Molecular Dynamics Simulation of Lipid-Modified Signaling Proteins

Abstract

In this chapter, we provide a practical guide on how to plan, execute, and interpret atomistic and coarse-grained molecular dynamics (MD) simulations of lipid-modified proteins in model membranes. After outlining some key practical considerations when planning such simulations, we survey resources and techniques to obtain force field parameters for nonconventional amino acids, such as posttranslationally lipid-modified amino acids that are unique to this class of proteins. We then describe the protocols to build, setup, and run the simulations, followed by a brief comment on the analysis and interpretation of the simulations. Finally, examples of insights that could be gained from atomistic and coarse-grained MD simulations of lipidated proteins will be provided, using RAS proteins as illustrative examples. Throughout the chapter, we highlight the main advantages and limitations of simulating RAS and related lipid-modified G-proteins in biomimetic membranes.

1. Introduction

Biological membranes are the fundamental structural frameworks of the cell and have many diverse functions. They serve as a selectively permeable protective envelope of the cell, compartmentalize various organelles that have specialized functions, and provide a two-dimensional structural framework to organize signaling proteins and lipids. A wide variety of signaling proteins, such as the guanine triphosphate hydrolyzing family of enzymes (GTPases or G-proteins), are targeted to the plasma membrane (PM) or various internal membranes via lipid-based membrane-targeting motifs. Among these motifs, the most common are generated by posttranslational modification of glycine and cysteine amino acids by acyl and prenyl groups. The quintessential example of acetylated and prenylated G-proteins is the RAS family of small GTPases.

RAS proteins function as intracellular molecular switches to control a wide variety of signal transduction pathways. All RAS proteins, including HRAS, NRAS, and KRAS that represent the three most common RAS isoforms in humans, cycle between active and inactive states by binding to GTP and GDP. Active RAS interacts with multiple downstream effectors to drive cell growth and proliferation. Somatic mutations that impair hydrolysis of GTP to GDP stabilize RAS in its activated form, causing sustained proliferative signaling and oncogenesis. In fact, RAS genes are found mutated in nearly 30% of all cancers and are the primary drivers of many lethal cancers such as pancreatic, colon, and lung cancer [1]. Despite years of studies, however, pharmacologically targeting RAS has proved extremely tricky. A discussion of the challenges in drugging RAS is beyond the scope of this chapter but can be found in our recent review [2].

An aspect of the biology of RAS and related G-proteins that is gaining increasing attention for its therapeutic prospect is their dynamic localization to and fluctuations at membrane surfaces. Unfortunately, both the process of membrane targeting and the dynamics of the complex are difficult to study experimentally due to the resolution limit of current techniques. In contrast, the exponential growth in computational power over the last decade has begun to enable a detailed characterization of these phenomena using atomic-level molecular simulations. In particular, invaluable insights have been gained into the dynamics of RAS in complex with biomimetic membranes using molecular dynamics (MD) simulations. The results were instrumental in bringing RAS back to the forefront of anticancer drug discovery by, for example, showing how interactions of RAS with lipids, partner proteins, or ligands can be modulated by allosteric transitions and conformational selection [3-11].

In this chapter, we provide a practical guide on how to plan, setup, and conduct MD simulations of lipidated proteins in various lipid bilayers. Our discussion is largely based on our own experiences over the last decade, but we also highlight key contributions from other laboratories. We use RAS proteins for illustration because they are the best characterized and most representative example of proteins that are targeted to membranes by lipid-based motifs. The chapter is organized as follows: we first provide an overview of practical considerations to keep in mind during the planning stage of a simulation, followed by a survey of resources and approaches to obtain force field parameters for nonconventional amino acids. We then describe protocols for system construction and setup, and for running and analyzing MD simulations of lipid-modified proteins. This is followed by examples of insights that could be gained from atomistic and coarse-grained MD simulation of lipidated proteins using RAS proteins as an example. Strengths and limitations of MD simulations to study RAS and related proteins are highlighted throughout the chapter.

