Molecular dynamics: a powerful tool for studying the medicinal chemistry of ion channel modulators

Abstract

Molecular dynamics (MD) simulations allow researchers to investigate the behavior of desired biological targets at ever-decreasing costs with ever-increasing precision. Among the biological macromolecules, ion channels are remarkable transmembrane proteins, capable of performing special biological processes and revealing a complex regulatory matrix, including modulation by small molecules, either endogenous or exogenous. Recently, given the developments in ion channel structure determination and accessibility of bio-computational techniques, MD and related tools are becoming increasingly popular in the intense research area regarding ligand–channel interactions. This review synthesizes and presents the most important fields of MD involvement in investigating channel–molecule interactions, including, but not limited to, deciphering the binding modes of ligands to their ion channel targets and the mechanisms through which chemical compounds exert their effect on channel function. Special attention is devoted to the importance of more elaborate methods, such as free energy calculations, while principles regarding drug design and discovery are highlighted. Several technical aspects involving the creation and simulation of channel–molecule MD systems (ligand parameterization, proper membrane setup, system building, etc.) are also presented.

The most important fields of molecular dynamics involved in investigating the interactions between chemical compounds and ion channels are reviewed, and a selection of case studies and their approaches is analysed and put into context.

1. Introduction

The molecular dynamics (MD) field has evolved during the past decades to become one of the most developed areas of theoretical studies in molecular biology.^1–6 It allows investigators to perform computer simulations on molecular motions and interactions with respect to time, largely increasing the cost-effectiveness of applied biological research. MD can contribute to revealing novel features of molecular interactions, especially when used in conjunction with experimental designs.^7–9 This has become a hallmark of biomolecular computation, whereby the computational structure–function investigation of protein and DNA motion is widely employed to gain information about the living world, its molecular machines, and the intricate ways chemical compounds interact with biological macromolecules.^10–12

Technical advances in the field of general-purpose graphics processing units (GP-GPU) and improved simulation algorithms have been key drivers of the development of MD computational methods.^13–20 Contrasting the past, when impressive supercomputers accelerated MD simulations, a modest workstation with low energy consumption is in present sufficient to allow researchers to perform nanosecond simulations in a matter of hours.^21–23 Additionally, nowadays, a plethora of tools exist that aid computational biologists in the preparation, simulation and analysis of molecular dynamics experiments. Tasks such as proper membrane placement, ligand parameterization and full simulation system building can be performed with good accuracy in seconds to minutes, given proper inputs are used in the respective tools, accelerating MD-based studies and reducing costs associated with time and computational power.^24,25 A significant amount of knowledge is required for performing and running accurate MD simulations by computational biologists; this knowledge applies on various aspects including system preparation, basic simulation settings, sampling enhancements, path calculation designs, analysis considerations, etc. Each of these implies a considerable amount of expertise regarding various fields from structural biology, statistical thermodynamics, or computer science. The past two decades marks the construction and constant development of applications, web servers and other tools that semi-automate the whole process related to running and analyzing MD simulations. With a good knowledge grasp of basic computational protocols and concepts, taking advantage of user-friendly interfaces with interactive usage, even bench researchers should be able to implement simple in silico methods complementing their own work, guided by specific software.^20,26–28 Despite their intrinsic simplicity and rough approach, these novice-friendly tools can lead bench researchers to asking the right questions related to their experiments.^27,28

Among biological macromolecules, ion channels are complex integral transmembrane proteins with distinct functional and structural characteristics, facilitating passive ion movement along the electrochemical gradient,^29–31 while their function is finely tuned by a wide array of regulators.^32–36 Developments of protein structure determination methods, including solid state nuclear magnetic resonance (ssNMR), X-ray crystallography, and especially of cryo-electron microscopy (cryo-EM), allow identification of structural models of transmembrane proteins, including ion channels, whose three-dimensional (3D) architectures have been notoriously difficult to elucidate.^37,38 They also enabled, in recent years, the emergence of computational biology methods as widely preferred when investigating ion channels and other transmembrane proteins.^39–43 These insightful methods, including molecular modeling, docking, and MD simulations, allow researchers to investigate multiple aspects of ion channel function at once, such as small or large motions, ion conduction and its regulation, interaction with lipids, modulators or other small and large molecules, with particular implications in medicinal chemistry research.^44–48

Fig. 1 reveals the past and recent trends in ion channel research, on one hand, and in MD applied on ion channels, on the other hand, as approximated by Medline keywords usage. The past decade marks a clear rise in the number of studies published using MD, as compared to general studies on ion channels, thus indicating the large applicability and involvement of MD protocols in the study of ion channels.

Fig. 1. The trend in ion channel research (green, left axis) vs. the trend in the area of molecular dynamics (MD) applied on ion channels (orange, right axis) in the period 1971–2018. The data used to build the chart were generated using the Medline (PubMed) trend tool.⁴⁹.

In this review, the focus is placed on methods and achievements regarding classical all-atom MD applied to examining channel–molecule interactions, which are synthesized and presented, alongside current protocols and technical aspects related to proper use of MD simulations in such complex systems (Fig. 2). Section two discusses, in brief, general aspects of the MD method. Following, various general applications of MD simulations and their advances in the discovery of ligand binding sites and poses, and local and global ligand binding-induced effects on ion channel function and structure. Section six approaches more elaborate aspects of the MD toolbox in the form of relative binding free energy calculations. Section seven discusses technical and practical considerations concerning the proper protocols employed in creating and running MD simulations of channel–molecule systems. Section eight is dedicated to MD analysis tools, usually performed after running MD simulations, while the final sections conclude.

Fig. 2. Illustration depicting common protocols governing investigation of ligand–channel complexes through MD simulations; free selectivity filters represent simulations whereby ion flow is not controlled; partly occupied selectivity filters signify low-conductance, while fully occupied depict increased conductance.

