ezGeno

ezGeno automates the optimization of deep-learning neural network architectures for 1D genomic data to improve prediction of transcription factor binding and tissue-specific enhancer activity.

Key Features:

• Automated Network Tuning: Automatically tunes neural network parameters and architectures for genomic tasks.
• Search Space Customization: Searches three explicit search spaces—number of filters, dilation factors, and connectivity between layers—to optimize model structure.
• Versatile Input Compatibility: Accepts 1D genomic inputs including genomic sequences, histone modifications, DNase feature data, and their combinations.
• Systematic Architecture Exploration: Systematically searches through network configurations to identify effective model structures for given tasks.
• Comparative Performance: Demonstrates higher average AUC than a one-layer DeepBind model and greater efficiency compared with AutoKeras in benchmark comparisons.

Scientific Applications:

• Transcription Factor Binding Prediction: Optimizes parameters and network structures on genomic sequence data to enhance TF binding prediction and to highlight relevant sequence segments.
• Tissue-Specific Enhancer Activity Prediction: Integrates sequence data and DNase feature data to improve prediction of tissue-specific enhancer activities relative to manually designed models.

Methodology:

ezGeno employs deep learning with automated neural network architecture search by systematically searching configurations across three search spaces: number of filters, dilation factors, and inter-layer connectivity.