Functionally Interacting Proteins : Analyses And Prediction

Abstract

Functional interaction of proteins is a broad term encompassing many different types of associations that are observed amongst proteins. It includes direct non-covalent interactions where the interacting proteins physically associate using an interface. There are also many protein-protein interactions where the proteins concerned are not involved in direct physical interactions but affect each other’s functions. Central focus of this thesis is to understand the various aspects of functionally interacting proteins. Chapter 1 of this thesis provides an introduction to functional interactions between proteins and discusses the key developments available in the literature. This chapter discusses the different types of functional associations observed commonly between proteins. Various approaches developed over time to elucidate such interactions have also been discussed. This chapter highlights how functional interactions between proteins have been helpful in understanding different cellular processes such as organization of metabolic pathways. The chapter emphasizes the importance of functional interactions between proteins, providing a motivation for development of methods with enhanced accuracy and sensitivity for the prediction of functional interactions. In this thesis, domain families which are found to co-exist in multidomain proteins have been used to understand and subsequently predict functional associations amongst proteins. Domains in proteins typically serve as modules associated with specific functions. There exist proteins with a single domain which describes the entire function of a protein, while there also exist proteins containing multiple domains, where various domains in unison describe the complete function of the multidomain protein. Therefore, by virtue of “guilt by association” domain families found together in multidomain proteins are functionally linked. This forms the basic premise for understanding functional association amongst proteins and is explained in great detail in the Introduction chapter. Using domain families which co-occur in multidomain proteins as the basis for functional association has many merits. First, as stated before, constituent domain families act as effective descriptors of function(s) of proteins. For example, members of SH3 domain family mediate protein-protein interactions by binding to regions with polyproline conformation irrespective of the multidomain protein in which it occurs. Thus, studies of domain families co-existing in multidomain proteins act as an accurate resource of functional associations between proteins. Also, assignment of domains to a protein relies on homology detection which has achieved a high level of reliability, thus, resulting in reasonably accurate prediction of functions. Such approaches enable exhaustive coverage of many diverse proteins including many multidomain proteins leading to detection of large numbers of functional associations between domains of multidomain proteins. Given the advantages attributed to functionally linked domain families in further understanding of functional associations, it is imperative to exhaustively enumerate all possible pairs of functionally linked domain families in multidomain proteins and study their various properties. This aspect is covered in the second chapter of the thesis. In the second chapter, analysis of domain families which co-occur in multidomain proteins, termed as 'tethered domain families', has been reported. For this analysis, a large dataset of multidomain proteins was considered from a diverse set of fully sequenced genomes from many eukaryotic and prokaryotic organisms. In every multidomain protein, all possible pairs of unique domain family pairs have been considered and they are assumed to be under the same functional/evolutionary constraint. Thus, from the entire dataset of multidomain proteins, all possible pairs of tethered domain families are obtained. For a given domain family, the number of other uniquely tethered families is referred to as the tethering number of a domain family. Therefore, tethering number of a domain family is an indicator of the diverse functional contexts in which a particular domain family is involved. Further analysis was carried out to understand various other attributes of domain families and its relation to tethering number. The results are summarized in the following points: 1) Distribution of tethering numbers of domain families in the entire dataset is found to be highly heterogeneous. Nearly 88% of domain families (10783 out of 12249 domain families) have tethering number of 10 or less and only 78 domain families show more than 100 unique associations. Further analysis reveals bias in functions of families showing high and low tethering numbers. The domain families with high tethering numbers are involved in processes such as signaling and protein-protein interactions. The domain families with low tethering numbers are often found to be involved in metabolic processes. 2) Differences are also observed in the type of organisms containing the domain families and their tethering numbers. Typically, domain families with high tethering numbers are ubiquitously found across almost all the kingdoms of life. In contrast, most of the domain families exclusively found in a kingdom have low tethering numbers. Furthermore, for the ubiquitously occurring domain families with high tethering numbers, the number of associations made and the type of associations are not strictly conserved across the kingdoms. Thus, the tethering preferences of such domain families vary across the kingdoms depending on their function. For instance, the protein kinase domain family which is a key regulator of signaling processes in eukaryotes, has a high tethering number in eukaryotes (270), and low tethering number in prokaryotes (96). 3) Tethering number of domain families is found to be correlated with the number of members (population) comprising a family. A Pearson correlation coefficient of 0.78 at a p-value ≤0.001 is obtained for the correlation between tethering number of domain families and their population. 