GRN inference algorithms

Use the pinned runtime and pre-import Dask configuration in SKILL.md. Calls below assume prepared inputs and a managed client; wrap process creation and inference in a main guard. Signatures were checked against PyPI 0.1.6 and current upstream algorithm source.

Choose the estimator

Both fit a regression for each target, exclude that target from its predictors, and emit positive-importance TF-target pairs. GRNBoost2 is not restricted to one-level decision stumps: its default kwargs leave scikit-learn's tree depth default in effect. GENIE3 uses one random forest per target, not an ensemble of random forests. Timing depends on observations, targets, candidate TFs, tree settings, and workers; there is no fixed 10,000-observation applicability cutoff.

Agreement between algorithms measures sensitivity to estimator choice. It does not independently validate regulatory causality. Scores across algorithms should not be compared with one shared numerical threshold.

Public signatures

grnboost2(expression_data, gene_names=None, tf_names="all",
          client_or_address="local", early_stop_window_length=25,
          limit=None, seed=None, verbose=False)

genie3(expression_data, gene_names=None, tf_names="all",
       client_or_address="local", limit=None, seed=None, verbose=False)

• gene_names: required for ndarray/CSC input; ignored in favor of DataFrame columns.
• tf_names: list of predictor names, or None/"all" for all genes.
• client_or_address: existing distributed.Client, scheduler address, or None/"local" to create a local cluster. Existing clients remain caller-owned.
• limit: top N links globally. Choose a positive integer or None.
• seed: regressor seed; record it and all dependency versions.
• early_stop_window_length: positive window size for mean out-of-bag improvement; this is not a held-out experimental validation set.

from arboreto.algo import grnboost2, genie3

network = grnboost2(expression_data=expression_matrix, tf_names=tf_names,
                    seed=42, limit=5000, client_or_address=client)
# Prefer a fresh managed client for each run if repeated runs cancel futures.
comparison = genie3(expression_data=expression_matrix, tf_names=tf_names,
                     seed=42, client_or_address=other_client)

Explicit estimator settings

grnboost2 and genie3 do not accept arbitrary scikit-learn kwargs. Use diy with regressor_type="GBM", "RF", or "ET" and a parameter dictionary:

from arboreto.algo import diy
from arboreto.core import SGBM_KWARGS, RF_KWARGS

custom_gbm = diy(
    expression_data=expression_matrix,
    regressor_type="GBM",
    regressor_kwargs={**SGBM_KWARGS, "n_estimators": 100,
                      "max_depth": 5, "learning_rate": 0.1},
    tf_names=tf_names, seed=42, client_or_address=client,
)
custom_rf = diy(
    expression_data=expression_matrix,
    regressor_type="RF",
    regressor_kwargs={**RF_KWARGS, "n_estimators": 1000, "max_features": "sqrt"},
    tf_names=tf_names, seed=42, client_or_address=other_client,
)

Both custom profiles ran on the synthetic fixture. ExtraTrees is source-verified but was not separately executed. Copy default dictionaries before overriding; do not mutate module globals. An XGBoost helper exists in core source, but XGBoost is not an implemented supported estimator here.

Arboreto uses scikit-learn's GradientBoostingRegressor.fit(..., monitor=...) and oob_improvement_; the current official API still documents these, but the executed compatibility matrix here is scikit-learn 1.5.2, not all newer releases. See current estimator API and Arboreto core.