Comparative multi-task deep learning models for protein-nucleic acid interaction prediction: Unveiling the superior efficacy of the PNI-MAMBA architecture.

Protein-nucleic acid interactions (PNI) play crucial roles in various life processes, including gene expression regulation, DNA replication, repair, recombination, and RNA processing and translation. However, accurately predicting these interactions remains challenging due to their complexity. This paper proposes a deep learning-based multi-task learning framework for predicting protein-nucleic acid interactions. The integrated framework comprises four independent deep learning models: PNI-FCN, PNI-Transformer, PNI-MAMBA, and PNI-MAMBA2. PNI-FCN leverages fully connected neural networks, PNI-Transformer utilizes Transformer networks, and both PNI-MAMBA and PNI-MAMBA2 are built upon Mamba network architectures. A novel binding site attention mechanism is introduced to capture key binding site information. The multi-task learning objective function combines the binary classification cross-entropy loss with a binding site loss to guide the model's focus on critical regions. Experiments on merged DNA and RNA datasets demonstrate the effectiveness of the proposed framework in accurately predicting protein-nucleic acid interactions and identifying binding PNI sites. Notably, the architectural framework leveraging PNI-MAMBA (s)-encompassing both PNI-MAMBA and PNI-MAMBA2-demonstrates superior overall performance, thereby enhancing both the accuracy and robustness of predictions. This work offers significant insights into the underlying molecular mechanisms and lays a strong foundation for the development of targeted therapeutic interventions.