4) Tethering numbers of domain families are also found to be well correlated with sequence and functional diversity within families. Thus, domain families with high tethering numbers comprise of members showing diversity in both sequence and functions. Thus, the work presented in second chapter provides a framework for understanding the tethering preferences of domain families. The use of tethered domain families to identify functional association amongst proteins is the central theme of third and fourth chapters of this thesis. The use of tethered domain families for the prediction of functionally interacting proteins originates from the initial idea of “Rosetta stone” approach, which was proposed by Ouzounis and coworkers and Eisenberg and coworkers in 1999. Rosetta stone approach demonstrated the use of fused genes in predicting functional interaction. It stems from the observation that in many organisms, genes corresponding to proteins acting in a metabolic pathway are found fused in another organism. Thus, enumeration of 'fused genes' in a template database could provide a good basis for prediction of functionally interacting proteins in target organisms in which the homologous genes are not found to be fused. The method has been shown, by others, to work quite effectively in prokaryotes, especially in the identification of interactions between metabolic proteins. Chapter 3 of this thesis explores the idea of “Rosetta stones” at the level of domain families, by considering tethered domain families as analogs to the fused genes. In this analysis, tethered domain families derived from multidomain proteins comprises the template dataset. If members of two domain families occurring in a multidomain protein are found to occur independently in two different proteins in the target organism then an interaction is predicted between these two proteins (collection of such predicted interactions is henceforth referred as TEDIP database, Tethered Domain-based Interaction Prediction). During this analysis, care is taken such that none of the proteins in the template dataset belongs to the target organisms. The entire analysis has been conducted on 6 model organisms which act as the target dataset where functional interactions between proteins are predicted. The effectiveness of tethered domain families in functional interaction prediction is compared with two other datasets 1) all experimentally known interactions and 2) interactions predicted on the basis of their homology with interacting domain families with known structure. Subsequently, an attempt has been made to answer these questions: 1) how effective is the information on tethered domain families in predicting functional linkages amongst proteins operating in pathways in eukaryotic organisms? 2) what is the false positive rate of the predictions? The above mentioned datasets show very little overlap in the coverage of functional interactions. This is largely attributed to insufficient sampling and inherent bias existing in each of the methods. The TEDIP datasets in the six organisms led to an average three-fold more functional interaction predictions in cellular pathways than the other two datasets. Nearly 90% of the predicted interactions derived from tethered domain families are amongst proteins across different pathways. In yeast, more than 60% of such interactions were found to be overlapping with a recent large scale genetic interaction screen based on synthetic lethality especially performed for metabolic proteins, thus establishing the effectiveness of this approach in understanding pathway crosstalk. Along with efficacy in identifying functional interactions, an assessment based on co-localization, co-expression and overall functional similarity based on Gene Ontology (GO) terms was carried out. It was found that the TEDIP predictions and experimentally found interactions show poor correspondence with co-expression and co-localization data (10% and 20% respectively for the two methods). Additionally, it was found that functional similarity between predicted interacting proteins in TEDIP dataset is low (5%) and is comparable to experimentally known interactions that shows 10% similarity in functions based on a scoring function for GO term similarity. From Chapter 3, it was concluded that the use of tethered domain families is effective in exhaustive enumeration of functionally associated proteins. However, the low co-expression and functional similarity measures are a cause for concern. On the one hand, co-expression and GO functional similarity have been found to be weak predictors of functional interactions, explaining the low values obtained for both predictions in the TEDIP datasets and experimentally known interactions. On the other hand, the poorer values shown for predictions in the TEDIP datasets suggest that further improvement in prediction accuracy is possible. Chapter 4 explores the use of machine learning in improving the accuracy of functional interaction prediction based on TEDIP dataset. In Chapter 4, two distinct machine learning approaches have been employed on a training dataset derived exclusively from yeast. Since the objective of the work is to improve the accuracy of prediction of functional interactions, the GO based functional similarity measures have been used to define positive and negative datasets. Thus, in the training dataset, positive interactions comprises of protein pairs which show high GO similarity in functions as defined in chapter 3 and 10% of this data overlaps with experimentally known interactions, while the negative dataset consists of protein pairs with no or insignificant similarity in their functions and additionally do not show similarity to any experimentally known interactions. Two machine learning approaches, namely Support vector machine (SVM) and Random forest, have been used on this training dataset. Use of two distinct approaches helps in addressing the weakness, if any, of these methods. Fourteen carefully chosen features have been utilized during the training process to aid in the process of distinguishing potentially correctly