Partitioning into phosphatidylcholine – cholesterol membranes: liposome measurements, coarse-grained simulations, and implications for bioaccumulation †

Membrane – water partitioning is an important physical property for the assessment of bioaccumulation and environmental impact. Here, we advance simulation methodology for predicting the partitioning of small molecules into lipid membranes and compare the computational predictions to experimental measurements in liposomes. As a step towards high-throughput screening, we present an automated mapping and parametrization procedure to produce coarse-grained models compatible with the Martini 3 force ﬁ eld. The methodology is general and can also be used for other applications where coarse-grained simulations are appropriate. This article addresses the e ﬀ ect on membrane – water partitioning of adding cholesterol to POPC (1-palmitoyl-2-oleoyl-sn -glycero-3-phosphocholine) membranes. Nine contrasting neutral, zwitterionic and charged solutes are tested. Agreement between experiment and simulation is generally good, with the most challenging cases being permanently charged solutes. For all solutes, partitioning is found to be insensitive to membrane cholesterol concentration up to 25% mole fraction. Hence, for assessment of bioaccumulation into a range of membranes (such as those found in ﬁ sh), partitioning data measured in pure lipid membranes are still informative.


Introduction
A global regulatory paradigm shi is currently underway to reduce or even eliminate the use of animal testing for assessing the safety of chemicals. 1Changes to legislation banning animal testing for cosmetic safety have been introduced in many countries. 2Historically, legislation has been focussed on in vivo studies to understand chemical toxicity endpoints (hazard) which are then contextualized within an assessment of risk through consideration of exposure. 3Integral to this regulatory assessment when considering environmental impacts is the need to assess and understand the potential for bioaccumulation of a chemical in sh. 4 In the European Union under REACH regulations, for example, chemicals manufactured at more than 100 tonnes per year require an assessment of bioaccumulation. 5This is typically characterized using metrics such as the whole-body bioconcentration factor (BCF), historically using in vivo studies.However, recent efforts now provide tangible options for replacing in vivo studies with computational approaches to assess bioaccumulation. 4,6or organic chemicals, the BCF is positively correlated with the octanol-water partition coefficient K OW (which refers to the charge-neutral form of a solute) and the octanol-water distribution coefficient D OW (which accounts for all relevant chargestates at a given pH). 7,8However, for many ionizable compounds these constants are difficult to measure or predict.Surface-active compounds (surfactants) tend to accumulate at the interface between hydrophobic and hydrophilic phases and form aggregates in solution, making experimental determination challenging. 9Furthermore, octanol-water partitioning does not fully reect the interactions of a molecule with the phospholipid bilayers in biological membranes.This is particularly notable for ionizable chemicals, which can have hydrophobic, polar and charge-charge interactions with the chemical. 10uch simple models for estimating BCF using octanol-water partitioning and one-compartment partitioning are suitable only for non-polar substances that are not biotransformed.In silico bioconcentration models such as BIONIC 6 and others 11 use hepatic and whole-body biotransformation clearance rates derived from in vitro methods combined with partitioning information to account for biotransformation and ionization.The membrane-water partition coefficient K MW and pHdependent membrane-water distribution coefficient D MW are used by these models as key descriptors of sorption to phospholipids instead of only considering their octanol-water counterparts K OW and D OW .1][22] Each has its strengths and weaknesses.
Liposomes or lipid vesicles provide the closest analogue to a cell membrane and there are several analytical methods for measuring the partitioning of a compound into a liposomewater system. 23However, it can be difficult and time-consuming to prepare and store the liposomes.
Solid supported lipid membranes (SSLM) 24 consist of lipid bilayers, non-covalently bound to silica beads.These SSLM beads are commercially available, which allows for simpler and more consistent experiments.The results have been shown to correlate well with liposome experiments for drug-like molecules. 25n immobilized articial membrane (IAM) chromatography, a lipid monolayer is bound covalently to a HPLC column, and the retention factor is measured as in other HPLC methods. 26his method is efficient, and has been used to estimate lipophilicity in several ecotoxological studies. 15,27However, the environment in an HPLC column does not perfectly represent the structure of a real lipid membrane.In particular, the interactions of charged species with the silica surface can affect the results, and conversion of retention factors to partition or distribution coefficients requires empirical correction factors. 10,22Additionally, IAM columns generally consist of onecomponent lipid membranes, limiting their use in investigating how membrane composition affects solute partitioning.
