Chromatin accessibility prediction via a hybrid deep convolutional neural network.

Area of Science:

• Genomics
• Computational Biology
• Genetics

Background:

• Most genetic variants linked to human diseases are in non-coding DNA regions, posing interpretation challenges.
• Accurately annotating cell-type-specific regulatory elements from large-scale sequencing data using machine learning remains difficult.
• Developing interpretable models to learn DNA sequence signatures is crucial for identifying causative genetic variants.

Purpose of the Study:

• To develop Deopen, a hybrid framework utilizing a deep convolutional neural network.
• To automatically learn the regulatory code of DNA sequences and predict chromatin accessibility.
• To identify causative non-coding variants associated with diseases.

Main Methods:

• Proposed Deopen, a hybrid framework based on a deep convolutional neural network.
• Trained the model to learn DNA sequence codes and predict chromatin accessibility.
• Compared Deopen's performance against existing methods for classification and regression tasks.

Main Results:

• Deopen demonstrated superior performance in classifying accessible DNA regions and predicting DNase-seq signals.
• Visualized convolutional kernels revealed identified sequence signatures matching known motifs.
• Successfully identified causative non-coding variants in a breast cancer dataset analysis.

Conclusions:

• Deopen accurately learns DNA sequence signatures and predicts chromatin accessibility.
• The model effectively identifies causative non-coding variants, aiding disease gene discovery.
• Deopen has broad applications in human genome annotation and identifying disease-associated variants.