A Bayesian Approach to High-Throughput Biological Model Generation

With the availability of hundreds and soon thousands of complete genomes, the construction of genome-scale metabolic models for these organisms has attracted much attention. Manual work still dominates the process of model generation, however, and leads to the huge gap between the number of complete genomes and genome-scale metabolic models. The challenge in constructing genome-scale models from existing databases is that usually such a directly extracted model is incomplete and contains network holes. Network holes occur when a network is disconnected and certain metabolites cannot be produced or consumed. In order to construct a valid metabolic model, network holes need to be filled by introducing candidate reactions into the network. As a step toward the high-throughput generation of biological models, we propose a Bayesian approach to improving draft genome-scale metabolic models. A collection of 23 types of biological and topological evidence is extracted from the SEED [1), KEGG [2], and BiGG [3] databases. Based on this evidence, we create 23 individual predictors using Bayesian approaches. To combine these individual predictors and unify their predictive results, we build an ensemble of individual predictors on majority vote and four classifiers: naive Bayes classifier, Bayesian network, multilayer perceptron network and AdaBoost. A set of experiments is performed to train and test individual predictors and integrative mechanisms of single predictors and to evaluate the performance of our approach.