CNNKSCEC: a deep learning-based framework for chromatin loop prediction with multi-source feature integration

Area of Science:

• Genomics and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Chromatin forms a complex 3D structure in the cell nucleus, with loops being key organizational units.
• Accurate chromatin loop prediction is vital for understanding gene regulation and disease.
• Current prediction methods struggle with noise, data imbalance, and multi-omics integration.

Purpose of the Study:

• To develop an advanced deep learning framework, CNNKSCEC, for improved chromatin loop prediction.
• To address limitations of existing methods by integrating multi-source data and sophisticated feature extraction.
• To enhance the understanding of gene regulation and disease mechanisms through accurate loop identification.

Main Methods:

• Developed CNNKSCEC, a deep learning framework utilizing multi-source feature fusion.
• Integrated Hi-C and DNase-seq data into a dual-channel feature matrix.
• Employed a three-stage iterative feature extraction: dual-branch convolutional module (CNNC), SCConv module (SRU, CRU), and ECHybridAddition module (ECA, CBAM attention).
• Utilized a fully connected layer for classification and density-based clustering for filtering false positives.

Main Results:

• CNNKSCEC demonstrated superior performance compared to existing chromatin loop prediction methods.
• The framework effectively extracts multi-scale features and enhances data representation.
• Achieved high accuracy in predicting candidate chromatin loops.

Conclusions:

• CNNKSCEC offers a robust and effective deep learning solution for chromatin loop prediction.
• The integration of multi-omics data and advanced attention mechanisms significantly improves prediction accuracy.
• This framework advances the study of 3D genome organization and its role in biological processes and diseases.