Skip to main content

​
How It Works

Most users never need to think about the MCP internals, but it helps to know why the workflow can take a few minutes.

​
Asynchronous job flow

Synthesize Bio MCP does not complete the entire analysis in a single instant response. Instead, it uses a background job flow:
  1. Claude starts the analysis and receives a job ID immediately.
  2. Claude checks the job status while the workflow continues on the Synthesize Bio platform.
  3. When the run is complete, Claude returns the finished report and any structured result references.
This design keeps the integration responsive even when the underlying analysis takes several minutes.

​
Analysis stages

Each run moves through three major stages:
  1. resolve_sample_metadata Interprets the natural-language prompt and extracts the sample metadata needed to run the comparison.
  2. GEM model Runs gene expression model inference for the requested groups and modality.
  3. Differential expression Performs statistical testing with Welch’s t-test and Benjamini-Hochberg false discovery rate correction.

​
Behind the scenes

The MCP uses several internal tools behind the scenes:
  • resolve_sample_metadata extracts the sample metadata needed for the comparison.
  • analyze_gene_expression starts the job.
  • get_analysis_results checks progress and returns the completion result, including a fenced JSON block of gene-level results and the platform dataset link.
  • get_counts_data_url provides a download URL for the raw counts data.
In practice, Claude usually handles this flow for the user. The important thing to remember is that a long-running analysis is expected behavior, not a failed request.

​
Result format

Completed runs return a Markdown summary that contains:
  • the analysis metadata (prompt, modality, group counts, significance summary),
  • a link to the Synthesize Bio platform dataset for the run, and
  • a fenced ```json block with up to 1,000 of the most significant differentially expressed genes (under a top-level results array), so an LLM can parse the block directly to drive downstream analysis or chart widgets (e.g. a volcano plot with x = log2FoldChange, y = -log10(padj)).
While a run is still in flight, get_analysis_results returns short progress messages instead. Continue to Usage Examples for prompt patterns that work well.