Computational approaches are attractive alternatives to experimental methods, due to their efficiency, ability to overcome some of the practical limitations in experimental approaches, and lower cost.Several methods for calculating K MW exist, with varying levels of accuracy and complexity.Simple fragment-based approaches 8,28 and polyparameter linear free energy relationships (pp-LFER)where partition coefficients are related to solute descriptors via an empirically tted linear equation 29,30 are accessible and easy to apply but have limited coverage of chemical domain.As pp-LFERs require measured data to create the model, they oen do not cover materials for which it is difficult to measure K MW , such as ionizable chemicals.More comprehensive approaches include COSMOmic, 31 which is an extension of the COnductor-like Screening MOdel for Realistic Solvation (COSMO-RS) 32 to consider micelles and membranes.In COSMOmic, a lipid bilayer-water system is represented as layers of isotropic solvents with differing compositions.Membrane-solute interactions are calculated by applying COSMO-RS to each layer to calculate the change in free energy across the system.COS-MOmic works well for most neutral organic molecules, 33 but has issues with some families of molecules like peruorinated surfactants. 19Both pp-LFERs and COSMOmic can struggle to model ionic compounds, 34 although COSMOmic can be improved using an additional empirical membrane potential. 35olecular dynamics (MD) is a general-purpose simulation method where a system is studied by propagating Newton's equations of motion.Atomistic MD, where molecules are modeled at the atomic level by 'ball-and-spring' representations, has been applied to lipid membranes and their interactions with small molecules, 36 and gives good results for K MW . 37owever, a notable disadvantage of atomistic MD compared to the other computational methods discussed here is its high computational cost, which prevents its use in high-throughput screening.This cost can be alleviated using coarse-grained (CG) MD, where groups of atoms are combined into interaction sites known as 'beads'.The Martini force eld is a prominent CG model for biological systems. 38,39It uses a building-block approach, where the interactions between CG beads are dened by selecting from a list of predened bead types.We have recently developed an efficient algorithm for automatic parametrization of solute models in Martini version 2, 38 and have applied it to membrane-water partitioning. 40The most recent version of the force eld, Martini 3, has more exibility in its bead types, 41 and as been shown to capture the thermodynamic properties of small organic molecules well. 42A more detailed explanation of our latest Martini scheme is given in Computational methods below.
Real cell membranes are more complex than the onecomponent model membranes usually used to estimate K MW and D MW .They contain multiple lipid types, as well as sterols, proteins, and other biomolecules. 43In particular, cholesterol is prevalent in most mammalian and sh cell membranes, and is essential for their biological function. 44Lipid content (including cholesterol) varies considerably between membranes.This is as true between different cell, tissue and organelle types within an organism as between species.Further temporal and geographic variability is also observed with organisms where compensatory lipid mechanisms exist to maintain membrane functions (such as uidity, ion transport and permeability) under changing stress and environmental factors such as diet and temperature. 45,46Cell membranes in sh typically contain 2 to 38 mol% cholesterol; [45][46][47][48][49] for example, liver cells and erythrocytes in trout lie in the range 15-22 mol%. 50][54][55] While some experimental data for partitioning into cholesterol-containing membranes do exist, few studies have looked at a wide chemical space using consistent methodology.The aim of this study is to help address this gap and gain greater understanding of the inuence of the presence of cholesterol on partitioning.We have, therefore, determined experimental liposome-water partitioning values in both pure 1-palmitoyl-2oleoyl-glycero-3-phosphocholine (POPC) and POPC-cholesterol membranes, for a varied set of chemicals using cholesterol levels of 0-25 mol%, which align with those typical of sh membranes.We have also updated our CG simulation protocol for Martini 3 and assessed it against the experimental measurements.The conclusions from this study will be useful in informing chemical risk assessment methodology, notably for parametrizing bioaccumulation models, and whether including realistic membrane compositions is likely to be benecial over single-component phospholipid bilayers.The results will also help determine whether K MW and D MW measures of partitioning sufficiently capture the interactions of chemicals in a phospholipid-cholesterol model.

