Towards 3D Modeling of Interacting TM Helix Pairs Based on Classification of Helix Pair Sequence

Abstract

Spatial structures of transmembrane proteins are difficult to obtain either experimentally or by computational methods. Recognition of helix-helix contacts conformations, which provide structural skeleton of many transmembrane proteins, is essential in the modeling. Majority of helix-helix interactions in transmembrane proteins can be accurately clustered into a few classes on the basis of their 3D shape. We propose a Stochastic Context Free Grammars framework, combined with evolutionary algorithm, to represent sequence level features of these classes. The descriptors were tested using independent test sets and typically achieved the areas under ROC curves 0.60-0.70; some reached 0.77.

Probabilistic grammatical model for helix‐helix contact site classification

“…Genetic Algorithms (GA) [37,42,109]‐ [115] or Genetic Programming (GP) [116,117]). A standard EM‐based approach using the Inside‐Outside algorithm [118] is highly sensitive to the choice of initial parameters and usually results in complex grammars, which are difficult to interpret [113].…”

“…Practically, two main approaches are used for learning probabilities of grammar rules: Expectation‐Maximization algorithms (EM) [ 12 , 13 , 15 , 108 ] and evolutionary methods (e.g. Genetic Algorithms (GA) [ 37 , 42 , 109 ]‐ [ 115 ] or Genetic Programming (GP) [ 116 , 117 ]). A standard EM‐based approach using the Inside‐Outside algorithm [ 118 ] is highly sensitive to the choice of initial parameters and usually results in complex grammars, which are difficult to interpret [ 113 ].…”

“…In the second layer, a generic set of grammar rules is provided. Preferably, the rules are constrained based on expert knowledge of general dependencies between residues in protein sequences of interest [ 37 , 42 ]. The main purpose of this step is to limit the GA search space in the next layer.…”

“…Based on our preliminary study [ 42 ] the properties of normalized van der Waals volume [ 125 ] and information value for accessibility with average fraction of 35% (called later accessibility ) [ 126 ] were chosen as the basis for the grammars. They are representatives of physiochemical properties and hydrophobicity categories in AAindex, respectively [ 122 ].…”

“…Moreover, probabilistic context‐free grammars have been applied to analysis of ligand binding sites [37] and antimicrobial peptides [38] showing that produced parse trees could provide biological insight into the modeled structures. Finally, whatever their complexity level, grammatical models have proved to be useful for protein structure annotation [37,39] and prediction [32,35,40]‐ [42] and to designing synthetic peptides [38,43]. …”