DeepGRN

DeepGRN predicts transcription factor binding sites across cell types using attention-based deep neural networks that integrate genomic DNA sequences and experimental data such as DNase-Seq coverage.

Key Features:

• Attention modules: Integrates a single attention module and a pairwise attention module to learn long-range dependencies within sequential genomic data.
• Input data: Leverages genomic DNA sequences and experimental data from parallel sequencing technologies, such as DNase-Seq coverage.
• Prediction scope: Predicts transcription factor binding sites across various cell types.
• Benchmarking: Evaluated on ENCODE-DREAM In Vivo Transcription Factor Binding Site Prediction Challenge datasets, achieving higher unified scores for six of thirteen targets compared to the top four methods in the challenge.
• Attention-weight correlation: Correlates attention weights with DNase-Seq coverage and motifs to identify informative inputs.
• Visualization: Provides visualization techniques for attention modules to aid interpretation of learned patterns across input types.
• Long-range dependency learning: Uses attention mechanisms to capture long-range dependencies in sequential data, similar to applications in natural language processing.

Scientific Applications:

• Transcription factor binding site prediction: Identifies potential DNA binding sites for transcription factors across different cell types.
• Regulatory sequence interpretation: Relates model attention to motifs and DNase-Seq signals to highlight informative regulatory features.
• Benchmark evaluation: Serves as a comparative method in challenges such as the ENCODE-DREAM In Vivo Transcription Factor Binding Site Prediction Challenge.
• Gene regulation studies: Supports analyses aimed at understanding gene regulation mechanisms across cell types.

Methodology:

Uses attention-based deep neural networks with a single attention module and a pairwise attention module applied to genomic DNA sequences and DNase-Seq coverage; correlates attention weights with DNase-Seq signals and motifs, visualizes attention modules, and evaluates performance on ENCODE-DREAM datasets using unified scores.