Cell-permeable chameleonic peptides: exploiting conformational dynamics in de novo cyclic peptide design

Abstract

Membrane-traversing peptides offer opportunities for targeting intracellular proteins and oral delivery. Despite progress in understanding the mechanisms underlying membrane traversal in natural cell-permeable peptides, there are still several challenges to developing designed membrane-traversing peptides with diverse shapes and sizes. Conformational flexibility appears to be a key determinant of membrane permeability of large macrocycles. We review recent developments in the design and validation of chameleonic cyclic peptides, which can switch between alternative conformations to enable improved permeability through cell membranes, while still maintaining reasonable solubility and exposed polar functional groups for target protein binding. Finally, we discuss the principles, strategies, and practical considerations for rational design, discovery, and validation of permeable chameleonic peptides.

Introduction

Advances in high-throughput screening methods, rational design, and formulation strategies have enabled the development of peptide-based lead candidates against diverse therapeutic targets considered non-ideal for conventional drug modalities [1–5]. However, the limited cell permeability, oral bioavailability, and pharmacokinetics of peptides pose challenges to their translation from research tools to therapeutics [6,7]. Consequently, there has been a renewed focus on identifying general principles for developing peptides with enhanced cell penetration, oral bioavailability, and physiologic stability, while retaining reasonable solubility and target binding specificity. Naturally occurring cell-penetrating peptides exploit direct translocation, endocytosis, and passive diffusion for cell permeability (reviewed in ref [8]). The underlying chemical and structural requirements for leveraging these routes rely on size, shape, polarity, and chemical composition. The role of amphipathic and polycationic cell-penetrating peptides in endocytosis has been reviewed extensively [8–12]. Here we focus on passive permeability and, specifically, on the role of conformational switching in passive membrane traversal.

Generally, peptides with low polar surface area (PSA) show an improved passive permeability [6,13–16]. Low PSA can be achieved by masking backbone amides through internal hydrogen bonding, N-methylation, or shielding by bulky nonpolar amino-acid side chains [7,16–21]. Many research groups over the last two decades have focused on exploring the limits of passive permeability and defining the features that promote passive diffusion [14,18–22]. The rates of passive permeability depend on size, polarity, and conformational flexibility [6,13,15,23]. The constraints on these properties are conceptually similar to those of small molecules, defined by Lipinsky’s rule of five (Ro5) [24]. However, peptides are generally larger and have more hydrogen bond donors and acceptors than the limits prescribed by the Ro5. Features governing their permeability are more complex [7,8]. Although most of these studies focused on small macrocycles composed of 4–7 amino acids, recent work by Lokey and coworkers has identified larger membrane-traversing peptides with molecular weights (MW) greater than 1000 Da [25,26]. We recently described a de novo design approach for accurately designing membrane-traversing macrocycles with diverse shapes and sizes between 6–12 amino acids and MW up to 1400 Da, which have conformations with backbone amides forming internal hydrogen bonds [27].

Unfortunately, the same features that increase passive permeability also lead to excessive hydrophobicity, resulting in poor solubility and decreased target selectivity [28]. To address this problem, chameleonic peptides are under development that can adopt significantly different conformations in the aqueous and membrane environments: the more polar conformation inside the cell with an exposed backbone allows for hydrogen bonding interactions with the solvent or the target of interest, and a less polar state within the membrane with all (or most) polar groups satisfied or shielded favors membrane traversal. Cyclosporin is the prototypical example, with chameleonic character that enables it to overcome the permeability barriers [29]. Multiple structural and in silico molecular dynamics (MD) studies of the 11-amino acid macrocycle cyclosporin in different solvents and membranes have highlighted the critical role of alternate conformations and conformational dynamics in its permeability [30,31]. However, developing new chameleonic membrane-traversing peptides in medium-to-high throughput will require structure-guided computational tools, high-throughput methods for parallel permeability measurements, and methods for structural characterization of multiple conformations.