2. Methods

2.1. During Planning: Practical Considerations About System Size and Model Resolution

In a typical MD simulation, the position and momenta of each particle in the system is computed by integrating Newtonian equations of motion at discrete time intervals. This results in a series of snapshots that can be used to track the state of the system throughout the simulation and to characterize its dynamical or equilibrium properties. These capabilities and its high (atomic or semi-atomic) resolution made MD a powerful tool to study complex systems in great detail, such as monomeric or multimeric RAS in the membrane. A detailed discussion of MD is beyond the scope of this chapter and can be found elsewhere (e.g., [12, 13]). For our purposes, it suffices to note that there are two important issues to carefully consider during the planning stage of MD: system size and resolution. Because MD calculates the pairwise interactions between nearly all the particles in the system, it has an O(N²) complexity (actual number could be slightly lower due to the use of cutoff schemes for long-range vdW and electrostatic interactions). Therefore, performing MD on large systems at atomic resolution can be resource intensive.

1. One way to mitigate this challenge is to choose a model system with minimal complexity but sufficient to address a given set of specific questions, and simulating it for a sufficiently long duration to ensure convergences of the results. For example, in the case of RAS, a typical consideration for tradeoff between complexity and sampling efficiency would be to ask whether one should simulate the minimal membrane-targeting motif (the lipid anchor), the hypervariable region (HVR), or the full-length protein (Fig. 1). The answer almost always depends on the questions to be addressed by the simulation. The smaller system size made possible by the use of the isolated lipid anchor or the HVR allows for conducting multiple relatively long simulations with minimal computational cost. The goal of such a simulation could be to study the role of individual residues on these segments for membrane insertion. Depending on the simulation length and the resolution of the model (see below), another goal of simulating the isolated lipid anchor or the HVR could be to study their lateral dynamics and self-assembly in bilayers. Simulating the larger system associated with the full-length RAS comes at a higher computational cost, but would allow one to additionally study the external as well as internal dynamics and interactions of the catalytic domain with partner proteins or lipids.

2. A related issue is compositional complexity, which is especially important in membrane simulations because the more diverse the lipid species in the simulation box the longer it takes to achieve equilibration. Here, too, one needs to carefully balance complexity with resource and time requirements to achieve the desired goal. A rule of thumb may be to choose—guided by prior knowledge of the system—the minimum number of lipid species required to address the question at hand. We will reiterate this issue later in this chapter in reference to simulating RAS proteins.

3. Another way of reducing computational cost is to use a coarse-grained (CG) model, wherein a group of physically connected non-hydrogen atoms (typically about four) are represented by a single-reaction center. This reduces the degrees of freedom in the molecules as well as the number of particles to be simulated, and averages out the high-frequency fluctuations. As a result, CG models allow one to simulate larger systems for longer durations [12]. Coarse-grained MD simulation is particularly relevant to study long timescale processes such as RAS lateral dynamics and clustering in domain-forming bilayers.

Fig. 1.

Domain architecture and membrane targeting of lipid-modified small GTPases using the KRAS protein as example. The conserved catalytic domain (residues 1–166) has two lobes: Lobe 1 (residues 1–86) and Lobe 2 (residues 87–166) and harbors switch 1 (residues 25–40) and switch 2 (residues 60–75) regions that are involved in nucleotide (shown in orange), and effector and regulator binding. The hypervariable C-terminus (HVR, residues 167–185) consists of a flexible linker region and a farnesylated (yellow) polybasic (red) lipid anchor. A portion of hypothetical membrane is shown as a gray surface to illustrate tethering of lipid-modified proteins to one leaflet of a membrane. Actual organization of the catalytic domain and the HVR on membrane surfaces is more complicated, highly dynamic, and will depend on sequence and structural details

2.2. Before Getting Started: Force Field Parameters for Posttranslationally Modified Amino Acids

In this section, we use RAS proteins as an example to provide a brief description of force field-related issues relevant for simulating lipidated proteins in both the all-atom (AA) and coarse-grained (CG) models. Force field parameters for AA-MD simulation of standard amino acids, lipids, and nucleic acids have been optimized over the decades and included in all of the popular force field distributions including CHARMM [14], AMBER [15-17], and GROMOS [18]. Similarly, parameters for the CG-MD simulation of standard amino acids and common lipids are available from various sources. Therefore, our discussion below focuses on the non-standard, post-translationally modified amino acids. Moreover, for brevity, we limit our discussion to the CHARMM force field [14] for AA-MD and the MARTINI [19] parameter set for CG-MD simulations, but the basic concepts discussed here also apply to other popular force fields.

