RNA-binding proteins (RBPs) play a significant part in several biological processes in the living cell, such as gene regulation and mRNA localization. The research indicates that the mutation of RBPs will lead to some serious diseases. Several deep learning methods, especially the model based on convolutional neural network (CNN), have been used to predict the binding sites. However, these methods only use single-scale filters to extract a fixed length of motifs features, which restricts the performance of prediction. For the sequence data, different sizes of filters may learn different biological information of the RNA sequence. Therefore, a deep multi-scale attention network (DeepMSA) based on convolutional neural network is proposed to predict the sequence-binding preferences of RBPs. DeepMSA extracts features by multi-scale CNNs and integrates these features with an attention model to predict the RBPs and binding motifs. Experiments demonstrate DeepMSA outperforms several state-of-the-art methods on the invivo and invitro datasets. The results indicate that attention can make the model learn the consistent pattern of candidate motifs, which can provide some important guiding advice for RBP motifs.

RNA 结合蛋白 (RBP) 在活细胞的多个生物过程中发挥重要作用，例如基因调控和 mRNA 定位。研究表明，RBPs的突变会导致一些严重的疾病。几种深度学习方法，特别是基于卷积神经网络（CNN）的模型，已被用于预测结合位点。然而，这些方法仅使用单尺度滤波器来提取固定长度的模体特征，这限制了预测的性能。对于序列数据，不同大小的过滤器可以学习到不同的RNA序列生物信息。因此，一个深度的多尺度提出了基于卷积神经网络的注意力网络（DeepMSA）来预测 RBP 的序列绑定偏好。DeepMSA 通过多尺度 CNN 提取特征，并将这些特征与注意力模型相结合，以预测 RBP 和绑定基序。实验表明，DeepMSA 在体内和体外数据集上的表现优于几种最先进的方法。结果表明，注意力可以使模型学习候选模体的一致模式，这可以为 RBP 模体提供一些重要的指导建议。