, piloting, planning, research, no-data), including institutions like MIT / Broad, Nictiz, and Peter Mac, motivating cross-border federation.
- Access to specific data sources is managed in Keycloak.
- The researcher obtains a token via OpenID Connect from Keycloak, which is then configured into DuckDB.
- DuckDB sends requests with the correct authorization headers to a domain proxy that checks the user's access rights.
- The proxy then forwards DuckDB, via signed URLs, to the actual data living in Parquet files on Google Cloud Storage.
More detail
This design keeps hosting mostly static while enforcing per-user authorization at request time: the proxy validates the Keycloak token before vending signed URLs to the Parquet files.
- GenOnco integrates AI to make the data even more accessible and easier to use.
- The approach is deliberately privacy-preserving: sensitive data is never shared with AI providers.
- The AI has access only to the shape of the data, not the data itself.
- It can write queries but not see the results, and make diagrams but not see the underlying values.
More detail
The AI panel is shown generating an oncoprint-style overview: it inspects available database views and schema, writes the WITH ... SQL, reads Mosaic spec docs, and renders the diagram, all without ever seeing the actual row values.
- GenOnco is still in early stages and being tested with users.
- Hartwig is talking with parties interested in becoming a domain for GenOnco with their own data.
- Follow Hartwig on GitHub (hartwigmedical) or LinkedIn (hartwig-medical-foundation) for updates.
More detail
The closing slide thanks the audience and points to Hartwig Medical Foundation's GitHub and LinkedIn for staying tuned on GenOnco's progress.
Key Takeaways
- Hartwig Medical Foundation runs whole genome sequencing for cancer patients via the open-source OncoAnalyser pipeline, feeding a research database of almost 8,000 patients.
- The database is one of the world's larger cancer genome databases and is free for research, powering 300+ research groups and nearly 200 publications.
- GenOnco is a browser-native, local-first data explorer powered by DuckDB WASM, complementing the traditional data request process.
- Visualizations use the Mosaic framework: YAML-specified, DuckDB-native, interactive diagrams like gene-by-tumor-type heatmaps and oncoprints.
- Sensitive cross-border data is federated in the browser rather than through a central middleman, sidestepping the need to combine data centrally.
- Authorization uses Keycloak and OpenID Connect tokens configured into DuckDB, with a domain proxy vending signed URLs to Parquet files on Google Cloud Storage.
- AI helps write queries and build diagrams while seeing only the data's shape, never the actual patient values, preserving privacy.