MTMixG-Net: mixture of Transformer and Mamba network with a dual-path gating mechanism for plant gene expression prediction

Accurate prediction of plant gene expression is essential for elucidating the regulatory mechanisms underlying plant development and stress adaptation. Traditional experimental approaches such as microarrays and RNA sequencing have provided valuable insights but remain limited in capturing the complexity and diversity of genomic regulation. Recent advances in deep learning have shown promise, yet existing models often struggle to generalize across species and to efficiently model long-range dependencies within genomic sequences. To address these challenges, we propose MTMixG-Net, a novel deep learning framework that integrates Transformer and Mamba architectures with a gating mechanism for enhanced gene expression prediction. MTMixG-Net consists of three main modules: the mixture of Transformer and Mamba encoder (MTMixEnc), the dual-path gating mechanism (DPGM), and the residual CNN chain (ResCNNChn). The MTMixEnc combines the self-attention capacity of Transformers with the state-space efficiency of Mamba to capture multi-scale regulatory dependencies while maintaining low computational complexity. The DPGM adaptively refines feature selection through dynamic gating, allowing the model to focus on the most informative representations. Finally, the ResCNNChn leverages a sequence of residual CNN blocks to extract high-level features and further boost predictive accuracy. We validate MTMixG-Net on multiple plant genomic datasets, demonstrating its superior accuracy and computational efficiency compared to existing methods. Our results highlight the potential of MTMixG-Net as a powerful tool for advancing plant genomics research and crop improvement strategies.