An assessment of machine and statistical learning approaches to inferring networks of protein-protein interactions

Abstract

Protein-protein interactions (PPI) play a key role in many biological systems. Over the past few years, an explosion in availability of functional biological data obtained from high-throughput technologies to infer PPI has been observed. However, results obtained from such experiments show high rates of false positives and false negatives predictions as well as systematic predictive bias. Recent research has revealed that several machine and statistical learning methods applied to integrate relatively weak, diverse sources of large-scale functional data may provide improved predictive accuracy and coverage of PPI. In this paper we describe the effects of applying different computational, integrative methods to predict PPI in Saccharomyces cerevisiae. We investigated the predictive ability of combining different sets of relatively strong and weak predictive datasets. We analysed several genomic datasets ranging from mRNA co-expression to marginal essentiality. Moreover, we expanded an existing multi-source dataset from S. cerevisiae by constructing a new set of putative interactions extracted from Gene Ontology (GO)-driven annotations in the Saccharomyces Genome Database. Different classification techniques: Simple Naive Bayesian (SNB), Multilayer Perceptron (MLP) and K-Nearest Neighbors (KNN) were evaluated. Relatively simple classification methods (i.e. less computing intensive and mathematically complex), such as SNB, have been proven to be proficient at predicting PPI. SNB produced the "highest" predictive quality obtaining an area under Receiver Operating Characteristic (ROC) curve (AUC) value of 0.99. The lowest AUC value of 0.90 was obtained by the KNN classifier. This assessment also demonstrates the strong predictive power of GO-driven models, which offered predictive performance above 0.90 using the different machine learning and statistical techniques. As the predictive power of single-source datasets became weaker MLP and SNB performed better than KNN. Moreover, predictive performance saturation may be reached independently of the classification models applied, which may be explained by predictive bias and incompleteness of existing "Gold Standards". More comprehensive and accurate PPI maps will be produced for S. cerevisiae and beyond with the emergence of largescale datasets of better predictive quality and the integration of intelligent classification methods.

Computational prediction of protein interaction networks through supervised classification techniques

“…Experimental methods can only identify a subset of the interactions that occur in an organism. Therefore, coverage (i.e., the area of the genome covered by protein pairs) of the interactome (the collection of all the PPI that occur within a cell) is limited (Browne et al, 2006). Methods such as the yeast-two hybrid system exhibit high False Positive (FP) and false negative interaction rates.…”

“…Classifiers exhibit systematic bias (i.e., a method produces solutions that highly favour a limited number of specific situations) or are based on major assumptions (e.g., independence between datasets). Such potential bias and assumptions may lead to systematic prediction errors (Browne et al, 2006). In this study we evaluated three computational classification techniques to integrate diverse sources of information for PW and MB prediction.…”

“…RF does not assume feature independence and is robust against noise and missing values. In a study by Browne et al (2006) NB, Multi-Layer Perceptron (MLP) and K-Nearest Neighbour (KNN) were evaluated in the task of PPI prediction using the PW approach. The classifier NB achieved the 'highest' predictive performance by obtaining an area under Receiver Operating Characteristic (ROC) curve (AUC) value of 0.99.…”

“…Five of the features 2.1-2.5 were used in a previous study by Lu et al (2005). The GO-driven PPI dataset was obtained from our previous study (Browne et al, 2006) and the PE dataset integrated for PW prediction was acquired from a recent study by Collins et al (2007). Both the MIPS Functional Catalogue (FUNCAT) and GO-driven semantic similarity (GOSEM) are high quality annotation datasets and independent from the GS; however, they are limited in terms of coverage of the interactome.…”

A Hybrid Rule-Induction/Likelihood-Ratio Based Approach for Predicting Protein-Protein Interactions

“…Some interesting observations are drawn in [22] regarding this data integration for protein interaction prediction. A detailed analysis of this data integration using different classifiers is researched in [5].…”