Liposome preparation and measurements
Liposomes were prepared based on the small-volume extrusion method for the preparation of large, unilamellar liposome vesicles. 56Cholesterol was added where necessary, to make up liposomes with 0, 10-15 and 20-25 mol% cholesterol.Liposome solutions were stored at 4 °C for up to 2 weeks before use.See ESI † for additional details of liposome preparation.
The following nine molecules were tested for partitioning to cover a range of ionization states under pH-neutral conditions, to reect their predominant state in the environment: 3-nitroaniline, diclofenac, sulfobetaine 3-10 (SB3-10), dodecyldimethylamine oxide (DDAO-C12), propranolol, dihexylamine, octyl sulfate, decyl sulfate and dodecyl sulfate.An additional seven moleculestryptophan, urea, diethyl adipate, hexylamine, dodecylamine, trihexylamine, and tridecyl sulfatewere also tested, but for reasons explained in the Discussion section, difficulties were encountered for these cases and the study proceeded with the remaining nine.The materials were clustered into four sets according to their suitability for analysis under the same analytical method, and tested in these groups.This procedure is described in the ESI.† The liposome-water partitioning was determined in a 96well plate rapid equilibrium dialysis (RED) device (by Thermo Fisher Scientic) based on a protein-binding protocol. 57Each well in the RED device contains a donor cell (red cell) and a receptor (white cell) separated by a 8 kDa molecular weight cut-off cellulose membrane, through which the liposomes cannot permeate.The liposome solutions were diluted to approximately 4 mM with PBS solution (pH 7.4) and dosed with the test chemicals at three different nominal concentrations of 10, 20 and 50 mM.Analysis was carried out by liquid chromatography coupled to electrospray ionization mass spectrometry (LC-MS/MS) using an Agilent 1200 LC system coupled to an Agilent 6460 Triple Quadrupole Mass Spectrometer.Further details of the method are detailed in the ESI.†

Calculation of experimental liposome-water partitioning
K MW is the ratio of the concentrations of a particular species between the membrane and water phases, where C M is the concentration of solute per unit mass of membrane (conventionally measured in kg) at equilibrium, and C W is the concentration of solute per unit volume (dm −3 ) of the coexisting aqueous solution.On the condition that the control samples achieved equilibrium at the end of the incubation period, i.e., the measured concentration in the receptor cell is equal to the measured concentration in the donor cell, it can be assumed that the measured concentrations in the sample receptor cells are equal to the free concentration in the donor cells.K MW can be calculated based on the free and nominal concentration measurements, C free and C nom , as follows: V well is the total volume of the well, n memb is the amount of solute in the membrane, and m is the total mass of POPC (and cholesterol, if present) added to the well.This calculation assumes (1) all the POPC and cholesterol formed liposomes, and (2) all of the solute is either free in the red and white cells, or bound to the liposome (see Discussion).
3 Computational methods