predicted interactions from incorrect predictions. Out of 14 features, some of the features chosen for the analysis are involved in quantifying the extent of similarity between the template proteins containing the fused domain families and the target protein pairs predicted to interact. The analysis also incorporates graph theory based parameters which are derived from a domain family based graph. In such a graph, each of the domain families which are involved in forming multidomain proteins represents the nodes and an edge is constructed between domain families which are found to co-exist in at least one multidomain protein. Graph theory based parameters such as clustering coefficient, degree and topological overlap have been employed. These are useful in down weighting appropriately the domain family pairs showing large number of associations which are expected to be promiscuous in their functions. These features also enable in identifying domain family pairs which are functionally related. Apart from the above mentioned features, coevolution and phylogenetic profiling of tethered domain families is also utilized to identify functionally related domain family pairs. Utilizing all these features in training, the machine learning approach yielded an accuracy of 94% using SVM and 92% using Random forest against the training data. Furthermore, the importance of using all these features has been addressed by performing principle component analysis, training both SVM and Random forest by removing one feature at a time and by quantifying the sensitivity by using only one feature. All of these suggest that the features used provide non-redundant information and contributed significantly to the classification. The models so generated were finally used on all the predicted functional interactions after the removal of the training dataset in yeast. The true positives observed were 56% using SVM and 63% using Random forest with around 80% of the interactions common between the two methods. Further analysis has been carried out on these interactions by first imparting a confidence score to these interactions using support vector regression that provides a probabilistic measure for SVM classification. Based on a cutoff of 0.5, 62455 interactions in total were termed as high confidence interactions. Further analysis was carried out for the high confidence interactions. Out of these, in 2855 interactions, both the proteins predicted to interact could be associated with a pathway in KEGG database. In-depth case studies have been performed on this dataset of 2855 interactions. Literature mining suggested that many known cross-pathway interactions such as between TCA and glycolysis are captured as high confidence interactions using TEDIP dataset. A few other case studies of high confidence interactions with supporting literature evidence are also presented in the chapter. These predictions could further aid in experimental characterization of pathway cross-talk between important metabolic and signaling pathways. So far, the thesis discussed analyses involving functional interactions and their prediction. In the subsequent chapters, analyses pertaining to two different types of functional interactions are discussed. Chapters 5 and 6 involve analyses incorporating metabolic proteins in diverse pathways in the pathogenic organism Plasmodium falciparum. Chapter 5 attempts to improve the coverage of the repertoire of metabolic proteins in P.falciparum while in Chapter 6 interactions and pathways prevalent in different stages in the life cycle of the parasite are deciphered and discussed. Apart from functionally interacting proteins in metabolic pathways, physically and transiently interacting proteins have been analyzed and discussed in Chapters 7 and 8. In Chapter 5, metabolic proteins participating in pathways in Plasmodium falciparum have been analyzed. P.falciparum is the causative agent of malaria, a disease which affects large populations in the subtropical regions. P.falciparum genome is atypical and is rich in Adenine/Thymine pairs, and there is presence of large stretches of amino acid repeats encoded in protein coding regions. Various sequence-related features of P.falciparum proteins when compared with those of other organisms show extensive divergence. All of these have made reliable function prediction, by homology to other proteins with known functions, daunting. Like other proteins in P.falciparum, metabolic proteins have also diverged significantly from their functional counterparts in model eukaryotes such as yeast. Metabolic pathways play an important role in the survival of the organism and hence are amenable towards the identification of proteins susceptible to drugs, thereby combating pathogenesis. Chapter 5 of the thesis aims at furthering knowledge pertaining to metabolic proteins by first quantifying the extent of divergence observed in the already characterized metabolic proteins. This knowledge is further used in identification of potential metabolic proteins which are not identified as proteins involved in metabolic pathways by other annotation efforts undertaken for P.falciparum. In the first part of the chapter, the extent of divergence in the sequences of metabolic proteins in P.falciparum has been determined by comparing the P.falciparum proteins with their functional counterparts from 34 completely sequenced unicellular eukaryotic organisms. Comparison of domain architectures between the P.falciparum proteins with their functional counterparts reveals that in nearly 54% of metabolic pathways, proteins show nearly the same domain architecture as the other functional counterparts. Inversion, deletion and duplication of domains are observed in rest of the proteins. Further analysis reveals that P.falciparum proteins are longer than their functional counterparts. It was also observed in nearly 15% of the cases, the domains are characterized by the presence of large non-conserved or plasmodium genus specific inserts within the domain assigned regions. There is also prevalence of unassigned regions in the N- and C- terminal regions in P.falciparum proteins when compared with their