Multiscale Simulation of Stratum Corneum Lipid Mixtures: Effects of Ceramide Headgroups on Structural Organization and Hydrogen Bonding Networks

Abstract

The barrier function of the outermost layer of human skin, the stratum corneum (SC), arises from its multilamellar lipid matrix composed primarily of ceramides (CERs), cholesterol (CHOL), and free fatty acids (FFAs). Coarse-grained (CG) and atomistic molecular dynamics simulations have been used to study self-assembled multilayers comprising CERs NS, NP, AS, and AP, in pure CER systems and mixtures of CERs with CHOL and FFAs. Equilibrated CG configurations were reverse-mapped to recover atomistic details and analyzed to extract structures and hydrogen bonding. Simulations of pure CERs agreed with experimental trends: phytosphingosine CERs (NP and AP) exhibited more C=O hydrogen bonds, consistent with lower amide I FTIR frequencies, than their sphingosine counterparts (NS and AS). Likewise, non-hydroxy CERs (NS and NP) exhibited more C=O hydrogen bonding than their α-hydroxy analogs (AS and AP). CER mixtures with CHOL and FFA showed reduced C=O hydrogen bonding compared to pure CERs, though this effect depended on water content. Hydroxyl location was critical: OH on the phytosphingosine base increased C=O hydrogen bonding, whereas the α-hydroxy on the acyl chain reduced it. In CER NP:AP mixtures with CHOL and FFA, simulations reproduced the experimental repeat distances of the NP-rich and AP-rich systems despite differences in hydrogen bonding. Simulations of multicomponent mixtures resembling the SC model of Bouwstra demonstrated the dominant effect of chain-length distribution, rather than CER hydrogen bonding, on permeability. This work shows how multiscale modeling integrated with experiments can uncover molecular mechanisms linking composition and SC barrier structure to interpret experimental results.

1. Introduction

The stratum corneum (SC), the outermost layer of the skin, serves as a vital barrier, protecting against environmental influences, dehydration, and infection. A key component of this barrier is the complex lipid matrix, which is organized into two coexisting crystalline lamellar phases, known as the long and short periodicity phases (LPP and SPP) to describe their respective repeat distances of ~13 nm and ~5 nm.^1-3 The lamellar phases are composed primarily of ceramides (CERs), cholesterol (CHOL), and free fatty acids (FFAs). Among these, CERs play a crucial role due to their structural diversity and unique ability to form highly ordered lamellae, which are essential for SC integrity and permeability. CERs consist of a fatty acid linked to a sphingoid base by an amide bond. CERs with different combinations of fatty acid and sphingoid base chains, such as CER NS (non-hydroxy sphingosine ceramide), NP (non-hydroxy phytosphingosine ceramide), AP (alpha-hydroxy phytosphingosine ceramide), and AS (alpha-hydroxy sphingosine ceramide) shown in Figure 1, play critical roles in the molecular architecture and function of human SC and are commonly represented in synthetic lipid models of the SC lipid matrix.⁴ Together, these four CER subclasses account for approximately half of the CER molecules identified by liquid chromatography/mass spectrometry in the SC from human forearms.^5-9

Figure 1:

Molecular structures of CERs NS, NP, AS, and AP, which are composed of a fatty acid linked to a sphingoid base by an amide bond. Both chains can vary in length. The structures shown have a 24-carbon acyl chain and a 18-carbon sphingoid base chain.

Experimental studies of films composed of a single CER subclass—NS, NP, AP, or AS—have revealed distinct lamellar behaviors arising from differences in headgroup hydroxylation, either on the sphingoid base (S or P) chain or the fatty acid (N or A) chain. These structural variations affect intermolecular interactions, packing density, and overall lamellar stability. For example, using Fourier transform infrared (FTIR) spectroscopy and differential scanning calorimetry, Rerek et al.¹⁰ observed notable differences among the CER subclasses. CER NP exhibited the strongest hydrogen bonding, indicated by a low amide I peak frequency (1612 cm⁻¹), corresponding with high carbonyl (C=O) hydrogen bonding. It also had the highest order-disorder transition temperature (T_m) of 115°C but displayed lower (hexagonal) packing order, as evidenced by a single CH₂ scissoring mode peak at 1467 cm⁻¹. Similarly, CER AP showed strong C=O hydrogen bonding (amide I peak at 1620 cm⁻¹) and hexagonal packing, but with a slightly lower T_m of 93°C. In contrast, CER NS displayed a higher (orthorhombic) packing order, indicated by the splitting of the CH₂ scissoring mode into two peaks, and had a T_m of 93°C (the same as CER AP). However, its C=O hydrogen bonding was weaker, as reflected by a higher amide I peak frequency of 1631 cm⁻¹. CER AS also exhibited orthorhombic packing but had the weakest C=O hydrogen bonding (highest amide I peak at 1637 cm⁻¹) and the lowest T_m of 80°C among the four CER subclasses. These findings, along with similar observations from Garidel et al.^11,12 are summarized in Table S2 of the Supporting Information.

Together, these results suggest that films of the phytosphingosine (P) type CERs (NP and AP) achieve stability primarily through strong hydrogen bonding between their headgroups, resulting in a less ordered hexagonal tail packing arrangement. In contrast, the sphingosine (S) type CERs (NS and AS) rely on a tightly packed orthorhombic tail arrangement for stability. Interestingly, the additional hydroxylation of the fatty acid in the alpha-hydroxy (A) type CERs (AS and AP), compared to their non-hydroxy (N) type counterparts (NS and NP), results in weaker amide I-related hydrogen bonding. This effect may be due to the additional hydroxyl group expanding the headgroup region and disrupting optimal hydrogen bonding interactions.¹⁰ Clearly, each CER subclass uniquely influences the packing arrangements of the lipid lamellae in these systems, which in turn affects their structural and functional properties. Observations from other experimental methods, including electron microscopy, neutron and X-ray scattering, and permeability measurements have also revealed differences in packing, phase behavior, structure and barrier function associated with the various CER subclasses.^4,13-15

MD simulations are a useful tool for testing experimental hypotheses and investigating characteristics inaccessible to experimental methods.¹⁶ In this paper, we compare MD simulation results with experimental observations from synthetic SC lipid mixtures of different compositions. This allows us to evaluate the accuracy of the new coarse-grained (CG) models for CERs NP, AP and AS, and to gain a molecular-level perspective into the effect of the different CER headgroups on molecular interactions, organization, and hydrogen bonding networks.

Studying the structure-function relationship of SC lipids computationally requires SC lipid models that accurately represent the complex lipid interactions and can self-assemble into lamellae during a MD simulation. Simulations in which the lipids self-assemble are essential because it mimics the formation of lamellar phases in the experimental systems studied. Furthermore, it has been shown that lipids extracted from native SC spontaneously self-assemble onto substrates as multilayer membranes. Significantly, these self-assembled membranes exhibit the same lamellar and lateral structures and phase behavior observed in native SC.^17,18

While atomistic simulations generally provide accurate interactions and detailed insights, their large computational cost renders them unsuitable for studies of large systems and long timescales, making self-assembly of mixed lipid systems with atomistic models essentially unattainable. CG models, which simplify molecular details while retaining key features, offer an effective solution for exploring SC lipids on larger, biologically relevant scales. However, CG simulations of CERs are challenging due to the need for accurate models that capture the chemical diversity among the different CER subclasses. This need was met using the flexible framework of the multistate iterative Boltzmann inversion (MS-IBI) method¹⁹ to develop transferable CG models for CERs NS, NP, AP, and AS.^20-22 To regain the atomic-level resolution available in atomistic simulation, a multiscale approach is used that integrates the ability of the CG simulations to self-assemble large, multilayer SC lipid membranes with atomistic simulations, by reverse-mapping the self-assembled CG multilayers to recover atomic-level structure.^13,23,24

Here, we describe applications of the CG CER models developed and the multiscale approach used to study a range of systems, including each of the four CERs (NS, NP, AP and AS), either alone or in ternary mixtures with CHOL and FFA, as well as multicomponent mixtures of CERs and/or FFAs with CHOL. These systems were selected because experimental measurements, including FTIR, differential scanning calorimetry, neutron and X-ray scattering, and permeability, are available from studies of SC lipid model membranes with the same or closely related compositions of the simulations, enabling both model validation and new insights into the experimental observations. These atomistic insights, often inaccessible through experiments alone, can clarify how lipid composition influences membrane structure and behavior.

2. Methods

2.1. Coarse-grained Simulations

CG simulations were conducted using Multistate-Iterative Boltzmann Inversion (MS-IBI) ¹⁹ developed models for four CER subclasses^20,22 (CERs NS, NP, AS and AP) along with CHOL^24,25, FFA^22,26,27, and water.^28-30 FFA is modeled as fully protonated, saturated, and uncharged. The atom-to-bead mappings for all lipid molecules are shown in Figure S1. All bonded and nonbonded interaction parameters are available in a publicly accessible GitHub repository.²¹ Consistent with the experimental systems, the CER sphingoid base chains contained 18 carbon atoms (C18), while the CER acyl chains and the FFA tails were 24 carbons (C24), unless stated otherwise.

In each CG simulation, SC lipids were randomly packed at a density of 0.8 g/cm³ between two water layers (1.0 g/cm³) using mBuild from the MoSDeF software library.^31-38 The systems were then self-assembled into six-leaflet multilayer stacks using HOOMD-Blue (version 2.9.7) with periodic boundary conditions applied in all three dimensions.^39,40 A leaflet refers to a plane of lipid headgroups connected to lipid tails oriented in the same direction (see Figure S2). In bilayer structures, two leaflets are arranged such that headgroups face outward and tails point inward, toward the bilayer center. Because CERs can adopt either a hairpin conformation (with both tails pointing in the same direction) or an extended conformation (with the tails pointing in opposite directions), a CER molecule can exist in either a single leaflet (if the CERs are in the hairpin conformation) or two leaflets (if the CERs are in the extended conformation).

