163 Enhanced sampling of peptides and proteins with a new biasing replica exchange method

Classical MD simulations (cMD) are limited by the sampling of relevant states of the peptides. Replica exchange (REMD) methods aim to search the conformational space of proteins more efficiently (reviewed in Ostermeir & Zacharias, 2013). We have developed a Hamiltonian REMD method that takes advantage of an intrinsic property of proteins, the specific Φ ? dihedral angle combinations along the polymer backbone. By employing a coupled two-dimensional biasing potential the energy barriers along the polymer backbone are reduced more effectively than by a previous approach based on a one-D biasing potential (Kannan & Zacharias, 2007). Thus, adjacent amino acids along the polymers backbone can easily switch between favourable regions in the Ramachandran plot. Additionally, energy barriers of rotameric states of amino acid side chains of proteins are also biased in the replica runs. The method improves the sampling of conformational substates of proteins at a modest number of replicas (nine replicas in the standard set-up with one replica running without biasing potential) compared to much larger numbers necessary in the case of standard temperature (T)-REMD simulations. A further improvement is achieved by a dynamical adjustment of the penalty potential levels in the replicas such that high exchange rates and improved mixing of conformations between different replicas are guaranteed. The biasing potential (BP)-REMD method turns out to be suitable to speed up both the folding of spaghetti-like test peptides and the refinement of loop decoy structures. Starting from extended structures, an α-helical oligo-alanine and β-hairpin chignolin and the Trp-cage protein fold more rapidly in near-native structures than in cMD simulations. The BP-REMD simulations not only accelerate the folding process of test proteins but also enlarge the variety of sampled configurations in conformational space. Since flexible parts of the protein can be penalized selectively, this method provides a precise tool to investigate regions of interest of the protein.     

Enhanced sampling of peptide and protein conformations using replica exchange simulations with a peptide backbone biasing-potential

During replica exchange molecular dynamics (RexMD) simulations, several replicas of a system are simulated at different temperatures in parallel allowing for exchange between replicas at frequent intervals. This technique allows significantly improved sampling of conformational space and is increasingly being used for structure prediction of peptides and proteins. A drawback of the standard temperature RexMD is the rapid increase of the replica number with increasing system size to cover a desired temperature range. In an effort to limit the number of replicas, a new Hamiltonian-RexMD method has been developed that is specifically designed to enhance the sampling of peptide and protein conformations by applying various levels of a backbone biasing potential for each replica run. The biasing potential lowers the barrier for backbone dihedral transitions and promotes enhanced peptide backbone transitions along the replica coordinate. The application on several peptide cases including in all cases explicit solvent indicates significantly improved conformational sampling when compared with standard MD simulations. This was achieved with a very modest number of 5-7 replicas for each simulation system making it ideally suited for peptide and protein folding simulations as well as refinement of protein model structures in the presence of explicit solvent.     

Exploring the energy landscape of protein folding using replica-exchange and conventional molecular dynamics simulations

Two independent replica-exchange molecular dynamics (REMD) simulations with an explicit water model were performed of the Trp-cage mini-protein. In the first REMD simulation, the replicas started from the native conformation, while in the second they started from a nonnative conformation. Initially, the first simulation yielded results qualitatively similar to those of two previously published REMD simulations: the protein appeared to be over-stabilized, with the predicted melting temperature 50-150K higher than the experimental value of 315K. However, as the first REMD simulation progressed, the protein unfolded at all temperatures. In our second REMD simulation, which starts from a nonnative conformation, there was no evidence of significant folding. Transitions from the unfolded to the folded state did not occur on the timescale of these simulations, despite the expected improvement in sampling of REMD over conventional molecular dynamics (MD) simulations. The combined 1.42 micros of simulation time was insufficient for REMD simulations with different starting structures to converge. Conventional MD simulations at a range of temperatures were also performed. In contrast to REMD, the conventional MD simulations provide an estimate of Tm in good agreement with experiment. Furthermore, the conventional MD is a fraction of the cost of REMD and continuous, realistic pathways of the unfolding process at atomic resolution are obtained.     

