BiDNAMamba: Pre-trained Bidirectional State Space Model for Motif Analysis in DNA

Promoter identification is an important field of genomics research and holds a significant impact on understanding the mechanism of transcriptional regulation. The recent advancements of computational methods have markedly enhanced classification accuracy. Nonetheless, the desire for higher accuracy drives the proposal of more innovative methods. In this paper, we propose BiDNAMamba, a novel state space model (SSM) for promoter motif identification from DNA sequences and offers the flexibility to easily expand scan path to accommodate various application scenarios. Compared with other pre-training computational algorithms, BiDNAMamba owned a significant superiority in terms of pre-training consumption. In addition, BiDNAMamba achieved excellent promoter identification performance, with an absolute improvement in the area under the receiver operating characteristic of 0.74%. Related interpretability analysis revealed the reasoning behind decisions made by the BiDNAMamba and filled the gap in visualization of SSM-based model. With its unique advantages and excellent performance, BiDNAMamba is poised to become a backbone in the realm of genomics foundation models. The code will be available at https://github.com/GaryinDeep/BiDNAMamba.