Simulations were performed on four sets of six-leaflet multilayer systems: (a) pure CER NS, NP, AS, and AP containing 1800 lipids; (b) each of the four CERs mixed with CHOL and FFA C24 (1800 total lipids for CERs AS and AP, and 2200 lipids for CERs NS and NP); (c) CER NP and CER AP with 1:2 and 2:1 molar ratios combined with CHOL and FFA C24 (1800 total lipids); and (d) seven- or eleven-component mixtures of CERs, CHOL and FFAs (2200 total lipids) representing variations in the compositions of the experimental SCS.^41-43 Comparative simulations of systems containing either 1800 or 2200 lipids showed no significant differences in structural parameters (Table S1 and related text in the Supporting Information). Simulations of pure CERs were repeated four times; simulations of all other systems were repeated three times. Each simulation included 10 CG water beads per lipid, equivalent to 40 water molecules per lipid as used in the atomistic simulations, described below.

The self-assembly simulations followed protocols reported in previous studies^13,20 and provided in Section S1.2 of the Supporting Information. Briefly, after a short equilibration period, a shape annealing procedure is employed to accelerate self-assembly. In this approach, the cross-sectional area of the simulation box is gradually expanded to 2.5 times the target area—estimated based on the number of lipids, the desired number of leaflets, and the estimated area per lipid (APL)—and then compressed back to the target area, while maintaining a constant simulation box volume.

Following shape annealing, the simulations were continued under physiological conditions (305 K and 1 bar) for 150 ns - 400 ns depending on the system being studied as detailed in the Section S1.2. Structural stability was confirmed by a constant simulation box width. The final 100 ns of each simulation were used as the NPT production run, with trajectory frames saved every 0.5 ns to yield 200 frames for analysis. In total, the CG systems are simulated for 1-2 μs.

2.2. Multiscale Modeling

Equilibrated CG lipid assemblies were reverse mapped to the atomistic level to determine hydrogen bonding, neutron scattering length density (NSLD) profiles, and other atomic-level properties. Atomistic resolution is recovered from the final frame of the CG production runs using the reverse-mapping procedure presented in Nadaban et al.¹³ The mBuild^33,44 toolkit acts as the translator to seamlessly bridge the CG and atomistic models and the HOOMD-Blue and GROMACS simulation engines.^39,40,45 A pre-assembled atomistic structure is constructed by taking the center-of-mass position of each CG lipid molecule, the orientation of its tail(s) (pointing in the +z or −z at the average tilt angle of all lipids from the CG simulation), and the conformation of each CER molecule (i.e., whether the tails are located in the same or adjacent leaflets) from the CG system. CG water beads located in the inner four leaflets are retained and replaced by four atomistic water molecules per bead, placed randomly with the same center-of-mass position as each replaced CG water bead. To avoid high energy atomic overlaps, a water molecule that contacts another molecule is shifted 1.5 Å in the x-direction. We note that this protocol contrasts with that of Shamaprasad et al.,²⁴ in which the CG water beads in the inner leaflets were removed.

After reverse-mapping, a brief AA simulation is performed using GROMACS⁴⁵ 2020.6 with a 1-fs timestep and periodic boundary conditions in all three dimensions. The simulations used the CHARMM36^46,47 force field, CHOL parameters from Cournia et al.⁴⁸ the TIP3P water model,⁴⁹ and CER headgroup parameters from Guo et al.⁵⁰ and Frame et al.²⁰ Topology and force field files are provided in a GitHub repository.²¹ The AA simulation begins with energy minimization using the steepest descent algorithm for up to 200,000 simulation steps to eliminate any atomic overlaps (i.e., steric clashes) that may have been introduced during reverse mapping. This is followed by 15 ns of NPT equilibration at 305 K and 1 bar using the Nosé-Hoover thermostat⁵¹ and the Parrinello-Rahman barostat⁵² with semi-isotropic pressure control. The final 2 ns of the equilibration serves as the production run, with the trajectory snapshots recorded every 10 ps to produce 200 frames for analysis.

2.3. Analysis

Leaflet-pair thickness was determined from the peak-to-peak distance in the mass density profile. Lipid tilt angle representing the average orientation of all lipid tails, was calculated as the angle between the bilayer normal and the principal (long) axis of lipid tail (i.e., the axis with the least resistance to rotation)⁵³, derived from the eigenvector corresponding to the minimum eigenvalue of the inertia tensor.⁵⁴ The nematic order parameter (S₂) was obtained from the largest eigenvalue of the nematic tensor as described in the supporting information of Wilson.⁵⁵ Interdigitation, $\lambda$, was calculated using the following equation⁵⁶:

$$\lambda =4{\int }_{z_{min}}^{z_{max}}\frac{{\rho }_{top}(z){\rho }_{bot}(z)}{{({\rho }_{top}(z)+{\rho }_{bot}(z))}^2}dz$$

where ${\rho }_{top}(z)$ and ${\rho }_{bot}(z)$ are the mass density profiles of lipids with headgroups above (top) and below (bot) the midpoint of the simulation box, and $z_{min}$ and $z_{max}$ are the minimum and maximum $z$-coordinates of the simulation box. CER molecules were considered to be in an extended conformation when the angle between the directional vectors associated with the centers of mass of the acyl and sphingosine tails was greater than 90°. All these structural metrics can be calculated for both the atomistic and CG systems.

Hydrogen bonding interactions were calculated from the equilibrated reverse-mapped multilayer simulations following the procedure presented in Nadaban et al.,¹³ using the GROMACS⁴⁵ hbond module. A hydrogen bond was identified based on the geometric criteria: of a donor-acceptor distance ≤ 0.35 nm and donor-hydrogen-acceptor angle ≤ 30°. Intramolecular hydrogen bonds were not considered because these are unlikely given the geometric criteria and the spatial arrangement of hydrogen bonding sites. Hydroxyl and amine groups were treated as donors, and oxygen and nitrogen atoms were considered acceptors. Hydrogen bonds involving the carbonyl (C=O) and amine (N-H) groups of the CERs are expected to influence the frequency of the amide I (1600–1700 cm⁻¹) and amide II (1500–1600 cm⁻¹) FTIR vibrational modes, respectively.

Hydrogen bonds were identified in each trajectory frame, averaged over the 200 frames within each independent simulation replicate, and then averaged across either three or four replicates. To best reflect the experimental conditions, hydrogen bonding was analyzed only within the four inner leaflets of the six-leaflet multilayers. This approach excludes contributions from lipid-water hydrogen bonds in the outer leaflets, where headgroups are exposed to bulk water (see Figure S2). Figure 2 shows the molecular structures of the lipids, highlighting the atoms involved in hydrogen bonding. Additional computational details are provided in Section S2.2 of the Supporting Information.

Figure 2:

Molecular structures with the hydrogen bonding sites identified on the headgroups of the four CER subclasses, along with CHOL, FFA, and water. For the CERs, these include O4 and N1 (associated with the amide I and II FTIR frequencies, respectively), and the four hydroxyl groups (O80, O84, O88 and O7). CHOL, FFA, and water hydrogen bonding sites are O3, O25 and O27, and O, respectively. The CG models of the CERs include four OH beads (see Figure S1), corresponding to O80 (OH1), O84 (OH2), O88 (OH3) and O7 (OH4).

Simulation trajectory analyses were performed using the Python libraries MDTraj⁵⁷ and SciPy.⁵⁸ Self-assembled CG simulations were reverse-mapped and equilibrated atomistically before analysis, except for the CER NP:AP mixtures at a 1:0.5:1 molar ratio of CERs:CHOL:FFA for which only CG simulations were performed. Simulation results are averaged over all leaflets in each frame and then across all frames from the simulation. All results, except tilt angle, are reported as the average and standard deviation calculated from replicate simulations with different randomized initial configurations. For tilt angle, which is calculated for each lipid, the standard deviation is reported as the square root of the variance, where the variance is calculated from the tilt angle of the molecules in each leaflet, of each frame, of each replicated simulation and then pooled across all leaflets, frames and replicates. Statistical significance calculations comparing two values were performed using the unpaired t-test with the one-tailed significance level set at the designated probability p, typically <0.05.

3. Results and Discussion

Multiscale, multilayer simulations are presented and compared with experimental results for lipid systems of increasingly complex composition. We focus first on hydrogen bonding in the simplest systems: pure CERs and ternary mixtures of a single CER subclass (NS, NP, AS, or AP) with cholesterol CHOL and one FFA. These comparisons with experimental FTIR data serve as an initial and crucial step for testing the validity of the newly developed CER CG models. The complexity increases in Section 3.2, where we examine lipid systems containing both CER NP and CER AP alongside CHOL and FFA C24. This allows us to further test the models' ability to capture the unique interactions in multi-CER systems and to explore the molecular basis for observed structural phenomena, such as lamellar thickness and lipid tail ordering. Finally, Section 3.3 addresses the most complex, multicomponent systems, specifically lipid mixtures with as many as eleven components designed to closely mimic the composition and behavior of the SC lipid matrix. This analysis provides a molecular basis for the mechanisms underlying the experimental observations of the effect of CER composition and FFA chain-length distribution on permeability.

3.1. Hydrogen Bonding of Pure CER and CER Mixtures with CHOL and FFA

Key experimental observations related to hydrogen bonding strength are summarized and then compared with our multiscale, multilayer simulation results to validate the new CER models. Further analyses of the simulation data elucidate the molecular mechanisms behind the observed behaviors. Finally, the new multilayer simulations are compared with previously published bilayer simulations of the same systems.

Experimental observations

Moore and Garidel studied the hydrogen bonding of CERs NS, NP, AS, and AP, both individually and in ternary mixtures with CHOL and FFA.^10-12,59-64 They did this by analyzing shifts in the amide I and amide II FTIR frequencies, which are related to hydrogen bonding with the C=O and N-H groups, respectively.⁶⁵ The FTIR results, along with ordered-disorder transition temperatures (which they also measured), are summarized in Table S2 in the Supporting Information. In these studies, CERs NP and AP, each with acyl and phytosphingosine chains containing 18 carbons, were either pure or in equimolar mixtures with FFA C18 and CHOL.^10,63 Mixtures containing CERs AS and NS consisted of FFA C16 with bovine brain CERs, which had acyl chain lengths ranging from C18 to C24.^10,12,62,63 In the pure CER studies, CERs AS and NS were from bovine brain CERs, except for the CER NS study from Rerek et al.,¹⁰ which used synthetic CER NS. Like CERs NP and AP,¹¹ the synthetic CER NS had 18-carbon acyl and sphingosine chains.^60,61 All lipid systems were hydrated.