Computational design of conformationally-switching peptides

In silico methods that calculate the difference in solvation free energy for a predicted low-energy conformer (or ensemble of conformers) going from a high-dielectric to low-dielectric solvents have been used to predict the cell permeability of macrocycles [32–34]. Monovich and coworkers combined such methods with analysis of intermolecular hydrogen bonds to design and synthesize orally bioavailable thioether-cyclized macrocycles between 6–8 amino acids [35,36]. Accurate low-energy conformations required for such permeability predictions can be generated using de novo design approaches, MD simulations, or deep-learning-guided models [37–39].

De novo peptide design methods offer opportunities to systematically evaluate structural features that could improve permeability. We have previously described computational methods for accurately designing peptides composed of canonical and noncanonical amino acids using Rosetta [38]. This approach relies on extensive in silico backbone sampling followed by the design of amino acid sequences that minimize the energy—as calculated by the Rosetta energy function—for each selected backbone. We recently expanded this method to design 6–12 amino acid membrane-traversing macrocycles that included N-methylated amino acids; a subset of which were designed to switch conformations in a solvent-dependent manner (Figure 1A-C) [27]. Generally, computational design methods rely on sampling structures with a single deep well in their energy landscapes [38]. In contrast, for chameleonic peptide design, we focused on generating designs with two or three distinct isoenergetic states in their structure-energy landscapes (Figure 1C). Other key criteria is to have one of the conformations with all backbone amides hydrogen-bonded or N-methylated for favorable membrane permeability [27]. We relied on cis-trans isomerization of peptide bonds for achieving multiple conformations and switching in our designed macrocycles. However, other strategies could also be employed to drive conformation-switching, including metal binding [40], light-sensitive cross-linkers [41], pH-dependent changes to sidechain protonation, or oligomerization. These switching mechanisms may involve different timescales and require appropriate experimental methods for validating multiple states. Improvements in the accuracy of energy functions are needed to further increase confidence in the calculations of isoenergetic states and improve the predictability of peptide designs with multiple local energy minima.

Figure 1. Computational Design of Chameleonic Peptides.

A. Proposed method for passive permeability in chameleonic peptides. Macrocyclic peptides exist as an equilibrium of conformations. When contacting a hydrophobic cell membrane, the equilibrium shifts towards the lowest PSA conformation, providing passive permeability. B. Computational design of macrocyclic peptides begins by generating thousands, and sometimes millions, of designs. This library of designs is filtered by scoring metrics. C. Energy prediction landscapes help identify chameleonic peptides. Traditional design calls for a “funnel” shape, resulting in a single low energy minima. Chameleonic design looks for two or more deep energy minima. D. Transwell assays provide in vitro data on permeability. Peptides added to the donor well diffuse across the membrane and into the acceptor well. Mass spectrometry measures the amount of crossing to estimate permeability. E. Macrocyclic peptides that demonstrate permeability are structurally validated using NMR and X-ray crystallography. NMR is capable of capturing the signatures of multiple peptide conformations as they interconvert in solution. In some cases, different solvents can be used to shift the equilibrium towards one conformation or another. Similarly, different crystallography solvents can be used to capture static images of chameleonic peptides in alternative conformations.

MD simulations provide an orthogonal approach to Rosetta-based methods and enable studies of chameleonic character [42–46]. Lin and coworkers have used MD simulations in their successful structure prediction and design of chameleonic peptide macrocycles [42,47], including peptides with N-methylated amino acids to promote membrane permeability [48]. MD trajectories can also predict the relative populations of conformers in solution as was done for cyclosporin [37,42,43,49]. In principle, MD can be used to model the conformations of peptides in different solvents or membranes with different compositions and predict permeability [50]; however, such calculations or simulations are computationally very intensive and cannot currently be used to screen thousands of design models [43]. Going forward, approaches that combine MD simulations with machine learning (ML) could provide the basis for rapid and accurate design of chameleonic cell-permeable macrocycles [37].

