A dual-scale approach toward structure prediction of retinal proteins

Abstract

We propose a dual-scale approach to predict the native structures of retinal proteins (RPs) by combining coarse-grained (CG) Monte-Carlo simulations and all-atom (AA) molecular dynamics simulations to pack their transmembrane helices correctly. This approach has been applied to obtain the structures of five RPs, including bacteriorhodopsin (BR), halorhodopsin (HR), sensory rhodopsin I (SRI), sensory rhodopsin II (SRII), and (bovine) rhodopsin. The proposed CG model predicts a reasonably good structure of RPs in days using a desktop computer, which also gives clear physical picture for the packing, tilting, and orientation of transmembrane helices. A high-resolution protein structure is obtained from the AA molecular dynamics simulations by refining the predicted CG structure. The root mean square deviation in coordinates of backbone atoms from the X-ray structure is 1.89 Å for HR, 1.92 Å for SRII, 2.64 Å for BR, and 5.54 Å for rhodopsin. Reasonable predictions of HR structure can be obtained by this approach in the case of using predicted secondary structures with certain alignment error. Since the crystal structure of SRI is not available in the protein data bank, the predicted structure of SRI from our dual-scale approach is compared to that obtained from homology modeling.

Introduction

Membrane proteins (MPs) play key roles in living cells, such as ion channels, drug receptors, and information transfers (White and Wimley, 1999). Functionally normal MPs are vital to survival and their defects lead to many known diseases. The clinical importance of MPs is demonstrated by the fact that more than 50% of known drugs are targeting on MPs (Moreau and Huber, 1999), which are also responsible for the uptake, metabolism, and clearance of these pharmacologically active substances. Although analyses show that more than a quarter of proteins coded in genomes are MPs (Gerstein, 1998, Wallin and von Heijne, 1998, Krogh et al., 2001), due to difficulties in crystallizing MPs, less than 200 unique structures have been derived (data obtained from http://blanco.biomol.uci.edu/Membrane_Proteins_xtal.html). Therefore, there exist great incentives for computational and theoretical studies of MPs (Milik and Skolnick, 1992, Chen, 2000, Floriano et al., 2000, Chen and Chen, 2003, Dobbs et al., 2002, Kokubo and Okamoto, 2004, Ou et al., 2007).

Retinal proteins (RPs) include MPs found in the purple membrane of Halobacterium salinarium (Zheng and Herzfeld, 1992), each with different functions: bacteriorhodopsin (BR) is a proton pump (Pebay-Peyroula et al., 1997), halorhodopsin (HR) is a chloride pump (Kolbe et al., 2000), while sensory rhodopsin I (SRI) and sensory rhodopsin II (SRII) (Royant et al., 2001) are photosensoric proteins. The two ion pumps, BR and HR, convert light energy for the bacteria to synthesize ATPs. The two photosensors, SRI and SRII, direct the bacteria toward optimal light conditions and to avoid exposure to photooxidative conditions. RPs are the focus of much interest and have become a paradigm for MPs in general and transporters in particular (Oesterhelt and Stoeckenius, 1973, Oesterhelt, 1976, Henderson, 1977). Their structure and function have been analyzed in great detail using a variety of experimental techniques. Structurally, they have a topology of seven transmembrane (TM) helices arranged in two arcs, an inner one containing helices B, C, and D and an outer one comprising helices E, F, G, and A. Between helices B, C, F, and G, there is a TM pore, which accommodates a retinal to separate the extracellular half channel from the cytoplasmic half channel. Since the general structure of RPs has a relatively simple topology and has been studied extensively using various experimental techniques, they serves as excellent model systems for constructing a physical model to predict the structure and thermodynamics of MP folding, particularly for seven transmembrane (7TM) receptors.

