4 Protein structure prediction based on structural class annotations
4.2 Related Work and Motivation
All fragment-based protein structure prediction methods described in the literature review are sequence-dependent since fragments are extracted from templates
77
selected using sequence-based information. However, it has also been proposed to create databases of fragment models, which are chosen independently of their amino acid compositions to constitute conformation assemblies (Baeten et al., 2008; Kolodny, Koehl, Guibas, & Levitt, 2002; Vanhee et al., 2011). Fragments are only defined by their ‘shape’ and substituted into the query sequence at positions where amino acids can conform to those shapes. Although such techniques have not been competitive against sequence-dependent predictors, they have shown interesting results in modelling loops (Kolodny et al., 2002; Vanhee et al., 2011).
A promising approach has been the integration of spatial constraints within standard fragment-based systems. So far, this has been performed using predicted contact maps – a matrix that represents the approximate value of the distance between each pair of amino acids (Kosciolek & Jones, 2014; Mao, Tejero, Baker, & Montelione, 2014; M. Michel et al., 2014; Mirco Michel, Menéndez Hurtado, Uziela, & Elofsson, 2017; Ovchinnikov et al., 2017; Ovchinnikov, Kim, et al., 2016; Ovchinnikov, Park, et al., 2016; Ramelot et al., 2009; S. Wu, Szilagyi, & Zhang, 2011). However, since accurate prediction of a contact map currently relies on the availability of a relatively large protein family (ideally more than 1000 homologous protein sequences) (Skwark, Raimondi, Michel, & Elofsson, 2014), their usage is not suitable for all protein targets. Moreover, low quality contact maps lead invariably to poor models, since incorrect constraints prevent appropriate exploration of the native structure conformation space. As a conclusion, there is a need for the design of alternative constraints to fragment- based protein structure prediction.
Although fragment assembly methods have been ranked as the most successful techniques for free-modelling predictions, yet many issues remain and need to be addressed (Dill & MacCallum, 2012). First, successful attempts to produce accurate conformations have been mainly restricted to targets whose lengths are less than 100 residues (Xu & Zhang, 2012) due to the enormous search space, even though protein fragments are used instead of individual amino acids. Second, even for small proteins, processing times are prohibitive for the typical user; Rosetta, for instance, needs on average 150 CPU days per target (S. Wu et al., 2007). Third, despite effective use of Monte Carlo simulations along with fragment replacements, a structure’s global energy minimum is likely to be missed. In addition, the design of the most appropriate force field is still open research question as current ones often fail to recognise native structures (Jooyoung Lee et al., 2009; Xu & Zhang, 2012). Finally, the large number of
78
decoys produced by most of those methods constitutes an additional barrier to identification of native-like conformations since there is no straightforward correspondence between free energy values and similarity to a native structure. As a consequence, design of model quality assessment programs has become an active research area of its own (R. Cao, Wang, Wang, & Cheng, 2014; Konopka et al., 2012).
4.3 Protein Structural Class Classifications
Categorisation of protein structural classes was first introduced by Levitt and Chothia in 1976 (Levitt & Chothia, 1976) when proteins were found to belong to one of four classes: (1) all-alpha proteins; (2) all-beta proteins; (3) alpha + beta proteins where beta strands tend to be segregated and are likely to form antiparallel beta sheets; (4) alpha / beta proteins where alpha helices and beta strands are rather mixed and therefore polypeptide chains are expected to contain parallel beta sheets. Two decades later, Chothia and co-workers established a manually curated online database the Structural Classification Of Proteins (SCOP) (Murzin, Brenner, Hubbard, & Chothia, 1995). The first level of its hierarchy was initially divided into five classes: the original four and a ‘multi-domain’ class. Later on two further classes were added, namely ‘Membrane and cell surface proteins and peptides’ and ‘Small proteins’ (SP) (Lo Conte, Brenner, Hubbard, Chothia, & Murzin, 2002). However, currently only the “small proteins” class exists in the database besides the original four (Andreeva et al., 2014).
Two years after the initial release of SCOP, an alternative database, CATH – named after the first four levels of its hierarchy: Class, Architecture, Topology and Homology – was established (Orengo et al., 1997). Since this showed that there was no clear separation between alpha + beta and alpha/beta proteins (Berman et al., 2000; Michie, Orengo, & Thornton, 1996), CATH has been based on only 4 classes: (1) mostly alpha; (2) mostly beta; (3) alpha beta and (4) Few secondary structures (FSS) (Sillitoe et al., 2015). Despite differences between SCOP and CATH, a comparative study (Csaba, Birzele, & Zimmer, 2009) has shown the top level of both hierarchies, i.e. ‘Class’, is relatively consistent in comparison to the remaining levels since it is defined according to high level structural features.