Coarse-grained models
CG models were built using the Martini 3 force eld. 41The POPC model was taken from the standard Martini 3 library of lipids.The cholesterol model was adapted from the virtual-site Martini 2 model, 58 by changing the bead types to their closest Martini 3 equivalents, based on existing Martini 3 solvent models.Solute models were parametrized for the nine compounds studied experimentally in this work, and additionally for hexylamine.The models were constructed using a new version of our cg_param tool, 40 to take advantage of the Martini 3 force eld.The mapping algorithm for dividing molecules into CG beads is unchanged from the previous version, but the subsequent assignment of bead types now reects the expanded choice in Martini 3. The version of cg_param used in this work supports the range of beads from C1 to P6, but no subtypes or halogenated beads.The regular, small and tiny (R, S, T) bead sizes available in Martini 3 are automatically assigned using the maximum number of bonds spanning the fragment.Bead types are now assigned based on experimental DG OW of the fragment where available, falling back on the predictions of the ALOGPS algorithm 59,60 only when direct measurements are not available.Certain neutral fragments now have their optimal bead assignments hard-coded, with ethyl, propyl and butyl fragments assigned as TC2, SC2 and C2 respectively.
The mapping of anionic charged groups based on pre-dened fragments has been updated, with the SO 4 − group assigned as a Q2 bead, and the SO 3 − group as SQ4.These assignments were based on matching existing experimental (SSLM) log K MW data for octyl sulfate and octyl sulfonate, 19 analogously to the approach in our Martini 2 version. 40he tetramethyl ammonium fragment NC 4 + was assigned as Q2, and the acetate fragment C 2 O 2 − as SQ5n, using existing bead assignments from the Martini 3 ions library. 41Other primary and secondary ammonium fragments from the automated mapping were parametrized manually using the general guidelines for charged bead assignments from the Martini 3 ESI. 41appings of the solute molecules tested in this work, and their bead assignments, are shown diagrammatically in Fig. 1.The bead assignments are also listed in ESI † and the updated cg_param is available at https://github.com/cgkmw-durham/cg_param_m3/tree/martini3_v1.While this new version of the algorithm takes advantage of several important new features in Martini 3, full exploitation of Martini 3 is a larger piece of work which is still underway.

Coarse-grained simulations
The insane tool was used for building initial congurations of all POPC-cholesterol systems. 61The water phase consisted of Martini 3 regular water (4 molecules per bead), with 0.15 molar NaCl.All bilayers contained a total of 162 molecules (2 monolayers of 9 × 9 molecules), at varying ratios of POPC and cholesterol.Partition coefficients were calculated for all solutes using membranes containing 0, 10-15 and 20-25 mol% cholesterol, to match the individual experimental liposome measurements.A wider range of concentrations was used to validate the cholesterol-dependence of membrane properties, and of partitioning for a subset of solutes.
All simulations in this study were carried out using Gromacs 2021.1. 62and the new-RF set of Martini simulation parameters. 63he leapfrog integrator, with a 20 fs timestep, was used.The velocity-rescale thermostat and Parrinello-Rahman barostat were used to keep the simulation at 310 K and 1 bar pressure (for equilibration and constant pressure simulations).The reaction eld method, 64 with a relative permittivity of 15.0 and a cutoff of 1.1 nm, was used to calculate electrostatic interactions.Van der Waals interactions were cut off at 1.1 nm using the potential shi method, which shis the potential to zero at the cutoff.LINCS was used to constrain bond lengths within CG ring systems, with an order of 12 and 2 iterations.These LINCS View Article Online parameters are recommended for cholesterol simulations to avoid articial temperature gradients. 65he same umbrella sampling setup was used as in our previous work. 40Two solute molecules were simulated in each umbrella window.The simulations were set up so that while one molecule is in the center of the membrane, the other is in the water.For neutral molecules, the umbrella windows were at 0.1 nm intervals between the center of the membrane and 5.0 nm away from the center, and the molecule was restrained using a harmonic potential with a 1000 kJ mol −1 nm −2 force constant.For charged molecules, umbrella windows were at 0.05 nm intervals, using a 2000 kJ mol −1 nm −2 force constant.The distance between a solute and the membrane center was dened using the cylinder method, to remove the effect of undulations in the membrane. 66he umbrella sampling simulations were analyzed in the same way as in our previous work. 40The weighted histogram analysis method (WHAM) was used to estimate the probability P of nding the solute at a given position z i between the center of the membrane (z 0 = 0.0 nm) and the water (z n = 5.0 nm).The water and bilayer parts of the prole were determined using a hard cut-off, z R which was determined separately for each prole.z R was dened as the point where the RMSD of the free energy prole between z n and that position rst increases above 0.1 kJ mol −1 .log K MW was then calculated using: V(z n ) is the volume of a layer of the simulation box, M is the mass of one leaet of the lipid bilayer, and N A is Avogadro's constant.K MW is in units of dm 3 kg −1 .Strictly speaking, K MW refers to a particular charge state (neutral or ionized) of a given solute.In contrast, the distribution coefficient D MW takes into account the possibility that more than one charge state may be relevant at a given pH, and the overall partitioning is a weighted sum of state-specic K MW values: here f(Q) and f(N) are the mole fractions of the ionized and unionized forms of the molecule at the relevant pH.
The pK a and f values at pH = 7.4 (the pH of our experimental measurements) for the molecules in this study are given in Table 1.In each case, the simulations have been conducted using a xed charge state corresponding to the dominant species in aqueous solution, as indicated in the nal column of the table.For compounds that are charged at pH 7.4, the experiments are effectively measuring D MW rather than K MW , but for convenience both K MW and D MW values are referred to as K MW throughout.In practice, the distribution of charge states in the membrane may differ from that in water.Methods for constant-pH Martini simulationswhere the explicit charge state of a molecule can change during a simulationare under development. 67,68Such techniques may provide scope for further renement of K MW predictions in the future and make it possible to extend the approach to species where no single charge state dominates.

