Application of Bayesian networks to two classification problems in bioinformatics

The application of machine learning techniques to bioinformatics problems has become increasingly popular in recent years. Of particular interest are probabilistic graphical models since they provide a concise representation for inferring models from data. Current applications include the learning of gene regulatory networks (Friedman, 2004) and protein function prediction. Bayesian networks are the subset of probabilistic graphical models that can be expressed as directed acyclic graphs (Jordan, 1996; Jensen, 2001). They use a combination of domain knowledge and data to provide a framework that can be used to model the relationships between sets of variables in a probabilistic manner. They are a particularly powerful tool in bioinformatics since they can handle incomplete data sets allowing the building of a model from a training set with (different) missing data, or facilitating the prediction of a variable’s state/value based on a restricted set of evidence. For example, an important variable may be unknown during testing, yet marginal probabilities can still be calculated. Causal relationships between variables can also be learnt giving us information on the contribution of each variable to the prediction, and any independencies between variables can be exploited to reduce the complexity of the model. In this work, we use Bayesian networks for two classification tasks in bioinformatics. Our results show an improvement over previous machine learning methods applied to these problems. This improvement may be due to the Bayes nets’ ability to capture the interactions between multiple variables. Preliminary results show that the framework of probabilistic graphical models provides a good basis for modelling bioinformatics data.