The investigation of the secondary structure propensities and free-energy landscapes of peptide ligands by replica exchange molecular dynamics simulations

The conformational states of two peptide sequences that bind to staphylococcal enterotoxin B are sampled by replica exchange molecular dynamic (REMD) simulations in explicit water. REMD simulations were treated with 52 replicas in the range of 280–501 K for both peptides. The conformational ensembles of both peptides are dominated by random coil, bend and turn structures with a small amount of helical structures for each temperature. In addition, while an insignificant presence of β-bridge structures were observed for both peptides, the β-sheet structure was observed only for peptide 3. The results obtained from simulations at 300 K are consistent with the experimental results obtained from circular dichroism spectroscopy. From the analysis of REMD results, we also calculated hydrophobic and hydrophilic solvent accessible surface areas for both peptides, and it was observed that the hydrophobic segments of the peptides tend to form bend or turn structures. Moreover, the free-energy landscapes of both peptides were obtained by principal component analysis to understand how the secondary structural properties change according to their complex space. From the free-energy analysis, we have found several minima for both peptides at decreased temperature. For these obvious minima of both peptides, it was observed that the random coil, bend and turn structures are still dominant and the helix, β-bridge or β-sheet structures can appear or disappear with respect to minima. On the other hand, when we compare the results of REMD with conventional MD simulations for these peptides, the configurations of peptide 3 might be trapped in energy minima during the conventional MD simulations. Hence, it can be said that the REMD simulations have provided a sufficiently high sampling efficiency.     

Starting‐structure dependence of nanosecond timescale intersubstate transitions and reproducibility of MD‐derived order parameters

Factors affecting the accuracy of molecular dynamics (MD) simulations are investigated by comparing generalized order parameters for backbone NH vectors of the B3 immunoglobulin‐binding domain of streptococcal protein G (GB3) derived from simulations with values obtained from NMR spin relaxation (Yao L, Grishaev A, Cornilescu G, Bax A, J Am Chem Soc 2010;132:4295‐4309.). Choices for many parameters of the simulations, such as buffer volume, water model, or salt concentration, have only minor influences on the resulting order parameters. In contrast, seemingly minor conformational differences in starting structures, such as orientations of sidechain hydroxyl groups, resulting from applying different protonation algorithms to the same structure, have major effects on backbone dynamics. Some, but not all, of these effects are mitigated by increased sampling in simulations. Most discrepancies between simulated and experimental results occur for residues located at the ends of secondary structures and involve large amplitude nanosecond timescale transitions between distinct conformational substates. These transitions result in autocorrelation functions for bond vector reorientation that do not converge when calculated over individual simulation blocks, typically of length similar to the overall rotational diffusion time. A test for convergence before averaging the order parameters from different blocks results in better agreement between order parameters calculated from different sets of simulations and with NMR‐derived order parameters. Thus, MD‐derived order parameters are more strongly affected by transitions between conformational substates than by fluctuations within individual substates themselves, while conformational differences in the starting structures affect the frequency and scale of such transitions. Proteins 2013. © 2012 Wiley Periodicals, Inc.     

Crystal contact-free conformation of an intrinsically flexible loop in protein crystal: Tim21 as the case study

BackgroundIn protein crystals, flexible loops are frequently deformed by crystal contacts, whereas in solution, the large motions result in the poor convergence of such flexible loops in NMR structure determinations. We need an experimental technique to characterize the structural and dynamic properties of intrinsically flexible loops of protein molecules.MethodsWe designed an intended crystal contact-free space (CCFS) in protein crystals, and arranged the flexible loop of interest in the CCFS. The yeast Tim 21 protein was chosen as the model protein, because one of the loops (loop 2) is distorted by crystal contacts in the conventional crystal.ResultsYeast Tim21 was fused to the MBP protein by a rigid α-helical linker. The space created between the two proteins was used as the CCFS. The linker length provides adjustable freedom to arrange loop 2 in the CCFS. We re-determined the NMR structure of yeast Tim21, and conducted MD simulations for comparison. Multidimensional scaling was used to visualize the conformational similarity of loop 2. We found that the crystal contact-free conformation of loop 2 is located close to the center of the ensembles of the loop 2 conformations in the NMR and MD structures.ConclusionsLoop 2 of yeast Tim21 in the CCFS adopts a representative, dominant conformation in solution.General significanceNo single powerful technique is available for the characterization of flexible structures in protein molecules. NMR analyses and MD simulations provide useful, but incomplete information. CCFS crystallography offers a third route to this goal.     

