Source: https://github.com/aipoch/medical-research-skills

When to Use

• You have a bulk RNA-seq expression matrix and want to infer transcription factor (TF) → target gene regulatory edges.
• You have single-cell RNA-seq data (after normalization/aggregation as needed) and want to recover putative regulatory interactions.
• You need GRN inference that can scale to large datasets using parallel/distributed execution.
• You want to compare gradient-boosting–based GRN inference (GRNBoost2) versus random-forest–based inference (GENIE3).
• You need a reproducible, scriptable pipeline to generate a ranked network edge list from expression data.

Key Features

• GRN inference from gene expression data using GRNBoost2 (gradient boosting) or GENIE3 (random forest).
• Scalable execution via Dask, from a single machine to multi-node clusters.
• Command-line workflow for generating a GRN edge list from a tabular expression matrix.
• Algorithm guidance and comparison: see references/algorithms.md.
• Distributed setup notes: see references/distributed_computing.md.

Dependencies

• arboreto
• dask
• distributed
• pandas
• scipy
• scikit-learn

Example Usage

Run GRN inference from an expression matrix (TSV) and write the inferred network to an output file:

python scripts/infer_network.py \
  --input expression_data.tsv \
  --output network.tsv \
  --algo grnboost2

To use the alternative algorithm:

python scripts/infer_network.py \
  --input expression_data.tsv \
  --output network.tsv \
  --algo genie3

Implementation Details

• Input/Output

  • Input: a gene expression matrix (e.g., TSV) where rows typically represent samples/cells and columns represent genes (exact expectations depend on scripts/infer_network.py).
  • Output: a ranked edge list representing inferred regulatory relationships (TF → target) with an importance/weight score.

• Algorithms

  • GRNBoost2: uses gradient boosting to estimate feature importance of candidate regulators for each target gene; generally preferred for larger datasets due to speed and scalability.
  • GENIE3: uses random forests to compute regulator importance per target gene; a classic baseline for GRN inference.
  • For a detailed comparison and practical guidance, refer to references/algorithms.md.

• Parallel/Distributed Execution

  • Computation is parallelized with Dask, enabling scaling from local multi-core execution to distributed clusters.
  • Cluster configuration and deployment considerations are documented in references/distributed_computing.md.

• Key Parameters

  • --algo: selects the inference method (grnboost2 or genie3), affecting runtime and model behavior.
  • Additional runtime/cluster parameters (if exposed by the script) typically control Dask scheduling, worker counts, and resource usage.