ML-based structure prediction networks for larger proteins have seen tremendous success over the last few years [51,52]. Although these deep learning methods may not generalize to small peptides as they are trained on larger protein structures composed mostly of natural L-amino acids [51,52], recent work by Gulsevin and Meiler suggests some limited predictive value for predicting small, rigid polypeptide structures [53]. Approaches that combine information and training from small molecules and proteins could guide ML methods for designing new peptides with desired properties, such as membrane permeability and oral bioavailability. Data from computationally-intensive QM calculations or MD simulations can also be used as training sets to develop ML models that can predict the states populated by a given amino acid sequence without running computationally expensive molecular simulations for each new peptide [37]. It remains to be seen how the accuracy of such methods translates to sequences that differ substantially from the training set in their chemical and structural diversity. Another potential approach is to skip structure prediction and train models that predict a property, such as membrane permeability, from the primary sequence alone. However, it may be challenging to deconvolute the diverse mechanisms and structural principles that underlie permeability from such sequence-driven property prediction models. Given the recent transformative success of generative models for larger proteins [54], an important area to explore will be the development of such models—simultaneously trained on proteins, small molecules, and peptides—to hallucinate polypeptides with desired chemical, structural, and functional properties.

Experimental methods for testing permeability:

There are several approaches for measuring cell penetration of peptides by tagging them with fluorescent dyes or transcriptional modulators [55], and tracking the intracellular localization by fluorescence microscopy [56,57], flow cytometry [58], fluorescence correlation spectroscopy [59], or some transcriptional readout [57,60]. However, there are concerns about the throughput of these assays and the influence of the tag on the permeability, solubility, and overall structure of the peptide being tested. Newer methods that use smaller tags [61,62] and rely on fluorescence measurements or cell sorting address some of these concerns, and enable higher throughput in permeability measurements. Alternatively, mass spectrometry (MS)-based transwell assays, such as Parallel Artificial Membrane Permeation Assay (PAMPA), enable tag-free measurement of permeability across a lipid membrane barrier between the donor and acceptor wells [63] (Figure 1D). MS-based quantification allows for high-throughput permeability measurements of thousands of peptides; however, PAMPA can be only be used for measuring passive membrane traversal. In contrast, transwell Caco-2 and MDCK assays use intestinal epithelial cells and canine kidney cells as permeation barriers [64] to appropriately account for efflux in permeability measurements, and are very useful in predicting oral bioavailability in vivo. Overall, these diverse set of established methods for measuring permeability and traversal across different cell types provide the basis for developing design principles for accessing specific cell and tissue types.

Dynamic structural analysis of chameleonic peptides

Structure determination can provide validation and new insights critical for de novo protein design (Figure 1E). Often chameleonic peptides can be crystallized from solvents of different polarities, and 3D conformations determined using X-ray crystallography. However, the crystal lattice may favor particular conformations, and it may not be possible to capture the full range of conformations present in the solution. Solution NMR studies can provide accurate structures of multiple conformations in dynamic equilibrium, as well as additional information about populations, relative free energies, and rates of interconversion between conformational states. Particularly for studies of multiple conformations, many research groups have turned to NMR studies in polar and nonpolar solvents.

NMR is especially useful for assessing intramolecular hydrogen bonds, which are essential for the characterization of chameleonic states. These include measurements of amide proton temperature coefficients by variable temperature NMR, amide hydrogen/deuterium exchange (HDX) rates, and in the case of a faster exchange, amide proton saturation transfer [65,66]. In some cases, hydrogen-bond analysis alone may provide the key insights needed for guiding design.

Solvent-dependent conformational changes are easiest to identify by 1D ¹H NMR in cases where two or more distinct populations interconvert slowly on the chemical shift timescale (e.g., longer lived than ~ 10 ms for typical ¹H chemical shift differences between conformers at 600 MHz), resulting in at least double the number of expected resonances (Figure 2A). Chameleonic peptides involving peptide bond isomerization typically have slow enough exchange rates to provide a discrete set of resonances for each peptide bond conformer. Relative populations of different conformations at equilibrium can typically be determined by integration of peaks, and conformational switching can be monitored by comparing spectral data recorded in different solvents and temperatures.

