Computational structure modeling for diverse categories of macromolecular interactions

Abstract

Computational protein-protein docking is one of the most intensively studied topics in structural bioinformatics. The field has made substantial progress through over three decades of development. The development began with methods for rigid-body docking of two proteins, which have now been extended in different directions to cover the various macromolecular interactions observed in a cell. Here, we overview the recent developments of the variations of docking methods, including multiple protein docking, peptide-protein docking, and disordered protein docking methods.

Introduction

Protein-protein interactions (PPIs) are involved in many essential biological processes in a cell. To understand the mechanisms of PPIs’ functions, tremendous efforts have been made to determine the tertiary structures of protein complexes using experimental methods. To supplement experimental approaches, computational methods have been developed that build predicted structures of protein-protein complexes. Computational prediction models do not always have an atomic-detailed accuracy; however, models with a lower accuracy are still able to provide useful information such as interface residues and overall docking orientation depending on the application. Furthermore, as we describe later, computational modeling and experiments can be combined.

Conventionally, the main development of protein docking methods is for modeling pairwise interaction of proteins that have more or less rigid structures. Although the focus of this article is to overview computational protein docking approaches that are outside of the conventional type, it would be appropriate to provide a quick summary of the conventional protein docking methods. Computational protein-protein docking methods can be roughly categorized into two types, template-based and ab initio (template-free) docking. Template-based docking approaches are increasingly applicable as experimentally determined complex structures have been rapidly accumulated in the Protein Data Bank (PDB) [1]. The algorithms range from those which use global templates [2–4] to those which use local interaction patterns of complexes [5–8].

In cases where no appropriate template is available, docking models can be constructed by an ab initio method, which constructs a pool of docking models by essentially exhaustively exploring candidate poses with different mutual rotation angles and translations. Then, a scoring function is used to select the most plausible models out of the pool. Various protein structure representations [9,10] and docking conformation search algorithms [11–13] were employed in ab initio methods. Scoring functions developed include those which are based on the statistics of atom interactions in known protein structures (statistical knowledge-based potentials), which give favorable scores to structure models that follow the distribution of atom pair distances and angles of native protein structures [14,15], scoring functions based on numerical optimization, which optimize parameters (energy values) so that near-native models can be selected [16], and re-parameterization of molecular mechanics force field [17]. There are also attempts to compute the binding free energy of protein docking models using molecular dynamics (MD) simulations [18,19]. Recently, 3D deep learning was applied for docking model selection [20]. A task that remains difficult in pairwise protein docking is to consider conformational changes of proteins that may occur upon docking. Methods used for generating docking conformations include Monte-Carlo (MC) approaches [21], elastic network models (ENMs) that consider proteins as mass-and-spring networks to simulate flexibility [13,22,23], and MD simulations [24,25].

The progress of docking methods has been periodically monitored by a community-wide protein docking experiment, the Critical Assessment of PRedicted Interactions (CAPRI). In the recent 7^th edition of the community-wide CAPRI assessment [26], targets were not limited to pairwise interactions, but have included a variety of interesting modeling targets, including large multimeric protein complexes, protein-peptide complexes, artificial protein complexes, and protein-oligosaccharide interactions. Modeling a complex is easy if a known protein complex structure can be used as a template. Oppositely, the level of difficulty increases if an available template has significant difference with the target complex, or a suitable template for a complex is not available. Another general lesson from CAPRI is that biological information, such as known interface residues, can drastically reduce the possible docking poses and lead to correct modeling. In the 7^th CAPRI assessment, there were eight protein-protein targets, among which five were dimers and three were multimers. Models of acceptable quality or better, were obtained for a total of six protein-protein complexes, three dimer and all three multimer targets. Participants were also able to model all three peptide targets and all five polysaccharide targets. Readers are referred to the recent assessment reports from the CAPRI organizers [26,27]. For more detailed information about recent protein docking development, refer to review articles [28,29].

Multiple protein docking

Starting from pairwise protein docking, docking methods have been expanded into several directions to reflect different types of interactions observed in a cell (Figure 1). One obvious need for an extension is methods for docking multiple subunits into a complex.

Figure 1.

Extending protein docking from pairwise to diverse categories. Methods for pairwise docking include LZerD ClusPro, SwarmDock, PatchDock, HADDOCK, and ZDOCK. For multiple docking, tools such as CombDock, Multi-LZerD, M-ZDOCK, SymmDock, and SAM are available. Membrane protein docking tools include Memdock and MPDock. Peptide-protein docking tools include PIPER-FlexPepDock CABS-Dock, HPepDock, AutoDock CrankPep, MDockpep, and pepATTRACT. For docking with a disordered protein, IDP-LZerD is available. Methods for docking a protein with DNA/RNA include FoldX, RnaX, HDOCK, and NPDock.