1. AA force field parameters for the most common lipidations, such as palmitoylation, myristoylation, and farnesylation, are now available in the CHARMM36 force field [20] (http://mackerell.umaryland.edu/charmm_ff.shtml). CHARMM-GUI (http://www.charmm-gui.org) may represent the easiest pipeline to set up AA-MD simulations using these parameters. However, there may be situations where simulations need to be setup from scratch or the desired lipidated amino acid is not yet parameterized. Parameters compatible with an earlier version of the CHARMM force field (CHARMM27) have been developed for farnesylated and palmitoylated cysteine using analogy to phospholipid and cysteine parameters [21, 22]. Missing parameters for lipidated amino acids that are compatible with the latest force fields can be prepared using either an ab initio-based or a semiempirical approach. The ab initio approach entails geometry optimization using quantum mechanical (QM) calculations to obtain partial charges, bond lengths, and angles. Such an approach has been used to obtain parameters for farnesyl [23] using the QM software Gaussian [24], and a more recent study described how to do so using NWCHEM [25] in a manner that can serve as a useful guide to parameterize other lipidated amino acids [26]. The semiempirical approach uses already parameterized analogous molecules as a reference to assign parameters to new molecules, followed by iterative optimization steps to reproduce experimental observations. This technique is most successfully implemented in the CHARMM Generalized Force Field (CGenFF) toolkit. CGenFF is primarily designed to generate parameters for small molecule ligands [27], but with little extra effort it could be used to generate parameters for farnesylated [28, 29] or other lipidated amino acids. Also, the force field ToolKit (ffTK) plugin in VMD (Visual Molecular Dynamics: https://www.ks.uiuc.edu/Research/vmd/) employs the same basic principle to generate parameters for small molecules and may be used to parameterize lipidated amino acids [30]. A detailed tutorial for ffTK can be found at https://www.ks.uiuc.edu/Research/vmd/plugins/fftk/.

2. In addition to lipid modification, mono- or poly-ubiquitination is common in many proteins including RAS, which has been shown to be ubiquitinated at multiple lysine residues such as Lys104, Lys147, and Lys170 [31]. These modifications may affect the dynamics and/or interactions of RAS with lipids, but there are no ready-to-use parameters for ubiquitinated lysine. Lysine ubiquitination entails formation of an isopeptide bond between the C-terminal glycine of ubiquitin and the lysine side chain. We recently used analogy to parameterize ubiquitin-modified Lys [84]. Briefly, a modified lysine residue, Lyq, was defined with the backbone NH2 atom type of Lys replaced by NH and the positive charge on the HZ2 atom redistributed to the neighboring ε-CH2 atoms. A covalent bond was then defined between the carboxy-terminal carbon of a modified glycine, Glq, and the side chain NH atom of Lyq. The bond length and corresponding angles were defined by analogy to backbone peptide bonds.

3. As in the atomistic models, efforts have been made to develop parameters for the CG-MD simulation of proteins harboring lipidated amino acids. A notable example in this regard is the MARTINI model. In this CG model, on average four covalently linked heavy atoms and associated hydrogens are represented by a single-interaction center or bead. Depending on the polarity of the atoms involved, the bead may be polar, intermediate, apolar, or charged (see http://cgmartini.nl/ for more detail). Parameters for palmitoylated and farnesylated cysteines compatible with an earlier version of MARTINI were developed using analogy and used to simulate relatively large RAS/membrane complexes for up to 40 μs, yielding invaluable insights into the mechanisms of RAS clustering and domain partitioning [32-34]. Following an update of the force field (MARTINI 2.2) to more accurately describe, for example, membrane partitioning of peptides [35], newer parameters for lipidated amino acids including prenylated cysteine have been developed [36]. The protocol described in ref. [36] can be used to parameterize those not already parameterized.