Figure 2. Solvent-Dependent Conformational Switching.

A. 1D ¹H-NMR data for chameleonic cyclic peptide D9.16 collected in CDCl₃ and DMSO-d₆ at 20 oC and 800 MHz. B. 2D ROESY (250 ms mixing time) data in CDCl₃ with schematic showing how CYANA was used to model two conformations simultaneously from the complete ROESY peak list, chemical shifts, and ³J_HNHA -based dihedral restraints, using tandem cyclic peptide sequences representing the between the trans-trans (55%) and cis-trans (45%) conformations (lower case residues are D-amino acids, * denotes N-methylated amino-acid positions, cP is cis L-Pro-8, and (xxx)₃ denotes a 9-residue sterically-transparent linker). Molar percentages given in parenthesis were determined from averaged peak integrations of the 1D spectra. C. Design models and experimentally-determine structures (X-ray and NMR) for different conformational states of D9.16. The design models are in the top row. The middle row shows the NMR structure in polar solvent DMSO-d₆ and the X-ray crystal structure, which both match the trans-cis design model (upper right) with 3 unsatisfied hydrogen bonds. In non-polar solvent (CDCl₃), two slowly-exchanging states in equilibrium were observed (bottom panel), a trans-cis state that matches the structures observed in DMSO-d₆, X-ray crystallography, and the trans-cis design model, and a trans-trans state that matches closely the trans-trans design model, with all backbone NHs satisfied. These studies show that the population of the trans-cis state observed in the X-ray crystal structure decreases from 100% in polar solvent to 45% in non-polar solvent, as 3 NH groups are buried by intra-peptide hydrogen bonds, providing the basis for conformational-switch-mediated membrane permeability.

When there is only a single set of chemical shifts(Figure 2A), NMR structure determination can be pursued with traditional approaches using restraint-satisfaction structure-generation methods and software such as CYANA [67], CNS [68], Xplor-NIH [69], and Surflex [70]. NMR-derived data typically includes chemical shift assignments, NOESY/ROESY distance restraints, and dihedral angles constraints from ³J coupling constants. Residual dipolar coupling (RDCs), amide proton temperature coefficient, and amide hydrogen exchange data are also used, as reviewed recently [66]. For chameleonic peptide D9.16 in DMSO, a standard CYANA protocol was used to interpret and calibrate NOE-based constraints and to model an ensemble of low energy conformers that satisfy the NOE and dihedral restraints (Figure 2B) [27].

In the case of slow-exchanging multiple conformations, structure analysis is further complicated because the data for both conformations are present in NMR data collected from a single sample (Figure 2B). Chemical shifts and NOE/ROE restraints need to be separated for each set of slow-exchanging isomeric conformers. If the ratios of two switching conformations are close to unity, NOEs for both forms can be deconvoluted from NOESY or ROESY peak lists simultaneously, as was done for cyclic peptide D9.16 in CDCl₃ (Figure 2). Difficulties arise when one conformation is significantly less populated than the other, as the subsequent calibration of NOE peak volumes (or intensities) into distance restraints needs to be appropriately population-weighted. However, it is often impossible to determine structures for low abundance conformations, where through-bond (TOCSY) and through-space (NOESY) cross peaks may be obscured by the more abundant conformer.