2. The MD method – an overview

MD simulations are one of the most powerful methods currently used in medicinal chemistry and computational biology. The MD methods are often regarded as a computational “microscope”, through which scientists are able to “see” and investigate atomistic interactions arising between various types of molecules and macromolecules. Reduced to their basics, the MD algorithm supposes numerical solvation of the classical equations of motion for a given set of particles over a specified period of time. The resulting output is represented by a trajectory, to be further analyzed. Various simulation software packages aid in performing these calculations, such as AMBER,⁵⁰ NAMD,¹⁶ and GROMACS.⁵¹

Two particular challenges still exist in the field, ones that are gradually overcome, as new frontiers are reached in molecular dynamics techniques, algorithms, and computational power. These refer to the timescales attainable by the simulations, and their reliability in accurately modeling the behavior of atoms.⁵² For the latter, the discussions usually revolve around the accuracy of the force fields employed in MD. Simply put, a force field is a set of parameters governing the interaction between particles, which should be as realistic as possible.⁵³ There are various force fields describing the motion of small molecules, which will be analyzed in more detail in section 7.

Adequate sampling of the system under study can be hindered by free energy barriers. Additionally, timescales covered by MD simulations may not suffice to mimic timescales of biomolecular motions. In such cases, free energy calculations could rely on unconverged simulations, with a chance of error. Elaborate derivations of the MD algorithm allow for more efficient explorations of the phase space, including here a plethora of enhanced sampling and biasing methods: orthogonal space random walk (OSRW),⁵⁴ adaptive biasing force (ABF), Bennett acceptance ratio (BAR), conformational flooding, metadynamics,^55,56 milestoning, umbrella sampling,^57–60 replica exchange (RE), parallel tempering (PT) and self-guided molecular Langevin dynamics (SGLD).^61–63 The former seven methods facilitate sampling by introducing additional bias or force potential on a selected number of degrees of freedom, termed collective variables.⁵⁵ The latter three methods do not require a priori information about the systems.⁶³ Complex biomolecular systems may include a combination of two enhanced sampling methods, including replica exchange umbrella sampling or bias-exchange metadynamics (BE-MetaD). The metadynamics method itself has various types: metadynamics with parallel tempering (PTMetaD), metadynamics with multiple walker (MW-MetaD), funnel metadynamics, parallel bias metadynamics (PBMetaD) and infrequent metadynamics (InMetaD).⁶³ Readers are referred to more in-depth treatments on the technical nature of enhanced sampling techniques and their applications in free energy calculations,^{14,15,63–65} which exceed the scope of this focused review.

3. MD simulations of ion channels exhibiting small conformational changes upon ligand binding

Deciphering the mechanisms of ligand binding to their biological targets is an active field of research in molecular biology, and one of the most important application of bio-computational methods.^66–68 As docking, experimental site-directed mutagenesis, or ligand chemical modification can only pinpoint the molecular features of binding, MD tools may serve as a way to discriminate between various binding hypotheses or to extend experimental observations, thus inferring the way ligands position themselves, bind to their appropriate pockets, and modify channel function. For this to succeed, trajectories from accurate and faithful MD simulations must be analysed. Such proper simulations rely on controlling and reducing intrinsic and/or systematic errors that arise due to many factors including improper sampling, force field inaccuracies, errors in system preparation, errors due to simulation setup, improper analysis, among many others.^69,70

Additionally, in the field of MD-based approaches on ligand–channel interactions, attention should be drawn to two other factors that impact the success of an MD experiment. The first one is the need to correlate the results and observations from the computational, MD-based approach, to experimentally measured facts. The second factor concerns with performing the computational study in conjunction with previous knowledge from the literature, in order to generate good lead hypotheses, followed by their testing under the computational “microscope”. When studying life processes through theoretical approaches instead of direct, experimental ones, inclusion of artefacts or inadequate modeling may hinder our observations.⁷¹ In this way, by considering the two factors, the disadvantages and drawbacks of in silico methodologies are greatly reduced, and the results gain significant scientific confidence.

The complexity of ligand binding processes and their investigation generally requires adapting experimental and/or computational protocols to each particular features of ligand–channel interaction in study. The links between applicable MD tools and the expected scientific results stream along a continuum, making delineating, subtyping or grouping MD tools applied to solving various biological problems a difficult challenge. Nevertheless, by subgrouping, with some unavoidable overlapping, this and the next two sections review three kinds of approaches. The current section concerns with application of MD methods to elucidate ligand binding modes, when the binding site does not undergo significant structural changes due to ligand binding. Section four deals with the opposite: cases in which the binding site exhibits larger conformational motions, either in apo, holo, or both forms. Section five presents case studies of ion channel MD simulations with included ion flow simulation, by applying an external membrane potential, to reveal previously unobserved phenomena in ligand-induced effects.

The importance of the two aforementioned factors needed for proper ligand–channel MD-based studies can be best illustrated by recent advances in the mechanisms of capsaicin action on the transient receptor potential vanilloid member 1, or TRPV1. The transient receptor potential (TRP) channels are mainly involved in an organism's environmental sensation, eliciting chemoreception, nociception or thermoception. These cation-selective channels manifest diverse regulation mechanisms and possess an immense therapeutic potential.^72–74 TRPV1 mediates the heat sensation and is activated by capsaicin, the spiciness-eliciting bioactive molecule found in chili peppers.⁷⁵ The binding location of capsaicin to TRPV1 had been determined by cryo-EM, but, due to insufficient resolution, the exact positioning of the molecule could not be resolved.⁷⁶ Umbrella sampling has been successfully employed to study capsaicin behavior in biomembranes⁷⁷ and the study also included a part of the transmembrane portion of TRPV1. This part hosts the binding site for capsaicin and MD revealed how capsaicin interacts with external sites of the protein to form encounter complexes, which theoretically precede full entry of capsaicin to its binding pocket. Docking and MD of capsaicin in the “tail-up, head-down” has shown the likeliness of this pose.

