DeepMEL

DeepMEL applies deep learning to predict enhancer activity and dissect the genomic regulatory code of enhancers using cross-species chromatin accessibility data.

Key Features:

• Deep learning with explainability: Uses explainable deep learning models to predict enhancer activity from chromatin accessibility data.
• Cross-species profiling: Trained and validated on chromatin accessibility data from 26 melanoma samples across six species.
• Orthologous enhancer identification: Detects orthologous enhancers among distantly related species where traditional sequence alignment methods fail.
• Variant-level inference: Pinpoints specific nucleotide substitutions that contribute to enhancer turnover.
• Enhancer architecture and TF analysis: Analyzes enhancer architectures and identifies transcription factor binding sites within core regulatory complexes across melanoma cell states, distinguishing roles in nucleosome displacement and enhancer activation.
• Benchmarking: Demonstrated superior predictive accuracy in the CAGI5 challenge on the melanoma-specific enhancer of the IRF4 gene.
• Candidate prediction and mutation prioritization: Enables prediction and optimization of candidate enhancers and prioritization of enhancer mutations for downstream studies.
• Generalizability: Computational strategy applicable to other cancer types and normal cell states for studying noncoding genome variation and informing targeted gene-therapy approaches.

Scientific Applications:

• Enhancer function and evolution: Dissect enhancer function and evolutionary turnover by linking nucleotide changes to regulatory activity.
• Functional orthology and conservation: Identify orthologous enhancers and study functional conservation across species where sequence alignment fails.
• Regulatory architecture mapping: Map transcription factor roles in core regulatory complexes and cell-state–specific enhancer activation in melanoma.
• Variant prioritization for disease: Prioritize enhancer mutations relevant to melanoma progression and other cancer types.
• Design of regulatory elements: Predict and optimize candidate enhancers for experimental validation and therapeutic engineering.

Methodology:

Train explainable deep learning models on cross-species chromatin accessibility profiles (26 melanoma samples, six species) to predict enhancer activity, identify orthologous enhancers where sequence alignment fails, analyze enhancer architectures and transcription factor binding sites, and benchmark performance in the CAGI5 challenge on the IRF4 enhancer.