N4-acetylcytidine (ac4C) modification is crucial for gene expression regulation. However, existing prediction methods are limited by one-dimensional sequence encodings, which inadequately capture the higher-order spatial and global statistical features of RNA. Moreover, while capsule networks have shown promise in biological sequence analysis, their primary capsule layer lacks an adaptive feature selection capability, leading to the introduction of redundant information and thus compromising model performance. To this end, we propose CapsBAM-ac4C. The model first utilizes Chaos Game Representation (CGR) to encode and normalize RNA sequences into two-dimensional images, which effectively represent global structures while eliminating scale shifts from variations in sequence length or base composition. The core innovation of the model lies in the integration of a dual capsule-level attention mechanism (CapsChannel and CapsSpatial) into the capsule network architecture. This mechanism adaptively enhances key features and suppresses noise from both the feature dimension and spatial position perspectives. Systematic evaluations on three independent test sets, covering sequence lengths of 201 nt and 415 nt as well as balanced and imbalanced sample distributions, demonstrate that CapsBAM-ac4C significantly outperforms state-of-the-art methods across multiple key metrics, including ACC, MCC, and AUROC. This study not only provides a highly accurate and robust tool for ac4C site identification but also highlights the immense potential of graph-based features and capsule attention mechanisms in the analysis of RNA modifications. For user convenience, we developed a user-friendly web server, which can be accessed for free at http://zhulab.org.cn/CapsBAM-ac4C/.
Shi et al. (2026) studied this question.