Novel algorithms for protein sequence analysis

Each protein is characterized by its unique sequential order of amino acids, the so-called protein sequence. Biology__s paradigm is that this order of amino acids determines the protein__s architecture and function. In this thesis, we introduce novel algorithms to analyze protein sequences. Chapter 1 begins with the introduction of amino acids, proteins and protein families. Then fundamental techniques from computer science related to the thesis are briefly described. Making a multiple sequence alignment (MSA) and constructing a phylogenetic tree are traditional means of sequence analysis. Information entropy, feature selection and sequential pattern mining provide alternative ways to analyze protein sequences and they are all from computer science. In Chapter 2, information entropy was used to measure the conservation on a given position of the alignment. From an alignment which is grouped into subfamilies, two types of information entropy values are calculated for each position in the MSA. One is the average entropy for a given position among the subfamilies, the other is the entropy for the same position in the entire multiple sequence alignment. This so-called two-entropies analysis or TEA in short, yields a scatter-plot in which all positions are represented with their two entropy values as x- and y-coordinates. The different locations of the positions (or dots) in the scatter-plot are indicative of various conservation patterns and may suggest different biological functions. The globally conserved positions show up at the lower left corner of the graph, which suggests that these positions may be essential for the folding or for the main functions of the protein superfamily. In contrast the positions neither conserved between subfamilies nor conserved in each individual subfamily appear at the upper right corner. The positions conserved within each subfamily but divergent among subfamilies are in the upper left corner. They may participate in biological functions that divide subfamilies, such as recognition of an endogenous ligand in G protein-coupled receptors. The TEA method requires a definition of protein subfamilies as an input. However such definition is a challenging problem by itself, particularly because this definition is crucial for the following prediction of specificity positions. In Chapter 3, we automated the TEA method described in Chapter 2 by tracing the evolutionary pressure from the root to the branches of the phylogenetic tree. At each level of the tree, a TEA plot is produced to capture the signal of the evolutionary pressure. A consensus TEA-O plot is composed from the whole series of plots to provide a condensed representation. Positions related to functions that evolved early (conserved) or later (specificity) are close to the lower left or upper left corner of the TEA-O plot, respectively. This novel approach allows an unbiased, user-independent, analysis of residue relevance in a protein family. We tested the TEA-O method on a synthetic dataset as well as on __real__ data, i.e., LacI and GPCR datasets. The ROC plots for the real data showed that TEA-O works perfectly well on all datasets and much better than other considered methods such as evolutionary trace, SDPpred and TreeDet. While positions were treated independently from each other in Chapter 2 and 3 in predicting specificity positions, in Chapter 4 multi-RELIEF considers both sequence similarity and distance in 3D structure in the specificity scoring function. The multi-RELIEF method was developed based on RELIEF, a state-of-the-art Machine-Learning technique for feature weighting. It estimates the expected __local__ functional specificity of residues from an alignment divided in multiple classes. Optionally, 3D structure information is exploited by increasing the weight of residues that have high-weight neighbors. Using ROC curves over a large body of experimental reference data, we showed that multi-RELIEF identifies specificity residues for the seven test sets used. In addition, incorporating structural information improved the prediction for specificity of interaction with small molecules. Comparison of multi-RELIEF with four other state-of-the-art algorithms indicates its robustness and best overall performance. In Chapter 2, 3 and 4, we heavily relied on multiple sequence alignment to identify conserved and specificity positions. As mentioned before, the construction of such alignment is not self-evident. Following the principle of sequential pattern mining, in Chapter 5, we proposed a new algorithm that directly identifies frequent biologically meaningful patterns from unaligned sequences. Six algorithms were designed and implemented to mine three different pattern types from either one or two datasets using a pattern growth approach. We compared our approach to PRATT2 and TEIRESIAS in efficiency, completeness and the diversity of pattern types. Compared to PRATT2, our approach is faster, capable of processing large datasets and able to identify the so-called type III patterns. Our approach is comparable to TEIRESIAS in the discovery of the so-called type I patterns but has additional functionality such as mining the so-called type II and type III patterns and finding discriminating patterns between two datasets. From Chapter 2 to 5, we aimed to identify functional residues from either aligned or unaligned protein sequences. In Chapter 6, we introduce an alignment-independent procedure to cluster protein sequences, which may be used to predict protein function. Traditionally phylogeny reconstruction is usually based on multiple sequence alignment. The procedure can be computationally intensive and often requires manual adjustment, which may be particularly difficult for a set of deviating sequences. In cheminformatics, constructing a similarity tree of ligands is usually alignment free. Feature spaces are routine means to convert compounds into binary fingerprints. Then distances among compounds can be obtained and similarity trees are constructed via clustering techniques. We explored building feature spaces for phylogeny reconstruction either using the so-called k-mer method or via sequential pattern mining with additional filtering and combining operations. Satisfying trees were built from both approaches compared with alignment-based methods. We found that when k equals 3, the phylogenetic tree built from the k-mer fingerprints is as good as one of the alignment-based methods, in which PAM and Neighborhood joining are used for computing distance and constructing a tree, respectively (NJ-PAM). As for the sequential pattern mining approach, the quality of the phylogenetic tree is better than one of the alignment-based method (NJ-PAM), if we set the support value to 10% and used maximum patterns only as descriptors. Finally in Chapter 7, general conclusions about the research described in this thesis are drawn. They are supplemented with an outlook on further research lines. We are convinced that the described algorithms can be useful in, e.g., genomic analyses, and provide further ideas for novel algorithms in this respect.Leiden University, NWO (Horizon Breakthrough project 050-71-041) and the Dutch Top Institute Pharma (D1-105)UBL - phd migration 201