Background: Protein-Protein Interactions (PPIs) play a key role in various biological processes. Many methods have been developed to predict protein-protein interactions and protein interaction networks. However, many existing applications are limited, because of relying on a large number of homology proteins and interaction marks.

Methods: In this paper, we propose a novel integrated learning approach (RF-Ada-DF) with the sequence-based feature representation, for identifying protein-protein interactions. Our method firstly constructs a sequence-based feature vector to represent each pair of proteins, via Multivariate Mutual Information (MMI) and Normalized Moreau-Broto Autocorrelation (NMBAC). Then, we feed the 638- dimentional features into an integrated learning model for judging interaction pairs and non-interaction pairs. Furthermore, this integrated model embeds Random Forest in AdaBoost framework and turns weak classifiers into a single strong classifier. Meanwhile, we also employ double fault detection in order to suppress over-adaptation during the training process.

Results: To evaluate the performance of our method, we conduct several comprehensive tests for PPIs prediction. On the H. pylori dataset, our method achieves 88.16% accuracy and 87.68% sensitivity, the accuracy of our method is increased by 0.57%. On the S. cerevisiae dataset, our method achieves 95.77% accuracy and 93.36% sensitivity, the accuracy of our method is increased by 0.76%. On the Human dataset, our method achieves 98.16% accuracy and 96.80% sensitivity, the accuracy of our method is increased by 0.6%. Experiments show that our method achieves better results than other outstanding methods for sequence-based PPIs prediction. The datasets and codes are available at https://github.com/guofei-tju/RF-Ada-DF.git.

背景：蛋白质-蛋白质相互作用（PPI）在各种生物过程中起着关键作用。已经开发出许多方法来预测蛋白质-蛋白质相互作用和蛋白质相互作用网络。然而，由于依赖于大​​量的同源蛋白和相互作用标记，许多现有的应用受到限制。

方法：在本文中，我们提出了一种新颖的基于序列特征表示的集成学习方法（RF-Ada-DF），用于识别蛋白质-蛋白质相互作用。我们的方法首先通过多元互信息（MMI）和标准化的Moreau-Broto自相关（NMBAC）构建了一个基于序列的特征向量来表示每对蛋白质。然后，我们将638维特征馈入一个集成的学习模型中，以判断交互对和非交互对。此外，该集成模型将随机森林嵌入AdaBoost框架中，并将弱分类器转变为单个强分类器。同时，我们还采用双重故障检测以抑制训练过程中的过度适应。

结果：为了评估我们方法的性能，我们对PPI进行了一些综合测试。在幽门螺杆菌数据集上，我们的方法达到了88.16％的准确度和87.68％的灵敏度，我们的方法的准确度提高了0.57％。在酿酒酵母数据集上，我们的方法达到了95.77％的准确度和93.36％的灵敏度，我们的方法的准确度提高了0.76％。在人类数据集上，我们的方法达到98.16％的准确性和96.80％的灵敏度，我们的方法的准确性提高了0.6％。实验表明，与其他出色的基于序列的PPI预测方法相比，我们的方法取得了更好的结果。数据集和代码位于https://github.com/guofei-tju/RF-Ada-DF.git。