In a later attempt to shed light on the precise binding position of capsaicin and to resolve two conflicting potential conformations, Darré and Domene⁷⁸ performed MD simulations combined with analysis methods to reveal the compound indeed attaining a “tail-up, head-down” positioning. In this particular study, a series of molecular dockings were performed prior to MD, to reveal starting points in analysis. When analyzing the potential binding poses of capsaicin to TRPV1 through metadynamics, two collective variables were defined, related to two binding pocket residues (Y511 and T550), thus enacting improved conformational sampling of capsaicin.⁷⁸ Aided by free energy calculations and by aggregating data from experimental research, the investigators provided a larger landscape of potential binding interactions, given the flexibility of the binding pocket-lining residues.

Another interesting recent study⁷⁹ showed how MD can be used in conjunction with alanine-scanning mutagenesis to reveal the binding mode, using as ligands enantiomers of bupivacaine and ropivacaine, and as target the Kv1.5 channel. The binding affinity of ropivacaine was much smaller, and, additionally, the two compounds manifested different interaction patterns related to the binding cavity residues of the Kv1.5 channel. Docking on a homology model of the channel was performed prior to 100 ns MD simulations of the channel embedded in a palmitoyl oleoyl phosphatidyl choline (POPC) membrane. Two proposed binding pockets were investigated: the central cavity and the side pockets. Small differences were seen in the binding modes in the central cavity. However, in line with the alanine-scanning, ropivacaine interacted with residues from the S4 segment, S4–S5 linker and the S5 and S6 segments of a neighbouring subunit. On the other hand, during MD simulations, bupivacaine occupied a different binding location in the side pocket, interacting with residues of the S5 segment of the same subunit. As such, Kiper and coworkers were able to rationalize a new binding mode for anaesthetics to the Kv1 channels.⁷⁹

In a different conflictual picture of ligand binding, studied in the author's lab, MD simulations helped identify a completely novel binding mode, not previously observed nor predicted.⁸⁰ In this case, the target was represented by the human ether-à-go-go related gene (hERG) channel, the pharmaceutical anti-target facilitating ventricular action potential depolarization in the heart, while possessing puzzling binding modes of the different pore-blocking chemical compounds.^81–83 Leading from two potential binding modes of the potent blocker clofilium, as presented in the literature, we sought to use the new cryo-EM hERG model,⁸⁴ combined with MD simulations, in order to investigate the known binding modalities (Fig. 2A, left). During MD simulations, however, clofilium adopted a novel positioning (Fig. 3A), which was analyzed through interaction energetics and compared to experimental mutagenesis data. The novel ligand placement was in better agreement with experimental data, especially with mutagenesis data of a small stretch of residues, placed at the bottom of the central cavity. Further scrutiny revealed an interesting correlation between a blocker's binding capacity and its resistance to mutations on the respective channel site. As anti-hERG screening must be performed in novel drug discovery experiments, this added insights to the much-needed steps in reducing hERG-related toxicity.⁸⁰

Fig. 3. A structural view of two illustrative ligand–channel interactions A. 3D structure of the hERG channel, retrieved from the SWISS-MODEL Repository¹⁰⁵ and based on its cryo-EM structure (PDB ID 5va1 (ref. 84)) used to illustrate the application of MD simulations to reveal a novel binding mode for clofilium.⁸⁰ The inset depicts the conformations of clofilium, which are translated and overlayed in the central cavity of the channel. Only two opposing subunits are shown, for clarity, one in transparent purple and the other one in rainbow solid. Lipid membrane approximate boundaries are shown in orange dotted lines. The two conflicting binding modes of clofilium are shown in red and green-dashed, respectively. During MD simulations, the “anchored channel block” in green changed its conformation to a stable, but different position (blue-dashed), which was put into experimental context. The red-dashed starting position could not be maintained.⁸⁰ B. Structural model of the TREK-2 channel (PDB ID 4xdk¹⁰³) complexed with norfluoxetine (green ball-and-stick and green arrow). Membrane location and coloring of the two subunits of the homodimer are similar to the one in A. Norfluoxetine binds to the channel, when helices M2 (solid purple) and M4 (red orange) from different subunits are splayed apart, in the “down” state. The green arrow also marks the relative location of lipid penetrance in TREK channels, while the red arrow points to the relative position of the binding sites of TKDC to TREK channels.

4. Exploration of conformational dynamics of ligand–channel systems through MD

In many cases, deciphering the binding mode of a small molecule to an ion channel is not enough to yield information about the precise mechanism through which the molecule exerts its effects, as one may need to investigate the conformational changes associated with ligand binding. Ion channels are characterized by a wide array of structural motions, which define their dynamic nature and tune ion flux, while allowing prompt response to diverse physical and molecular stimuli.⁸⁵ In the latter category, understanding the true molecular features of the interactions between small molecules and ion channels usually necessitates knowing how the intrinsic high structural plasticity of the ion channel under scrutiny either alters, or is altered by, the nature of chemical interactions.⁸⁶ Ion channel conformational motions in response to diverse inputs (such as phosphorylation, heat, mechanical motion, allosteric effects, binding of modulators etc.) occur with precision, to facilitate quick response, as ion movement across membranes occurs rapidly and its fine and prompt regulation is fundamental to cell and tissue homeostasis. Additionally, different ion channel families and subfamilies of ion channels are properly specialized. Two complex processes make ion channels quite interesting to study: ion selectivity and gating.^87,88 Ion selectivity is coordinated by perfectly designed ion conduction pathways and filters.⁸⁹ Gating allows opening and closing of a channel and three main gating classes can be inferred.⁹⁰ Ligand-gated, or type I channels, are marked by opening due to transmitter binding. The channels included in the second type of gating open and close according to pH. The larger class III gating type is attributed to channels that respond to diverse physical stimuli such as membrane voltage, tension or temperature. Ion channels are inherently dynamic, similar to other proteins manifesting crucial rapid conformational changes needed for their function.³⁸ These conformational changes can occur with various speeds, mostly on the millisecond timescale.^83,91 In this sense, accurate theoretical (such as molecular docking, MD, molecular modeling) and experimental (such as NMR spectroscopy, chemical alterations) methods are widely employed to understand and gather insights on the intricate atomic interactions governing ion channel modulation and its implications in disease.^44,92,93