As described in the introduction, in experiments pure sphingosine CERs (NS and AS) exhibit weaker C=O hydrogen bonding (based on shifts in the FTIR amide I peaks) than their phytosphingosine counterparts (NP and AP, respectively) consistent with the additional hydroxyl group on the phytosphingosine chain. However, as also described in the introduction, C=O hydrogen bonding (inferred from the amide I peak shifts) does not always increase with additional hydroxyl groups. In fact, pure non-hydroxy CERs (NS and NP) exhibited stronger hydrogen bonding than their α-hydroxy counterparts (AS and AP, respectively), despite the presence of an additional hydroxyl group on the fatty acid chain in the latter. Furthermore, pure CER NS displayed split peaks for both amide I and II bands, whereas pure CER AS did not. Assuming the CERs predominantly adopt the hairpin (rather than extended) conformation, Moore et al.⁶² hypothesized that the split peaks of CER NS signify inter-leaflet hydrogen bonding (i.e., transverse interaction between headgroups in opposing leaflets), while the single peak of CER AS reflects intra-leaflet hydrogen bonding (lateral cohesion within the bilayer cross-section).

Building on these findings, Rerek et al. compared hydrogen bonding strengths inferred from amide I peak shifts in pure CER systems with equimolar mixtures of the same CER with CHOL and FFA.⁶³ They observed weaker hydrogen bonding strengths in the ternary mixtures of CERs NP and AP, but similar hydrogen bonding in the ternary mixtures of CERs NS and AS compared to their pure counterparts. In a subsequent study, Garidel et al.¹² also found similar hydrogen bonding strengths for CERs NS and AS in both pure and ternary systems, consistent with Rerek et al. However, in contrast to Rerek et al., they observed no difference in hydrogen bonding strength for CER NP between the two systems. CER AP mixtures were not included in the Garidel et al. study.¹²

Table 1 summarizes the relative hydrogen bonding strengths of CERs inferred from the amide I vibration frequency shifts observed in the Moore and Garidel studies. Hydrogen bonding strengths inferred from amide II frequency shifts are also listed. While the amide I peak, which predominantly arises from the C=O stretching mode, will shift with the strength of hydrogen bonding to the C=O group, amide II shifts are less straightforward to interpret. The amide II mode includes contributions from both N-H bending, which can be affected by hydrogen bonding, and C-N stretching, which is not directly involved in such interactions. As a result, N-H hydrogen bonding interactions may not be the sole cause of amide II frequency shifts.^65,66

Table 1:

Relative values of the hydrogen-bond numbers from the simulations compared with relative hydrogen-bond strength deduced from the amide I and II FTIR peak frequencies observed by Moore, Garidel, and colleagues (summarized in Table S2) in experiments of pure CERs and CER mixtures with CHOL and FFA.^a,b

System		Amide I associated		Amide II associated
		C=O H-bond numbers from simulation	H-bond strength from experiments	N-H H-bond numbers from simulation	H-bond strength from experiments
Pure CERs		NP > NS***	NP > NS	NP > NS**	NP > NS or NP ≈ NS c
		AP > AS***	AP > AS	AP ns > AS	AP > AS
		NP > AP*	NP > AP	NP < AP *	NP > AP
		NS > AS***	NS > AS	NS < AS *	NS > AS
CER mixtures with CHOL and FFA		NP > NS*	NP > NS	NP > NSns	NP > NS or NP≈ NSd
		AP > AS ***	AP ≈ AS	AP > AS**
		NPns > AP	NP > AP	NP < AP***
		NS > AS**	NS > AS	NS < AS ***	NS > AS
Pure CER compared to CER mixture with CHOL and FFA	NP	pure > mix***	pure > mix or pure ≈ mixe	pure > mix **	pure ≈ mix
	AP	pure > mix*	pure > mix	pure > mix*
	NS	pure > mix **	pure ≈ mix e	pure ≈ mix	pure ≈ mix
	AS	pure > mix **	pure ≈ mix e	pure ≈ mix	pure < mix

^aStatistical significance of the system with the larger number of hydrogen bonds with either C=O or N-H groups is designated as

^*p < 0.05,

^**p < 0.01,

^***p < 0.001, or not significant (ns).

^bBold and italicized text designates systems in which the hydrogen-bond numbers from the simulations are inconsistent with the experimental observations from Moore, Garidel, and colleagues of hydrogen-bond strength deduced from either amide I or amide II FTIR peak frequencies.

^cDepends on which amide II results are used for pure CER NS. A single peak at ~1545 cm⁻¹ from Garidel et al.¹² is smaller than 1556 cm⁻¹ for CER NP. However, the single peak at 1550 cm⁻¹ from Rerek et al.¹⁰ and the average value of 1556.5 for two peaks at 1545 and 1568 cm⁻¹ from Moore et al.⁶¹ (also from Mendelsohn et al.⁶⁰ as well as similar results from Rerek and Moore⁷⁰) are similar to 1556 cm⁻¹ for CER NP.

^dDepends on which amide II results are used for the CER NS mixture (single peaks at 1548 cm⁻¹ from Moore et al.⁶¹, Moore and Rerek⁵⁹ and Garidel et al.,¹² or the average value of 1556.5 cm⁻¹ for two peaks at 1548 and 1565 cm⁻¹ from Mendelsohn et al.⁶⁰) compared with ~1558 cm⁻¹ for CER NP.

^eGaridel et al.¹² reported in Table 2 of their paper that the hydrogen-bond strength determined from the amide I peak was similar for pure and ternary mixtures of CERs NP, NS and AS. They did not measure ternary mixtures of CER AP.

Molecular simulations of the experimental systems

To compare with the experimental FTIR results described above, six-leaflet multilayer systems of pure CER and CER:CHOL:FFA mixtures with a 1:0.5:1 molar ratio were self-assembled. Previous work²⁰ showed that acyl tail length (C24 or C16) does not affect the number of hydrogen bonding interactions in simulations of CER NS alone, with CHOL, or with both CHOL and FFA.^23,67 Thus here, CERs with acyl chain lengths of C24 were simulated as these are more common in experimental SC lipid model membrane studies. To be consistent with this choice, FFA chain lengths in the CER:CHOL:FFA mixtures were also C24. The molar ratio of CHOL in simulations of the CER:CHOL:FFA mixtures was half of that typically used in the experiments because experimental studies with equimolar amounts of CER, CHOL, and FFA include separate phase crystalline CHOL that disappears when the CHOL amount is reduced by about half.^{13,15,41,43,68,69} Based on these observations, the composition of the lamellar phase in the experimental SC lipid model membranes is likely to be close to the 1:0.5:1 CER:CHOL:FFA molar ratio used in the simulations.

Because lipids were self-assembled into six leaflets, rather than pre-assembled and divided equally among the leaflets, the compositions of the four inner leaflets differed slightly from the overall 1:0.5:1 molar ratio for CER:CHOL:FFA. Table 2 lists the actual CHOL:CER and FFA:CER ratios of the self-assembled four inner leaflets as well as the fraction of CERs in the extended conformation (approximately 36% and 29% in the mixed and pure systems, respectively); detailed numbers provided in Table S4. Interestingly, differences in the distribution of CERs in the inner leaflets (where they can be extended) compared with the outer leaflets (where they are restricted to the hairpin conformation) was minimal. For the pure CERs, there are fewer lipids (by 4% or less) in the inner four leaflets than would be seen in a uniform distribution, except for CER AP, which has ~6% more lipids. In the ternary mixtures, there are also fewer CERs in the inner leaflets by a small amount (< 3% ) except for CER AS, which had almost 10% fewer CERs in the inner leaflets.

Table 2:

Composition and CER conformation of the four inner leaflets in configurations from reverse-mapped atomistic simulations of the self-assembled six-leaflet membranes for pure CER and CER:CHOL:FFA mixtures with a 1:0.5:1 molar ratio; mean ± standard deviation is calculated from three (CER mixtures) or four (pure CERs) independent simulations.

System composition and CER configurationa	NP		AP		NS		AS
	Mixture	Pure	Mixture	Pure	Mixture	Pure	Mixture	Pure
Water molecules /lipid b	0.34 ± 0.02*	0.26 ± 0.02	0.28 ± 0.03	0.23 ± 0.05	0.34 ± 0.07*	0.13 ± 0.03	0.34 ± 0.04*	0.19 ± 0.04
CHOL/CER	0.53 ± 0.02		0.53 ± 0.02		0.53 ± 0.02		0.58 ± 0.01
FFA/CER	0.93 ± 0.10		1.01 ± 0.10		0.94 ± 0.07		1.22 ± 0.05
Fraction of Extended CERs	0.38 ± 0.01	0.29 ± 0.01	0.39 ± 0.03	0.31 ± 0.01	0.32 ± 0.03	0.29 ± 0.03	0.34 ± 0.03	0.25 ± 0.01
CER number in inner leaflets	586 ± 19	1189 ± 22	467 ± 11	1276 ± 29	581 ± 14	1158 ± 11	433 ± 7	1155 ± 16
% CER excess (deficit) c	(0.2 ± 0.01)	(0.9 ± 0.01)	(2.7 ± 0.06)	6.3 ± 0.14	(1.0 ± 0.02)	(3.5 ± 0.03)	(9.8 ± 0.16)	(3.8 ± 0.05)

^aThe number of each lipid type and water, and the number of extended CERs are listed in Table S4 in the Supporting Information.

^bAsterisks designate that the amount of water in the CER mixture was statistically significantly larger than in the corresponding pure CER simulation (p < 0.05).

^cPercentage of excess or deficit (in parentheses) in the number of CER molecules in the four inner leaflets relative to the number of CERs in four leaflets if equally distributed among the six leaflets of the membrane (i.e., four inner leaflets with equal distribution, contain 1200 CER molecules in the pure CER systems, 587 CER molecules in the mixtures with CERs NP and NS, and 480 CER molecules in mixtures with CERs AP and AS).