Improved side‐chain torsion potentials for the Amber ff99SB protein force field

Recent advances in hardware and software have enabled increasingly long molecular dynamics (MD) simulations of biomolecules, exposing certain limitations in the accuracy of the force fields used for such simulations and spurring efforts to refine these force fields. Recent modifications to the Amber and CHARMM protein force fields, for example, have improved the backbone torsion potentials, remedying deficiencies in earlier versions. Here, we further advance simulation accuracy by improving the amino acid side‐chain torsion potentials of the Amber ff99SB force field. First, we used simulations of model alpha‐helical systems to identify the four residue types whose rotamer distribution differed the most from expectations based on Protein Data Bank statistics. Second, we optimized the side‐chain torsion potentials of these residues to match new, high‐level quantum‐mechanical calculations. Finally, we used microsecond‐timescale MD simulations in explicit solvent to validate the resulting force field against a large set of experimental NMR measurements that directly probe side‐chain conformations. The new force field, which we have termed Amber ff99SB‐ILDN, exhibits considerably better agreement with the NMR data. Proteins 2010. © 2010 Wiley‐Liss, Inc.     

Folding simulations of Trp‐cage mini protein in explicit solvent using biasing potential replica‐exchange molecular dynamics simulations

Replica exchange molecular dynamics (RexMD) simulations are frequently used for studying structure formation and dynamics of peptides and proteins. A significant drawback of standard temperature RexMD is, however, the rapid increase of the replica number with increasing system size to cover a desired temperature range. A recently developed Hamiltonian RexMD method has been used to study folding of the Trp‐cage protein. It employs a biasing potential that lowers the backbone dihedral barriers and promotes peptide backbone transitions along the replica coordinate. In two independent applications of the biasing potential RexMD method including explicit solvent and starting from a completely unfolded structure the formation of near‐native conformations was observed after 30–40 ns simulation time. The conformation representing the most populated cluster at the final simulation stage had a backbone root mean square deviation of ～1.3 Å from the experimental structure. This was achieved with a very modest number of five replicas making it well suited for peptide and protein folding and refinement studies including explicit solvent. In contrast, during five independent continuous 70 ns molecular dynamics simulations formation of collapsed states but no near native structure formation was observed. The simulations predict a largely collapsed state with a significant helical propensity for the helical domain of the Trp‐cage protein already in the unfolded state. Hydrogen bonded bridging water molecules were identified that could play an active role by stabilizing the arrangement of the helical domain with respect to the rest of the chain already in intermediate states of the protein. Proteins 2009. © 2008 Wiley‐Liss, Inc.     

Dynameomics: Data‐driven methods and models for utilizing large‐scale protein structure repositories for improving fragment‐based loop prediction

Protein function is intimately linked to protein structure and dynamics yet experimentally determined structures frequently omit regions within a protein due to indeterminate data, which is often due protein dynamics. We propose that atomistic molecular dynamics simulations provide a diverse sampling of biologically relevant structures for these missing segments (and beyond) to improve structural modeling and structure prediction. Here we make use of the Dynameomics data warehouse, which contains simulations of representatives of essentially all known protein folds. We developed novel computational methods to efficiently identify, rank and retrieve small peptide structures, or fragments, from this database. We also created a novel data model to analyze and compare large repositories of structural data, such as contained within the Protein Data Bank and the Dynameomics data warehouse. Our evaluation compares these structural repositories for improving loop predictions and analyzes the utility of our methods and models. Using a standard set of loop structures, containing 510 loops, 30 for each loop length from 4 to 20 residues, we find that the inclusion of Dynameomics structures in fragment‐based methods improves the quality of the loop predictions without being dependent on sequence homology. Depending on loop length, ～25–75% of the best predictions came from the Dynameomics set, resulting in lower main chain root‐mean‐square deviations for all fragment lengths using the combined fragment library. We also provide specific cases where Dynameomics fragments provide better predictions for NMR loop structures than fragments from crystal structures. Online access to these fragment libraries is available at http://www.dynameomics.org/fragments .     

A novel approach to segregate and identify functional loop regions in protein structures using their Ramachandran maps

The loops which connect or flank helices/sheets in protein structures are known to be functionally important. However, ironically they also belong to the part of protein whose structure is least accurately predicted. Here, a new method to isolate and analyze loop regions in protein structure is proposed using the spatial coordinates of the solved three‐dimensional structure. The extent of dispersion among points of successive amino acid residues in the Ramachandran map of protein region is utilized to calculate the Mean Separation between these points in the Ramachandran Plot (MSRP). Based on analysis of 2935 protein secondary structure regions obtained using DSSP software, spanning a range from 2 to 64 residues, taken from a set of 170 proteins, it is shown that helices (MSRP < 17) and strands (MSRP < 64) stand effectively demarcated from the loop regions (MSRP > 130). Analysis of 43 DNA binding and 98 ligand binding proteins revealed several loop regions with clear change in MSRP subsequent to binding. The population of such loops correlated with the magnitude of backbone displacement in the protein subsequent to binding. Can changes in MSRP quantify the temporal oscillations in dihedral angles among structured/unstructured regions in proteins? Molecular dynamics simulations (10 ns) revealed that deviations in MSRP among different snapshots in the trajectory were at least twofold higher for unstructured proteins in comparison with ordered proteins. The above results validate the use of MSRP parameter as a tool to identify and investigate functionally active loops and unstructured regions in protein structures. Proteins 2010. © 2009 Wiley‐Liss, Inc.     