As previously mentioned, MD is especially helpful when structural information is limited to site-directed mutagenesis studies, or when the binding pose of a drug is known, but its binding features or mode-of-action are poorly understood. When advances are made and the precise coordinates of the ligand (such as a drug) atoms are approximated through methods such as X-ray crystallography or cryo-EM, MD simulations of a ligand–ion channel system can shed light on the molecular features of drug action. This has been accomplished, for instance, in deciphering the action of blockers of the TWIK-related K⁺ (TREK) subfamily channels. The TREK channels are members of the two-pore domain (K_2P) family,^94,95 the latter comprising 15 different K⁺ channels in mammals, further classified in six subfamilies and sharing significant diversity, tissue distribution, and therapeutic interest.^96–98

The K_2P family was thought to include only constitutively open channels,⁹⁴ until the cryo-EM identification of TASK-1 (through X-ray diffraction⁹⁹) and TASK-2 (through cryo-EM¹⁰⁰) structures revealed formation of a lower gate. K_2Ps have particular properties, assembling as dimers and being controlled by multiple intra- or extracellular stimuli.⁹⁴ Their transmembrane helices can switch between an “up” state to a “down” state, relative to the membrane. This particular phenomenon is similar to the C-type inactivation, as its effects are transmitted upstream towards the selectivity filter, being best characterized in the TREK channels.^101,102 The “down” state corresponds to a less conductive channel, and can be stabilized by membrane compression or modulator (in this case, an inhibitor) binding.^103,104 The “down” state also correlates with the appearance of transmembrane fenestrations, lateral openings tunnelling the inner cavity to the lipid membrane and representing a common binding site for inhibitors.^103,104

Starting from experimental knowledge, MD tools allowed researchers to reveal the mechanisms of blocker action on TREK channels. While studying the effects of membrane tension on the TREK-2 channel through MD simulations, Aryal and coworkers also performed simulations of a norfluoxetine-bound channel,¹⁰⁶ an inhibitor whose binding coordinates had been revealed through X-ray diffraction (Fig. 3B).¹⁰³ During MD simulations, the blocker prevented fenestration closure due to membrane stretch and, implicitly, prevented rearrangement of the fenestration-lining helices towards an “up” state. This was in excellent agreement with functional studies.

At this point, it is important to include a discussion about the MD simulations of lipid–channel interactions. Apart from the interactions arising at the interface between membrane lipids and transmembrane core of the channels (e.g. see section 7), there are cases when lipids play more intricate roles in channel function. Multiple and diverse binding sites for a special phospholipid, phosphatidyl inositol 4,5-bisphosphate (PIP2), have been identified for various transmembrane proteins, including ion channels.¹⁰⁷ Its role in ion channels is vast, and its action can be exerted both indirectly (i.e. influencing channel expression at the plasma membrane) and directly, through molecular interaction. This leads to multiple responses, usually enabling stabilization of channels in a specific conformational state to influence gating parameters.^108,109 Through MD, researchers have been able to reveal how PIP2 migrates between domain interfaces of KCNQ and hERG channels, explaining its physiological role at the structural level. Moreover, for the calcium-permeable transient receptor potential melastatin member 8 (TRPM8), cryo-EM revealed very intricate interactions forming the basis of allosteric coupling between chemical agonists and PIP2, explaining its role from a structural point of view.¹¹⁰ Apart from PIP2, membrane lipid molecules may even protrude inside the cavity of ion channels, as evidenced from computational and microscopical investigations of the TREK channels (Fig. 3B).¹¹¹

Molecular dynamics of the TREK-1 channel formed the basis of successful discovery of a novel inhibitor, providing an excellent example of novel binding pocket finding in drug discovery using MD (Fig. 2B). In this case study, Ma and coworkers devised a drug discovery strategy, starting from careful assessment of conformational changes during the transition from the “down” to the “up” state.¹¹² They accurately observed that an intermediate state exists, which correlates with a novel druggable pocket. The novel drug-binding region was then subjected to virtual screening, whereby the binding affinity of a large collection of small molecules (potential drug candidates) is predicted through computational protocols.^113,114 Several “hits” were then put to experimental test in patch-clamp measurements; a novel inhibitor, with low-micromolar binding affinity was thus discovered.¹¹²

Bio-computational methods, including MD, served another case study in the K_2P family, by revealing a novel binding site and mode-of-action of specific blockers, enabling important observations after their prior discovery by experimental screenings¹¹⁵ (Fig. 2B). The putative binding site of a TREK blocker, named TKDC, was revealed by docking methods to probably locate in the extracellular cap of the TREK channels (Fig. 3B). Further site-directed mutagenesis and chemical modifications confirmed the binding site. Through MD, TKDC's action mode was brought to light, as it enacted a conformational shift of an α-helix that physically obstructed the ion conduction pathway. Inhibition of another member of the K_2P family, the TWIK-related, acid-sensitive K⁺ channel member 1 (TASK-1), was also investigated by means of MD simulations. The binding site of the TASK-1 inhibitor bupivacaine was also determined by site-directed mutagenesis combined with molecular docking. Next, similarly to previous studies, MD revealed the way bupivacaine prevents fenestration closure while interacting with all but two false positives of the amino acids whose mutation experimentally removed or reduced inhibition.¹¹⁶