To enable direct comparison with the experimental multilayers in the studies of Moore and Garidel, the number of intermolecular hydrogen bond interactions at 305 K was calculated for the four inner leaflets of the simulated six-leaflet structures. Consistent with the experimental conditions, these inner leaflets do not contact bulk water (Figure S2). However, their headgroup regions contain small amounts of water: approximately 0.3 water molecules/lipid in the CER/CHOL/FFA mixtures and between 0.13 (CER NS) and 0.26 (CER NP) water molecules/lipid in the pure CER systems (Table S2). Except for CER AP, the water content in the CER mixtures was statistically significantly greater than in the corresponding pure CER system. Although the simulated water amounts are lower than the experimental estimates of approximately one or two water molecules per lipid for a fully hydrated SC lipid model membrane,^41,71 they are not unreasonable considering the limitations of the experimental measurement methods used.

Figure 2 identifies the hydrogen bonding sites on the four CER headgroups, as well as on CHOL, FFA and water. The CER AP headgroup has six hydrogen bonding sites; CERs NP and AS each have five, and CER NS has four. CHOL and FFA molecules have one and two hydrogen bonding sites, respectively. The number of bonds of each of the hydrogen bonding pairs with CER molecules were calculated for the four CER subclasses, either alone or in mixtures with CHOL and FFA. Table 3 reports the total number of normalized hydrogen bonding interactions by C=O, N-H, and other CER sites with lipids and water molecules located in the headgroup region of the inner four leaflets. Detailed numbers of hydrogen bonds by each bonding pair associated with C=O and N-H sites (normalized by the number of CER molecules) are reported in Tables S5 and S6, respectively. Table S7 lists the number of hydrogen bonds for each bonding pair involving CER sites other than C=O and N-H.

Table 3:

Number of hydrogen bonds with CER molecules normalized by the number of CER molecules for C=O (related to the amide I FTIR signal), N-H (related to the amide II FTIR signal), and other hydrogen bonding sites. Results are reported for the four inner leaflets of simulations of pure CER and CER:CHOL:FFA mixtures with a 1:0.5:1 molar ratio; mean ± standard deviation is calculated from three (CER mixtures) or four (pure CERs) independent simulations. See Section S2.2 and Tables S5-S7 for the complete atom pair breakdown.

		NP		AP		NS		AS
	H-bonds	Mixture	Pure	Mixture	Pure	Mixture	Pure	Mixture	Pure
CER-lipid	C=O	0.776 ± 0.019	0.890 ± 0.007	0.769 ± 0.021	0.837 ± 0.026	0.720 ± 0.029	0.787 ± 0.010	0.668 ± 0.018	0.735 ± 0.010
	N-H	0.474 ± 0.007	0.504 ± 0.010	0.544 ± 0.004	0.532 ± 0.014	0.464 ± 0.010	0.489 ± 0.006	0.521 ± 0.009	0.517 ± 0.013
	Other	1.155 ± 0.019	0.714 ± 0.021	1.784 ± 0.002	1.231 ± 0.036	0.850 ± 0.021	0.510 ± 0.013	1.396 ± 0.055	0.946 ± 0.026
	Total a	2.271 ± 0.031	1.906 ± 0.039	2.969 ± 0.022	2.424 ± 0.078	1.904 ± 0.027	1.613 ± 0.015	2.472 ± 0.072	2.031 ± 0.034
CER-water	C=O	0.325 ± 0.01	0.176 ± 0.018	0.206 ± 0.015	0.121 ± 0.019	0.291 ± 0.050	0.112 ± 0.021	0.244 ± 0.018	0.107 ± 0.028
	N-H	0.129 ± 0.006	0.069 ± 0.008	0.107 ± 0.004	0.06 ± 0.009	0.102 ± 0.017	0.040 ± 0.008	0.122 ± 0.002	0.05 ± 0.013
	Other	0.524 ± 0.028	0.288 ± 0.027	0.562 ± 0.02	0.242 ± 0.042	0.401 ± 0.048	0.155 ± 0.023	0.592 ± 0.049	0.183 ± 0.036
	Total a	0.977 ± 0.040	0.531 ± 0.049	0.876 ± 0.027	0.413 ± 0.072	0.794 ± 0.113	0.289 ± 0.029	0.958 ± 0.066	0.322 ± 0.055
CER-lipid + CER-water	C=O	1.101 ± 0.009	1.066 ± 0.016	0.975 ± 0.026	0.958 ± 0.017	1.012 ± 0.027	0.899 ± 0.015	0.912 ± 0.007	0.841 ± 0.021
	N-H	0.603 ± 0.002	0.573 ± 0.002	0.651 ± 0.004	0.592 ± 0.01	0.566 ± 0.007	0.529 ± 0.003	0.643 ± 0.008	0.567 ± 0.006
	Other	1.678 ± 0.017	1.002 ± 0.021	2.347 ± 0.022	1.474 ± 0.028	1.251 ± 0.059	0.665 ± 0.013	1.988 ± 0.035	1.129 ± 0.015
	Total a	3.249 ± 0.011	2.438 ± 0.03	3.845 ± 0.042	2.837 ± 0.054	2.698 ± 0.092	1.901 ± 0.015	3.430 ± 0.033	2.353 ± 0.024

^aEqual to the sum of the hydrogen bond numbers for other sites (i.e., hydrogen bonds not involving C=O and N-H) and C=O and N-H sites minus the number of CER-CER N1-O4 hydrogen bonds, which are included in the numbers for both the C=O and N-H sites (see Section S1.1).

Comparison of hydrogen bonding in simulations and experiments

Figure 3 compares the CER-normalized number of C=O and N-H hydrogen bonds between lipids for the four CER subclasses, both alone and in ternary mixtures with CHOL and FFA. Statistical significance is indicated for differences between (i) pure CERs and their mixtures, (ii) phytosphingosine (P-type) and sphingosine (S-type) CERs (alone or in mixtures), and (iii) non-hydroxy (N-type) and alpha-hydroxy (A-type) CERs (alone or in mixtures). Table 1 compares these simulation results with the hydrogen bonding strengths inferred from shifts in the amide I and II FTIR peak frequencies.

Figure 3:

Hydrogen bonds formed by the C=O group (related to the amide I) and the N-H group (related to the amide II) of the CER molecules in the inner four leaflets, normalized by the number of CER molecules in those same leaflets. Results are shown for each CER subclass, both in pure systems and in ternary mixtures with CHOL and FFA at a 1:0.5:1 molar ratio. Blue bars with diagonal lines represent pure CER systems; solid red bars represent ternary mixtures. Asterisks indicate statistical significance between the indicated system pairs: *p < 0.05, **p < 0.01, or ***p < 0.001.

Overall, the hydrogen bonding with C=O from simulation aligns well with the experimental amide I peaks measured by FTIR. As inferred from amide I peak shifts in the experiments, pure CERs showed significantly greater C=O hydrogen bonding for P-type CERs (NP and AP) compared to their S-type counterparts (NS and AS, respectively), and for N-type CERs (NS and NP) compared to their A-type counterparts (AS and AP, respectively). The hydrogen bonding numbers for C=O in the CER mixtures were also generally consistent with experimental amide I data, except for CER AP, which exhibited higher C=O hydrogen bond numbers than CER AS in the simulations, but similar amide I peak vibration frequencies in the experiments.

In simulations of all four CER subclasses, pure CERs exhibited greater C=O hydrogen bonding than their mixtures with CHOL and FFA. This contrasts with experimental observations from Rerek et al.⁶² for CERs NS and AS, and from Garidel et al.¹² for CERs NS, AS, and NP, which reported similar amide I vibrational frequencies in pure and mixed systems of the same CER subclass. A possible explanation for this discrepancy is the lower water content in the pure CER simulations compared to mixtures—0.13 and 0.19 water molecules per lipid for pure CERs NS and AS, respectively, versus 0.34 in their mixtures. Lower water levels reduce CER–water hydrogen bonding, allowing for more CER–lipid hydrogen bonding. Thus, increasing the water levels in the pure CER simulations to match those in the mixtures might reduce C=O (amide I-related) hydrogen bonding in the pure systems, yielding results that are more consistent with the experimental observations for CERs NS, AS, and NP.

In contrast with the C=O and amide I results, the correspondence between simulated N-H hydrogen bond numbers and hydrogen bonding strength inferred from amide II peak shifts is much less consistent. For instance, simulations and experiments yielded opposite conclusions when comparing pure CERs NP and AP, as well as CERs NS and AS in both their pure and mixed systems. In some cases, agreement between simulations and experiments depends on which amide II results are considered (e.g., comparisons of CERs NP and NS in pure and mixed systems). Notably, unlike the amide I band, which is primarily a C=O stretching mode and is therefore directly influenced by hydrogen bonding, the amide II band, arises from both N-H bending, which depends directly on hydrogen bonding, and C-N stretching, which does not. Consequently, discrepancies between simulation results and hydrogen bonding estimates based on amide II peak shifts are not necessarily concerning, as the latter may include effects other than just N-H hydrogen bonding strength.^65,66

Additional insights from the hydrogen bonding results

Beyond the comparison of simulated C=O and N-H hydrogen bonding with amide I and II FTIR results, the simulations provide additional insights that are not accessible from experiments alone. The hydrogen bond numbers for CER molecules with CERs, CHOL, and FFA, summarized in Table 3 from the detailed numbers provided in Tables S5-S7, reveal patterns that extend our understanding of hydrogen bonding behavior. For example, as Figure 4A shows, adding an OH group to the CER acyl chain (site O7; OH4) reduces the number of C=O hydrogen bonds, while adding an OH group to the sphingoid base chain (site O88; OH3) increases them. The detailed results in Table S5 clarify these effects on specific hydrogen bonds. Both OH3 (O88) and OH4 (O7) reduce C=O hydrogen bonding with OH1 and OH2 (sites O80 and O84). However, OH3 forms enough additional bonds to compensate for the loss of C=O hydrogen bonds, whereas OH4 does not. As a result, OH3 leads to a net increase in C=O hydrogen bonds, while OH4 causes a net decrease.

