Extending GRADIS: Supervised Learning of Gene Regulatory Networks by Inclusion of Binding Motifs on Promoter Regions

Project Overview

This Master's project extends the GRADIS (Gene Regulatory Network Inference using GRAph DIStance-based Supervised learning) framework by incorporating transcription factor binding motif information from promoter regions. The approach enhances gene regulatory network (GRN) inference in Escherichia coli by integrating expression data with binding site evidence from multiple biological databases.

Project Structure

The project is organized into two main components:

1. Data Gathering and Analysis (databases.ipynb)

• Data Sources: Integration of three databases:

  • RegulonDB: Curated knowledge of transcriptional regulation in E. coli
  • Ecocyc: Encyclopedia of E. coli genes and metabolism
  • TEC Database: Transcription profiles for E. coli

• Consensus Database Construction:

  • Identified transcription factor (TF) binding sites within 500 basepairs upstream of genes
  • Established gene-TF relationships across all three databases
  • Created a consensus database with five columns: genes, TF, regDB, ecocyc, TEC
  • Database agreement indicators (1 = relationship present, 0 = absent)

2. GRADIS Implementation and Extension (gradis_ex.ipynb)

GRADIS Framework Implementation:

• Input Data: E. coli expression profiles (805 samples) and gold standard validated regulations
• Feature Engineering:
  • Clustered expression data to 50 centroids using K-means
  • Scaled expression data and constructed edge-weighted complete graph
  • Generated adjacency matrix using Euclidean distance metric
  • Created feature vectors by horizontally stacking matrix rows
• Labels: Binary classification (1 = known regulation, 0 = unknown regulation)

Data Imbalance Handling:

• Dataset: 152,280 total rows (2066 known regulations, ~150,000 unknown)
• Systematic training approach with balanced sampling
• Unknown class divided by number of known regulations for balanced training sets

Extension Methodology:

1. Initial GRADIS Prediction: Train SVM model on original feature vectors
2. Negative Prediction Analysis: Investigate TF-target relationships predicted as negative
3. Database Integration Filtering:
   • Filter 1: Relationships with agreement in ≥2 databases
   • Filter 2: Relationships with agreement in ≥1 database
4. Iterative Model Refinement: Train separate SVM models after each filtering step

Evaluation:

• Random selection of negative predictions matching known regulation size
• Train-test split (80-20) for final model evaluation
• Performance metrics: AUC (Area Under ROC Curve) and AUPR (Area Under Precision-Recall Curve)

Results

Method	AUC	AUPR
GRADIS (Baseline)	0.89	0.91
GRADIS + Filter 1 (≥1 database agreement)	0.95	0.96
GRADIS + Filter 2 (≥2 database agreement)	0.94	0.96

Key Findings

• Integration of binding motif information significantly improves GRN inference accuracy
• Database consensus filtering effectively identifies false negatives in predictions
• Even single-database agreement provides substantial performance improvement
• The framework demonstrates robust handling of severe class imbalance

Future Work

• Application to eukaryotic systems with more complex regulatory mechanisms
• Extension to other bacterial species
• Incorporation of additional data types (e.g., ChIP-seq, DNase-seq)