Intrinsic α‐helical and β‐sheet conformational preferences: A computational case study of alanine

A fundamental question in protein science is what is the intrinsic propensity for an amino acid to be in an α-helix, β-sheet, or other backbone dihedral angle (-ψ) conformation. This question has been hotly debated for many years because including all protein crystal structures from the protein database, increases the probabilities for α-helical structures, while experiments on small peptides observe that β-sheet-like conformations predominate. We perform molecular dynamics (MD) simulations of a hard-sphere model for Ala dipeptide mimetics that includes steric interactions between nonbonded atoms and bond length and angle constraints with the goal of evaluating the role of steric interactions in determining protein backbone conformational preferences. We find four key results. For the hard-sphere MD simulations, we show that (1) β-sheet structures are roughly three and half times more probable than α-helical structures, (2) transitions between α-helix and β-sheet structures only occur when the backbone bond angle τ (N–C_α–C) is greater than 110°, and (3) the probability distribution of τ for Ala conformations in the “bridge” region of-ψ space is shifted to larger angles compared to other regions. In contrast, (4) the distributions obtained from Amber and CHARMM MD simulations in the bridge regions are broader and have increased τ compared to those for hard sphere simulations and from high-resolution protein crystal structures. Our results emphasize the importance of hard-sphere interactions and local stereochemical constraints that yield strong correlations between -ψ conformations and τ.     

Antibodies as a model system for comparative model refinement

Predicting the conformations of loops is a critical aspect of protein comparative (homology) modeling. Despite considerable advances in developing loop prediction algorithms, refining loops in homology models remains challenging. In this work, we use antibodies as a model system to investigate strategies for more robustly predicting loop conformations when the protein model contains errors in the conformations of side chains and protein backbone surrounding the loop in question. Specifically, our test system consists of partial models of antibodies in which the “scaffold” (i.e., the portion other than the complementarity determining region, CDR, loops) retains native backbone conformation, whereas the CDR loops are predicted using a combination of knowledge‐based modeling (H1, H2, L1, L2, and L3) and ab initio loop prediction (H3). H3 is the most variable of the CDRs. Using a previously published method, a test set of 10 shorter H3 loops (5–7 residues) are predicted to an average backbone (N? Cα? C? O) RMSD of 2.7 Å while 11 longer loops (8–9 residues) are predicted to 5.1 Å, thus recapitulating the difficulties in refining loops in models. By contrast, in control calculations predicting the same loops in crystal structures, the same method reconstructs the loops to an average of 0.5 and 1.4 Å for the shorter and longer loops, respectively. We modify the loop prediction method to improve the ability to sample near‐native loop conformations in the models, primarily by reducing the sensitivity of the sampling to the loop surroundings, and allowing the other CDR loops to optimize with the H3 loop. The new method improves the average accuracy significantly to 1.3 Å RMSD and 3.1 Å RMSD for the shorter and longer loops, respectively. Finally, we present results predicting 8–10 residue loops within complete comparative models of five nonantibody proteins. While anecdotal, these mixed, full‐model results suggest our approach is a promising step toward more accurately predicting loops in homology models. Furthermore, while significant challenges remain, our method is a potentially useful tool for predicting antibody structures based on a known Fv scaffold. Proteins 2010. © 2010 Wiley‐Liss, Inc.     

