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

Chromatin in the cell nucleus adopts a complex three-dimensional (3D) structure shaped by folding and interactions, with chromatin loops serving as fundamental organizational units. Accurate loop prediction is essential for understanding gene regulation and disease mechanisms. However, existing chromatin loop prediction methods still face challenges in noise handling, data imbalance, and multi-omics integration.

In this study, we present CNNKSCEC, a deep learning-based framework for chromatin loop prediction via multi-source feature fusion. The model integrates Hi-C and DNase-seq data into a dual-channel feature matrix as input. It employs a three-stage iterative feature extraction framework consisting of a dual-branch convolutional module (CNNC), a SCConv module combining SRU and CRU, and an ECHybridAddition module integrating both ECA and CBAM attention mechanisms. This design enables iterative multi-scale feature extraction and enhances the feature representation capability of the input matrix. Finally, the model uses a fully connected layer for classification, generating candidate chromatin loops with prediction scores, and filters out false candidates through density-based clustering. In the experiments, we compare CNNKSCEC with existing chromatin loop prediction methods, and the results demonstrate that the approach outperforms other methods overall in terms of performance. The code is available from https://github.com/zhengbingzi/CNNKSCEC.git.