Figure 4:

Number of C=O, N-H, and other CER-lipid hydrogen bonds per CER in the inner four leaflets for each CER subclass, either alone (solid lines) or in mixtures with CHOL and FFA (dashed lines). The CER subclasses are ordered left-to-right (A) by increasing amide I hydrogen bond strength in the FTIR experiments (CERs AS, NS, AP, NP), and (B and C) by increasing numbers of hydrogen bonds other than C=O and N-H calculated from the simulations (CERs NS, NP, AS, AP).

In contrast with C=O, the number of hydrogen bonds between other sites (i.e., other than the C=O and N-H groups) increases when a OH group is added to the sphingoid base chain (OH3; O88) and increases even more when an OH group is added to the fatty acid chain (OH4; O7 site) as shown in Figure 4B. Thus, CER AP, with OH groups added to both the sphingoid base chain and the fatty acid chain, has the largest number of hydrogen bonds with other sites followed by CERs AS, NP and NS in descending order. Interestingly, N-H hydrogen bonding increases in the same CER order as the other hydrogen bonds, although the amount of increase is small (Figure 4C).

The number of hydrogen bonds with other sites is larger for CER mixed with CHOL and FFA than in the pure CER systems, which is the reverse of the C=O hydrogen bonding (Figure 4B). Also, the number of other hydrogen bonds exceeds the number of C=O hydrogen bonds for mixtures of all four CER subclasses. In contrast, for the pure CER systems, the number of other hydrogen bonds is larger than the C=O hydrogen bonds for CERs AS and CER AP, which contain the OH4 group (O7 site), but not for CERs NS and NP, which do not (Figure 4B).

The hydrogen bonding behaviors of the C=O and N–H sites reveal distinct trends in saturation, interactions within CERs, and hydrogen bonding preferences with FFA and CHOL. C=O hydrogen bonding sites, including bonds with both lipids and water, are fully saturated or nearly so (i.e., each CER forms approximately one C=O hydrogen bond) in pure CER NP, as well as in FFA and CHOL mixtures with CERs NP, NS and AP, and exceed 0.84 in the pure and mixture systems (Table 3). In contrast, the N-H hydrogen bonding sites with both lipids and water in both pure and mixed systems are between half and two-thirds of their one bond per CER capacity (Table 3).

C=O and N-H sites exhibit distinct preferences for certain CER hydrogen bonding partners. The primary bonding partners for C=O are O84 (OH2) and O80 (OH1), with O84 being the most common, except in the CER AS mixture, where O80 forms 5% more bonds than O84 (Table S8). Hydrogen bond numbers for C=O with N-H (O4-N1 bond pair) are similar to those for C=O with O84 (O4-O84 bond pair) in CERs AP and AS but significantly lower in CERs NP and NS. For N-H (Table S9), hydrogen bonds with C=O (N1-O4 pair) are the most common, followed closely by O80 (N1-O80 pair), except in pure CER NS, where the N1-O80 pair slightly outnumber those with C=O (N1-O4 pair). For both C=O and N-H, bonding with OH3 (O88) and OH4 (O7) is significantly less frequent than bonding with their primary partners: O84 and O80 for C=O (O4) and O4 and O80 for N-H (N1).

In CER mixtures, both C=O and N-H sites favor FFA over CHOL, partly because the mixtures include twice as much FFA as CHOL. As noted earlier, this ratio was chosen to reflect the experimental composition of the SPP-type lamellar phase in experimental studies of equimolar lipid mixtures where approximately half of the CHOL phase separates. Table S5 shows that the total number of O4 hydrogen bonds with CHOL (site O3) and FFA (site O25) remains largely unaffected by the CER headgroup. However, the bond distribution consistently favors FFA over CHOL with an ~2:1 ratio, ranging from 1.9 to 2.8 for CERs NS, AS, and NP. For N-H (site N1), which can hydrogen bond with two FFA sites (O25 and O27), the preference of FFA over CHOL is even greater at ~3.3:1 for CERs NS and NP, which lack the OH4 (site O7), and greater still at ~5.1 for CERs AS and AP, which contain OH4 (Table S6). The total number of N-H hydrogen bonds with CHOL and FFA differs only slightly between CERs lacking OH4 (NS and NP) and those containing it (AS and AP), at 1.6 vs. 1.8 per CER. Thus, the presence of OH4 in the CER headgroup increases the ability for FFA and N-H hydrogen bonding, while OH3 has no effect.

These observations highlight the importance of performing simulations using molecular compositions that match those of the experimental lamellar phase, rather than the overall experimental mixture. For example, simulations that use equimolar CER, CHOL, and FFA—ignoring the phase separation of crystalline CHOL that occurs experimentally—model a different lamellar phase than the actual SPP-like phase. In particular, a simulation of 1:1:1 CER:CHOL:FFA would overrepresent the presence of CHOL in the lamellar phase, likely leading to a reduced role for CER-FFA hydrogen bonding than actually occurs. Also, at least in terms of hydrogen bonding, CER subclass matters.

Overall, the presented hydrogen bonding results and analysis illustrate the distinct hydrogen bonding characteristics of each CER subclass in pure and mixed systems. The unique amide I and II-related bonding patterns observed across the CER subclasses underscore the importance of specific molecular interactions that likely affect the stability and organization of lipid lamellae, with important implications for the structural roles of these lipids in the SC.

Comparison with other molecular dynamics simulations

Badhe et al.⁷² studied hydrogen bonding of six CER subclasses, including AP, NP, and AS, with C24 acyl chains and C18 sphingoid base chains in simulations of pure CER bilayers using the united atom GROMOS-Notman force field¹⁶. While they did not simulate CER NS in their work, Badhe et al. did examine CER NdS, which differs from CER NS only by the presence of a double bond in the sphingosine backbone, making it unsaturated. In their bilayer simulations, CER AP formed the most hydrogen bonds between CER molecules (4.4 per CER), while CER NdS had the fewest (2.0 per CER). These trends agree with the results presented here for six-leaflet multilayers, where CER AP also formed the most hydrogen bonds, and CER NS, like CER NdS in their study, formed the fewest. However, unlike the present study of leaflets isolated from bulk water, Badhe et al. found that CER NP in bilayers formed more hydrogen bonds with lipids than CER AS (3.4 vs. 2.8 per CER). This reversal is likely due to CER AS engaging in more hydrogen bonding with interfacial water at the O7 site (OH4), thereby reducing its hydrogen bonding with adjacent CERs. In contrast, CER NP, with its O88 site (OH3) positioned further from the water interface, experienced less competition from water. Supporting this hypothesis, Badhe et al. observed that CER-water hydrogen bonds outnumbered CER-CER bonds for CERs AS and NdS by about one per CER, whereas the numbers for CERs NP and AP were similar, which could be due to the positioning of OH3 (O88 site).

As expected, hydrogen bonding with water is significantly lower in the inner leaflets of a six-leaflet membrane (0.3 to 0.5 per CER) compared to the bilayer simulations by Badhe et al.⁷² (3.6 to 4.3 per CER). For comparisons with experimental systems, multilayer simulations that analyze inner leaflets are preferable, as they better represent the conditions of experimental studies, which contain many layers without direct water contact. A puzzling aspect of Badhe et al.’s findings is that the total number of hydrogen bonds, including both CER-CER and CER-water interactions, exceeds the total number of available hydrogen bonding sites in CERs AP, NP, AS, and NdS by approximately 40%, which is physically impossible. In contrast, the present study finds that hydrogen bonding sites in pure CER systems are only about 50% saturated (Table 3). This discrepancy indicates that the absolute numbers reported by Badhe et al.⁷² must be overestimated (perhaps double counted), even though the overall trends may still be reliable.

3.2. Mixtures of CERs NP and AP with CHOL and FFA

Schmitt et al.⁷³ examined CER:CHOL:FFA C24 mixtures with a molar ratio of 1:0.7:1, where the CER component was CER NP C24 and CER AP C24 with either a 1:2 (AP-rich) or 2:1 (NP-rich) molar ratio. The NP-rich ratio more closely resembles compositions observed in native SC. Both mixtures equilibrated to form an SPP along with separate phase crystalline CHOL. Using neutron diffraction with selective deuteration of the terminal methyl group of the acyl chains of either the CER NP or CER AP, Schmitt et al.⁷³ determined the positions of the terminal methyl groups within the SPP lamellar unit cell. For both mixtures, the distance from the headgroup to the end of the acyl chain attached to that headgroup was shorter for CER AP compared with CER NP, despite the repeat distance, measured by neutron diffraction, being the same (5.45 nm). Furthermore, the distance between the headgroup and terminal methyl group was shorter for both CER AP and CER NP in the AP-rich mixture compared with the NP-rich mixture (Figure S3).

Based on these findings, the assumption that all CERs adopt the hairpin conformation, and an idealized geometric model (i.e., perfectly linear lipid tails), Schmitt et al. hypothesized that the CER acyl chains (along with all other lipid tails) exhibit a larger tilt angle in the AP-rich mixture compared with the NP-rich mixture, while the CER acyl chains show less interdigitation with each other in the AP-rich mixture compared with the NP-rich mixture.⁷³ Increased tilt was inferred from the shorter distance between the headgroup and terminal methyl group in the AP-rich mixture. Less interdigitation was required in the AP-rich mixture to maintain the same repeat distance as the NP-rich mixture while accommodating greater tilt. However, if we accept their idealized model and corresponding hypothesis, the tilt angle in the AP-rich mixture needs to be ~25° larger than in the NP-rich mixture to account for the observed shortening of the headgroup-to-terminal-methyl distance (see Figure S3 and the accompanying text). To explain why the CER AP terminal methyl group was closer to the headgroup plane than that of CER NP in both mixtures, Schmitt et al. hypothesized that the hydroxyl group on the CER AP acyl chain causes it to pull this tail closer to the headgroup region.⁷³

To further test the CG models, six-leaflet multilayers of AP-rich (1:2 molar ratio of CER NP:CER AP) and NP-rich (2:1) mixtures, based on Schmitt et al. but with reduced CHOL (1:0.5:1 CER:CHOL:FFA) to account for the phase-separated crystalline CHOL observed in their experiments, were self-assembled.⁷³ Mixtures at the experimental composition of 1:0.7:1 CER:CHOL:FFA were also studied. This second set of simulations allows the examination of the effect of incorporating extra CHOL in the lipid lamellae and to compare our simulation results with other simulation studies that used the experimental composition. Table 4 reports the thickness, tail tilt angle, interdigitation of all lipid tails, and the S₂ order parameter of the middle leaflet pair, along with the fraction of CERs in the inner four leaflets that were in the extended conformation calculated from the CG simulations. The same metrics are also reported from the reverse-mapped AA simulations performed at the experimental composition (1:0.7:1 CER:CHOL:FFA) as part of the hydrogen bonding analysis described below.

