Predicting protein-protein interactions in unbalanced data using the primary structure of proteins

Abstract

BackgroundElucidating protein-protein interactions (PPIs) is essential to constructing protein interaction networks and facilitating our understanding of the general principles of biological systems. Previous studies have revealed that interacting protein pairs can be predicted by their primary structure. Most of these approaches have achieved satisfactory performance on datasets comprising equal number of interacting and non-interacting protein pairs. However, this ratio is highly unbalanced in nature, and these techniques have not been comprehensively evaluated with respect to the effect of the large number of non-interacting pairs in realistic datasets. Moreover, since highly unbalanced distributions usually lead to large datasets, more efficient predictors are desired when handling such challenging tasks.ResultsThis study presents a method for PPI prediction based only on sequence information, which contributes in three aspects. First, we propose a probability-based mechanism for transforming protein sequences into feature vectors. Second, the proposed predictor is designed with an efficient classification algorithm, where the efficiency is essential for handling highly unbalanced datasets. Third, the proposed PPI predictor is assessed with several unbalanced datasets with different positive-to-negative ratios (from 1:1 to 1:15). This analysis provides solid evidence that the degree of dataset imbalance is important to PPI predictors.ConclusionsDealing with data imbalance is a key issue in PPI prediction since there are far fewer interacting protein pairs than non-interacting ones. This article provides a comprehensive study on this issue and develops a practical tool that achieves both good prediction performance and efficiency using only protein sequence information.

Computational Approaches to Predict Protein Interaction

“…However, there are very limited techniques developed to confirm that two proteins do not interact. Recently, several studies have addressed this problem in evaluating computational methods of identifying protein interactions (Yu, Chou et al 2010;Yu, Guo et al 2010). This issue is still in a chaos stage and there is no perfect solution that fit everyone's requirements.…”

“…Yu et al reported that using SVM to perform a complete interaction analysis on human genome may take years (Yu, Chou et al 2010). In this regard, efficient ML algorithms with acceptable accuracy are reasonable alternatives to SVM.…”

“…In this regard, efficient ML algorithms with acceptable accuracy are reasonable alternatives to SVM. The relaxed variable kernel density estimation (RVKDE) algorithm (Oyang, Hwang et al 2005) has been practically used in recent interaction studies (Chang, Syu et al 2010;Yu, Chou et al 2010). The time complexity of RVKDE is an order faster than SVM.…”

Computational Tools and Databases for the Study and Characterization of Protein Interactions

“…Other approaches include SVMs that combine sequence profiles and other sequence-based information such as spatially neighbouring residues (Koike and Takagi 2004;Res et al 2005;Chen and Li 2010), a RF that integrates physicochemical properties of residues, evolutionary conservation and amino acid distances (Chen and Jeong 2009), and a naive Bayesian classifier trained to integrate position-specific scoring matrix and predicted accessibility (Murakami and Mizuguchi 2010). Finally, other sequence-based methods have been developed to improve prediction by tacking issues such as the problem of unbalanced data in protein sets (Yu et al 2010), i.e. the interface accounts for a small proportion of the exposed residues so the number of negative cases (non-interface residues) is much larger than the number of positive cases (interface residues) or improving the sampling (Engelen et al 2009) in evolutionary trace-based (Lichtarge et al 1996) methodologies.…”

Domain‐mediated protein interaction prediction: From genome to network

“…Limitations of these machine learning methods are similar to the ones described above, notably including the lack of well-defined true negative examples. For instance, Yu et al evaluated the effect of positive-to-negative ratio in training and test sets for SVM based methods and found that it had considerable effect on classifier accuracy [83]. Lastly, use of sequence signatures to refine high-resolution PRM-mediated interaction networks must avoid duplicate counting of the domain-motif interaction knowledge already used to generate the original network.…”