2.3. Getting Started: System Construction and Simulation Setup

With the appropriate force field parameters in hand, one can begin to build an initial configuration of the system to setup the simulation.

1. Preparing the protein: For most lipidated proteins including RAS, the structure of the highly dynamic lipid anchor and a number of residues in its vicinity is unavailable in the Protein Data Bank (PDB). In fact, almost all available RAS structures consist only of the catalytic G-domain. Therefore, the structure of the lipid anchor and the linker connecting it to the catalytic domain need to be modeled, typically as an extended structure. This can be done using any of the many different structure prediction algorithms such as MODELLER [37], the Molefacture plugin in VMD [38], the Build Structure functionality in UCSF-Chimera [39], or PyMol [40]. If needed, the resulting structure can be energy-minimized and subjected to a short MD simulation using standard protocols. Then, it is ligated to the structure of the catalytic domain downloaded from the PDB to build the full-length protein structure (see for example ref. [22]); missing residues in flexible regions of the PDB structure, if any, can be modeled using the tools mentioned above. Alternatively, the full-length protein structure may be directly built using homology modeling programs such as MODELLER (https://salilab.org/modeller/tutorial/). The isolated lipid anchor or the full-length protein can then be prepared for simulation after appropriate patching for disulfide bonds (if any) or chain termini (negative, positive, or neutral). For example, RAS has an oxymethylated C-terminus, which can be added to the structure by selecting CT1 as a C-terminal patch. The resulting structure may then be subjected to energy-minimization and a short equilibration to ensure integrity.

2. Preparing the bilayer: In parallel, a bilayer patch of the desired size and lipid composition is prepared either from a pre-equilibrated solvated bilayer or from scratch (e.g., using CHARMM-GUI (http://www.charmm-gui.org/?doc=input/membrane.bilayer) [41, 42] or similar other resources). The size of the bilayer patch or the number of lipids is determined by the nature (e.g., isolated lipid anchor versus full-length) and the number of the protein to be simulated. In general, the bilayer should be large enough to accommodate the full extent of anticipated conformational fluctuations of the protein in a manner that avoids boundary effects. The nature of the protein to be simulated and its target membrane in the cell also dictate the choice of the lipid composition in the bilayer. For example, the PM has a different lipid composition from the generally more curved internal membranes such as the endoplasmic reticulum. However, the amino acid composition is of the protein that interacts with lipids is perhaps more critical for the choice of lipids in the model membrane. For example, while all three most common human RAS proteins are targeted to the inner leaflet of the PM, the prenylated polybasic lipid anchor of KRAS, but not the palmitoylated and prenylated HRAS or NRAS, preferentially interacts with anionic phospholipids such as POPS (1-palmitoyl-2-oleoyl-sn-glycero-3-phospho-L-serine), which is enriched at the inner leaflet of the PM. Therefore, KRAS is typically simulated in an anionic membrane such as a POPC/POPS (20–30%) bilayer while NRAS and HRAS can be simulated in a pure POPC bilayer. Once the appropriate bilayer model is constructed, it can be subjected to standard energy-minimization steps to relieve bad atomic contacts and equilibrated through a series of restrained and unrestrained MD runs to obtain an equilibrated bilayer model satisfying the expected measures of bilayer structural integrity, such as bilayer thickness and area per lipid.