All-atom simulations
All-atom simulations were also carried out, using the CHARMM36 force eld, 76 to help validate the CG membrane properties.Gromacs input les were generated using the CHARMM-GUI. 77The same set of cholesterol concentrations and system sizes were used as for the CG simulations.
For each bilayer, equilibration was carried out using the default series of parameters from the CHARMM-GUI, which involves a minimization followed by a series of short equilibrations in which position restraints on the membrane are removed, and the time-step increased.A production run lasting 100 ms was then carried out using the default CHARMM-GUI production parameters.A 2 fs time step was used with the leapfrog integrator.The temperature was kept constant at 310 K using the Nosé-Hoover thermostat, with a coupling constant of 1.0 ps.The pressure was kept constant at 1.0 bar with the Parrinello-Rahman barostat, with a coupling constant of 5.0 ps.Electrostatic interactions were calculated using the particle mesh Ewald (PME) method, with a 1.2 nm short-range cutoff.Van der Waals interactions were cut off by switching the forces to zero between 1.0 and 1.2 nm.The LINCS algorithm, with an order of 4 and 1 iteration, was used to constrain all bonds containing hydrogens.

Validation of membrane models
The Martini 3 POPC/cholesterol system in this work is closely based on its widely used predecessor 58 in Martini 2. We have validated the present version by comparing density and lipid ordering to results from all-atom simulations.Results were obtained at 0, 15 and 30 mol% cholesterol to cover the range at which partition coefficients were calculated, as well as at 40, 50 and 60 mol% cholesterol to assess the transferability of the CG models to membranes with higher cholesterol content.
The condensing effect of cholesterol on the lipid structure was examined by comparing the area per lipid, A lipid , with respect to cholesterol concentration (Fig. S1 †).For the pure POPC membrane, both CG and atomistic models agree well 2.47* 0.000 1.000 0.000 0 Diclofenac 72 4.15 0.999 0.001 0.000 −1 Propranolol 72 9.42* 0.000 0.009 0.991 +1 Hexylamine 73 10.64* 0.000 0.001 0.999 +1 Dihexylamine 74 11.31* 0.000 0.000 1.000 +1 SB3-10 (ref.74)  1.275 0.000 1.000 0.000 0 DDAO-C12 (ref.74) 4.79* 0.000 0.998 0.002 0 Octyl sulfate 75 1.9 1.000 0.000 0.000 −1 Decyl sulfate 75 1.9 1.000 0.000 0.000 −1 Dodecyl sulfate 75 1.9 1.000 0.000 0.000 −1 with the experimental value 78 of 0.643 nm 2 .The all-atom simulations show a steady decrease in A lipid with cholesterol concentration up to 40% cholesterol, aer which the gradient decreases.The Martini 3 model shows a steady decrease at a comparable rate to the all-atom simulations, but the gradient is uniform over the concentration range.Additionally, the ordering of the lipid molecules in the membranes was determined by calculating the bond order parameter for each interbead bond in the CG molecule.For more detail on the simulated condensing and ordering effect of cholesterol, see ESI. † The liquid structure of the lipid bilayers was investigated by calculating 2D radial distribution functions (RDFs) of the lipid tails of POPC (see ESI †).Overall, the qualitative trend of increased ordering on adding cholesterol is followed by the Martini models of POPC/cholesterol mixtures.At around 40 mol% cholesterol there is a transition to the liquid-ordered phase from the liquid-disordered/liquid-ordered coexistence region, which is captured by atomistic simulations but not Martini 3 models.However, physiological cholesterol concentrations in sh are mostly less than 40 mol% and our experimental measurements are all below 40 mol%.Hence, an accurate description of the transition region is not required for the present study.Nevertheless, there is scope for further reparametrization of the bead types and bonded structure of the lipid and sterol models in Martini 3 in order to improve the phase behavior of mixed bilayers in the future. 79

