regCNN

regCNN: Convolutional Neural Network for Genome-Wide Cis-Regulatory Module Identification

regCNN predicts cis-regulatory modules (CRMs) genome-wide by integrating base-resolution local patterns between epigenetic marks and transcription factor binding site (TFBS) motifs using a convolutional neural network architecture.

Key Features:

• Base-Level Pattern Integration: Captures local interactions between epigenetic profiles and TF binding motifs at single-base resolution.
• Deep Learning Architecture: Applies a convolutional neural network to model combinatorial TFBS composition and regulatory patterns.
• Performance Metrics: Achieves 84.5% accuracy and 92.5% area under the receiver operating characteristic curve (auROC).
• Model Optimization: Improves auROC by 4.7% over pure multi-layer perceptron models and exceeds other tools by at least 11.3% in auROC.
• Length Normalization: Uses a resizing window hyperparameter to maintain stability across variable CRM lengths.

Scientific Applications:

• Transcriptional Regulatory Network Construction: Identifies CRMs to support modeling of gene regulation in metazoans.
• Genetic Disorder Research: Facilitates genome-wide CRM mapping for analysis of regulatory variants.

Methodology:

regCNN encodes epigenetic marks and TF binding site motifs at base resolution and applies convolutional layers to detect local combinatorial patterns. A resizing window hyperparameter standardizes input lengths, enabling robust CRM prediction across genomic regions. Model performance is evaluated using accuracy and auROC metrics.