SSAL：一种基于序列与结构特征及注意力模型的circRNA-RBP相互作用位点预测方法

SSAL: A Prediction Method for circRNA-RBP Interaction Sites Using an Attention Model Based on Sequence and Structural Features

摘要: 预测环状RNA (circRNA)与RNA结合蛋白(RBP)之间的相互作用位点，对于揭示疾病调控机制及开发新型治疗靶点具有重要意义。随着基因组范围内circRNA结合事件数据的日益丰富，计算模型已成为高效预测circRNA-RBP相互作用位点的主流工具。然而，如何通过有效提取circRNA的多尺度特征来提升预测准确率，仍是该领域面临的核心挑战。为解决这一问题，本研究提出了一种名为SSAL的深度学习模型，旨在实现大规模数据集上circRNA-RBP相互作用位点的精准预测。该模型的核心模块包括：首先，系统性地提取circRNA的序列特征与二级结构特征；随后，利用注意力机制对多尺度序列特征进行融合。为增强模型的稳定性和泛化能力，本文构建了一个集成学习框架，通过整合多个子模型的预测结果，有效缓解了单一模型固有的误差与随机性。为验证SSAL的性能，本文在14个大规模circRNA数据集上进行了全面评估，并将其与当前主流方法进行了对比。实验结果表明，SSAL的平均曲线下面积(AUC)达到97.66%，不仅充分证实了其在效率与鲁棒性方面的优势，且在预测准确率上均优于所有对比方法。

Abstract: Predicting the interaction sites between circular RNA (circRNA) and RNA-binding proteins (RBPs) is of significant importance for deciphering disease regulatory mechanisms and developing novel therapeutic targets. With the increasing accumulation and availability of genome-wide circRNA binding event data for computational analysis, computational models have become mainstream tools for efficiently predicting circRNA-RBP interaction sites. However, enhancing prediction accuracy by effectively extracting multi-scale features of circRNA remains a key challenge in this field. To address this issue, this study proposes a deep learning model named SSAL, which can accurately predict circRNA-RBP interaction sites on large-scale datasets. The core modules of this model include: first, systematic extraction of circRNA sequence features and secondary structure features; Subsequently, an attention mechanism was employed to fuse multi-scale sequence features. To enhance model stability and generalization capability, we constructed an ensemble learning framework that integrates predictions from multiple sub-models, effectively mitigating the errors and randomness inherent in individual models. To validate the performance of SSAL, we conducted comprehensive evaluations on 14 large-scale circRNA datasets and compared it with current mainstream methods. The results demonstrate that SSAL achieved an average AUC of 97.66%, not only fully confirming its advantages in efficiency and robustness but also surpassing all comparison methods in prediction accuracy.