Partitioning data
The experimental and simulated log K MW values for all molecules in the pure POPC bilayer are shown in Fig. 2.There is good agreement between the two approaches for many of the solutes, both neutral and charged, indicating that the CG solute models are a good representation of the molecules in question.Nevertheless, there are some outliers among the charged molecules, highlighting the greater difficulty in parametrizing this class of solute.Where a well-validated bead assignment was available for the charge-bearing fragment, including SO 3 − , SO 4 − and NC 4 + , good results were achieved.
However, the primary and secondary ammonium fragments were challenging due to the lack of specic bead assignments in the literature.For these fragments, we followed the general guidelines for ammonium fragments by the Martini 3 developers. 41Using this approach, the dihexylamine model achieved a very good match with experiment.However, propranolol was signicantly underestimated with respect to the experimental values.This indicates that whilst the model works well for many compounds with charged moieties, including SO 3 − , SO 4 − and NC 4 + , further development is still required before consistently good models can be achieved for ammonium-containing compounds.Fig. 3 shows how the experimental and simulated log K MW values vary with cholesterol concentration.The experimental results are separated in order to illustrate the effect of the initial dosing concentration.Each experimental point is the average of three concentration measurements taken from the same experiment, and the error bars are the standard deviations of those three measurements.Each simulation point is the calculated value from a single set of umbrella-sampling simulations.The error bars are the maximum observed standard deviation when 10 replicate simulations were carried out for two test cases (dodecyl sulfate and diclofenac) at all cholesterol concentrations.The resulting uncertainty of 0.10 in log K MW has been taken as representative for all the simulation points.
The change in log K MW with cholesterol concentration is generally small from both experimental and computational methods.There is most oen a small overall downward trend on adding cholesterol in the experimental results, although there is variation from molecule to molecule.In the experimental measurements, the difference between results at different dose concentrations is oen larger than the difference with respect to cholesterol.There is no consistent trend with respect to dose concentration across all molecules, suggesting that the choice of dosing does not introduce any systematic bias in the results.Nevertheless, the differences between the three batches of results for each molecule indicate that a statistical analysis of the data from any one dosing may not reect the true uncertainty in the result.We therefore take the mean of all individual measurements at all three dose concentrations as the most reliable value of the solute partitioning at a given cholesterol concentration.This gives values for each solute and cholesterol combination based on a minimum of 4 and a maximum of 9 measurements.These consolidated gures and their uncertainties, along with the corresponding simulation predictions, are summarized in Table 2. Additionally in Table 2 are simulation data for hexylamine, which has a log K MW value too low to be measurable by this experiment (see Discussion).The simulated log K MW results presented here differ from those in our previous work using the Martini 2 force eld to model POPC/cholesterol and DMPC/cholesterol membranes. 40In that paper, log K MW of aromatic molecules generally decreased on adding cholesterol, whereas it increased by a small amount for many aliphatic molecules.This effect comes from the way