3. Building and solvating the protein–membrane system: The protein (or peptide) and membrane models prepared as described above can be assembled into a single system using, for example, the protein structure file generator (PSFGEN) program (https://www.ks.uiuc.edu/Research/vmd/plugins/psfgen/) and VMD. This is typically followed by a series of manual or automated translations and rotations of the protein relative to the bilayer so that the soluble portion remains away from the bilayer, and the lipid anchor is placed near the bilayer surface oriented in such a way as to increase the likelihood of its membrane insertion. At this stage, the system can be solvated (e.g., using the Solvate plugin in VMD and the TIP3P water model) and ions and, as needed, salt are added to obtain a charge-neutralized system having the desired ionic strength (such as the physiologic ionic strength of 150 mM). In principle, the resulting system can now be energy-minimized, equilibrated, and simulated following standard protocols. In practice, however, the available search space and the barrier at the lipid–water interface make the spontaneous insertion of the lipid anchor into the bilayer costly in terms of both resource and time. A quicker work around is to slowly pull one or more of the lipid chains on the protein (e.g., using a soft harmonic potential applied to the terminal carbon and the bilayer center) during an MD run until a portion of the hydrocarbon chain (about five carbon atoms) reaches the hydrophobic core of the bilayer [21]. This would allow for the hydrophobic lipid anchor to cross the energy barrier at the lipid–water interface so that, once released, the rest of the anchor inserts spontaneously. This protocol has proven highly effective in a number of previous simulations. Alternatively, a portion of the lipid anchor can be manually inserted into the bilayer using a series of translations and rotations, coupled with deletion or displacement of water and/or lipid molecules that overlap with protein atoms, followed by multiple steps of energy-minimization and equilibration runs. Some of these procedures can nowadays be done more easily using CHARMM-GUI [43]. These include building a bilayer of the desired lipid composition, adding a lipid modification to an otherwise complete protein structure (using the Add Lipid Tail option), and placing the protein in proximity to or in the membrane. One could also orient a desired surface of the protein toward the bilayer or position it at a desired distance from the membrane. Finally, if the protein were to be tethered only to one leaflet, the surface area of that leaflet may become larger than the opposing leaflet, causing membrane bending. This can be avoided fairly easily by adding the appropriate number of extra lipids in the protein-free leaflet or by removing from the leaflet housing the protein until the monolayer areas are equalized.

4. Converting to a CG- model: A fully atomistic protein–membrane-water system constructed as described above and, preferably, subjected to an AA-MD simulation can serve as a starting structure for a CG-MD simulation. For the MARTINI CG model, this can be achieved through AA-to-CG mapping using the martinize.py tool available in the MARTINI website (http://cgmartini.nl/). The CG-MD simulation can directly commence from this step if the target is the minimal membrane anchor, which in most lipidated proteins is devoid of a secondary structure. For the full-length protein, elastic network-based restraints need to be applied on the secondary structure elements to maintain the integrity of the tertiary structure. This can be done using the ElNeDyn utility [44]. Because a CG approach is primarily chosen to gain access to more complicated processes such as lipid domain formation, the bilayer lipid composition is typically more complex than in an AA model. In fact, the bilayer may be made up of two to tens of lipid species [45, 46] and can be simulated alone for reference or with the protein of interest bound. Moreover, it is often desirable to have multiple copies of the same protein (e.g., to increase sampling or to study self-assembly), or to have different proteins of varying numbers (e.g., to study the formation of signaling complexes). Construction of such a complicated system can be achieved using the insane utility, which allows the user to generate bilayers of essentially any lipid composition (provided that the parameters exist) and to insert the protein (s) of interest to any desired depth [47].

2.4. Running the Simulation and Analyzing the Data

Once the system is properly setup, the production phase of the simulations can be conducted following standard protocols. The AA-MD simulation can be run with NAMD [48], AMBER [49], or GROMACS [50] just to name a few, while the MARTINI-based CG-MD is typically run with GROMACS. In each case, no significant system-specific adjustments are required during the production phase of simulating lipidated proteins, with the possible exception of the need to regularly monitor insertion and equilibration of the lipid anchor in the bilayer. Similarly, most of the protocols and tools used to analyze MD trajectories of other protein–membrane complexes are directly applicable to lipidated proteins. Additional analysis techniques that are somewhat specific to surface-bound proteins, such as defining appropriate reaction coordinates to study orientations of the catalytic domain on membrane surfaces, can be found in several recent reports [51-54]. Limitations associated with force fields or the sampling of phase space that are typical of other protein–lipid complexes also apply to RAS and related proteins. A possible exception here may be the need to run multiple copies for as long as possible. This is because, first, the simulations are often started from structures that lack the flexible region connecting the catalytic domain with the lipid anchor. This requires extensive sampling to achieve convergence and internal consistency of the results, for example, in terms of conformations sampled and distributions of residue-lipid interactions. Second, this class of proteins is highly dynamic and tends to adopt many diverse transient conformations that may not be captured in a single-short simulation. Third, there are accumulating evidences suggesting that many lipidated proteins, especially the RAS family of small GTPases, may form transient oligomers for which experimental data for benchmarking is scarce or entirely unavailable. Therefore, simulation of the oligomerization process of lipidated proteins on membrane surfaces, often conducted using CG-MD for sampling efficiency, should ideally be coupled with new experiments however limited or indirect, and the results interpreted with care.

