CHASOS: A Novel Deep Learning Approach for Chromatin Loop Predictions

Chromatin loop prediction is integral to understanding the mechanisms of gene regulation, cellular homeostasis, and disease pathogenesis. Previous prediction methods usually suffer from heterogeneities among multiple feature data, which restricts the chromatin loop prediction accuracy. Moreover, existing approaches are often bottlenecked due to the sequencing cost and the availability of samples to be sequenced. In this paper, we propose a novel method called CHASOS (CHromatin loop prediction with Anchor Score and OCR Score) to achieve accurate chromatin loop predictions. In particular, CHASOS constructs deep learning models that integrates multi-receptive field large kernel convolutional modules and scale modules to generate new features used for dealing with heterogeneities among multiple feature data. Using the extracted features, CHASOS constructs a gradient boosting tree model to accomplish the chromatin loop predictions without introducing extra sequencing. Experiment evaluations confirm that CHASOS outperforms existing approaches, showcasing superior generalization ability and enhanced prediction performance, particularly in the presence of feature data heterogeneity.