View Article Online
interactions between differently sized beads are handled in Martini 2. The Martini 2 models for all aliphatic models were built using R beads, while cholesterol and all ring compounds are modeled using S beads.R-S interactions are not scaled relative to R-R interactions in Martini 2, resulting in an overestimation of the attraction between aliphatic molecules and cholesterol.This was a recognized shortcoming 80 of Martini 2 and has been rectied in Martini 3 by proper scaling of the Lennard-Jones interactions between different bead sizes. 41The results presented in this article benet from this improvement in the force eld.
The driving force for reduced solute partitioning into cholesterol-containing membranes is generally understood to be changes in membrane ordering on the addition of cholesterol.Cholesterol increases the lipid ordering for membranes composed of phospholipids, but the extent of this ordering depends on the nature of the acyl chain.Fully saturated lipids show a signicantly greater increase in ordering than those with one or two unsaturated acyl chains. 81In the case of POPC, which has one fully saturated chain, there is a moderate ordering effect, as shown in our atomistic results, which explains the small decrease in partitioning in the experimental results.This underestimation of ordering in the CG simulations could indicate why this change in partitioning, where experimentally detectable, is also underestimated.Nevertheless, the effect of cholesterol on partitioning even in the experimental POPC system is small.
Experimental liposome measurements are well established for estimating log K MW and are considered the closest representation of a real biological environment within a controlled experimental setting.As such, they are extremely valuable when assessing the bioaccumulation of important chemicals.The history of their use gives a good understanding of the method's limitations, such as the difficulties in measuring values for certain molecules.Some chemicals which we intended to study could not be assessed experimentally.Urea, which is very hydrophilic (and so would be expected to have a high concentration in the water phase) could not be quantied using LC-MS, highlighting the difficulty of nding general analytical methods when studying a varied set of molecules.Dodecylamine, on the other hand, was insoluble in the dosing solution.Even though it most likely can accumulate in the bilayer, it was not possible to accurately determine its concentration in the aqueous phase.
Finally, the amount of hexylamine in the liposomes at equilibrium could not be determined accurately.This problem occurs when the bound concentration is smaller than the variability of the bound concentration measurement, and sets a lower limit on the log K MW values that can be reliably calculated using a particular analytical method.This limitation does not exist in simulations, because the umbrella sampling technique can force a molecule to sample unfavorable environments.This is one of the areas where simulation can complement experiment.For example, simulations may be validated using experimental results for the more soluble members of a chemical series, and then trends with respect to, for example, alkyl chain length can be examined using simulation.In the experimentally marginal case of hexylamine, the CG model shown in Fig. 1(d) returns log K MW values in the range 0.29 to 0.37, as listed in Table 2.Such low values correspond to a difference between dosed concentration and free concentration in the liposome system of less than 3%, which is indeed below the resolution of the present experiments.Hence, the simulations correctly predict that an experimental measurement of log K MW would be challenging for hexylamine and, in the process, provide a prediction of what the value would be.
It is also important to bear in mind the assumptions associated with liposome experiments.The calculated log K MW is based on measuring the free aqueous concentration in the dialysis receptor cell and the assumption that any remaining solute must be in the liposome.Concentrations during control experiments (with no liposome) show a loss of material in some cases (recovery less than 80%).Stability assessment in the same matrix and under the same conditions were carried out as part of the method development for all chemicals and did not show signicant degradation, which rules this out as the main loss.It is proposed that losses result from binding to the RED plate.In the case of diethyl adipate, tryptophan, trihexylamine and tridecyl sulfate, the recovery was substantially lower (less than 25%) in the reference compound or the solute itself, and we have therefore been unable to obtain reliable log K MW values for these molecules.Scaling of the measured concentrations to the respective control values gave unphysical results where the scaled concentration was higher than the initial dosed concentration.There may be a more complex three-way equilibrium between the water, liposome and the binding of solute to the experimental equipment but it is not clear how best to account for this.

