DeepGMAP

DeepGMAP predicts gene regulatory regions from genomic sequences using deep learning trained on epigenomic data to annotate functional elements in non-protein-coding regions.

Key Features:

• Deep Learning Architecture: Employs various neural network architectures trained on epigenomic data to map potential gene regulatory regions across the genome.
• FRSS Convolution Layers: Implements forward- and reverse-sequence scan (FRSS) convolution layers that integrate information from both forward and reverse DNA strands to enhance prediction accuracy.
• Training and Benchmarking: Trains models on epigenomic datasets and compares performance across different architectures.
• Overfitting Mitigation: Assesses data structures highlighted in previous studies to address sources of overfitting and improve model generalizability.
• Visualization of Learned Patterns: Incorporates visualization methods to examine model-learned sequence patterns and interpret biological relevance of predicted regulatory regions.

Scientific Applications:

• Regulatory Region Prediction: Predicts locations of gene regulatory regions from sequence and epigenomic signals to support genome annotation without immediate experimental validation.
• Non-coding Genome Annotation: Enables functional interpretation of non-protein-coding genomic regions by identifying putative regulatory elements.
• Functional Genomics: Supports elucidation of gene regulatory networks by providing candidate regulatory regions for downstream analysis.
• Personalized Medicine: Informs variant interpretation by mapping potential regulatory regions that may modulate gene expression.
• Evolutionary Biology: Aids comparative analyses by identifying conserved and species-specific regulatory elements across genomes.

Methodology:

Training deep learning models on epigenomic data, comparing performance across architectures, using FRSS convolution layers to process forward and reverse DNA strand information, assessing data structures to mitigate overfitting, and applying visualization methods to examine learned sequence patterns.