The construction of a general model for MP structure prediction remains to be one of the great challenges (Bowie, 2005). A traditional approach for predicting MP structures is the homology modeling, which relies on the identification of one or more known protein structures likely to resemble the structure of the query sequence. Many investigators have used the structure of BR or bovine rhodopsin as a template to build models of 7TM receptors (Pardo et al., 1992, Davies et al., 1996, Baldwin, 1998, Herzyk and Hubbard, 1998). However, due to the low sequence identity (less than 30%) between most 7TM receptors and rhodopsin (or BR), it should be noted that their arrangement of TM helices could be very different, which leads to an incorrect prediction of their structures. In fact, the average sequence identity of 99% of human 7TM receptors to bovine rhodopsin is lower than 20% (Archer et al., 2003). Another template based method for structure prediction is the threading method (Zhang et al., 2006, Yarov-Yarovoy et al., 2006), whose success depends on the completeness of the library of solved structures in the protein library. An alternative method for structure prediction of MPs without using homology has been developed by Goddard III and coworkers (Trabanino et al., 2004, Kalani et al., 2004). The disadvantage of this method is that an experimental electron density map of MPs is required as an initial input. The ab initio approach is based on the global minimization of a physical potential energy function, which thus far has had limited success for small proteins (Simons et al., 1997, Liwo et al., 1999, Zhang et al., 2003). We note that most previous studies of MP structures (such as homology modeling and threading) show little interest on the thermodynamic hypothesis of protein folding, which is the main focus of this article. In addition, most previous methods in studying the structure of large proteins often require extensive computation. Our study intends to obtain a reasonably good structure within limited computational time by a dual-scale approach.

It has been suggested that folding of many integral MPs can be understood based on the two-stage model: Independently stable helices are formed in lipid bilayers in the first stage, and the helices interact with others to form a functional MP in the second stage (Popot and Engelman, 1990, Popot and Engelman, 2000, Booth and Curran, 1999, Pappu et al., 1999). In this article, based on the two-stage model, we demonstrate the feasibility of predicting the native structure of RPs by a dual-scale approach of computer simulations. Five RPs, including BR, HR, SRI, SRII, and (bovine) rhodopsin are tested without using their 3D structure information in the protein data bank (PDB) (Berman et al., 2000). Their secondary structures are considered to be known here for simplicity, and the derivation of such information is demonstrated in the Supplementary information (supplementary information, 2008).¹ Using possible secondary structures of HR from our secondary structure prediction algorithm, predicted 3D structures of HR are compared to its crystal structure to further validate this dual-scale approach toward a more general structure prediction algorithm. Since most MPs contain large number of amino acids (the average length of MPs of known structures is about 400 amino acids), it is extremely difficult to study their folded structures and folding dynamics by computer simulations at atomic resolution. Therefore, we first construct a coarse-grained (CG) protein model for helical bundle MPs (HBMPs), which includes most dominant physical interactions of the system. A detailed description of our CG model is given in Section 2. The lowest-energy state structure of HBMPs can be identified using the parallel tempering (PT) algorithm (Hansmann, 1997), as described in Section 3. PT is an efficient algorithm in finding the ground state, but the dynamic information of MP folding is missing. On the other hand, Monte-Carlo (MC) simulations are not so efficient to find the ground state, but can be used to obtain thermodynamic information of MP folding. In Section 4, we delineate our dual-scale simulation methods. At a low resolution scale, CG MC simulations are performed to find the ground state structure identified by PT simulations and to obtain folding dynamics from a random initial structure. At a high resolution scale, an all atom (AA) representation of our lowest energy CG structure of RPs is refined to give the predicted 3D structures of RPs using AA molecular dynamics (MD) simulations. In Section 5, we discuss results from our computer simulations of four RPs. The predicted packing, tilting, and orientation of helices are found to be consistent with experimental data by comparing the native structures of RPs with our predicted structures. The root mean square deviation (RMSD) in the tilting angles is 4.8 degrees for HR and is 3.2 degrees for SRII, while RMSD in the orientation angles is 6.8 degrees for HR and 23 degrees for SRII. Section 6 gives our conclusion.