Demonstrated use cases

These are reproducible against the demo's Drug Discovery sample. The goal in each case is to refine a large gene list into a focused set of candidate targets. They are written so you can adapt the pattern to a different portal: substitute your own catalogue and field names, and keep the shape of the conversation. If a request is ambiguous, the model asks a targeted follow-up before querying (see the reseaher workflow).

Use case 1: iterative refinement

Iterative refinement narrows a large candidate set into a focused one through cross-table filtering, one conversational step at a time. This example moves from the mutation catalogue to the expression catalogue.

1. Discover the data.

   What datasets are available?

   The model calls list-catalogues and returns the catalogues the server exposes.

2. Explore a catalogue's fields.

   What fields are available in the mutation catalogue?

   The model calls get-catalogue-fields and returns fields such as hugo_symbol, cancer_type, overall_mutation_frequency, is_oncogene, and is_tumor_suppressor_gene, each with its type and valid operators.

3. Run a first query.

   Show me BRCA genes with an overall mutation frequency above 10%.

   The model constructs a SQON filter from the field metadata and calls execute-query, returning the matching genes as a table.

4. Refine across a second catalogue.

   Which of these genes show elevated average expression?

   The model cross-references the expression catalogue using the gene list from the previous step and keeps the genes that also meet the expression criterion.

5. Narrow further.

   Of those, which are annotated as oncogenes?

   The model adds an is_oncogene filter, leaving a short, focused candidate list.

Each step approves and runs one query, and each follow-up operates on the previous result set, so a researcher builds a precise cohort without writing a single filter by hand.

Use case 2: pathway-informed filtering

Pathway-informed filtering uses the same refinement pattern to assemble a candidate gene set, then relates those genes to biological pathways, the question that feeds deeper analysis.

1. Assemble a candidate set by mutation burden and cancer type:

   What are the most mutated genes in colorectal cancer?

   An open-ended request like this is usually ambiguous, so the model clarifies intent (which catalogue, which threshold) before proposing a filter such as mutation_frequency > 10% AND hotspot = true for the COADREAD cohort, then runs it once you approve.

2. Add an expression criterion to keep genes that are also active:

   Which of these genes show elevated expression levels?

3. Ask the pathway question to relate the focused set to known biology:

   What pathways are these genes involved in?

What is reproducible today, and what is not. The gene-refinement pattern (steps 1 and 2) is reproducible against the demo, as long as you stay within each catalogue's loaded coverage. The demo's mutation sample is BRCA-only and does not include a hotspot field, so the colorectal (COADREAD) example above illustrates the full Drug Discovery Portal rather than the demo sample; to reproduce it locally, run the same shape of conversation against BRCA and the fields the demo actually exposes (see use case 1). Step 3 is the hand-off point: dedicated pathway querying (a companion MCP server over MSigDB, the Molecular Signatures Database) is a forthcoming capability, so today the workflow produces the focused gene list that pathway analysis will consume.