CombDock is a pioneering method that performs multiple protein docking [30]. The method first predicts pairwise interactions between subunits of a target protein complex using PatchDock [31]. Then, it performs a hierarchical combinatorial assembly of sub-complexes, starting from building sub-complex models of three subunits, then four subunits, until models of the full-subunit complex are build, by each time combining models of smaller sub-complexes. At each stage, top-scoring models are selected to reduce the number of possible combinations. Another multiple protein docking method, Multi-LZerD [32], also starts from pre-computed pairwise models of all possible subunit pairs using a pairwise docking method, LZerD [10]. In Multi-LZerD, these pairwise docking models are assembled into full-subunit models using a genetic algorithm (GA). The GA constructs a population of different models at each stage (called a generation), which are modified to produce a new set of models in the next generation. At each generation, a population of models (~200) is selected by a scoring function, and the procedure is iterated until convergence. The two methods are applicable to complexes with up to ~8 subunits but would face difficulty for larger complexes because of a combinatorially growing docking conformational space. Also, the strategy of using pairwise models as starting may not be appropriate for larger complexes where many subunit pairs do not tightly interact. For assembling larger proteins, extra information of interacting and non-interacting subunit pairs or interface residues from experiments would be helpful. Multi-LZerD was later extended to predict the assembly pathway of multimeric complexes (Path-LZerD) [33]. The complex building and optimization process by the GA in Multi-LZerD can be considered as a simulation of the assembly process, from which the assembly order can be observed.

CombDock and Multi-LZerD do not require additional constraints for modeling and are able to model asymmetric hetero complexes. On the other hand, model construction is significantly easier for homo-symmetric complexes. M-ZDOCK [34], SymmDock [35], and SAM [36] perform an exhaustive docking conformation search restricted to specified point group symmetries. The first two methods are for complexes with a symmetric symmetry (C_n), whereas SAM can also handle higher-order symmetries, dihedral (D_n), tetrahedral (T), octahedral (O), and icosahedral (I) groups.

Membrane protein docking

Application to membrane proteins is an important sub-category of docking as a relatively lower fraction of structures are available in PDB despite their biological importance. Membrane proteins are also important targets for drug discovery [37]. Docking of membrane proteins typically starts by placing two membrane proteins in the correct topology relative to the membrane. Then docking conformations are sampled under the constraints that interacting residues from two proteins are at a similar height relative to the membrane and that the two proteins keep more or less perpendicular orientation to the membrane. Memdock [38] takes this approach. Membrane proteins are initially oriented using the Orientations of Proteins in Membranes (OPM) database and the PPM server [39], which determines the spatial position of a protein in the membrane by optimizing the transfer free energy from water to the lipid bilayer [40]. Another method, MPDock [41], combines an MC algorithm implemented in RosettaDock [17] with a membrane energy function and conformational sampling that account for the membrane environment. A benchmark dataset for membrane protein docking consisting of 37 membrane proteins was collected by the Bonvin group [42], which encourages future developments of novel docking algorithms for membrane proteins.

Peptide & intrinsic disordered protein docking

Protein-peptide docking, another class of docking, is important since protein-peptide interactions are part of various signaling pathways, and also due to the emergence of peptide therapeutics [43]. Likely reflecting their importance in therapeutics application, several notable new developments have been recently observed. In this problem setting, flexibility cannot be readily neglected since detailed atomic interactions are usually of interest in peptide docking. In this category, we also consider assembly involving intrinsically disordered proteins (IDPs), since their flexibility needs to be explicitly modelled.

A general protocol for peptide-docking is to generate an ensemble of peptide conformations, dock them to a receptor protein, and finally refine selected docked models. Thus, methods can be characterized by examining the techniques used in each step. To generate the initial ensemble of peptide conformers, pepATTRACT [44] and HADDOCK [45] use three simple ideal conformations of peptides while MDockPep [46] uses a homology modeling method, MODELLER [47], to model conformations. HPepDock [48], PIPER-FlexPepDock [49], and IDP-LZerD sample peptide conformations from a fragment library. For the docking step, a common choice is to use rigid-body docking, as adopted by HPepDock, HADDOCK, pepATTRACT, PIPER-FlexDock, and IDP-LZerD [50]. Some methods perform the first two steps, peptide conformation generation and docking, simultaneously. MDockPep uses Autodock [51], a popular protein-small chemical molecule docking program, for flexible peptide docking. AutoDock CrankPep [52] is a recent addition to the AutoDock program series that is specifically designed for peptide docking. Rosetta FlexPepDock [53] starts from an extended peptide conformation placed at a known binding site and folds the peptide on the receptor protein surface using a library of 3- to 5-mer peptide fragments. Anchordock starts from a folded peptide conformation that is placed at a known binding residue on the receptor surface and then optimizes the docked conformation using MD [54]. CABS-dock [55] performs MC simulation of peptide conformation sampling and docks using a coarse-grained protein representation. Most of the methods perform a structure refinement step at the last stage in the protocol.

