For the complete documentation index, see llms.txt. This page is also available as Markdown.

Running Analysis

You run an analysis by asking the agent in plain language. The agent inspects your files, chooses an established workflow, runs it, and explains each step. This page follows one single-cell RNA-seq dataset (GEO series GSE221575, colorectal cancer) from upload to annotated cell types.

Start by describing the data

Ask the agent to describe your uploaded files before you analyze them:

Describe the data

Agent response listing five samples and their metadata, read from an uploaded .tar archive

The agent also flags issues that affect later steps. Here it notes that the counts are raw barcode totals before QC, and that each sample has its own gene list, so the gene identifiers must be intersected before the samples are merged.

Check this description against what you know about the experiment. Correcting a misread sample label here is cheaper than after the analysis has run.

Run the analysis

Ask for the analysis you want. A single prompt can cover several steps:

QC, harmonize, and annotate the data

Cell-type annotation summary, caveats, and the numbered output folders

For cell-type annotation, the summary names the reference (a tissue-matched CellTypist model with majority voting), the number of subclusters, and composition differences between conditions. The caveat here is a high fraction of cells left unlabeled, attributed to stressed, high-mitochondrial samples. The step folders are 01_QC, 02_Harmonization, and 03_Annotation, and every file in them can be downloaded.

The progress indicator above the message box (for example Step 4 / 5) shows how far a multi-step run has got.

Reproducibility

Mithrl-1 runs standard analyses through skills: tested, versioned workflows that fix the method, parameters, and order of operations.

Two sessions side by side showing identical coarse annotation results

See the code behind a result

Every line of code the agent runs is recorded in the session's notebook.

You can also ask the agent to export the analysis as standalone code that runs outside Mithrl-1. In this example the agent was asked to build a two-sample subset, write a self-contained script replicating the harmonization and annotation pipeline, and check the script's output against the platform's:

Create a miniature dataset: comprising 2 samples; 1 normal 1 CMV, subsample to 3000 cells each. Run the miniature dataset through the harmonization and cell annotation pipeline you just did. Create a self-contained python code that perfectly replicates the harmonization -> cell type annotation pipeline performed such that the results from this and the skills based pipeline on the miniature dataset are identical. If the output is not identical (or within 99% concordant across all metrics), update the self-contained code until it does.

Concordance table between the platform pipeline and standalone code, with the notebook open

Match a reference figure

To reproduce a figure layout, upload an image of the figure and ask for one like it:

Please create a figure like the uploaded Screenshot png file

Reference figure: UMAP and marker dot plot
Uploaded reference figure
Generated figure matching the reference layout

The same approach works for reports: upload a sample report as a PDF and ask the agent to follow its layout.

Last updated

Was this helpful?