ChiMamba: Predicting Chromatin Interactions Based on Mamba

Chromatin interactions play a crucial role in establishing specific chromatin structures and regulating gene expression over distant genomic regions. However, the high cost of sequencing techniques limits the identification of chromatin interactions across diverse samples. Considering its significance, quite a few deep learning-based methods have recently emerged for computationally detecting chromatin interactions. In this study, we propose ChiMamba, a bidirectional selective state space model based on Mamba, which integrates genomic sequences and chromatin accessibility data to predict chromatin interactions. ChiMamba exhibits superior and stable performance across six human cell lines. Moreover, the bidirectional design further enhances Mamba’s predictive capabilities. Subsequent analysis suggests that Mamba may outperform Transformers as a framework in the biological field, which not only achieves higher precision but also demonstrates linear scalability in time complexity with increasing input length. These observations indicate that Mamba holds higher suitability for deployment on ultra-long genomic sequences and serves as a novel foundational model for biological data.