Fast de novo discovery of low‐energy protein loop conformations

In the prediction of protein structure from amino acid sequence, loops are challenging regions for computational methods. Since loops are often located on the protein surface, they can have significant roles in determining protein functions and binding properties. Loop prediction without the aid of a structural template requires extensive conformational sampling and energy minimization, which are computationally difficult. In this article we present a new de novo loop sampling method, the Parallely filtered Energy Targeted All‐atom Loop Sampler (PETALS) to rapidly locate low energy conformations. PETALS explores both backbone and side‐chain positions of the loop region simultaneously according to the energy function selected by the user, and constructs a nonredundant ensemble of low energy loop conformations using filtering criteria. The method is illustrated with the DFIRE potential and DiSGro energy function for loops, and shown to be highly effective at discovering conformations with near‐native (or better) energy. Using the same energy function as the DiSGro algorithm, PETALS samples conformations with both lower RMSDs and lower energies. PETALS is also useful for assessing the accuracy of different energy functions. PETALS runs rapidly, requiring an average time cost of 10 minutes for a length 12 loop on a single 3.2 GHz processor core, comparable to the fastest existing de novo methods for generating an ensemble of conformations. Proteins 2017; 85:1402–1412. © 2017 Wiley Periodicals, Inc.     

Monte Carlo folding of trans-membrane helical peptides in an implicit generalized Born membrane

An efficient Monte Carlo (MC) algorithm using concerted backbone rotations is combined with a recently developed implicit membrane model to simulate the folding of the hydrophobic transmembrane domain M2TM of the M2 protein from influenza A virus and Sarcolipin at atomic resolution. The implicit membrane environment is based on generalized Born theory and has been calibrated against experimental data. The MC sampling has previously been used to fold several small polypeptides and been shown to be equivalent to molecular dynamics (MD). In combination with a replica exchange algorithm, M2TM is found to form continuous membrane spanning helical conformations for low temperature replicas. Sarcolipin is only partially helical, in agreement with the experimental NMR structures in lipid bilayers and detergent micelles. Higher temperature replicas exhibit a rapidly decreasing helicity, in agreement with expected thermodynamic behavior. To exclude the possibility of an erroneous helical bias in the simulations, the model is tested by sampling a synthetic Alanine-rich polypeptide of known helicity. The results demonstrate there is no overstabilization of helical conformations, indicating that the implicit model captures the essential components of the native membrane environment for M2TM and Sarcolipin.     

Molecular dynamics simulations of Guanine quadruplex loops: advances and force field limitations

A computational analysis of d(GGGGTTTTGGGG)(2) guanine quadruplexes containing either lateral or diagonal four-thymidine loops was carried out using molecular dynamics (MD) simulations in explicit solvent, locally enhanced sampling (LES) simulations, systematic conformational search, and free energy molecular-mechanics, Poisson Boltzmann, surface area (MM-PBSA) calculations with explicit inclusion of structural monovalent cations. The study provides, within the approximations of the applied all-atom additive force field, a qualitatively complete analysis of the available loop conformational space. The results are independent of the starting structures. Major conformational transitions not seen in conventional MD simulations are observed when LES is applied. The favored LES structures consistently provide lower free energies (as estimated by molecular-mechanics, Poisson Boltzmann, surface area) than other structures. Unfortunately, the predicted optimal structure for the diagonal loop arrangement differs substantially from the atomic resolution experiments. This result is attributed to force field deficiencies, such as the potential misbalance between solute-cation and solvent-cation terms. The MD simulations are unable to maintain the stable coordination of the monovalent cations inside the diagonal loops as reported in a recent x-ray study. The optimal diagonal and lateral loop arrangements appear to be close in energy although a proper inclusion of the loop monovalent cations could stabilize the diagonal architecture.     

Reproducible polypeptide folding and structure prediction using molecular dynamics simulations

The folding of a polypeptide from an extended state to a well-defined conformation is studied using microsecond classical molecular dynamics (MD) simulations and replica exchange molecular dynamics (REMD) simulations in explicit solvent and in vacuo. It is shown that the solvated peptide folds many times in the REMD simulations but only a few times in the conventional simulations. From the folding events in the classical simulations we estimate an approximate folding time of 1-2 micros. The REMD simulations allow enough sampling to deduce a detailed Gibbs free energy landscape in three dimensions. The global minimum of the energy landscape corresponds to the native state of the peptide as determined previously by nuclear magnetic resonance (NMR) experiments. Starting from an extended state it takes about 50 ns before the native structure appears in the REMD simulations, about an order of magnitude faster than conventional MD. The calculated melting curve is in good qualitative agreement with experiment. In vacuo, the peptide collapses rapidly to a conformation that is substantially different from the native state in solvent.     

