Chemical selection

The method of chemical selection is described by Najjar et al. (2024). Forty-one salicylates were selected (Fig. 1). These were restricted to single substances (as opposed to mixtures), which were classified as salicylates according to the structure shown in Fig. 1. The exceptions to this are acetylsalicylic acid and capryloyl salicylic acid, which are aliphatic carboxylic acid structures (denoted in the figure with an asterisk). The common metabolic pathway for all these substances is hydrolysis to salicylic acid, catalyzed by carboxylesterases.

Fig. 1

Structure and hydrolysis of esters of salicylic acid. The code number and color relate to the identity of the salicylates used in Figs 7 and 8. The generic structure at the top of the figure does not include acetylsalicylic acid or capryloyl salicylic acid (denoted with an asterisk)

Skin absorption data

Measured in vitro data were available for 15 substances (see Supplementary Table 1), the values of which were used in the chemical grouping. Of these, data were available from the literature for 9 substances and the remaining values for 6 substances were generated as described below and the summary of the results are shown in Supplementary Table 2). These studies were all compliant with the Organization for Economic Cooperation and Development (OECD) Test Guidelines (OECD 2004b) and SCCS recommendations (SCCS 2010, 2023). Only data for substances applied in typical cosmetic formulations with an exposure duration of 24 h were considered (thus, data for substances applied in solvents such as acetone or studies with a shorter exposure were excluded from the current evaluation).

For the 26 substances without measured data, predicted skin absorption values were derived using the Local and Global in silico models (described below).

In vitro skin penetration studies

The in vitro skin penetration of six chemicals applied in a typical body lotion (representative as a typical oil-in-water formulation used for many products) was tested in frozen human skin according to Good Laboratory Practice and according to the OECD Test Guideline 428 (OECD 2004b) and OECD Guidance Document No. 28 (OECD 2004a) and following the SCCS Basic criteria (SCCS 2010).

Unlabeled chemicals were acquired from Takasago International Chemicals (Europe) S.A., El Palmar Murcia, Spain (isopropoxyethyl salicylate), Sigma-Aldrich (glycol salicylate), Givaudan Suisse AS, Vermier, Switzerland (cyclohexyl salicylate and (4Z)-hept-4-en-2-yl salicylate) and from Hallstar (butyloctyl salicylate). Isotridecyl salicylate was tested using radiolabeled substance only. The radiolabeled chemicals were from Eurofins Selcia Ltd. (Ongar, UK). The body lotion formulation is a typical oil-in-water but proprietary base formulation from Beiersdorf AG, Hamburg, Germany. The radioactivity content and homogeneity of the dose preparation was confirmed by analyzing sub-samples of solvent dilutions by liquid scintillation counting. The radiochemical purity of > 99% and stability over 24 h was confirmed by HPLC-Flow Scintillation Analysis. All substances were confirmed to be sufficiently soluble in the receptor fluid (phosphate buffered saline supplemented with 6% (w/v) polyoxyethylene (20) oleyl ether (Brij O20)) to achieve sink conditions.

The volatility of the substances (shown in Supplementary Table 2) was measured by applying a dose of final formulation (2 or 10 mg/cm2) on 12 glass slides which were placed into Franz cells. Subsequently, 3 cells were left unoccluded, 3 cells had one carbon filter, 3 cells had 3 carbon filters and 3 cells were fully occluded using parafilm. The cells were maintained at 32 °C for 24 h, after which slides, donor cells and film were washed with acetonitrile and the wash analyzed by liquid scintillation counting. Carbon filters were extracted with acetonitrile and the extract analyzed by liquid scintillation counting.

Discs of human dermatomed skin (thickness of 400 µm) from 4 donors (n = 3 per donor) were thawed and mounted on static Franz diffusion cells. Skin integrity was assessed according to electrical resistance across the sample and only skin discs with a resistance of > 5 kΩ were used (Brackin et al. 2024). The skin was unoccluded and maintained at 32 ± 1 °C. The exposure area was 2.54 cm2. The receptor chamber contained 500 µL receptor fluid, which was stirred using a magnetic stirrer.

An amount of 2 mg/cm2 formulation containing test substance at a concentration of 0.5% (cyclohexyl salicylate, (4Z)-hept-4-en-2-yl salicylate, isopropoxyethyl salicylate), 1% (butyloctyl salicylate, glycol salicylate) or 0.1% (isotridecyl salicylate) was applied to the skin surface. The concentrations of the substances were selected considering the following: In defense regulatory dossiers, the ingredient is usually tested at its maximum use concentration; however, there is a risk that the substance does not fully solubilize at this concentration in the considered testing formulation (vehicle), which would artificially decrease the calculated % absorption. When the substance is fully soluble in the vehicle, the % absorption is assumed to be concentration-independent. Therefore, to be able to extrapolate the % absorption to different use concentrations, a lower and soluble concentration was selected. The amount of formulation applied (2 mg/cm2) exceeded the intended consumer use condition (1 mg/cm2) used here to achieve an even distribution over the skin disk.