Table 4:

Thickness, tilt angle, interdigitation, and S₂ of the middle leaflet pair and the fraction of extended CERs in the inner four leaflets of self-assembled CG-multilayers for NP-rich and AP-rich mixtures at the experimental composition (1:0.7:1 molar ratio of CER:CHOL:FFA) and reduced CHOL composition (1:0.5:1) calculated from the CG and reverse-mapped all atom simulations; mean ± standard deviation from three independent simulations.

Property	CER:CHOL:FFA 1:0.5:1 (CG)		CER:CHOL:FFA 1:0.7:1 (CG)		CER:CHOL:FFA 1:0.7:1 (AA)
	NP-Rich	AP-Rich	NP-Rich	AP-Rich	NP-Rich	AP-Rich
Leaflet-pair Thickness (nm)	5.32 ± 0.03	5.34 ± 0.01	5.24 ± 0.02	5.24 ± 0.04	5.19 ± 0.02	5.24 ± 0.08
Tilt Angle (deg)	10.7 ± 1.8	9.4 ± 0.4	9.7 ± 1.4	9.9 ± 0.6	11.6 ± 1.4	11.0 ± 0.7
Interdigitation (nm)	1.13 ± 0.02	1.09 ± 0.01	1.09 ± 0.02	1.08 ± 0.01	0.71 ± 0.05	0.70 ± 0.03
S2	0.92 ± 0.03	0.94 ± 0.01	0.94 ± 0.02	0.94 ± 0.01	0.92 ± 0.02	0.93 ± 0.01
Fraction Extended CER
CER NP	0.36 ± 0.02	0.38 ± 0.02	0.38 ± 0.02	0.32 ± 0.01	0.37 ± 0.03	0.33 ± 0.02
CER AP	0.34 ± 0.07	0.36 ± 0.03	0.37 ± 0.07	0.39 ± 0.01	0.37 ± 0.06	0.39 ± 0.02
CERs NP and AP	0.36 ± 0.04	0.37 ± 0.01	0.37 ± 0.04	0.37 ± 0.01	0.37 ± 0.04	0.37 ± 0.02

As can be seen from Table 4, tail tilt, tail interdigitation (of all lipid tails), and S₂ of the middle leaflet pair, as well as the fraction of CERs in the extended conformation in the inner four leaflets were found to show no statistically significant differences between the NP- and AP-rich systems at either the low CHOL or experimental compositions. This differs from the hypothetical structures proposed by Schmitt et al., in which the AP-rich mixture exhibited larger tilt angles for the lipid tails and less interdigitation between the CER acyl chains compared to the NP-rich mixture. Interdigitation of the acyl tails of CER NP, CER AP, and FFA was also calculated for each lipid type individually and for their combinations (CER NP with CER AP, and both CERs with FFA). No statistically significantly differences were observed (Table S12).

Consistent with distances in the experiments, both mixtures had the same middle leaflet-pair thickness. At the low CHOL composition, which better approximates the CHOL content of the experimental lamellae, the thickness was ~5.33 nm, in close agreement with the experimental repeat distance of 5.45 nm. At the experimental composition, the thickness was 5.24 nm, which was a statistically significant, albeit small, decrease from the 5.33 nm thickness of the low CHOL composition, reflecting the higher proportion of shorter lipids (CHOL) relative to the longer ones (FFA and CERs). At the experimental composition, the fraction of extended CER was slightly higher for CER AP (0.39 ± 0.01) than for CER NP (0.32 ± 0.01) in the AP-rich mixture. However, in the NP-rich mixture at the same composition, and in both mixtures at the low CHOL composition, CER AP and NP had similar fractions in the extended conformation. Structural properties from the reverse-mapped all atom simulations closely matched those from the CG simulations, except for interdigitation, which was smaller in the reverse-mapped all atom simulations (~0.7 nm vs. ~1.1 nm). This outcome should be expected. The relatively gentle atomistic relaxation in the reverse-mapping protocol preserves the overall lipid organization from the CG simulations, resulting in similar values for most structural properties, whereas the finer resolution of atomistic tails allows more curling and kinking than the CG model can capture, leading to reduced interdigitation.

Recently, Badhe et al.^74,75 and Rivero et al.^76,77 described molecular simulations of lipid mixtures based upon the NP- and AP-rich experiments from Schmitt et al.⁷³. Both studies carried out atomistic simulations of pre-assembled triple bilayers of CERs NP and AP, at 1:2 and 2:1 molar ratios, mixed with CHOL, and FFA at the experimental composition (1:0.7:1 molar ratio), using the CHARMM36 AA force field.^74,76 All CERs were in the hairpin conformation. The Rivero et al. simulations included two water molecules per lipid in the headgroup region of the middle bilayer, which correspond to one water molecule per lipid in the four inner leaflets, whereas Badhe et al. included none. Additionally, Rivero et al. assumed that FFA was protonated, while Badhe et al. assumed FFA was dissociated with potassium cations added to neutralize the system. Rivero et al. used random walk molecular dynamics (RWMD)⁶⁷ to reduce the likelihood of correlation between the initial and final configurations. Both studies calculated bilayer thickness, quantities related to interdigitation, tail tilt, and hydrogen bonding, although using different approaches.

Badhe et al.⁷⁴ calculated repeat distances from simulated NSLD profiles that differed significantly between the AP-rich (4.0 nm) and NP-rich (4.8 nm) mixtures, in contrast to the much larger but consistent experimental repeat distance (5.45 nm) for the two mixtures. Interdigitation was assessed for all lipid tails using the product of the number densities from the upper and lower leaflets of the bilayer normalized by the maximum value. In a frequency versus z plot, this quantity shows a peak at the bilayer midpoint and a width approximating the extent of interdigitation, which was slightly larger for the AP-rich mixture in contradiction with the Schmitt et al. hypothesis. Tilt angle, defined as the angle between the bilayer normal and a vector connecting the first and fourth carbon atom of a four consecutive carbon atom segment, was calculated for the CER and FFA tails starting from the first carbon atom of each tail to its last carbon. For each four-carbon segment of each lipid tail, the average angle of the AP-rich mixture exceeded that segment in the NP-rich mixture, leading the authors to conclude that their simulations agreed with the Schmitt et al. hypothesis of larger tilt in the AP-rich mixture. However, they did not assess if their calculated tilt angle increase could account for the decreased headgroup-to-terminal methyl distance observed by Schmitt et al.⁷³

Unlike Badhe et al., Rivero et al. determined that the average bilayer thickness (using the reference atoms in the headgroup approach) of the middle bilayer was almost the same for the NP- and AP-rich mixtures (5.44 nm for NP-rich and 5.48 nm for AP-rich) and similar to the experiments.⁷⁶ They defined CER tilt as the angle between the bilayer normal and a vector connecting two carbon atoms at opposite ends of the C24 acyl tail: the second carbon from the carboxyl group and the second carbon from the terminal methyl (Figure S4). By this definition, the average tilt angle of acyl chains from both CERs was, as proposed by Schmitt et al., slightly larger in the AP-rich mixture (~10° compared with ~6° in the NP-rich mixture). However, this ~4° difference is far smaller than the ~25° tilt angle increase that would be required, based on the Schmitt et al.⁷³ idealized hypothesis, to account for the observed shortening of the headgroup-to-terminal methyl distance. Rivero et al. also concluded from density profiles of the lipids combined with measurements of the length of a vector drawn from OH1 (site O80 in Figure 2) of the CER headgroup to the terminal methyl of the acyl tail that there was “overlap” of CER acyl tails in the NP-rich mixture, but not, or to a much lesser extent, in the AP-rich mixture.

To summarize, Badhe et al.⁷⁴ reported different and smaller repeat distances for the two mixtures, in contrast with the experiments, where the repeat distances were identical and larger. They also measured more interdigitation of all lipids and larger tilt angle for the AP-rich mixtures. In contrast, Rivero et al.⁷⁶ measured bilayer thicknesses that were the same for the two mixtures and similar to the experiments, along with less interdigitation of the CER acyl tails and slightly more lipid tilt for the CER AP mixture, as hypothesized by Schmitt et al., although the increased tilt was far too small to support the different tilt hypothesis. In this study, no differences were observed in tail tilt, tail interdigitation and leaflet pair thickness (repeat distance) for the NP- and AP-rich systems and leaflet pair thickness was similar to the experimental value. Also, no statistically significant differences were observed in tilt angles from the reverse-mapped all atom simulations when these were calculated using the methods of Badhe et al.⁷⁴ and Rivero et al.⁷⁶ (data not shown). Thus, multilayer simulations modeling Schmitt’s synthetic lipids are not yet able to corroborate that differences in tilt and reduced interdigitation underlie the experimental observations for the AP-rich versus NP-rich systems.