The peptide docking methods described here have been shown to be able to dock sequences at most 16 residues long ab initio. The exception is IDP-LZerD [50], which has been shown to successfully dock IDPs of up to 69 residues long. In IDP-LZerD, 30 peptide conformations are generated for each of the 9-residue long fragment windows slid along the sequence of an input IDP, which are then independently docked to a receptor. Then, docked peptides are selected on the receptor surface so that they can be connected to the full-length IDP.

DNA/RNA protein docking

As in the case of protein-protein docking, a template-based approach works well as long as an appropriate template structure of a protein-nucleic acid complex exists in PDB [56]. Other existing methods include the application of rigid-body docking programs [57–60] and scoring functions that specifically target protein-DNA/RNA complexes [57,61–63]. For nucleotide docking, modeling flexibility is very important as nucleotides tend to be more flexible than proteins. HADDOCK was combined with ENM-based conformer generation to quickly search the conformational space [64], although the degree of the conformational changes considered by ENM was small, 1 to 2 Å root mean square deviation (RMSD) to the initial structures.

Regarding the nucleotide flexibility, FoldX [65] and RnaX [66] are of particular note. These two methods use fragment pair libraries of double-stranded (ds) DNA and RNA, respectively, bound to protein fragments to generate plausible traces of the nucleic acid bound to the target protein. In FoldX, peptide-DNA fragments collected from PDB have a peptide length of 6–12 amino acids and a dsDNA length of 4–8 base pairs. To model a protein-DNA complex, peptide-DNA fragments are selected and connected along the protein by considering the sequence similarity and evaluating with a force field and geometrical constraints. Similarly, in RnaX, a library of protein-RNA fragments with 6 amino acids and RNA fragments of 4–8 nucleotides were used. Thus, in principle, the methods are able to model protein-nucleotide complexes in the absence of a clear template. FoldX and RnaX have been tested on complexes with fairly long dsDNAs and RNAs, which are over 190 bps and 150 bases, respectively.

Data-assisted protein docking

In the last section, we briefly overview methods that use additional information to guide docking. For more details of methods in this category, readers are referred to a recent review [67]. Roughly, types of experimental data can be classified into three categories, those which provides interface and surface residue information, residue and atom distance data, and data that indicate overall shape (Figure 2). Mutagenesis [68] and hydrogen/deuterium (H/D) exchange [69] are experiments that can provide information on binding site residues. On the other hand, XL-MS and FRET provide distance constraints by identifying cross-linked residues (XL-MS) [70] or a donor and an acceptor pair that transfer the energy (FRET) [71,72] between interacting proteins. NMR provides restraints of interface residues with Chemical Shift Perturbations (CSP), orientation with Residual Dipolar Couplings (RDCs), and distance information with Nuclear Overhauser Effects (NOEs) [73]. Interface residues and residue distances information can be readily input into protein docking web servers, such as ClusPro [12], HADDOCK [74], HDOCK [75], LZerD [76], PatchDock [31], and PyDock [77] servers as constraints.

Figure 2.

Three categories of experimental data that can assist protein docking. From left, data that indicate residues at the interface or other surface, which include data from mutagenesis, H/D exchange. Middle, residue/atom distance data, which can be provided by FRET, XL-MS, and NMR. Right, complex shape information, which can be provided by SAXS and EM.

Notably, the same types of information, interface residues and contacting residue pairs, can also be predicted from the protein sequence and structure and have been successfully used for guiding docking. Protein docking interface residues can be identified by mutation patterns [78], common interaction sites of ligand proteins [79], or combinations of residue features [80]. PI-LZerD [81] uses predicted interface residue information with an assumption that prediction may not be entirely correct. There is a rapid development of prediction methods for contacting residues and distances between residues in protein structures [82]. Such methods have also been applied for predicting contacts across two proteins for guiding protein docking [83–85].

SAXS and cryo-electron microscopy (cryo-EM) provide low-resolution shape information of macromolecular complexes. A SAXS profile of the complex is used as a filter to select docking models that agree with it [86–89]. In cryo-EM, although near-atomic resolution density maps are more highlighted recently, there are still many maps determined at a lower resolution. These low-resolution maps are still useful for guiding docking [90–92].

Docking methods, including HADDOCK [74] and IMP [67], integrate different sources of experimental information to improve the docking process.

Future direction

We have discussed how the development of protein docking, which started from pairwise protein docking, expanded into various directions to be able to construct and predict major types of molecular interactions that occur in a cell. An interesting future direction would be to model and simulate macromolecular interactions in large scale biomolecular systems in a crowded environment of a cell [93]. Recently, substantial efforts have been made in the dynamics simulation of proteins in a crowded environment [94]. Work has also begun on protein docking methods that explicitly consider interaction in a crowded environment [95]. Uniting structure modeling and dynamics simulations of such macromolecular interactions integrated with large-scale experiments [96,97] could yield a clearer and more realistic picture for understanding the behavior of molecules in a living cell.