Receptor fluid samples were taken just prior to treatment and at 1, 2, 3, 4, 6, 8, 10, 12, 16, 20 and 24 h post application. The volume of fluid in the receptor chamber removed was immediately replaced by an equal volume of fresh receptor fluid. After 24 h, the skin was washed three times with a natural sponge pre-wetted with 3% Teepol L® in water. The sponges were digested in Goldisol® and made up to a recorded volume before analysis. The receptor chamber was washed with acetonitrile and the sample was analyzed by liquid scintillation counting.

The stratum corneum (SC) was removed by tape stripping using Scotch 3 M Magic Tape, to a maximum of 20 strips. Tape strips 1–5 were extracted individually, while the remaining 15 tape strips were combined in groups of five (6–10, 11–15 and 16–20) prior to overnight extraction in acetonitrile. The epidermis on the remaining skin disc was separated from the dermis using a heat separation technique (Hewitt et al. 2020). The epidermis and dermis were separately digested in Goldisol® and a sample of the digests was analyzed by liquid scintillation counting.

The bioavailable amount was calculated as the % of the applied dose recovered in the epidermis, dermis and receptor fluid.

In silico skin absorption models

The Global and Local models were used to predict the % absorption of substances applied in a body lotion formulation over 24 h (results are listed in Supplementary Table 1).

Global model

The Global model was developed in four steps: (1) prediction of skin absorption (cumulative mass over 24 h of a chemical absorbed into and across the skin (i.e., total skin + receptor fluid) from an aqueous vehicle, (2) correction for cosmetic vehicles, (3) correction for the amount in the SC, and (4) inclusion of chemical volatility. It was built on the following assumptions. First, despite application of a finite dose of formulation, a steady state can be achieved. This assumption is supported by the data showing that for a wide range of chemical and cosmetic vehicles, most of the chemical remains at the skin surface (Haque et al. 2016; Tampucci et al. 2018). Moreover, it is a conservative assumption as the flux for an infinite dose is always greater than the flux for a finite dose at the steady state. All cosmetic formulations used were treated as oil-in-water emulsions. It was assumed that only the fraction of chemical in the aqueous phase was able to partition with the SC.

Step 1: Starting equations for water as the vehicle

The equations (Eqs. 1 and 2) proposed by Cleek and Bunge (Bunge et al. 1995) to calculate the amount of chemical leaving the vehicle and entering the skin were used. Considering \(^\), the time to reach steady state,

for \(_\le ^\)

$$\frac} } }}} }}^ }} = 2K_ \sqrt t_ }}}$$

(1)

for \(_> ^\)

$$\frac} } }}} }}^ }} = \frac }} + K_ h\frac }} }}.$$

(2)

The Potts & Guy relationship (Potts and Guy 1992) (Eq. 3) was used to calculate the permeability coefficient (Kp) and the corresponding diffusion coefficient in the SC, \(_\), and chemical partition coefficient between SC and water, \(_\).

$$\log Kp \left( \right) = - 2.71 - 0.0061 MW + 0.71\log D_}}}$$

(3)

where t_exp = time of exposure, \(_\) = Cumulative amount of chemical that permeates the skin and in the Receptor fluid; \(h\) = thickness of the SC set at 10 µm; \(A\) = Exposure surface area; \(_\) = the initial chemical concentration in the vehicle; \(_\) = chemical partition coefficient between SC and water (calculated as follows: \(\text_= 0.71D}_\)); \(_\) = chemical diffusion coefficient in the SC (calculated as follow: \(\text\frac_} (cm/h) = -2.71-0.0061 MW\)). The LogD is used instead of LogP to correct for the ionization state of the permeant (Grégoire et al. 2009a).

B defines the ratio of permeability between SC and viable epidermis (Cleek and Bunge 1993).

For \(B\le 0.6\)

For \(B>0.6\)

$$t^ = \left( - c^ } } \right)\frac }} }}$$

With \(b = \frac\left( \right)^ - c\)

$$c = \frac }} \right)^ }}$$

Step 2: Adaptation for use with formulations—training set

These equations only apply for water as the vehicle; however, cosmetic vehicles are designed to solubilize non-aqueous soluble chemicals at a concentration higher than the maximum solubility in water. Thus, these equations were corrected to consider non-aqueous vehicles. The ratio of permeability between the SC and viable epidermis is independent of the vehicle as the vehicle does not modify their properties (Bunge et al. 1995). Potential penetration enhancement effects of the vehicle were excluded based on data showing that this is not the case for typical cosmetics formulations (Grégoire et al. 2009a).

