How to use the Arboreto gene network inference agent skill

What the skill makes the agent do

Arboreto's GRNBoost2 method fits stochastic gradient boosting regressions, and GENIE3 fits random forests. For each target gene, the method predicts its expression from candidate regulator genes, excluding the target itself. The result is a ranked list of regulator-target pairs with an importance score.

The skill adds the operational know-how: checking inputs, using a pinned Python 3.11 environment, running a bundled command-line wrapper, and checking stability across seeds. The scores are observational predictive associations, not proof of direct binding, activation, repression or causal regulation.

Install the skill from the ZIP or the Claude apps

Download the ZIP for the arboreto skill from this page. To install it for all your projects, unzip it into ~/.claude/skills/ so the arboreto folder (containing SKILL.md) sits inside that folder. To install it for one project only, unzip it into the project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

• Personal install: ~/.claude/skills/arboreto/
• Project install: .claude/skills/arboreto/ inside your project
• Apps: Settings → Capabilities → Skills

Set up the pinned Python environment

The skill pins a Python 3.11 environment because the latest PyPI release of Arboreto needs specific pinned dependencies to run inference. Create the environment with uv (or another virtualenv tool) and install the exact stack. Network access is needed only for this installation step.

Before importing Arboreto in your own code, set the legacy Dask dataframe backend, as the skill's example shows. Restart a notebook kernel if it already imported the newer backend.

• Use the pinned stack exactly; do not assume an unpinned install will run inference.
• Sparse inputs depend on the tested SciPy pin.
• The skill needs no API keys or credentials.

Trigger the skill with natural requests

You do not need special syntax. Describe the biology and the file in plain language, and name the method if you have a preference. The agent should load the skill when your request involves inferring gene regulatory networks from expression data.

Good requests state the data format, the regulator list and the goal. Vague requests such as 'analyse my genes' give the agent less to work with.

• Infer a candidate gene regulatory network from expression_data.tsv using GRNBoost2 and the TF list in tfs.txt.
• Compare GRNBoost2 and GENIE3 results on these two conditions and tell me which edges are stable across seeds.
• Check my expression matrix for the input problems Arboreto needs before running inference.
• Run a small pilot on 500 cells, then scale the worker count.

Run the bundled wrapper on your data

The skill includes scripts/basic_grn_inference.py. Run it from the skill directory with a TSV file that has gene headers and numeric rows. Add --index-col 0 only if the first column holds cell or sample identifiers. Without a TF file, every gene is treated as a candidate regulator, even though the output column is named TF.

The wrapper rejects duplicate headers, nonnumeric or nonfinite values, an empty TF overlap, invalid limits and empty results. It writes a headerless TSV with TF, target and importance columns. Start with one worker and increase it only after a successful pilot.

• Output is sorted by decreasing importance; zero-importance links are omitted.
• limit keeps the top N links globally, not N links per target.

Read and validate the output

Scores are useful for ranking, not as absolute confidence levels. GRNBoost2 scores can exceed 1 and are not on the same scale as GENIE3, and there is no universal importance cutoff. Define a selection rule for each run before comparing runs.

For consensus, count the fraction of all runs that keep each TF-target pair. A missing edge is not a score of zero to average over present rows. An empty target result can come from a failed regression, so check logs and target coverage. Agreement between GRNBoost2 and GENIE3 is a method sensitivity check, not independent biological validation.

• Archive each network with its seed, package versions, filters and identifier lists.
• Match preprocessing, sample sizes and gene sets before comparing conditions.
• Use motif, binding or perturbation evidence to support any candidate.

Safety notes and limits

This skill runs scripts, so read scripts/basic_grn_inference.py before executing it on sensitive data. The skill states that inference itself does not use the network and needs no credentials; network access is needed only during installation.

The skill does not do cell quality control, normalization or batch correction. You must document those steps yourself. The pySCENIC stages, such as motif pruning, regulon definition and AUCell, are separate steps that the skill did not test.

• Runs scripts: yes. Uses network during inference: no.
• Requires Python 3.11 and uv or another virtualenv tool.
• Not for causal claims from output alone.