Ligand binding-induced effects have also been studied through MD simulations in prokaryotic pentameric ligand-gated ion channels, specifically, the Gloeobacter violaceus ligand-gated ion channel (GLIC), as a model of eukaryotic counterparts,¹¹⁷ which play fundamental roles in neurotransmission.¹¹⁸ Removal of the anesthetic propofol from its crystal binding site in wild type (WT) or mutated (M205W) with subsequent comparative MD simulations revealed the mechanism through which propofol acts. The ligand-binding cavity from the M205W mutant, a mutation which potentiates propofol-induced effects, exhibits a small decrease in the volume upon modulator removal, as expected. In contrast, the WT GLIC showed a dramatic cavity volume decrease upon MD simulation of the apo system, as compared to the holo simulations. The cavity shrinkage was also associated with a drastic change in the distance from the nearest pore-lining helix. Aided by further experimental and computational mutagenesis, Heusser and collaborators created a model explaining the opposing actions of propofol acting on these specific ion channels, with strong implications in future drug design of allosteric modulators.¹¹⁷ Readers interested in finding more about the role of MD and other approaches on pentameric ligand-gated ion channels are referred to a more specific review paper.⁴⁵

5. MD simulations with external membrane potential

In certain cases, application of membrane potential as a constant in the MD simulation protocol can influence or shed light on the mechanisms of ligand-induced conformational or ion conduction changes. Such an approach may not be needed when, for instance, ion flow through the conduction pathway may not be altered by ligand binding and the ligand acts by physically interfering with ion passage in other locations. Thus, most papers reporting ligand-induced effects in ion channels do not include membrane potential in their methodological strategy. However, a series of results provided in special cases emphasize the importance of proper inclusion of external membrane potential in evaluating the interaction between small molecules and ion channels. Particularly important, evidence showed that the accuracy in conductance prediction in MD computation is very dependent on simulated timescales and the employed force field, with significant differences at the end of long MD simulations run with various force fields.⁵² Multiple replicas, longer timescales (i.e. hundreds of ns) and comparative employment of different force fields is optimum. Choice of membrane composition may also affect channel conductance as previously investigated.⁸⁵

Although simulating the effects of membrane tension in apo (no ligand bound) TREK-2 channels, this method helped reveal the mechanism through which the C-type gate functions in the selectivity filter gate of TREK channels. The higher channel conductance in the “up” state as compared to the alternative “down” state was revealed to be due to transient carbonyl flips occurring in the S3 position in the selectivity filter when the channel is in the “down” state.¹¹⁹ This carbonyl flip halts ion flux through the selectivity filter, by removing the obligatory electrostatic interactions between ions and all carbonyl moieties lining the selectivity filter. This result, stemmed from proper functional context simulations, changed the previous knowledge about carbonyl flips occurring in the (S0)/S1 site, which had been obtained without included membrane potential.¹⁰⁶

Schewe et al. demonstrated a universal mechanism underlying activation of multiple ion channels, presented as a “pharmacological master key”.¹⁰¹ The binding site of the TREK activator BL-1249 was crystallographically revealed to be placed in the central cavity, but only the position of its bromine atoms could be resolved. Further, MD solved the ligand spatial orientation and ion permeation rates were studied by including an external membrane potential. The computational observations were consistent with experimental ones. When combined with the results showing BL-1249 as a general ion channel activator, a holistic picture emerged of a universal mechanism of negatively-charged activators binding underneath the selectivity filter and acting by favoring electrostatic potential and allowing increased ion flow (Fig. 2B, upper left).

Powered by the Anton supercomputer developed by the DESRES group,¹²⁰ Song et al. performed MD simulations using timescales of 12–60 μs on the N-methyl-d-aspartate (NMDA) receptor under physiological transmembrane voltage.¹²¹ These long simulations allowed opening of the receptor and observing the close-to-natural ligand-binding process. The inhibitors, MK-801 and memantine, were introduced in the aqueous phase to travel along the ion permeation pathway, rather than being placed directly in their binding site. The two most stable poses of MK-801 during the simulations faithfully reflected the electron density observed in the crystal structure, which represented an intermediate pose.

6. Binding free energy methods of evaluation of ligand binding affinity to ion channels

Insights into the molecular interactions between a ligand and its biological target can be computationally appreciated by calculating its predicted experimental binding affinity and comparing it to actual, true values,¹⁴ which can be included in the experimental design (Fig. 2C). This can be accomplished using outputs from specially-designed MD simulations through a variety of methods, including linear interaction energy (LIE), free energy perturbation (FEP), or thermodynamic integration (TI). FEP and TI rely on specially designed MD simulations, which include, for instance, introduction of the coupling parameters; nonetheless, several preconditions must be also satisfied in LIE-related simulations.¹²² Readers are referred to key papers regarding the principles and practice of the methods of relative binding free energy computation.^14,123–125 Furthermore, these methods can also help in discriminating the subtle interaction mechanisms arising locally at the ligand binding site,⁶⁵ which is of utmost importance when the interacting residues exhibit large or complex motions, as in ion channels. This enables researchers to validate the discoveries, and has been performed, for instance, in the paper elucidating the true capsaicin binding mode.⁷⁸

Nevertheless, relative binding free energy methods, used for computing binding differences from congeneric molecules, are particularly useful when information is available for a series of either binding site residue mutations or ligand derivatives affinity.²⁰ The former application was included in a recent work conducted by Lim and coworkers on the binding of 2GBI to an open state model of the Hv1 voltage-gated proton channel, a channel facilitating proton efflux out of the cell and its organelles and contributing to sperm maturation and reactive oxygen species production.¹²⁶ A good MD starting position of 2GBI was obtained through docking and comparisons with known data from previous 2GBI mutations (Fig. 2C, right, presents a simplified experimental design). Next, authors performed relative binding free energies calculations for 2GBI binding to WT and six Hv1 mutants, obtaining a good overall correlation between experimental and theoretical affinities. An improved correlation was then obtained through careful treatment of positively-charged residues' movement during alchemical FEP, thus precisely confirming the binding position taken by 2GBI to the Hv1 proton channel.¹²⁶

