Investigation into the role of sequence-driven-features for prediction of protein structural classes Sundeep Singh Nanuwa and Huseyin Seker Abstract - There have been a number of techniques developed for the prediction of protein structural classes, however, they show various degrees of accuracies over different assessment procedures and, in particular, the role of sequence-driven- features (SDF) not rigorously investigated. Therefore, the aim of this study is to carry out the largest comprehensive and consistent investigation on approximately 1500 protein sequence-driven-features that form 65 subsets in order to develop a robust predictive model and identify how well these feature(s) are at predicting protein structural classes. For evaluation of the features, two high quality 40% (or less) homology datasets that contain over 7000 protein sequences were extracted from proteomic databases. As a predictive technique, an optimum K-Nearest Neighbour Classifier, namely multiple-K-NN (MKNN) was developed, which not only records MKNN results, but also a predictive accuracy for each K nearest neighbourhood for K=l to 11. In order to make the analyses consistent, three different cross-validation test procedures, 10-fold, leave-one-out and independent set, were used for all data sets and methods implemented. Over 5000 individual predictive results obtained, no firm consensus found on which features are highly associated with protein structural classes. However, interestingly, the best subsets of the features are found to be traditional AAC (48.62%) for 10-fold and (50.09 % ) for LOO, and dipeptide composition (85.91 %) for independent set. The results appear to suggest that the AAC features are one of the best two subsets over 65 different subsets. Interestingly, in particular, with pseudo-amino-acid composition (PseAAC), unlike other research results presented in the literature, this investigation finds that there is no statistical improvement obtained from the sequence-order effect aspect (lamda) of PseAAC, which averaged 39.15%. The results also suggest that most of its predictive power comes from the AAC part that averaged at 46.84 % , and the overall average predictive accuracy for PseAAC is 47.86%. This information appears to suggest that this feature set, which is claimed to better capture sequence order, yields almost no improvement and can be considered a redundant and noisy feature set. It should be noted that overall outcome of this comprehensive study sheds light not only in structural class prediction, but also other proteomic studies. I. INTRODUCTION P ROTEIN prediction is one of the most difficult and important fields within proteomics, mainly because the thousands of conformational changes in a protein makes it difficult to predict how it will fold into its secondary or Manuscript received July 5, 2008. Authors are with the Bio-Health Informatics Research Team at the Centre for Computational Intelligence, School of Computing, De Montfort University, Leicester. UK, LEI 9BH (e-mail: ssn@dmu.ac.uk. hseker@dmu.ac.uk) tertiary structure. Computational biology has a huge impact in this field because of relatively inexpensive computational power that has enabled vast amounts of data to be analysed relatively quickly. A protein is a biological molecule that carries out a specific function within the body, knowing and incorporating the structural class information can (1) improve prediction accuracy of secondary and tertiary structure prediction [1-4] and more significantly, (2) to bridge the gap between verified and unverified protein structures. The number of unverified protein structures is over 6 million [Release 39.0 of 22-July-2008 UniProtKB/TrEMBL] very different from how many have been verified, as of 12-August-2008 there are 52402 structures in Protein Data Bank (PDB). Levitt and Chothia [4] developed the standard for protein structural classes used in this study, which consists of four main types of protein structural classes: - 1. All-Alpha (a) - proteins with only small amount of strands 2. All-Beta - proteins with only small amount of helices 3. Alpha / Beta (a / - proteins that include both helices and strands and where strands are mostly parallel 4. Alpha + Beta (a + - proteins with both helices and strands and where strands are mostly anti-parallel There is substantial progress in protein structural class prediction [5-19], some of these studies use selected sequence features i.e. amino acid composition (AAC) only, which may not include crucial physiochemical properties and/or using poor quality datasets with low number of sequences at higher homology and all combined with inconsistent methodologies to arrive at often boosted results. The approach this project is taking, which is unique within the field, is to use approximately 1500 protein features extracted from the web server ProFEAT [20] as it is more of an interest to examine how additional and combination of features predict protein structural classes. Analysing these features using the predictive model multiple-k-nearest neighbourhood (MKNN) classifier and finally gathering results with three-test procedures. With the abovementioned approach, the projected outcomes are (1) identifying which of these sequence-driven-features is good at predicting protein structural classes and (2) a study that has used a consistent and comprehensive methodology throughout.