2.5. What Can We Learn from the Simulations? Some Illustrative Examples

In this section, we briefly review some of the important and likely broadly applicable observations from previous atomistic or coarse-grained MD simulations of RAS proteins conducted in our laboratory and those of others. Although not discussed here to save space, we note that MD simulations have been applied to a number of other RAS-like lipidated small GTPases [55-58].

1. AA-MD simulation of the RAS lipid anchors: RAS localizes to the PM via an S-farnesyl cysteine carboxylmethyl ester (common to all RAS proteins), complemented by palmitoylation of proximal cysteine(s) (in NRAS and HRAS) or a polybasic domain of six contiguous lysine residues (KRAS). Prior et al. [59] identified residues 175–185 as the minimal membrane-targeting motif of KRAS and residues 180–186 as those of HRAS or NRAS. In the cell, these peptides (referred to as tK, tH, and tN) are targeted to the PM in a manner that generally recapitulates the full-length cognate [60]. These motifs can therefore be used as model systems in MD simulations to study the structural and physiochemical factors underlying RAS insertion into and dynamics at membranes. In one of the earliest such studies, AA-MD simulation was conducted on tN in a 1,2-dimyristoyl-sn-glycero-3-phosphocholine (DMPC) bilayer [21]. It was found that the insertion of about five terminal carbon atoms of either the palmitoyl at Cys181 or the farnesyl at Cys186 (modeled by a hexadecyl) into the bilayer core is sufficient to cross the barrier at the bilayer–water interface. Once the barrier is crossed, the rest of the peptide spontaneously partitioned into the bilayer, with tN lipids immersing into the hydrophobic core dragging with them the hydrophobic side chains of Met182 and Leu184; hydrogen bonds between the peptide backbone and lipid head groups provided additional stabilization [21]. These observations were consistent with and helped explain data from a combined FTIR, solid-state NMR, and neutron diffraction spectroscopy analysis [61]. A similar study of tH yielded the same results: the farnesylated (Cys186) and palmitoylated (Cys181 and Cys184) cysteines plus Met182 are engaged in vdW interactions with lipid acyl chains while the peptide backbone and polar side chains form hydrogen bonds with lipid head groups [62]. Similarly, an AA-MD study of tK in a POPC/POPG (2-oleoyl-1-palmitoyl-sn-glycero-3-phosphoglycerol) bilayer showed that insertion of the farnesyl tail and electrostatic interactions of the lysine residues with the anionic POPG lipids together stably tether tK to the membrane [63]. This observation was supported by more recent extended simulations of tK in a physiologically more relevant POPC/POPS bilayer [64] and experiments in cells [64]. The key role of electrostatics in tK-membrane association was further demonstrated by the observation that its phosphorylation of Ser181 decreases membrane affinity [65].

   AA-MD simulations were also used to examine the energetics of tH-membrane interaction [66]. For example, potential of mean force calculations of tH insertion into a DMPC lipid bilayer using the adaptive biasing force (ABF) technique indicated a steep decline in the free energy profile after insertion of a few carbon atoms of the lipid anchor, corroborating an earlier observation [21, 62]. Follow up studies to investigate the relative role of individual lipid modifications and the thermodynamics of tH-DMPC interactions arrived at two important conclusions: (a) the palmitoyl modification at Cys181 contributes more to the insertion free energy than that at Cys184 [67], and (b) tH-DMPC binding is dominated by an enthalpy-driven hydrophobic effect resulting primarily from the vdW interactions between tH and membrane lipids [68].