A similar approach was followed by Boukharta et al. (Fig. 2C, left), but in this case, the binding affinity of sertindole and eight derivatives thereof had been used in comparisons.¹²⁷ The hERG channel represented the ion channel under study, and, after docking and MD simulations, the relative binding free energies were computed. A high positive correlation (i.e. r² = 0.6, r = 0.77, 0.7 < r < 0.9 (ref. 128)) was noticed between experimental and theoretical (generated through LIE method) inhibition coefficients, much improved compared to the docking affinity predictions. Alongside, the correlation could well discriminate between low- and high-affinity derivatives; also, a good Spearman rank coefficient of 0.85 was calculated, to confirm the successful identification of the binding modes. One of the analogues (numbered 3) did not attain a low-energy conformation during initial MD simulations, as it was trapped in a high-energy state. Free energy perturbation was employed, to model an alchemical transformation (compound 10 to compound 3). This revealed the correct binding mode of compound 3 and attributed its initial behavior to unfavorable interaction with cavity water molecules.¹²⁷

7. Some technical considerations concerning MD simulations of ion channels

Conducting an MD simulation to scrutinize the interactions between proteins and small, drug-like molecules, has traditionally been a daunting, labor-intensive task. Nowadays, much-needed semi-automation tools are available, making the endeavors of bio-computational researchers much easier. In the field of membrane proteins this is indeed true, as web servers (e.g. ref. 24 and 129), standalone software (e.g. ref. 130 and 131), tutorials (e.g. ref. 132–135), scripts (e.g. ref. 136 and 137), or other tools aid in the various tasks of building a full MD system with ease, making MD simulations available even to non-informaticians or experimentalists.

Simulating membrane potential for studying the electrophysiology of ion channels is usually performed straightforward in MD engines, using their specific parameters. Of note here is the Computational Electrophysiology (CompEL) protocol implemented in GROMACS.¹³⁷ CompEL performs this task by enacting a small imbalance of charges across the membrane, which in turn provides a potential difference. Two lipid membranes are included in the simulation, with replicas of the studied channel. These lipid membranes are stacked, and, using the periodic boundary conditions, two aqueous compartments are created. Depending on the positioning of the channels in the membrane (i.e. parallel or antiparallel), different approaches on ion flow may be investigated. Despite the large simulation box, attributed to the inclusion of two membrane lipids instead of one, the computational efficiency is unchanged; ion permeation rates are measured twice per simulation time. To avoid voltage depletion due to ion translocation, ion count is monitored and readily adjusted by ion/water position exchanges between compartments, as needed. MD system preparation for CompEL is straightforward, by performing few additional command-line steps; plus, the output is generated in a user-friendly fashion and includes ion permeation events, either in aggregated form, or, specifically, per ion and channel type, when discriminating between the two ion channel copies is needed, as in studying rectification effects.

Pure membranes consisting only of POPC molecules can be included in the simulations, especially if classical membrane builders, such as the Visual Molecular Dynamics (VMD) Membrane Builder module,¹³⁰ are used. However, choice of lipid membrane composition can be important, as biological membranes are typically composed of different lipid molecules in proportions varying with leaflet (inner vs. outer), membrane localization, or cell type.^138,139 This can lead to improper sampling, especially when lipophilic drugs and their mechanism of action are simulated,^77,140 or in special cases, where channel–membrane interactions has functional effects.^52,53,141 The activity of membrane proteins, including ion channels such as voltage-gated potassium channels, is influenced by the lipid bilayer distribution, which may either interact directly with the protein, or generate ionic or electrostatic imbalances.^53,142 As explained in detail elsewhere,^85,107 lipid composition is very important as proven by several experimental and computational studies of membrane proteins, in order to accurately model their functional parameters. When needed, more sophisticated programs aid the building of accurate heterogenous lipid membranes, including here the GROMACS-based MemBuilder.²⁵

However, CHARMM-GUI stands out as, probably, the most complete suite of utilities for semi-automatic creation of MD simulation systems. One of these, the CHARMM-GUI Membrane Builder^24,143 allows, at the time of this writing, creating a heterogenous lipid membrane in user-specified proportions, composed of either of 434 different lipid molecules. Its user-friendly on-line interface also facilitates precise membrane placement relative to the simulated protein, and its output files can be tailored to a wide variety of MD simulation programs, including NAMD, AMBER, or GROMACS.^143,144 Especially for our interest in channel–ligand simulations, CHARMM-GUI Membrane Builder even permits inclusion of small, drug-like molecules in the simulation systems, and proper management thereof,^24,145 as evidenced in Table 1. It consists of a six-step protocol, all performed on-line in a very intuitive framework. The initial file manipulation (step 0) enables protein modifications, such as mutations, setting protonation states, disulphide bonds, glycosylphosphatidylinositol (GPI) anchoring, and complex glycosylated moieties. Step 1 relates to the membrane orientation procedure, using various modifications that allow fine adjustments, as needed. In step 2, the membrane size and composition can be precisely altered and the cross-sectional lipid/protein area can be investigated as a check. In step 3 the researcher is shown the estimative system size and he may choose the salt to be included and ion concentration of the aqueous solution. Either protein insertion or replacement method may be selected for the algorithm of lipid/protein overlap. The final steps allow for specifying simulations conditions, including temperature, check of lipid ring penetration, or choosing of desired simulation program output. CHARMM-GUI Membrane Builder also features algorithms for highly mobile membrane-mimietic (HMMM) models, Nanodisc Builder, Monolayer Builder, Micelle Builder and Hex Phase Builder.²⁴ Other semi-automation tools, included in the CHARMM-GUI suite¹⁴⁶ but also in various standalone tools,^20,147 enable researchers to perform free energy calculations from MD trajectories.

