| The present study's main focus is to exploit diverse sources of biological knowledge and use machine learning as a technique to successfully distinguish between true and false membrane proteins' interactions from high-throughput experiments. The Support Vector Machine is trained to evaluate such interactions based on co-expression and Gene Ontology terms similarity features. In total, seven kernels were employed and their results were analyzed from five different perspectives: the predictions' features performance, their statistical significance, the different kernels overlaps, integrating the qualitative and quantitative information through a scoring scheme, and considering the biological significance of the best. The seven employed kernels had very high training performance rates. Their overall trend for a lower sensitivity than their positive predictive value could explain to some extent the partial inconsistency of the predicted interactions with the bait-dependent interactions set. Finally, the five most significant predicted interacting preys are suggested for further experimental validation. |