functional counterparts. Finally, it was also observed that metabolic proteins of P.falciparum show significantly low sequence similarity when compared with other functional counterparts. From this analysis, it can be clearly seen that metabolic proteins of P.falciparum have significantly diverged from such proteins in other organisms, thus making function prediction by homology very difficult. There are several steps in metabolic pathways in P.falciparum which are expected to be active based on experimental analysis. However, some of these proteins with expected functions have not been identified so far. One of the reasons for this apparent incompleteness is the high divergence observed in the metabolic proteins of P. falciparum. To overcome this limitation, in the second part of the chapter, a sensitive approach based on domain family assignment (MulPSSM), developed in-house, has been used to identify proteins which are potentially involved in metabolic pathways. The approach is based on reverse PSI–BLAST, where multiple sequence profiles for each family are used to search against sequence databases. This approach has been shown to be better or at-par with other remote homology detection procedures. Using this approach, 15 P. falciparum proteins have been identified which can potentially function as metabolic proteins and were not characterized in P.falciparum so far. All the proteins identified by the approach show low sequence similarity to other well characterized proteins and contain significant fractions of unassigned regions thus, making function recognition non-trivial. Supporting literature and other data is provided to demonstrate the robustness of the homology-based annotation of the identified pathway proteins. Chapter 6 is an analysis of the dynamic changes occurring in the metabolic network of P.falciparum during its life cycle. In this chapter, two aspects of P. falciparum metabolic proteins have been integrated and analyzed. First, the dataset of protein-protein interactions derived from experimental studies and second, the datasets of microarray analysis providing information on stage specific expression of P. falciparum genes corresponding to the metabolic proteins. As a first step, protein-protein interaction information for the metabolic proteins was gathered. A total of 810 interactions have been obtained, where one or both proteins are involved in a pathway. Subsequently, these interactions were compared with 14070 interactions involving metabolic proteins from free-living and non-pathogenic unicellular eukaryote yeast. Comparison across the two organisms shows wide discrepancy in the number of proteins involved in interactions and also the pathways in which they participate. Out of the 810 interactions in P.falciparum, 173 are found uniquely in plasmodium where both or one of the protein have no identifiable homolog in yeast. Insufficient sampling of interactions made by proteins in P.falciparum in comparison to yeast, is one of the reasons for the observed discrepancy. However, the differences due to the parasitic lifestyle of P.falciparum could also be a potential reason. Further analysis of the protein-protein interactions by the metabolic proteins revealed that a large fraction of interactions are made between a metabolic protein and a non-metabolic protein. For instance, interaction observed between glycolytic protein phospoglycerate kinase with MAP kinase. This trend is observed in both plasmodium and yeast where 65% and 77% of the interactions, respectively, involve proteins not directly participating in metabolic pathways. Further, interactions between proteins belonging to different pathways and lastly, interactions between proteins in the same pathway are uncovered. All of these interactions depict the different modes by which metabolic pathways are regulated through protein-protein interactions. Another aspect explored in this analysis is the stage specific expression of genes encoding these metabolic proteins. The analysis is especially relevant in the parasite because its entire life cycle is divided into seven distinct stages. Upon integrating the protein-protein interactions with the gene expression data, it became apparent that the trophozoite, schizont and gametocyte stages show large fractions of co-expressed genes encoding proteins involved in protein-protein interactions within metabolic pathways. The high preponderance of co-expressed genes encoding for interacting protein pairs in these stages is also consistent with metabolic requirement of plasmodium in the various stages. Glycolytic pathway is central to energy production in the parasite and is discussed at length in this chapter. Members of this pathway are involved in interactions with other glycolytic proteins (9 such interactions), they also interact with proteins involved in other pathways (30 interactions) and with proteins not involved directly in any metabolic pathway (75 interactions). Nearly 70% of the interactions made by the glycolytic proteins are encoded by genes found to be co-expressed across the various stages. Integration of gene expression data along with protein-protein interaction information for metabolic pathways such as the glycolytic pathway thus, highlights the complex mode of regulation underlying this pathway. The analysis carried out in this chapter emphasizes on the intricacies involved in the regulation of metabolic proteins in P.falciparum. Chapter 7 describes an in-depth analysis carried out to understand the basis for interaction specificity between small monomeric GTPases and their regulators, the Guanine nucleotide Exchange Factors (GEFs). Monomeric GTPases are involved in binding to guanine nucleotide. These proteins can bind to both GTP and GDP. However, transition from GDP bound to GTP bound form occurs with large conformational changes and requires binding of the GEFs. The conformational changes that arise due to the nucleotide exchange are required for the GTPases to bind to its various effectors. For the analysis carried out in Chapter 7, GTPases belonging to the Ras superfamily have been considered. The superfamily is further subdivided into 5 