Commonly used force fields for small molecules.

Force field (FF) for small (drug-like) molecules	Compatibility (families of force fields)	Automatic parameter assignment programs and algorithms	Observations
OPLS-all-atom (OPLS-AA),154 OPLS3 (ref. 155)	OPLS156	LigParGen157	—
CHARMM generalized force field (CGenFF)158–160	CHARMM160–162	ParamChem158,159	ParamChem generates penalty scores that need careful examination post-parameter assignment;158,159 ParamChem may be employed via CHARMM-GUI Membrane Builder24
CHARMM162,163	CHARMM160–162	MATCH164	The MATCH algorithm works by extending the CHARMM force field to any novel molecule;164 MATCH-provided files can be manually uploaded to CHARMM-GUI Membrane Builder, after atom type renaming
General AMBER force field (GAFF),165,166 GAFF2 (ref. 167 and 168)	AMBER50	Antechamber,166 PrimaDORAC169	Antechamber may be employed via CHARMM-GUI Membrane Builder;24 GAFF2 could be employed with caution170
Merck molecular force field (MMFF)171	CHARMM, GROMOS172	SwissParam172	SwissParam combines “bonded” parameters from MMFF and “non-bonded” parameters from the CHARMM22 force field;172 MMFF94S could be employed with caution170
GROMOS (specific force fields)173	GROMOS174	ATB,175 PRODRG176,177	Should be used cautiously, may need careful examination post-parameter assignment178
smirnoff99Frosst179	AMBER179	Open Force Field toolkit	Open Force Field toolkit may be used for simulation with OpenMM; ParmEd may be used for converting to other simulation formats (https://openforcefield.org/force-fields/force-fields/)
—	AMBER,50 CHARMM,160–162 OPLS156 and GLYCAM180	PyRED, through the R.E.D. Server interface181	“R.E.D. Server provides the resources […] required for charge derivation and force field library building to any computational biologist, who wishes to use MEP-based atomic charges in MD simulations”181

Prior to MD simulations of a ligand–channel system, a good quality protein model of the ion channel has to be generated. This first step is best accomplished using, when available, 3D structures of good resolution of target proteins, determined through experimental procedures and deposited on the Protein Data Bank.^148–150 Missing loops or amino acids can be included, with cautionary checks, with the aid of molecular modeling applications.^151,152 But, if the native structure of the channel had not been determined through experimental procedures, the task of homology (comparative) modeling is performed. Generally, such a method should be chosen with extreme caution, using good alignments and templates, with post-modeling tests to assess the quality of the protein model.¹⁵³

The second crucial step in properly simulating ligand–channel interactions is the ligand parameterization. Ligand parameterization involves employing a specific force field describing the potential energy function of all “non-bonded” (usually electrostatic and Lennard-Jones) and “bonded” (harmonic potentials for bonds, angles, dihedrals), governing the molecular geometry and structural integrity while reproducing observed physical properties. With considerable efforts, such force fields have been designed specifically for small molecules, and may only be used if other system components (e.g., lipids, proteins) have been parameterized using force fields derived in similar modes.^160,182Table 1 presents the commonly used force fields describing the atomistic behavior of small, drug-like, organic molecules and general considerations concerning their implementation. The force fields and algorithms associated are constantly evolving, as new MD methods are developed at a rapid pace.^179,183,184

Nonetheless, various benchmarks using different force fields, for small molecules^170,185 and all-atom^52,186 alike, reveal consistent discrepancies among the resulting simulations, generating concerns related to this particular challenging aspect surrounding MD simulations methodology.^52,187,188 Force field development for MD simulations is a matured field, yet new frontiers are sought by researchers.¹⁸⁹ As also depicted in the observations in Table 1, a cautionary stance should be employed when choosing the force field, evaluating previous knowledge from the literature and the specifics of the system to be simulated.

Despite the abundance of tools for semi-automation of MD protocols, researchers' own adaptations, analyses, and implementation thereof cannot be substituted. When prior information is available, the MD methodology should be appropriately adapted.^78,80,101 Additionally, in the previously-mentioned case studies, a number of situations arose where the authors had to intervene in order to properly modify the general workflow when studying ligand–channel interactions. These included careful analysis of trajectory equilibrium and its manipulation through free energy perturbation,^126,127 mixed theoretical–experimental procedures,¹¹⁵ or computational residue-by-residue mutational analysis.¹¹⁷

A discussion about the timescales involved in MD simulations is also important, as they should match the biological reality, to allow observing trajectory convergence, thermodynamic stability, and faithful interactions during the course of the simulation. Typically, in the case of protein–ligand interactions, the simulations are performed in nanosecond timescales, as the protein–ligand association step, normally taking place over the course of milliseconds to seconds, is performed instantaneously through docking or manual ligand placement.¹⁹⁰ Unbiased production simulations usually reach equilibrium in about 3 ns to 30 ns, as reported in several case studies.^78,80,127 Simulations may be extended on scales of 30 ns to 100 ns, to further validate stability.^80,101 However, longer simulation times (e.g. 500 ns) are needed to reveal ligand-induced conformational rearrangements,¹¹⁵ or ion flow rate modification.^101,119 Microsecond simulations may permit analyzing larger conformational changes and movements of the ligands under scrutiny.¹²¹ Multiple replicas allow increasing certainty of the results and distinguishing between true or false poses, thus offering scientific confidence of discoveries.¹⁹¹

8. MD analysis tools

Aside from the range of tools available for purposeful MD system building, a number of other pieces of software aid in analysing the post-simulation data. This section provides a non-exhaustive overview on the most popular software varieties implemented in analysis of the MD trajectories, with a focus on ligand–channel systems.

