尽管已有利用lncRNA和蛋白质的信息来预测lncRPI的方法,但仅利用蛋白质和RNA的序列特征来进行预测相互作用仍然是一个挑战,并且模型预测的准确性有待提高.因此,本文提出了一种融合卷积神经网路和自注意力机制的预测模型LPI-Attention(Long non-coding RNA based on self-attention mechanism),该模型采用了k-mer方法来编码RNA和蛋白质序列特征作为模型的输入,这种方法可以同时考虑两种序列的信息,从而提高了预测的准确性.此外,在密集型卷积模块中,使用两种尺度的特征提取,更好地捕捉局部和全局的信息.最后,将得到的特征输入自注意力循环网络层中,更好地处理序列数据的长期依赖关系,将得到的RNA、蛋白质二者特征信息融合成新的特征放入全连接层进行预测.实验结果表明,该模型不仅扩展了生物特征预测领域,而且可以学习RNA序列与蛋白质序列之间更多的相互作用关系,在预测RPIs方面表现优于大多数同类方法,在数据集RPIs1446、RPIs1807、RPIs488上的准确率分别达到91.7%、96.6%、91.6%.
On Protein-RNA Interaction Prediction Method Based on Self-attention Mechanism
Although there are methods that use information from long non-coding RNA(lncRNA)and protein to predict lncRPI,it remains a challenge to predict their interaction solely based on the se-quence features of protein and RNA,and the accuracy of model predictions needs improvement.There-fore,this paper proposes a prediction model called LPI-Attention(Long non-coding RNA based on self-attention mechanism),which combines convolutional neural networks and self-attention mecha-nism.The model adopts the k-mer method to encode the sequence features of RNA and protein as in-put to the model,which can consider the information of both sequences simultaneously and improve prediction accuracy.In addition,in the dense convolutional module,feature extraction is performed at two scales,which can better capture local and global information.Finally,the obtained features are in-put into the self-attention recurrent network layer,in which long-term dependencies in sequence data can be better handled,and the fused features of RNA and protein are put into the fully connected layer for prediction.Experimental results show that this model not only expands the field of biological fea-ture prediction but also learns more interactions between RNA sequences and protein sequences.It performs better than most advanced methods in predicting RPIs,with accuracy rates of 91.7%,96.6%,and 91.6%on datasets RPIs1446,RPIs1807,and RPIs488,respectively.
NETL
NSTL
万方数据