2. AA-MD simulation of full-length RAS proteins: The first AA-MD simulation of full-length RAS was performed on HRAS in complex with a DMPC bilayer, and the results were compared with simulations of tH and the HVR (the hypervariable region, residues 167–186) [22]. It was found that although the residues at the lipid anchor are the primary determinants of membrane binding, the rest of the protein is not a passive spectator. In particular, in the context of the full-length protein, a pair of positively charged residues at the HVR (Arg169 and Lys170) or the catalytic domain (Arg128 and Arg135) interacts with lipids directly but not simultaneously. The catalytic domain adopting two distinct orientations with respect to the membrane plane made this possible. During similar simulations of full-length KRAS in a DMPC bilayer, the catalytic domain was highly dynamic and did not adopt a stable orientation [51, 69]. This was because KRAS requires anionic lipids for a stable membrane binding, as demonstrated by more recent extended simulations of KRAS in anionic lipid bilayers [51, 54]. One of these studies, conducted in the more physiologically relevant POPC/POPS bilayer and involving over 10 MD runs of up to 1 μs each, identified multiple conformational states in which the catalytic domain interacts with the membrane [51]. These included orientation state 1 (OS1) in which the three C-terminal α-helices face the bilayer and orientation state 2 (OS2) in which β-strands 1–3 and α-helix 2 are near the bilayer. In the latter, the effector binding switch loops are mostly occluded from solvent (and hence partner proteins) by the membrane [51]. Subsequent 20 μs-long AA-MD simulations on three different mutants of KRAS showed that [52, 53]: (a) transitions between OS1 and OS2 occur through an intermediate orientation OS0, (b) the signaling-competent (non-occluded) S0 and OS1 orientations dominate the overall population of orientation states, and (c) the orientational dynamics is underpinned largely by intrinsic conformational fluctuations.

   While most AA-MD simulations including those previously conducted on RAS and discussed above were limited to tens of microseconds at best, a recent simulation of KRAS in a POPC and POPC/POPS bilayers achieved a millisecond aggregate time by running 290 copies of 5 μs-long simulations [54]. The overall conclusions of the study regarding protein–lipid interactions and KRAS membrane reorientation were consistent with the previous observations. However, the enormous data generated by the study enabled the quantification of the extent to which POPS slows down the dynamics of KRAS on the bilayer. It was found that, compared with a pure POPC bilayer, the presence of 30% POPS slowed down the translational diffusion of KRAS by a factor of ~2 and the likelihood of its disengagement from the membrane by a factor of 8. Other AA-MD simulations of membrane-bound RAS focused on characterizing protein–protein interaction interfaces (PPIs) involved in RAS dimerization [70-73] or its interaction with downstream effectors Raf [74] and PI3K [75].

3. CG-MD simulation of Ras proteins: Imaging and other cell biology experiments had shown that RAS proteins form spatially segregated nanoclusters on the membrane, but the structural details and physicochemical basis of this process remained elusive [60]. In one of the earliest CG-MD studies of RAS, 64 tH molecules were arrayed in a domain-forming bilayer made up of dipalmitoylphosphatidylcholine (DPPC) and dilinoleoylphosphatidylcholine (DLiPC) plus varying concentrations of cholesterol, and simulated at different temperatures for up to 40 μs effective time [34]. It was found that tH spontaneously assembles into clusters of 4–10 molecules. The clusters preferentially localized to the boundary between the liquid ordered (lo) and liquid disordered (ld) phases driven primarily by the affinity of the saturated palmitoyl and poly-unsaturated farnesyl moieties for DMPC and DLiPC, respectively. Using the same CG-MD model, subsequent studies investigated the effect of tH concentration and lipid composition on the formation [33] and stability [76] of tH nanoclusters. These simulations predicted a minimal concentration of tH below which nanoclustering may not occur and a saturating concentration beyond which increasing tH concentration may not result in changes in cluster size. Changes in cholesterol concentration did not affect the size of the nanoclusters but rather their stability through its impact on the stability of membrane domains. This is because, as noted above, tH nanoclusters tend to accumulate at lipid domain boundaries whose stability and size is a function of cholesterol content. Moreover, it was found that the accumulation of asymmetrically bound tH [32] and full-length HRAS [32, 77] nanoclusters at domain boundaries reduces the boundary line tension and stabilizes membrane curvature, largely through monolayer area expansion. This has been further demonstrated using the mesoscale dissipative particle dynamics simulation technique applied to monomeric and variously cross-linked simplified model systems that partitioned to the domain boundary of a small unilamellar vesicle [78].