Potential causes for the differences in structural parameters reported from the three studies include both the methods used to calculate these parameters and differences in the system initialization and simulation protocols. For example, the larger leaflet pair (bilayer) thickness from Rivero et al. compared with this study is consistent with their use of a headgroup-based reference atom method, which is known to give larger values than the distance between headgroup density peaks used here.¹⁶ Important differences in the systems simulated in the three studies include the presence (this study and Rivero et al.) or absence (Badhe et al.) of water on the outer boundary, the amount of water in the inner four leaflets (none for Badhe, ~ 0.4 waters per lipid in this study, and 1 water per lipid in Rivero et al.), and fully protonated (this study and Rivero et al.) or dissociated (Badhe et al.) FFA. Differences in the simulated configuration may also have affected the results. Rivero et al. and Badhe et al. used a stack of three pre-assembled bilayers with CERs in the hairpin conformation, whereas the self-assembled systems of this study contained approximately 40% of the CERs in the extended conformation and 0.4 water molecules per lipid in the inner four leaflets. Atomistic simulations of pre-assembled systems are prone to correlation with the starting configuration,¹⁶ although the use of the RWMD method makes this less likely in the simulations from Rivero et al. However, extended CER conformations are much less likely to be observed in the final configurations of both the Badhe et al. and Rivero et al. simulations. Recent experimental studies indicate that the majority of CERs are extended in SPP-forming SC lipid multilayer model membranes,^13,14,78,79 which may require a revision in the hypothesis proposed by Schmitt et al.⁷³

Badhe et al. and Rivero et al. also reported hydrogen bonding in the NP- and AP-rich mixtures. To compare with their results, the self-assembled CG simulations of mixtures at the experimental composition were reverse-mapped to recover atomistic details, from which hydrogen bonding was calculated for the inner four leaflets. The GROMACS⁴⁵ method of calculating hydrogen bonds was the same as that used by Badhe et al. In contrast, Rivero et al. used the TRAVIS^80,81 software to calculate the average number of hydrogen bonds from the number integral of the first peak of the first radial distribution function. Rivero et al. reported bonds for the central bilayer with the neighboring monolayers, i.e., the inner four leaflets. Because Badhe et al. did not include water on the outer boundaries of their system, their hydrogen bond results were not constrained to the inner leaflets, which were also dehydrated. The Rivero et al. simulations included one water molecule per lipid in the four inner leaflets. In our simulations, the 0.35-0.42 water molecules per lipid that were present after CG self-assembly were retained in the atomistic simulations.

Figure 5 compares our hydrogen bonding results with those from Badhe et al.^74,75 and Rivero et al.^76,77 Table 5 summarizes for the three studies the total number of hydrogen bonds between two lipids, between a lipid and water, and the sum of both. Section 3 of the Supporting Information and Table S10 provide additional details.

Figure 5:

Normalized number of hydrogen bonds formed by the lipid listed at the top of the four sections in the graph with the four lipid types and water in NP-rich and AP-rich mixtures (1:2 and 2:1 molar ratios of CER AP:CER NP, respectively) at the experimental composition (CER:CHOL:FFA molar ratio 1:0.7:1). Results from this study are compared with those reported by Badhe et al.^74,75 and Rivero et al.⁷⁶. Badhe et al. simulations contained no water. Data for CHOL, FFA, and water are not shown for bars marked with asterisks because Rivero et al. did not explain the normalization method used for hydrogen bonds involving lipids other than CERs (see Section S3 in the Supporting Information).

Table 5:

Number of hydrogen bonds between each lipid type with lipids or water, normalized by the number of molecules of that lipid type, from simulations of NP-rich and AP-rich mixtures (1:2 and 2:1 molar ratios of CER AP:CER NP, respectively) with CHOL and FFA at the experimental composition (CER:CHOL:FFA molar ratio 1:0.7:1). Results from this study are compared with those reported by Badhe et al.^74,75 and Rivero et al.^{76,77 a}

Lipid	Hydrogen bonded with	This study		Badheb		Riveroc
		NP-rich	AP-rich	NP-rich	AP-rich	NP-rich	AP-rich
CER AP	Lipids	3.52 ± 0.13	3.12 ± 0.10	3.83	3.50	3.79 ± 0.04	3.71 ± 0.14
	Water	1.06 ± 0.06	1.16 ± 0.03			1.92 ± 0.19	2.04 ± 0.06
	Lipids + water	4.58 ± 0.12	4.28 ± 0.07	3.83	3.50	5.71 ± 0.18	5.75 ± 0.09
CER NP	Lipids	2.77 ± 0.13	3.09 ± 0.24	3.10	3.19	2.82 ± 0.09	3.02 ± 0.12
	Water	0.90 ± 0.03	0.99 ± 0.17			1.52 ± 0.04	1.47 ± 0.17
	Lipids + water	3.67 ± 0.15	4.08 ± 0.36	3.10	3.19	4.33 ± 0.12	4.49 ± 0.05
CHOL	Lipids	0.80 ± 0.06	0.77 ± 0.02	1.14	1.14
	Water	0.31 ± 0.02	0.33 ± 0.04
	Lipids + water	1.11± 0.06	1.10 ± 0.02	1.14	1.14
FFA	Lipids	1.24 ± 0.06	1.23 ± 0.06	1.45	1.44
	Water	0.64 ± 0.05	0.71 ± 0.02
	Lipids + water	1.88 ± 0.11	1.93 ± 0.08	1.45	1.44
Water molecules/lipid in the inner four leaflets		0.31 ± 0.02	0.35 ± 0.03	0	0	1	1

^aValues from this study were calculated at 305 K using hydrogen bond counts (Table S10) and lipid compositions (listed in Table S11) from the four inner leaflets of reverse-mapped atomistic simulations of CG self-assembled six-leaflet membranes. Badhe et al. results are from dehydrated six-leaflet membranes (temperature not specified, likely 305 K to match experiments). Rivero et al. results are from the four inner leaflets of hydrated six-leaflet membranes at 303 K. Data from all three studies are reported as the mean ± standard deviation from three replicated simulations.

^bBadhe et al. ^74,75 reported values as mean ± standard deviation for CER NP, CER AP, CHOL, and FFA C24. The totals shown in this table were calculated from those mean values; standard deviations are not reported because individual data points were not available.

^cOnly results involving CER AP and CER NP are presented, as Rivero et al. ^76,77 did not specify the normalization scheme for the reported hydrogen bonding values not involving a CER as Lipid 1 (see Section S3 in the Supporting Information).

The total number of hydrogen bonds that each lipid type forms with lipids and water combined aligns with its number of hydrogen bonding sites: six for CER AP, five for CER NP, one for CHOL, and two for FFA. Except for CHOL, the total number of hydrogen bonds per designated lipid increased with increasing water content. Values from this study are higher than those reported by Badhe et al. for dehydrated mixtures, but lower than those from Rivero et al., which included more than twice as much water. In the Rivero et al. study, both CERs are close to their bonding capacity in both the NP- and AP-rich mixtures (~5.7 hydrogen bonds per CER AP and ~4.5 bonds per CER NP), with water hydrogen bonding accounting for ~2.0 bonds per CER AP and 1.5 bonds per CER NP. In both this study and that of Badhe et al., CHOL is at its hydrogen bonding capacity (CHOL data from Rivero et al. is incomplete).

The total number of CER-lipid hydrogen bonds was similar across the studies, although there is some variation in the distribution among the lipid types. In all cases, CER AP forms more hydrogen bonds with both CER AP and CER NP in the AP-rich mixture than in the NP-rich mixture. Similarly, CER NP forms more hydrogen bonds with both CERs in the NP-rich mixture. CHOL also follows this trend, forming more hydrogen bonds with CER NP than with CER AP in NP-rich mixtures, and fewer with CER NP in AP-rich mixtures.

The number of hydrogen bonds formed by FFA with either CHOL or FFA is relatively insensitive to the CER composition in both Badhe et al. and this study (hydrogen bonding results for FFA and CHOL with themselves are not presented for Rivero et al.). In general, FFA hydrogen bonding with both CERs and CHOL is greater in the Badhe et al. simulations than in this study; FFA-CER bonding was also greater in Badhe et al. than in Rivero et al. Badhe et al. observed no FFA–FFA hydrogen bonding, consistent with ionized FFA in their simulations.

As water content increases, CHOL tends to form fewer hydrogen bonds with other lipid types and more with water, as illustrated by CER NP hydrogen bonds with CHOL in both the NP- and AP-rich mixtures: the number of CER NP-CHOL bonds (shown in pink in the CER NP bars of Figure 5) are highest in the water-free Badhe et al. simulations, lowest in the Rivero et al. simulations with the most water, and intermediate in this study, which had about half as much water as in Rivero et al. CHOL–CHOL hydrogen bonds are essentially absent (no pink in the CHOL bars of Figure 5) in both Badhe et al. and this study (data not available for Rivero et al.), consistent with CHOL’s positioning between the more flexible CERs or FFAs, which promotes hydrogen bonds with these other lipids rather than with another CHOL.

Collectively, results of atomistic multilayer simulations of the same CER NP:AP:CHOL:FFA composition (1:0.7:1 molar ratio) from three different studies highlight the sensitivity of SC lipid simulations to the models used, system composition (e.g., the presence and amount of water, protonated or deprotonated FFA), initial system setup (e.g., hairpin or extended conformation), and simulation protocols.

3.3. Synthetic Lipid Mixtures that Mimic SC Lipid Behavior

Since CG models for CERs NP, AS and AP are now available, it is possible to simulate more representative mixtures of the SC. Specifically, the stratum corneum substitute (SCS) developed by Bouwstra et al. can now be examined. The SPP version of the SCS is an equimolar mixture of CER_mix, CHOL and FFA7, where CER_mix consists of five CERs (NS, NP, AP, and AS with C24 acyl chains, plus AP with C16 acyl chains) at molar ratios of 60:19:11:6:5. FFA7 is a mixture of seven FFAs (C16, C18, C20, C22, C23, C24, and C26) at molar ratios of 1.8:4.0:7.7:42.6,5.2:34.7:4.1, corresponding to a C22.4 average chain length.^41-43 Uchiyama et al.⁴² conducted a series of experiments comparing membranes of this SCS formulation with those of three variants: one in which CER NS C24 replaced CER_mix, another where FFA C24 replaced FFA7, and a third where CER NS C24 and FFA C24 replaced CER_mix and FFA7. The repeat distances of the four mixtures, determined from X-ray diffraction, were identical at 5.4 nm.⁴² However, membranes containing FFA7 exhibited higher permeability of ethyl p-aminobenzoate (E-PABA) than those with only FFA C24, whereas replacing CER_mix with CER NS C24 had minimal effect.

To better understand the molecular mechanisms underlying the experimental findings, a series of simulations investigating these four mixtures were performed. The study focused specifically on two factors affecting permeability: lipid tail order, measured by the nematic order parameter (S₂), and hydrogen bond strength, as represented by the total number of hydrogen bonds between lipid headgroups.

