MAMBF: A Multi-head Attention and Multi-branch Feature Fusion-based Deep Learning Model for Rapid Prediction of RNA Molecular Specific Hybridization Results

Abstract With the development of biochemistry, Ribonucleic acid (RNA) molecular hybridization has become a core technology for key measurement tasks in molecular recognition, target quantification, and other biochemical fields. Simulation software serves as an important auxiliary tool for designing hybridization experiments. However, the current simulation method suffers from severe concentration measurement inefficiency due to its exponential computational complexity, and this inefficiency becomes even more persistent especially in large-scale exploration or inspection tasks, making it challenging to obtain the hybridization analysis results within an acceptable time. To overcome this bottleneck for large-scale screening tasks, the main objective of this work is to propose a fast deep learning alternative that can directly predict hybridization outcomes without expensive traversal. The MAMBF designed for two tasks: perfect hybridization screening and evaluating the degree of non-specific hybridization. The model incorporates multi-sub feature extraction spaces, multi-head attention, multi-branch feature analysis, enables it to address the inefficiencies of recurrent neural networks in learning local features and alleviates the challenges faced in capturing feature dependencies. Using only a small number of samples, MAMBF can sufficiently learn multi-level features associated with hybridization and deliver outstanding performance that far outpaces comparative algorithms: It achieves an average accuracy of 99.3% in perfect hybridization screening and 87.1% in non-specific hybridization degree evaluation. For perfect hybridization screening task targeted in this work, MAMBF achieves approximately 18,000-fold speedup compared with the conventional simulation pipeline, resolving the trade-off between accuracy and computational cost. Additionally, validation through real experiments on RNA hybridization yielded results consistent with the model’s predictions, further verifying that its advantages in accuracy and efficiency are not just theoretical, but practical and implementable. It builds a bridge for the integration of artificial intelligence and automated auxiliary analysis of molecular measurement processes, providing a feasible auxiliary solution for enhancing the efficiency of pre-experiment simulation.