Assigning a protein structure to a specific class is not trivial. Whereas CATH uses an automated and explicit method (Michie et al., 1996), SCOP relies on manual inspection. Except for discrimination between ‘alpha/beta’ and ‘alpha + beta’, the critical criterion is the percentage of helix and strand content of the protein structure.
79
Many studies have been conducted to establish the best thresholds for classification, which led to a variety of values (K.-C. Chou, 1995; K.-C. Chou, Liu, Maggiora, & Zhang, 1998; P. Chou, 1989; Eisenhaber, Frömmel, & Argos, 1996; Klein & Delisi, 1986; Kneller, Cohen, & Langridge, 1990; Kurgan, Zhang, Zhang, Shen, & Ruan, 2008; Nakashima, Nishikawa, & Ooi, 1986). Eventually, a thorough comparative study established that the 15% helix and 10% strand thresholds are optimal – those are used by CATH - see Figure 4.1- even if overlapping regions exist between adjacent classes (L. A. Kurgan et al., 2008). Some instances of disagreement between CATH and SCOP structural class classification are mainly a result of the disagreement of domain classification in the first place, especially between ‘alpha+beta’ and ‘mainly beta’. This is due to two causes: the similarity of beta sheets in both classes and whether an alpha helix can be considered a part of the domain or simply a peripheral. Example of such a disagreement between SCOP and CATH is the haemagglutinin (PDBID: 1HGG). Whilst SCOP considers it as mainly-beta ignoring a helical part, CATH treats the whole conformation as one domain and classifies it under alpha-beta (Hadley & Jones, 1999). It is worth noting that CATH employs the distance between various secondary structure as a secondary criterion for classification to cope with this problem. Based on certain thresholds (H-H: 8Å, H-E: 10Å and E-E: 21Å), a secondary structure element can be considered then whether it is a part of the folding unit or not.
Figure 4.1: Scatter plot of helix and strand content percentages (X-axis and Y-axis respectively) for a large set of proteins classified as either all-alpha or all-beta classes. Taken from (Kurgan, Zhang, et al., 2008).
80
Since knowledge of a protein’s structural class based on its sequence may reveal crucial information concerning folding types and functions (K.-C. Chou, 2005b; K.-C. Chou & Zhang, 1995) and can be considered as a first step towards solving the structure prediction problem, sequence based class prediction has become an active research area (K.-C. Chou, 2011). Proposed approaches take advantage of either 1) machine learning techniques such as Support Vector Machines (SVM) (Anand, Pugalenthi, & Suganthan, 2008; Dehzangi, Paliwal, Lyons, Sharma, & Sattar, 2014; Hayat & Khan, 2012a), Artificial Neural Networks (Jahandideh, Abdolmaleki, Jahandideh, & Asadabadi, 2007), rough sets (Y. Cao et al., 2006), bagging (Dong, Yuan, & Cai, 2006), ensembles (Chen, Kurgan, & Ruan, 2008; Dehzangi, Paliwal, Sharma, Dehzangi, & Sattar, 2013; Hayat, Khan, & Yeasin, 2012; J.-Y. Yang, Peng, & Chen, 2010) and Meta-Classifiers (Cai, Feng, Lu, & Chou, 2006; Feng, Cai, & Chou, 2005); or 2) features that reveal class- related information like physiochemical-based information (Dehzangi et al., 2013; Z.-C. Li et al., 2008), pseudo amino acid composition (K.-C. Chou, 2000; Y.-S. Ding, Zhang, & Chou, 2007), amino acid sequence reverse encoding (Deschavanne & Tufféry, 2008; Mizianty & Kurgan, 2009), Position Specific Scoring Matrix (PPSM) profile (Hayat & Khan, 2012b) and structural based information including secondary structure prediction (Jones, 1999; Kurgan & Chen, 2007; Kurgan, Zhang, et al., 2008; Tian Liu & Jia, 2010). Detailed reviews can be found in (K.-C. Chou, 2005a; Kurgan & Homaeian, 2006). Although state-of-the-art tools, including SCPRED (Kurgan, Cios, & Chen, 2008), MODAS (Mizianty & Kurgan, 2009), RKS-PPSC (J.-Y. Yang et al., 2010), PSSS-PSSM (S. Ding, Li, Shi, & Yan, 2014), AADP-PSSM (Taigang Liu, Zheng, & Wang, 2010), SCEC (Chen et al., 2008), AATP (S. Zhang, Ye, & Yuan, 2012), AAC- PSSM-AC (Taigang Liu, Geng, Zheng, Li, & Wang, 2012) and PSSP-RFE (L. Li et al., 2014) report overall accuracy up to 90%, challenges remain, in particular with proteins with low sequence similarity and discrimination between alpha/beta versus alpha + beta classes (S. Ding et al., 2014). It is worth noting that most tools only deal with the four original SCOP classes which comprise around 90% of annotated domains (K.-C. Chou, 2005a).