Características inerentes a medidas de centralidade e uso de algoritmos de aprendizado de máquina para classificação de bridging nodes

Abstract

Protein-protein interaction (PPI) networks often carry expressive numbers of proteins and interactions in the order of thousands, possibly millions. The promising nodes in large PPI networks, which can be used in drug production, can be identified through exact methods such as bridging centrality, however, this can become computationally expensive to be overcome due to the complexity of these networks. As an alternative, the use of machine learning is suggested. The objective of this study was analyzed machine learning algorithms (ML) applied to the problem of determining bridging centrality in PPI networks, modeled by Complex Network, and identify the algorithm which offers results closest to generated by the exact algorithm taken as a reference, but with less computational effort. PPI networks were selected from nine different bacteria, on which a set of structural metrics were generated using Gephi. Then, each PPI file containing the generated metrics was submitted to the analysis of 15 algorithms selected for ML available in the WEKA. The prediction metrics files were submitted to the best predictive model identified and then the nodes were classified as weak or strong. Finally, the performance of the classifier was evaluated using the R software and the ROCR package, obtaining the ROC curve, the area under ROC curve (AUC), the corresponding accuracy and threshold. The best learning models identified correspond to the algorithms Bagging and Random Forest and the worst were the NaiveBayes and OneR. In general terms, the mean prediction accuracy of the nodes of the PPIs was 74.38% ± 5.84%, with a mean threshold of 96%, and mean AUC of 65.09 ± 4.48%. The nodes correctly predicted by the classifier were, on average, 24.33%, with 2.75% true positive and 21.58% true negative. On the other hand, 75.66% were incorrectly classified, being 23,67% false positive and 51.99% false negative. Comparing the 2.75% true positive with those identified by the exact algorithm, an average hit rate of 77.16% ± 20.23% was obtained. The predictive result generated by the ML process approached that obtained by the exact algorithm, presenting efficacy, however, with a considerable error rate. Thus, our results corroborate the literature knowledge about the use of ML in complex networks, that is, ML algorithms applied to complex centrality measures such as bridging centrality are not effective. The plugin implemented as one of the products of this work for GEPHI software, version 0.9.1 or higher, is available on sourceforge.net under the name of BridgingCentralityPlugin.