Identification of protein complexes from proteomic experiments is crucial to understand not only their function but also the principles of cellular organization. Advances in experimental techniques have enabled the construction of large-scale protein-protein interaction networks, and computational methods have been developed to analyze high-throughput data. In most cases several parameters are introduced that have to be trained before application. But how do we select the parameter values when there are no training data available? How many data do we need to properly train a method. How is the performance of a method affected when we incorrectly select the parameter values? The above questions, although important to determine the applicability of a method, are most of the time overlooked. We highlight the importance of such an analysis by investigating how limited knowledge, in the form of incomplete training data, affects the performance of parametric protein-complex prediction algorithms. Furthermore, we develop a simple non-parametric method that does not rely on the existence of training data and we compare it with the parametric alternatives. Using datasets from yeast and fly we demonstrate that parametric methods trained with limited data provide sub-optimal predictions, while our non-parametric method performs better or is on par with the parametric alternatives. Overall, our analysis questions, at least for the specific problem, whether parametric methods provide significantly better results than non-parametric ones to justify the additional effort for applying them.
Parametric and Non-parametric Gradient Matching for Network Inference