Peptoid conformational free energy landscapes from implicit-solvent molecular simulations in AMBER

To test the accuracy of existing AMBER force field models in predicting peptoid conformation and dynamics, we simulated a set of model peptoid molecules recently examined by Butterfoss et al. (JACS 2009, 131, 16798-16807) using QM methods as well as three peptoid sequences with experimentally determined structures. We found that AMBER force fields, when used with a Generalized Born/Surface Area (GBSA) implicit solvation model, could accurately reproduce the peptoid torsional landscape as well as the major conformers of known peptoid structures. Enhanced sampling by replica exchange molecular dynamics (REMD) using temperatures from 300 to 800 K was used to sample over cis-trans isomerization barriers. Compared to (Nrch)5 and cyclo-octasarcosyl, the free energy of N-(2-nitro-3-hydroxyl phenyl)glycine-N-(phenyl)glycine has the most "foldable" free energy landscape, due to deep trans-amide minima dictated by N-aryl sidechains. For peptoids with (S)-N (1-phenylethyl) (Nspe) side chains, we observe a discrepancy in backbone dihedral propensities between molecular simulations and QM calculations, which may be due to force field effects or the inability to capture n --> n* interactions. For these residues, an empirical phi-angle biasing potential can "rescue" the backbone propensities seen in QM. This approach can serve as a general strategy for addressing force fields without resorting to a complete reparameterization. Overall, this study demonstrates the utility of implicit-solvent REMD simulations for efficient sampling to predict peptoid conformational landscapes, providing a potential tool for first-principles design of sequences with specific folding properties.     

Toward better refinement of comparative models: predicting loops in inexact environments

Achieving atomic-level accuracy in comparative protein models is limited by our ability to refine the initial, homolog-derived model closer to the native state. Despite considerable effort, progress in developing a generalized refinement method has been limited. In contrast, methods have been described that can accurately reconstruct loop conformations in native protein structures. We hypothesize that loop refinement in homology models is much more difficult than loop reconstruction in crystal structures, in part, because side-chain, backbone, and other structural inaccuracies surrounding the loop create a challenging sampling problem; the loop cannot be refined without simultaneously refining adjacent portions. In this work, we single out one sampling issue in an artificial but useful test set and examine how loop refinement accuracy is affected by errors in surrounding side-chains. In 80 high-resolution crystal structures, we first perturbed 6-12 residue loops away from the crystal conformation, and placed all protein side chains in non-native but low energy conformations. Even these relatively small perturbations in the surroundings made the loop prediction problem much more challenging. Using a previously published loop prediction method, median backbone (N-Calpha-C-O) RMSD's for groups of 6, 8, 10, and 12 residue loops are 0.3/0.6/0.4/0.6 A, respectively, on native structures and increase to 1.1/2.2/1.5/2.3 A on the perturbed cases. We then augmented our previous loop prediction method to simultaneously optimize the rotamer states of side chains surrounding the loop. Our results show that this augmented loop prediction method can recover the native state in many perturbed structures where the previous method failed; the median RMSD's for the 6, 8, 10, and 12 residue perturbed loops improve to 0.4/0.8/1.1/1.2 A. Finally, we highlight three comparative models from blind tests, in which our new method predicted loops closer to the native conformation than first modeled using the homolog template, a task generally understood to be difficult. Although many challenges remain in refining full comparative models to high accuracy, this work offers a methodical step toward that goal.     

Structural comparison and classification of alpha‐helical transmembrane domains based on helix interaction patterns

Structural classification of membrane proteins is still in its infancy due to the relative paucity of available three‐dimensional structures compared with soluble proteins. However, recent technological advances in protein structure determination have led to a significant increase in experimentally known membrane protein folds, warranting exploration of the structural universe of membrane proteins. Here, a new and completely membrane protein specific structural classification system is introduced that classifies α‐helical membrane proteins according to common helix architectures. Each membrane protein is represented by a helix interaction graph depicting transmembrane helices with their pairwise interactions resulting from individual residue contacts. Subsequently, proteins are clustered according to similarities among these helix interaction graphs using a newly developed structural similarity score called HISS. As HISS scores explicitly disregard structural properties of loop regions, they are more suitable to capture conserved transmembrane helix bundle architectures than other structural similarity scores. Importantly, we are able to show that a classification approach based on helix interaction similarity closely resembles conventional structural classification databases such as SCOP and CATH implying that helix interactions are one of the major determinants of α‐helical membrane protein folds. Furthermore, the classification of all currently available membrane protein structures into 20 recurrent helix architectures and 15 singleton proteins demonstrates not only an impressive variability of membrane helix bundles but also the conservation of common helix interaction patterns among proteins with distinctly different sequences. Proteins 2010. © 2010 Wiley‐Liss, Inc.