Even when the data can be interpreted as a single conformation that satisfies all of the NMR-based distance restraints, the molecule may in fact adopt multiple conformations in fast exchange, resulting in population-weighted averaging for all the peaks in the NMR spectra. This is further complicated by the fact that peptide systems may provide only a small number of NOESY/ROESY peaks, resulting in a sparse network of distance restraints, which can be satisfied by a single-conformation model even when the underlying system is dynamic. One potentially powerful approach to this challenge is NMR Analysis of Molecular Flexibility in Solution (NAMFIS) [71], in which modeling is used to generate a ‘complete’ structural ensemble, and back-calculated NMR data (e.g. NOEs and J coupling) are used to rule in or rule out members of this predicted ensemble and to fit relative populations, as in recent studies of macrocyclic drugs [72,73]. In a similar way, advanced conformational sampling by the program ForceGen can be used to obtain an ensemble of NMR-validated structures [70,74]. A related approach by the Lokey lab used only ¹H and ¹³C chemical shift data, without any NOESY/ROESY data, to simultaneously determine the structures of both major and minor population cyclic peptides [75]. Here, Quantum Mechanics (QM) was used to back-calculate population-weighted chemical shifts for a family of MD-generated conformers, allowing identification of the conformational ensemble best-fitting the NMR data. Such MD and QM approaches are computationally expensive, but can exploit ergotic ensemble-averaged chemical shifts. However, it is important to ensure comprehensive conformational sampling when using such modeling methods.

Another challenge for NMR studies is determination of molecular topology and templates for non-canonical amino acids, such as N-methylated and D-amino acids, and cyclization. The Automated Topology Builder (ATB) server has been developed to address some of these problems [76]. Larger polypeptides and miniproteins have greater spectral overlap, making simultaneous structural analysis of multiple conformations even more challenging. This problem is addressed in proteins by biosynthetic enrichment with ¹³C and ¹⁵N, allowing for additional dimensional separation in 3D/4D NMR data collection. Unfortunately, in the case of synthetic chameleonic peptides containing non-canonical amino acids, it can be prohibitively expensive to obtain labeled precursors for incorporation into peptides.

Although there are many advantages in NMR studies of slowly-exchanging multiple conformations, these methods are generally slow and expensive. HDX-MS is another powerful approach for characterizing the solvent-exposure of amide protons, as highlighted by recent studies of cyclosporin [77,78]. Here, conformers of cyclosporin analogs were separated using liquid chromatography and further separated in the gas phase by ion mobility (IM) or shape differential mobility spectroscopy (DMS), prior to MS/MS detection [77,78]. HDX-MS was used to determine the amount of gas-phase HDX, which correlated with differences in intermolecular hydrogen bonding. A related approach is pulsed HDX-MS, which has been used to investigate amide hydrogen-bonding behavior in proteins and peptides (reviewed in [79]). These MS approaches can potentially be extended to high-throughput analysis of cyclic peptides, providing key insights prior to more extensive NMR studies.

Conclusions

Conformational flexibility has emerged as a key feature for peptides with enhanced passive permeability [6,73]. Chameleonic character provides a path towards addressing the critical challenge of balancing the overall hydrophobicity required to cross non-polar cellular membranes with reasonable solubility and polarity required to bind therapeutic targets. However, translating the idea of conformational flexibility to a broadly-applicable design principle or a prediction metric for membrane-traversing peptides will require designing and testing tens of thousands of peptides for permeability and conformational flexibility. Recent advances in structure-guided design methods have enabled precise design of membrane-traversing and orally bioavailable peptides [27,35]. These computational methods enable rapid enumeration of the structural and chemical diversity of chameleonic peptides; however, they will need to be integrated with high-throughput methods for measuring membrane permeability and chameleonic behavior. Advances in the peptide library screening approaches [2,80] and high-throughput methods for measuring permeability [26,61,62] have been leveraged to diversify validated membrane-traversing peptides and explore their mutational tolerance. Biophysical studies of the conformational dynamics of chameleonic behavior present unique opportunities and challenges, as outlined above. High-throughput HDX-MS studies can potentially be leveraged as an initial coarse-grained filter for structural flexibility prior to more extensive characterization with NMR and X-ray crystallography. Future studies that combine rational design with high-throughput experimental methods will be instrumental in generating extensive scaffold sets of membrane-traversing chameleonic peptides. While the application of machine learning for peptide design is still in its infancy [53], we will see a larger role going forward of such methods in both rational design and NMR data interpretation. The key will be to integrate these diverse approaches to develop chameleonic peptides that can bind to intracellular therapeutic targets with high potency and selectivity while retaining membrane permeability, solubility, and metabolic stability.