Bayesian Inference of Protein and Domain Interactions Using the Sum-Product Algorithm
In order to fully understand the functions of proteins in living organisms we must study their interactions and construct accurate interaction maps. Each protein can be composed of one or several peptide chains called domains
and each protein interaction can be seen as a consequence of an underlying interaction of two domains, one from each protein. Since high-throughput methods of measuring protein interactions, such as yeast two-hybrid assay, have high error rates, the structural similarities between proteins can be exploited to detect some of the experimental errors and predict new, unmeasured interactions. In this paper we solve the problem of Bayesian inference of protein and domain interactions by computing their likelihoods conditioned on the measurement results. We formulate the task of calculating these conditional likelihoods as a functional marginalization problem, where the multivariate function to be marginalized naturally factors into simpler local functions, and demonstrate how this equivalent problem can be solved using the sum-product algorithm. We note that such task is structurally similar to the decoding the low density parity check codes using the message passing algorithm. The robustness and accuracy of our approach is demonstrated by predicting protein and domain interactions on both real and artificial measurement data.