distinct families based on their functions. The 5 families are Ras, Ran, Rab, Arf and Rho. Members belonging to each of these families are involved in a wide array of cellular processes such as signaling and cytoskeletal remodeling. Members of each of these GTPase families bind to structurally distinct GEFs, and in some cases, multiple GEFs are involved in nucleotide exchange within a family. It is intriguing therefore, to understand how GTPases belonging to the same structural family maintain specificity across the highly dissimilar GEFs and this forms the main objective of this analysis. So far, 13 distinct complexes between GTPases and their cognate GEFs have been solved using X-ray crystallography. This set of structural complexes forms the starting point of the analysis. As a first step, pairwise structural comparison of the interfaces has made between various pairs of complex structures. Based on these comparisons, it is apparent that most of the interfaces in the GTPase and GEF complexes comprise of residue positions which are topologically not equivalent suggesting different modes of binding across these complexes. Further analysis was carried out to probe the extent of specificity underlying these complexes. This is achieved by determining interface residues which are found to be conserved in a family specific manner. Such residue positions have been obtained by using a statistically robust algorithm Contrast Hierarchical Alignment and Interaction Network (CHAIN) that extracts sequence patterns most distinguishing two sets of homologous sequences. The analysis indicated the presence of family specific residues at the GTPase and GEF interface. Such residues could be implicated in maintaining the specific interactions between the GTPases and the GEFs. The robustness in the specificity of the interactions was further interrogated by providing an energetic basis to the specificity in the interactions mediated by the cognate GTPases and the GEFs and also understanding how crosstalk is prevented across the non-cognate complexes. For each of the 13 cognate complexes, empirical interaction energies have been estimated using FoldX. The interaction energy is compared to non-cognate complexes which are obtained by swapping the interface residues of the cognate GTPase with the non-cognate GTPase residues. For most of the complexes, it was observed that the interaction energies for the cognate complexes are much lower than the non-cognate complexes. Energy values across the non-cognate complexes are usually indicative of reduced stability, thereby precluding such interactions from occurring. Such large energy differences between cognate and non-cognate interactions arise due to drastic substitutions at the interface patch due to difference in the charge or other stereochemical aspects of the amino acids. Both evolutionary and energy based analysis indicates the presence and importance of few family specific residues in the cognate complexes and also the presence of unfavorable residues in the non-cognate complexes thus preventing crosstalk. However, apart from changes at the interfaces, many positions outside the interface also undergo changes across the various homologs within the same family/subfamily of GTPase. Coevolutionary analysis of GTPase and GEFs from multiple eukaryotic organisms has been carried out in these complexes and it was observed that most of the coevolving positions are not found at the interface. Many of these residue positions are near the active site or near the interface. Identification of such coevolving positions, where residue variations in the GTPase are strongly coupled to the GEF, may provide initial clues to the possible allosteric path adopted in connecting the binding of GEF to the vast structural changes observed during GTP exchange in GTPases. Thus, the analysis provides a comprehensive framework to understand how interaction specificity has evolved between the GTPase and GEF complexes. Chapter 8 discusses another example of transient protein-protein interaction observed between proteins implicated in signaling process in Dictyostelium discoideum. The work reported in this chapter was carried out in collaboration with Prof. Nanjundaiah and coworkers from Molecular Reproduction and Developmental Genetics department, Indian Institute of Science. All the experimental analyses mentioned in this chapter were carried out by Prof. Nanjundaiah and coworkers and the author carried out all the computational analysis. Experimental analysis indicated the presence of a ribosomal protein S4 in D. discoideum which mediates interactions with CDC24 and CDC42. The protein is speculated to be a functional analog of yeast scaffolding protein Bem1. However, the exact structural and sequence features of the protein which can accommodate its non-ribosomal function as a scaffold by mediating protein-protein interactions are not clearly understood. With the aid of structural modeling, a 3-D structure was generated for the C-terminal regions of D. discoideum protein S4. The modeled structure, as in the template used for modelling, resembled the fold of SH3 domain which has been shown to be involved in protein-protein interactions. Structural and sequence analyses were carried out to evaluate the potential mode by which interactions could be mediated by this protein. The hypothesis generated was further corroborated by experimental analysis. Thus, both experimental and computational analysis provide evidence for the functional role of the ribosomal protein S4 from Dictyostelium discoideum as a scaffold. Chapter 9 summarizes the conclusions reached in various chapters of the thesis. The thesis embodies analyses probing various aspects of functional interactions between proteins. A frame work has been provided to elucidate functional interactions using tethered domain families in multidomain proteins. Further, the role of these functional interactions have been explored in different scenarios by exhaustively analyzing metabolic proteins and their regulation in pathogenic organism Plasmodium falciparum and by also analyzing two distinct types of transient protein-protein interactions.