Because the CG model represents fatty acid tails with TAIL beads of three-carbon segments and a terminal bead containing either two (TER2) or three (TER) carbons, it cannot explicitly model FFAs of lengths C20, C23, or C26. Therefore, the experimental FFA7 mixture was approximated using a CG five-component mixture (FFA5_CG) consisting of C16, C18, C21, C22, and C24 at molar ratios of 1.8:4.0:7.6:47.8:38.8. Except for substituting FFA C21 for FFA C20, FFA5_CG matches the 5-component perdeuterated FFA mixture (FFA5) that has been used instead of FFA7 in some experiments because perdeuterated FFA C23 and C26 were not available.^{15,41,42,82-88} When the self-assembled CG multilayers were reverse-mapped to AA multilayers, FFA C21 was mapped as FFA C20 (i.e., TER2 was mapped as CH₃), thereby replicating the experimental FFA5 substitute for FFA7. The average chain lengths of the FFA5 and FFA7 mixtures are essentially the same.⁸⁶ Also, because SCS membranes contain phase-separated crystalline CHOL in addition to the SPP, the simulated systems contained half the experimental CHOL amount to more closely represent the relative proportions of CER, CHOL, and FFA in the SPP lamellae.

Six-leaflet multilayers were self-assembled and then reverse-mapped to restore atomic details in order to calculate S₂ of the lipid tails and hydrogen bond counts between lipids, as well as the average leaflet-pair thickness and tilt angle. To better represent the conditions of the multilayer experiments, the average leaflet-pair thickness was calculated for lipids in the middle leaflet pair to exclude bulk water effects on tail ordering in the outer (top and bottom) leaflet pairs. Hydrogen bonds were counted in the inner four leaflets to capture interactions among all headgroups that are unaffected by bulk water. Average leaflet-pair tilt angle and S₂ were calculated for both the middle leaflet pair and the inner four leaflets. Results for S₂ and hydrogen bond counts for the inner four leaflets are presented in Figure 6; leaflet-pair thickness and tilt angle are listed in Table 6. A detailed breakdown of hydrogen bond counts with each lipid class and with water is provided in Table S14. Lipid and water compositions in the inner four leaflets are reported in Table S13.

Figure 6:

Simulation results for the four SCS variants (CER:CHOL:FFA, 1:0.5:1 molar ratio) with CER as either CER NS or CER_mix and FFA as either FFA C24 or FFA5, measured in the four inner leaflets of six-leaflet membranes: (A) nematic order parameter (S₂), and (B) total number of lipid-lipid hydrogen bonds. The SCS variants are color-coded by FFA type: shades of blue for FFA C24 and shades of red for FFA5. Bars representing CER NS systems are marked with circles, while CER_mix systems are marked with diagonal lines. Statistically significant differences between pairs of SCS variants are denoted with asterisks: *p < 0.07, **p < 0.01, or ***p < 0.001.

Table 6:

Structural properties of simulated six-leaflet membranes of the four SCS variants (CER:CHOL: FFA, 1:0.5:1 molar ratio).

Property	Leaflets in calculation	CER NS FFA C24	CERmix FFA C24	CER NS FFA5	CERmix FFA5
Leaflet-pair thickness (nm)	Middle pair	5.1 ± 0.1	5.2 ± 0.2	4.9 ± 0.1	4.7 ± 0.03
Interdigitation	Middle pair	0.74 ± 0.07	0.75 ± 0.01	0.72 ± 0.02	0.74 ± 0.05
Tilt angle (deg)	Middle pair	10.8 ± 0.5	12.1 ± 0.7	13.0 ± 0.7	13.6 ± 0.1
	Inner four	10.9 ± 0.4	12.3 ± 0.7	12.7 ± 0.7	13.6 ± 0.1
S2	Middle pair	0.930 ± 0.004	0.920 ± 0.005	0.893 ± 0.009	0.891 ± 0.003
	Inner four	0.925 ± 0.007	0.918 ± 0.004	0.898 ± 0.009	0.887 ± 0.005
Fraction CER extended	Inner four	0.323 ± 0.035	0.366 ± 0.012	0.348 ± 0.037	0.369 ± 0.032

As shown in Figure 6, S₂ was significantly higher in mixtures containing the same CERs but with FFA C24 instead of FFA5, while replacing CER_mix with CER NS in mixtures with the same FFA content had no significant effect. The increase in tail ordering with FFA C24 is expected, as uniform FFA chain lengths promote more consistent lipid packing. The lack of change with CER_mix replacement suggests that headgroup differences in these CERs do not significantly affect tail ordering.

The hydrogen bonding analysis revealed that variations in FFA chain length in mixtures of either CER_mix or CER NS caused only a small, non-significant decrease in the total number of hydrogen bonds, whereas replacing CER_mix with CER NS led to a statistically significant decrease. These results align with the smaller number of hydrogen bonding sites in CER NS, with two hydroxyl groups, compared with CER_mix, with an average of 2.5 hydroxyl groups (three for CERs NP and AS, and four for CER AP), and the absence of different hydrogen bonding numbers for FFA5 compared with FFA C24. Interestingly, compared to FFA C24, the chain length variation in FFA5 did reduce hydrogen bonding between FFA molecules by approximately 20%, but had no effect on FFA hydrogen bonding with CHOL, CERs, or water (Table S14).

Permeation experiments showed that membranes with FFA7 had higher E-PABA permeability than those with only FFA C24, while substituting CER_mix with CER NS C24 had little effect.⁴² Comparison with simulation results for S₂ and hydrogen bonding suggest that this increased permeability correlates with reduced tail ordering (lower S₂ for FFA5 vs. FFA C24) but not with differences in hydrogen bond numbers, which differed significantly between CER NS and CER_mix, but not between FFA5 and FFA C24.

Table 6 summarizes the structural properties from AA simulations of reverse-mapped CG self-assembled six-leaflet multilayers for the four SCS variants. In the simulations, the middle leaflet pair was slightly thinner (4.7 to 5.2 nm) than the 5.4 nm experimental repeat distance. Leaflet-pair thickness did not depend on whether the mixture contained CER NS or CER_mix, but it was larger with FFA C24 than with FFA5, consistent with the smaller average chain length of FFA5 (22.4 vs. 24 carbons). For fully extended tails with no tilt, this difference could correspond to about 0.4 nm, assuming 0.15 nm per CH₂ group. However, the thickness difference was statistically significant only for FFA5 with CER_mix, which exhibited the largest tilt angle, although the tilt angle is not large enough to make a measurable difference on thickness. Interestingly, both tilt angle and S₂ were the same for the inner four leaflets and the middle leaflet pair (Table 6) indicating that, at least in these simulations, the hairpin conformation of the outermost leaflets, imposed by bulk water, did not influence tilt angle or tail ordering of the inner leaflets.

Self-assembly produced nonuniform lipid distributions across the leaflets, which led to either enrichment or depletion of a lipid class in the four inner leaflets compared with the outer ones. While this could affect S₂ and hydrogen bonding along with leaflet-pair thickness, it is unlikely given the small deviations of the inner four leaflets from the uniform distribution (Table S13): (a) ~2% fewer lipids, (b) 3-8% fewer FFA molecules, and (c) ~4% extra CHOL in the mixtures containing only FFA C24. Interestingly, the inner four leaflets of the four SCS-type mixtures contain ~0.35 water molecules per lipid, all of it associated with the lipid headgroups. This is consistent with simulations of the mixtures summarized in Table 2, and comparable to the approximately one water per lipid observed experimentally.⁷¹ While lipid hydrogen bonding with these water molecules is significant, it is about 50% smaller than the number of lipid-lipid hydrogen bonds in all four SCS variants (Table S14).

4. Conclusions

Computational study of SC lipids requires models that can self-assemble into multiple layers while accurately capturing their complex interactions in order to avoid potential initialization bias. This was achieved herein using a multiscale approach that combined self-assembly of CG models with reverse-mapping, to recover atomistic-level structure, followed by atomistic simulations. Furthermore, to better reflect experimental multilayer lipids, hydrogen bonding and structural parameters were evaluated in the four inner leaflets or middle leaflet pair of the simulated six-leaflet multilayers to exclude contributions from lipid-water hydrogen bonds in the outer leaflets, where headgroups are exposed to bulk water. The studies, using newly developed CG models for CERs NP, AS, and AP together with the existing CG model for CER NS, of pure CERs, ternary CER:CHOL:FFA mixtures, and more complex multicomponent systems, provide molecular-level insights into several SPP-forming experimental SC lipid models that complement and extend experimental observations.

Analysis of hydrogen bonding in simulations of the different CER subclasses showed that the number and positions of hydroxyl sites influence CER-lipid interactions, in agreement with experimental amide I FTIR trends. Simulations of pure CERs reproduced experimental trends in amide I related hydrogen bonding: phytosphingosine CERs (NP and AP) exhibited more C=O hydrogen bonds consistent with a lower amide I peak frequency than their sphingosine counterparts (NS and AS), which was also the case for non-hydroxy CERs (NS and NP) compared with their α-hydroxy analogs (AS and AP). Mixing CHOL and FFA with CER modestly reduced the number of C=O hydrogen bonds in simulations, unlike the unchanged hydrogen bonding strengths deduced from the amide I peak frequencies in experiments—a difference that may arise from the lower water content of the simulations.

In CER NP and AP mixtures with CHOL and FFA, the simulated systems reproduced consistent experimental lamellar thicknesses for both the NP and AP-rich systems. More broadly, mixtures of CHOL with CER NS and FFA C24 yielded similar lamellar thicknesses to those containing a mixture of CER subclasses and/or FFAs with varying chain length, despite differences in hydrogen bonding and tail ordering. Comparisons of simulation and experimental results for these mixtures indicate that increased permeability of E-PABA is associated with reduced tail ordering from variation in FFA chain lengths, rather than with reduced hydrogen bonding associated with changes of the CER subclass in the mixture.

Overall, the findings of this work demonstrate multiscale modeling as a powerful framework for probing SC lipid organization and barrier function. By clarifying how CER subclass diversity and lipid composition shape structural features and hydrogen bonding, this work extends experimental interpretations and provides mechanistic insight into the molecular basis of skin barrier properties.