4. MD simulations for the identification of interaction hot spots: The examples described in the previous subsections illustrate the most widely used applications of MD in the study of lipid-modified proteins. Another appealing application is probe-based MD (pMD), a solvent mapping technique that uses small organic molecules as probes to identify cryptic ligand binding sites for therapeutic targeting. When this technique was applied to the soluble catalytic domain of KRAS [79], it was found to be effective in identifying all previously characterized allosteric pockets [6]. However, most of the commonly used probes, such as isopropyl alcohol, tend to partition into membranes, limiting the application of pMD only to soluble proteins. To overcome this, a modified version of pMD called pMD-membrane has been proposed [80] wherein the affinity of probe molecules for the bilayer core is reduced by modifying selected probe–lipid pairwise vdW interactions. This can be done fairly easily using the correction term NBFIX for simulations with the CHARMM force field. Typically, the well depth of the Lennard-Jones potential is reduced to about 0.01 kcal/mol, and the minimum interparticle distance increased to about 7 Å. This prevents the probes from penetrating the hydrophobic core of the bilayer while providing access to the side of the protein facing the bilayer. A pMD-membrane simulation starts by solvating the bilayer-bound protein with water containing ~20% probe and adding K⁺, Na⁺, or Cl⁻ ions to neutralize the system and to achieve a desired ionic strength. The system is energy-minimized and the probes homogenized in the bulk water using simulated annealing with the protein heavy atoms harmonically restrained to prevent unfolding. The rest of the simulation follows standard protocols and can be run in multiple copies for several hundred nanoseconds to achieve convergence. The combined trajectories can then be analyzed using various techniques, including a recently introduced technique based on surface topography maps [81], to determine probe occupancy in putative binding pockets. Application of pMD-membrane on KRAS bound to a POPC/POPS bilayer in various orientations [80, 81] suggested that the way in which the catalytic domain interacts with lipids significantly affects the probe accessibility of druggable pockets, suggesting a similar potential impact on drug binding.

3. Summary and Future Prospects

Elucidating the interaction of surface-bound lipid-modified proteins with membranes presents multiple difficulties to experimental techniques due to the complex dynamics of these systems in space and time. Molecular dynamics simulations conducted at different levels of detail can help fill this gap, as demonstrated in this chapter using RAS proteins as an example. Advances in computer power and the continuing development of force field parameters and analysis tools mean that MD simulations are poised to address a wide range of questions whose answer may lie in visualizing how lipidated proteins engage membrane lipids. We hope that the protocols for parameterization, system construction, system setup, and simulation described in this chapter will facilitate the application of MD in its various flavors to any lipid-modified protein monomer or multimer tethered to a model membrane whose lipid composition mimics the protein’s biological host membrane. Note that it is possible to draw some general principles by studying a few representatives from the different families of lipidated proteins, as we have attempted to show using RAS proteins. However, the substantial diversity of lipid-modified signaling proteins in sequence/structure, function, subcellular localization, posttranslational modification, and pathophysiology suggests that we have barely scratched the surface. Therefore, MD simulation will continue to play an indispensable role in shining light on the many hidden mysteries behind the seemingly simple domain architecture of lipidated signaling proteins. The following are two examples of interesting future applications: The PM is characterized by an asymmetric distribution of lipids [82]; in particular, the intracellular side is enriched by POPS and therefore negatively charged. Further, the differential distribution of ions across the PM generates a transmembrane potential. Both of these may modulate interaction of surface-bound proteins with lipids and/or the orientation, lateral dynamics, and assembly of the proteins on membrane surfaces. Therefore, simulating lipidated proteins tethered to an asymmetric bilayer using classical AA and CG models on the one hand and a polarizable force field (see ref. [83] for a review) on the other would be an appealing future work.