Protein function prediction and classification using uncertainty

The overall aim of this project is to investigate the use of Bayesian networks (Needham et al., 2006b) in integrating information, expressing relationships and making inferences or predictions on biological problems, motivated by data generation in genomics and proteomics. We have already successfully applied Bayesian networks to two problems in which we have previous experience. In the first instance, surface patch analysis was combined with a Bayesian network to predict protein-protein binding sites (Bradford et al., 2006) with a success rate of 82% on a benchmark dataset of 180 proteins, improving by 6% on previous work and well above the 36% that would be achieved by a random method. Interestingly, a comparable success rate was achieved even when evolutionary information was missing, suggesting that, in most cases, only chemical and physical surface properties are required for accurate prediction. Next, we used Bayesian networks to predict the functional consequences of missense mutations on proteins (Needham et al., 2006a). Exploiting the ability of the Bayesian network to handle missing data automatically, we found that structural information is significantly more discriminatory than evolutionary information in this classification task and on the dataset used. Indeed, the top three strongest connections with the class node in the network all involved structural nodes. We therefore derived a simplified Bayesian network that used just these three structural descriptors, with comparable performance to that of an all node network. Currently, we are using Bayesian networks to integrate heterogeneous data sources including sequence motif, protein-protein interaction (PPI) and gene expression data to assign functions (described by the Gene Ontology) to proteins of Arabidopsis thaliana. There are over 30000 unique gene products in Arabidopsis. However, 47% of these have an unknown molecular function, and 49% and 64% have yet to be assigned to a cellular compartment and biological process respectively. We aim to assign GO terms in all three of these functional categories. Bayesian networks are particularly suitable to this problem as they can handle the noisy and uncertain data, and relate the functional categories.