Improving prediction of the contact numbers of residues in proteins from primary sequences

Contact number is one kinds of one-dimensional features of proteins. Knowing the number of residue contacts in a protein is crucial to derive constraints useful in protein structure prediction. In this study, we evaluate and compare several methods and different features for contact number prediction. The experiments are performed on a non-redundant dataset containing 1109 proteins. The contact number prediction is formulated as a multi-class classification problem. Three-fold cross validation is used to get the performance of various methods with different combinations of features as input. The experimental results show that the profile feature containing evolutionary information of proteins can achieve better performance than simple amino acid sequences. Further performance improvement is achieved by including the predicted secondary structure and relative solvent accessibility as additional features. In all experiments, each tested method can improve the performance by more than 10 percent in comparison with the base-line method. The best Q score for two-class classification is 79.7%, which is higher than the best results reported in the literature by 2 percent. The results obtained here can provide valuable information for protein structure reconstruction, model quality assessment, etc.