Conclusions
In summary, we have presented membrane-water partitioning data for a varied set of charged and neutral molecules, calculated using experimental liposome measurements and coarsegrained molecular dynamics simulations.The two methods were found to agree well for most molecules, and where there were discrepancies there are clear reasons from either the simulation or experimental side.This adds weight to the use of simulation to complement other methods, for example for molecules where experiments are difficult to carry out, or to ll in gaps in the data for series of similar molecules, or by way of initial high-throughput screening.
Another key aim of this study was to determine whether it is appropriate to include cholesterol in model membranes for helping to determine bioaccumulation.Many of the solutes studied do show a slight reduction in partitioning for cholesterol-containing membranes compared to pure POPC.However, this change is small, and is generally less than the difference between, for example, measurements from different dose concentrations or methods for calculating K MW and D MW .For the purposes of a risk assessment, where several other factors in addition to bioaccumulation must be taken into account, this small change may not be signicant.Some chemicals, e.g., peruoroalkyl surfactants, are known to have specic affinity for other membrane components such as View Article Online certain proteins, and so require more specic consideration. 19owever, even for these chemicals, log K MW for pure membranes has been demonstrated to provide improved correspondence with BCF values. 19We conclude that K MW determined for pure phospholipid membranes is a suitable surrogate for partitioning into cholesterol-containing membranes characteristic of sh as part of an overall assessment of bioaccumulation.The present study highlights the need for further work in several directions.In this article we have considered the inuence of cholesterol on partitioning, but other membrane components also have the potential to inuence the uptake of ionizable compounds, despite being present at lower levels.For example, acidic phospholipids are able to attract positively charged molecules and repel negatively charged molecules. 82hile these phospholipids may be present only in small concentrations, their overall impact has not yet been fully quantied across a broad chemical space (including those chemicals considered in this study).Additionally, the agreement between experiment and simulation for 3-nitroaniline in this study is poor.Considering the large amount of experimental data for this compound, disagreement suggests that further development is needed for simulations of compounds containing nitro groups and zwitterions in general.Finally, another priority is the development of consistently good models for compounds containing ammonium, especially quaternary nitrogen.As log K MW for these molecules oen proves difficult to measure experimentally, the development of simulation models for such compounds is a focus for future research opportunities.View Article Online

Table 1
Charge states at pH = 7.4 of the ten solutes simulated in this study.References indicate the source of the pK a values.Values of pK a marked * refer to the conjugate base.The fractions f of negatively charged (Q−), neutral (N) and positively charged (Q+) species were calculated from the Henderson-Hasselbalch equation 69,70

Fig. 2
Fig. 2 Comparison of simulated and experimental partition coefficients log K MW for the pure POPC system, for all molecules studied (correlation coefficient R 2 = 0.69).Dotted lines indicate the region of 1 log unit deviation from a perfect match.Experimental values are averaged over all dose concentrations, and horizontal error bars indicate the standard deviation of all experimental measurements in pure POPC membranes for each solute.The vertical error bars are the standard deviation of log K MW obtained from 10 independently seeded simulations for representative solutes.

Fig. 3
Fig. 3 Membrane-water partitioning coefficient log K MW as a function of cholesterol concentration c chol for each molecule studied.Results are shown for Martini 3 simulations, and experiments with initial solute dosing concentrations of 10, 20 and 50 mM.The lines joining the points in each data set are included to facilitate comparison between sets.

Table 2 log
K MW in POPC membranes as a function of cholesterol content.Experimental values are averaged over all measurements a minimum of 4 and up to 9 for each cholesterol concentration.See ESI for raw data.Statistical uncertainties in the final digit(s) of the experimental results are given in parentheses.The statistical uncertainty in the simulation results is 0.10 Environmental Science: Processes & Impacts Paper Open Access Article.Published on 02 May 2023.Downloaded on 3/22/2024 5:57:12 AM.This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.