Arboreto

Arboreto

When to use

Use Arboreto to rank candidate regulator-target associations from expression measurements. GRNBoost2 fits stochastic gradient boosting regressions; GENIE3 fits random forests. Each target is predicted from candidate regulators, excluding itself. These are observational predictive associations, not proof of direct binding, activation/repression, or causal regulation.

The latest PyPI release checked is 0.1.6 (2021-02-09). Read the Docs still labels its documentation 0.1.5; the current GitHub source contains fixes that are not in the PyPI wheel. Do not assume a successful unpinned installation can run inference. See compatibility and distribution details.

Installation and compatibility

The following exact stack passed dense and CSC-sparse GRNBoost2, dense GENIE3, custom GBM/RF, and wrapper smoke tests on macOS arm64 with Python 3.11.11:

uv venv --python 3.11 .venv-arboreto
uv pip install --python .venv-arboreto/bin/python \
  'arboreto==0.1.6' 'dask[complete]==2024.7.1' 'distributed==2024.7.1' \
  'numpy==1.26.4' 'pandas==2.2.3' 'scikit-learn==1.5.2' 'scipy==1.13.1'

This is a bounded compatibility recipe, not a claim that current releases of all dependencies work. PyPI Arboreto builds an empty metadata graph that the newer Dask dataframe implementation rejects. For this pinned Dask version, select its legacy dataframe backend before importing Arboreto or dask.dataframe:

import dask
dask.config.set({"dataframe.query-planning": False})
from arboreto.algo import grnboost2, genie3

The bundled wrapper does this for Dask 2024.7.1. Restart an existing notebook kernel if it has already imported the newer dataframe backend. Sparse targets also use .A inside Arboreto 0.1.6; this attribute was removed in SciPy 1.14. Keep the tested SciPy pin for sparse inference. No monkeypatch to site-packages is required by this recipe.

Workflow

1. Select biologically comparable cells/samples; document normalization, filtering, batch handling, organism, identifier namespace, and expression layer.
2. Prepare rows = observations, columns = genes. Exclude sample IDs from expression values. Require unique gene names, numeric finite values, and a TF list with a nonempty overlap. All-zero/constant genes provide no useful targets.
3. Choose GRNBoost2 for an efficient starting analysis, GENIE3 for method comparison, or diy for explicit regressor settings. See algorithms.
4. Run a small subset first in the pinned environment, then scale worker counts to available memory. Keep the if __name__ == "__main__": guard in process-based scripts.
5. Inspect worker warnings and target coverage, save the full ranked network, and assess stability across seeds and resampled observations before prioritizing edges.

Run the bundled wrapper

From this skill directory, with a TSV containing gene headers and numeric rows:

.venv-arboreto/bin/python scripts/basic_grn_inference.py expression_data.tsv network.tsv \
  --tf-file tfs.txt --seed 777 --workers 2 --limit 5000

Add --index-col 0 only if the first column contains cell/sample identifiers. The wrapper rejects duplicate headers before pandas can rename them, nonnumeric or nonfinite values, empty TF overlap, invalid limits, and wholly empty results. It reports TF overlap and uses a fresh bounded Dask client that closes on error. The default is one worker; increase it after a successful pilot. Without a TF file, all genes are candidate regulators, even though the output column is named TF.

Output is a headerless TSV in TF, target, importance order. For downstream consumers that require column headers (including pySCENIC adjacency loading), write a separate copy with header=True rather than assuming every tool accepts the headerless upstream example format.

Minimal Python example

This synthetic example checks execution and output structure; it is not a biological benchmark. The same calls were tested with a 32-observation, four-gene fixture.

import dask
dask.config.set({"dataframe.query-planning": False})
import numpy as np
import pandas as pd
from arboreto.algo import grnboost2
from distributed import Client, LocalCluster

if __name__ == "__main__":
    rng = np.random.default_rng(123)
    values = rng.normal(size=(32, 4))
    values[:, 2] = 3 * values[:, 0] + rng.normal(scale=0.1, size=32)
    matrix = pd.DataFrame(values, columns=["TF1", "TF2", "G1", "G2"])
    with LocalCluster(n_workers=1, threads_per_worker=1,
                      dashboard_address=None) as cluster, Client(cluster) as client:
        network = grnboost2(expression_data=matrix, tf_names=["TF1", "TF2"],
                            seed=777, client_or_address=client)
    assert not network.empty
    assert not (network["TF"] == network["target"]).any()
    network.to_csv("network.tsv", sep="\t", index=False, header=False)

For real DataFrame, ndarray, CSC, and AnnData input conventions, read basic inference.

Interpret and validate output

Results are sorted by decreasing importance; zero-importance links are omitted. GRNBoost2 rescales feature importance by the fitted number of trees, so its scores can exceed 1 and are not on the same scale as GENIE3. There is no universal importance > 0.5 confidence cutoff. limit=N keeps the top N links globally; it does not limit target regressions or return N links per target.

For consensus, define a per-run selection rule first, then count the fraction of all runs retaining each TF-target pair. An edge missing from a run is not an observed score to average only over present rows. Archive individual networks, seeds, package versions, filters, and identifier lists. Match preprocessing, sample sizes and gene sets across conditions; differences in scores alone do not establish differential regulation. Use independent motif, binding or perturbation evidence to assess candidates. Agreement between GRNBoost2 and GENIE3 is method sensitivity analysis, not independent biological validation.

Upstream retries target-level regression failures and can return empty target results after warnings. A nonempty overall network does not prove every target fit succeeded. Check logs and expected target coverage; absence of an edge may reflect zero importance, filtering, missing predictors, or a failed regression.