CalibGRN

CalibGRN is a research framework for Gene Regulatory Network (GRN) inference using Calibrated Transformer models.
It supports both supervised and unsupervised workflows, integrates baseline methods, and adds novel calibration and attention regularization techniques for more reliable biological network predictions.

Features

• Synthetic & real GRN datasets (DREAM4, DREAM5, etc.)
• Baseline methods: ARACNE, Pearson, TIGRESS, GENIE3, GRNBoost2
• Transformer-based GRN inference
• Calibration methods:
  • Temperature Scaling
  • Edge-level Dirichlet Calibration
• Attention regularization with biological priors
• Unified evaluation: AUROC, AUPR, ECE, reliability diagrams
• Open-source Python package for reproducibility

Installation

git clone https://github.com/yourusername/calibgrn.git
cd calibgrn
pip install -e .

Project Structure

calibgrn/         # Main package source
notebooks/        # Experiment notebooks
tests/            # Unit tests
setup_env.py      # Environment setup helper
requirements.txt  # Dependencies
setup.py          # Install script

Goal

Develop a unified and reproducible framework for GRN inference that combines:

• Baseline algorithms for benchmarking
• Transformer architectures for improved sequence/context modeling
• Novel calibration techniques for trustworthy probability outputs
• Biological prior regularization for interpretability

Milestones

• Milestone 1: Baseline implementations (ARACNE, Pearson, TIGRESS, GENIE3, GRNBoost2)
• Milestone 2: Prototype Transformer model for GRN inference
• Milestone 3: Calibration methods (Temperature Scaling, Edge-level Dirichlet Calibration)
• Milestone 4: Attention regularization with biological priors
• Milestone 5: Ablation studies on synthetic data
• Milestone 6: Transfer to real-world datasets (DREAM4, DREAM5)
• Milestone 7: Final evaluation and comparison against baselines

Novel Contributions

1. Calibration-aware training loss: Combines BCE with a differentiable calibration penalty (e.g., soft ECE surrogate) to improve both accuracy and calibration.
2. Attention regularization with biological priors: Loss term nudging attention weights toward known TF→target likelihoods (e.g., motif presence).
3. Edge-level Dirichlet calibration: Models dependencies among outgoing edges from the same regulator, beyond independent scaling.

License

This project is licensed under the MIT License.