Protein tunnels, cavities, pores, pockets and other openings have special interest in drug design and discovery of new modulators, as they may bind chemical species in specific ways, that may be successfully employed for developing new drugs. A large variety of algorithms exist for identification, visualization and analysis of protein tunnels and channels, with representative examples of HOLE¹⁹² and CAVER^193,194 (see ref. 195 for an organized collection). These facilitate observing the dynamics of conformational changes associated with protein dynamics, and play important roles in probing the sizes and dynamics of selectivity filter, cavities, and pores of ion channels. In certain cases, such tunnels may act as binding sites for ligands, which act by obstructing ion flow.¹⁹⁶

The central cavity present in most ion channels provides a common binding site for ligands that may either activate¹⁰¹ or inhibit^80,94 channel function. Vestibules for binding of chemical modulators usually form at the boundaries of the central cavity, offering new advents in drug discovery and design.^83,112,197 To this respect, algorithms and software for identifying more or less cryptic and druggable binding sites and their subsequent analysis have been developed by the community, such as the more popular LIGSITE^CSC, Fpocket, CriticalFinder or metaPocket.^{195,198–201} The cavity-identification algorithm of Fpocket has been used, for instance, by Ma et al. to observe the transient pockets revealed during MD simulations in the intermediate states associated with conformational changes of the TREK-1 channel.¹¹²

Other tools that may be used post-simulation are included in toolboxes for running and visualizing MD trajectories. VMD performs this task, using inputs from a large variety of simulation engines; it may perform multiple post-simulation tasks, such as computing various parameters for simulation convergence, analyzing hydrogen bonds, or evaluating interaction energetics. Its functionality can be extended with the use of plugins, such as CaFe for end-point free energy methods¹³⁶ or VMD Store, an open-source plugin for VMD that retrieves VMD extensions posted on GitHub, for facile extension employment in the VMD environment.²⁰² AmberTools is part of the Amber suite, and consists in a series of packages used to carry out simulations with AMBER and analyze their outputs.^50,131 The MD-simulator GROMACS has built-in functionalities for analyzing the trajectories, including the package ngmx.⁵¹ MDAnalysis is a Python package widely used for handling and analyzing trajectories from CHARMM, GROMACS, AMBER, and NAMD.²⁰³ Apart from basic utilities, it has commands for visualization of the flow of lipids in the membrane bilayer patch and serves as back end for other packages. MDtraj has excellent performances in manipulating and converting trajectories from various formats.²⁰⁴

Post-MD, a crucial process is the investigation of the ligand–protein interactions. Visual inspection may not suffice; the previously-mentioned packages may be employed for use in this process, alongside other specifically-build programs. The Simulation Interactions Diagram (SID) is integrated with the Desmond MD engine from the Schrödinger molecular analysis platform. It allows interactive analysis of ligand–protein post-MD analysis.²⁰⁵ Snapshots from any ligand–channel simulations may be investigated using third-party tools, such as the Protein–Ligand Interaction Profiler (PLIP), which summarizes and visually presents all the molecular interaction between the given ligand to the target residues.²⁰⁶ Last but certainly not least, PlayMolecule.org stores several newly-designed web applications that help computational biologists in protein and membrane preparation pre-MD,^207,208 discovering novel binding pockets,²⁰⁹ or assessing energetics of ligand binding to its target.^210,211

9. Developments, trends and future directions

The research area surrounding MD applications on ligand–channel interactions has seen a steady, increasing interest in the past years, with significant published results stemming from accurate and complex MD protocols, proper integration with experimental data, and a large decrease in associated costs. By carefully aggregating, interpreting and implementing data from the literature in their MD simulations, medicinal chemistry scientists benefit from a complementing arsenal of MD methods to put experimental approaches into proper structural and functional context and to accurately validate and identify the way drug-like molecules exert their effects through binding to ion channels.

Algorithms and force fields alike, describing atomistic interactions, are given particular interest from the scientific community, which leads to continuous improvements in accurate modeling of the related bio-processes. This translates into the existence of a large diversity of MD simulations techniques, forbye associated tools, providing new opportunities for simulating more complex mechanistic processes, such as those underlying drug-channel binding and transduction of effects.

Combined with ever-decreasing costs of MD simulations and wide accessibility, taking advantage of the surge in the number of 3D structures of ion channels, this particular methodological approach is able to generate a vast, yet uncharted, exploration area in molecular structural biology and drug design. MD not only evaluates the findings of experimental screenings, but may also be used to infer new properties of biological molecules, thus having the potential to lead, in the future, more and more drug discovery processes.

10. Conclusions

With considerable success, proven by numerous case studies, MD methods can be used in every step of drug discovery and design, from identification of novel druggable regions to shedding light on the mode-of-action of potential drugs. MD simulations offer a powerful, yet affordable solution to reveal the mechanics of complex life and disease processes, such as the ones exhibited by ion channels and their chemical modulation. New and improved MD methods and tools, unprecedented timescales, and semi-automation protocols facilitate investigation of complex functions and chemical regulations of ion channels. All considered, MD methods should be taken into serious consideration in every research study involving investigation of ligand–channel complexes, including exploratory research as well as therapeutic targeting.

Conflicts of interest

There are no conflicts to declare.

Biography

Daniel Şterbuleac.

Daniel Şterbuleac, PhD, is Assistant Professor and Researcher at the “Ştefan cel Mare” University of Suceava, Romania. He obtained his PhD in Biology, from the “Alexandru Ioan Cuza” University of Iaşi, Romania, in 2020, studying the role of ion channels as potential targets in antitumour therapy, using a combination of abductive reasoning and computational biology methods. His current research interests include the computational structural biology of ion channels and cancer therapy through ion channel modulation, while sharing a strong passion for interdisciplinary approaches in the fields of environmental education and macroeconomics.