Objective Non-coding RNA (ncRNA) plays a crucial regulatory role in various biological processes. Numerous studies have shown that the functions of ncRNAs depend not only on their nucleotide sequences but also closely on their spatial conformations, particularly their secondary structures. Traditional methods for predicting RNA secondary structures often have low accuracy, usually around 50 to 70 percent, and their performance further declines when dealing with complex structures or long sequences. Although deep learning methods have improved prediction performance to some extent, they still face challenges such as high model complexity, difficulty in capturing long-range dependencies, and poor generalization in predicting complex structures. Therefore, it is necessary for RNA secondary structure prediction to develop new methods.Methods This study proposed an end-to-end prediction model RMDfold, which employed a feature extraction strategy combining residual Mamba (resMamba) and dense connections (Dense). The model framework consisted of three modules: one-dimensional (1D) modeling, feature mapping, and two-dimensional (2D) modeling. For the 1D modeling module, the model learned contextual dependencies among nucleotides from RNA sequences, providing the foundation for possible base-pairing; for the feature mapping module, the 1D features were projected into a 2D space to form a constraint matrix that represented potential base-pairing relationships; for the 2D modeling module, the model further learned the pairing patterns between nucleotides, and determined the unique pairing of each nucleotide through pairing constraints, to obtain the final secondary structure. For both 1D and 2D modeling modules, a four-layer Dense block composed of batch normalization, ReLU activation, and convolution was used to extract short-range features; and a dual-branch residual structure resMamba based on a state-space model was used to model long-range dependencies, thereby achieving effective integration of short- and long-range features. To validate the effectiveness of the proposed method, it was compared with Ufold, REDfold, TransUfold, and sincfold on the three public datasets RNAStralign, ArchiveII, and bpRNA-new.Results The proposed RMDfold method demonstrates superior performance compared with existing algorithms in structure prediction, pseudoknot prediction, sequence prediction across varying lengths, and model complexity analysis, while requiring fewer parameters and achieving faster inference. For the structure prediction, the method achieved F1, Matthews correlation coefficient (MCC), precision and recall of 0.973 5, 0.973 1, 0.975 6 and 0.972 3 on RNAStralign, 0.854 3, 0.855 6, 0.874 7 and 0.876 2 on ArchiveII, and 0.382 8, 0.401 5, 0.536 5 and 0.318 7 on bpRNA-new, respectively. For the pseudoknot prediction based on ArchiveII, the model achieved F1, MCC, precision and recall of 0.741 4, 0.743 3, 0.752 5 and 0.739 6. For the sequence prediction across different lengths, RMDfold maintained an accuracy of 0.70 for sequences ranging from 200 to 500 nt. In terms of model complexity, RMDfold required 2.886 7 M parameters and achieves an inference speed of 0.026 0 s.Conclusion RMDfold enables highly accurate prediction of RNA secondary structures. It helps to deeply and comprehensively reveal the central roles of RNA molecules in gene expression regulation, molecular recognition, and catalysis. and also provides important structure support for elucidating disease-related variant mechanisms, designing RNA-targeted drugs, and advancing research in evolution and comparative genomics.