DeepFun

DeepFun predicts the functional impacts of non-coding genetic variants at single-nucleotide resolution using convolutional neural network-based deep learning and integrated epigenomic annotations.

Key Features:

• Single-nucleotide resolution: Predicts variant effects at individual nucleotide positions in non-coding regions.
• Non-coding focus: Targets non-coding regions of the human genome that account for over 90% of variants identified by GWAS.
• Deep learning model: Employs a convolutional neural network (CNN) framework extensively evaluated for accuracy and reliability.
• Epigenomic integration: Integrates annotations from ENCODE and Roadmap to inform predictions.
• DNase I profiles: Incorporates 1548 DNase I accessibility profiles into the feature space.
• Histone mark profiles: Incorporates 1536 histone mark profiles into the feature space.
• Transcription factor profiles: Incorporates 4795 transcription factor binding profiles into the feature space.
• Tissue and cell type specificity: Profiles span 225 distinct tissues or cell types to enable tissue- and cell type-specific assessments.
• GWAS validation: Includes independent validations using datasets from various GWAS studies to assess predictive performance.
• Motif visualization: Facilitates visualization of potential sequence motifs around variants.

Scientific Applications:

• Variant prioritization in genetics: Prioritizes non-coding variants identified by GWAS for downstream analysis.
• Functional genomics: Interprets regulatory effects of non-coding variants across tissues and cell types.
• Disease mechanism studies: Associates non-coding variant impacts with disease-relevant tissues and pathways.
• Personalized medicine and complex trait analysis: Provides detailed functional insights to inform studies of individual-level variant effects and complex genetic architectures.

Methodology:

Convolutional neural network-based deep learning trained on a feature space constructed from ENCODE and Roadmap epigenomic annotations (1548 DNase I profiles, 1536 histone mark profiles, 4795 transcription factor binding profiles across 225 tissues/cell types), with independent validation using GWAS datasets and motif visualization.