Probabilistic Paths for Protein Complex Inference

Understanding how individual proteins are organized into complexes and pathways is a significant current challenge. We introduce new algorithms to infer protein complexes by combining seed proteins with a confidenceweighted network. Two new stochastic methods use averaging over a probabilistic ensemble of networks, and the new deterministic method provides a deterministic ranking of prospective complex members. We compare the performance of these algorithms with three existing algorithms. We test algorithm performance using three weighted graphs: a naïve Bayes estimate of the probability of a direct and stable protein-protein interaction; a logistic regression estimate of the probability of a direct or indirect interaction; and a decision tree estimate of whether two proteins exist within a common protein complex. The best-performing algorithms in these trials are the new stochastic methods. The deterministic algorithm is significantly faster, whereas the stochastic algorithms are less sensitive to the weighting scheme.