The amount leaving the vehicle and entering the skin (e.g. total skin + receptor fluid) (Eqs. 1 and 2) was corrected accordingly, considering vehicle as an oil-in water emulsion (Eq. 4).

$$\frac }}^ }} = \frac }} + \left( } \right) \times 10^ }} }}$$

(4)

where \(_\)= the concentration in the water phase of the model oil-in water emulsion

\(_^\)= the concentration of the chemical in the formulation

\(_\)= Fraction of the substance that is water-soluble

a = A coefficient that may represent a scaling factor related to how LogD affects absorption

\(D}_\)= the partition coefficient between octanol and water of the chemical at the pH of the vehicle

A training set of the model consisting of 101 individual data points corresponding to 36 chemicals tested in OECD TG 428 compliant ex vivo human skin studies was used to optimize parameters \(a\) and \(_\) in Eq. 4. Optimal performance for all formulations used in the training set was found with \(_\)=0.5 and \(a=1-0.032 log _\) (Grégoire et al. 2009a).

Step 3: Correction for dermal delivery

To calculate the dermal delivery, which only considers the amount in viable epidermis (VE) + dermis (D) + receptor fluid (RF), Eqs. 1 and 2 were corrected for the amount of chemical in the SC. The relative permeability of the SC to the VE is assumed to be independent of the vehicle since the vehicle does not modify the properties of VE and SC (Cleek and Bunge 1993) (Eq. 2). As VE is more hydrophilic than the SC, the resistance of the VE increases for lipophilic chemicals; hence, the permeability ratio, denoted as B in Eq. 5, increases with logP. Lipophilic chemicals do not easily enter the VE, leading to their accumulation in the SC.

The relative ratio of permeability is assumed to be proportional to the ratio between the SC and amount entering the VE and the absorbed amount (i.e. VE + D + RF) (Eq. 6).

$$B \cong \frac }} }}$$

(6)

The amount in the Total Skin + RF is equal to the sum of absorbed amount and the amount in the SC:

Combining Eqs. 6 and 7, the amount Total Skin + RF is corrected by a factor B to calculate the amount VE + D + RF (Eq. 8).

Step 4: Inclusion of chemical volatility

The initial model was refined to include the impact of volatility as a rate limiting step for skin absorption (Grégoire et al. 2009a). This is important for three of the esters of salicylic acid which exhibit some significant volatility (ethyl salicylate, isopropyl salicylate and methyl salicylate, see Supplementary Table 2).

The evaporation rate kevap (g.cm−2.h−1) was calculated using Eqs. 9, 10 and 11 (Kasting and Miller 2006).

$$k_ = \frac P_ MW}} RT}}$$

(9)

where MW (g.mol−1) is the molecular weight; PVP (torr) is the vapor pressure; R (L.atm.mol−1.K−1) is the universal gas constant; T (K) is temperature (set at 305 K) and kg (cm.h−1) is the mass transfer coefficient in air calculated according to Eq. 10.

$$k_ = \frac }} }}$$

(10)

with u the air velocity (in m.s−1) adjacent to the skin (Stempfer and Bunge 2005).

Thus, the evaporated amount is calculated using Eq. 11.

Input parameters and output values

The Global model uses MW and LogD (defined at a given pH, calculated using the LogP and pKa), pKa and the vapor pressure (values for the substances evaluated here are shown in Annex Table 1). Since exposure time, amount of formulation and pH of the formulation are variable parameters, the following scenario was used to predict the values shown in Table 4: 2 mg/cm2 in a formulation with a pH of 6.5 applied for 24 h on ex vivo human skin. For volatile chemicals, the concentration of chemical in the formula was arbitrarily set at 1%. Only three chemicals are sufficiently volatile to limit their skin absorption (ethyl salicylate, isopropyl salicylate and methyl salicylate).

The final model predicts (in vitro) dermal delivery (i.e., the percentage in the epidermis, dermis and the receptor fluid of an infinite dose of test substance (Gregoire 2011)).

Applicability domain of Global model

MW and lipophilicity (LogD, calculated using the LogP and pKa) are parameters that describe the chemical applicability domain of the Global model. The validation set included 289 individual datapoints from 73 chemicals (including 6 salicylates) in different formulations. The MW of this dataset ranged from 76 to 741 g/mol and the logD from – 3.27 to 9.4, with most chemicals being within the MW range of 100 to 400 g/mol and LogD – 3.27 to 6. There are no measured in vitro skin absorption data for chemicals with higher MW or LogD values; therefore, the applicability domain was restricted to the second highest MW at 563 g/mol and second highest LogP at 6.2. If the substance does not fit with one of these criteria, the prediction quality is affected:

Extrapolation can be conducted for: MW > 541 and MW < 741 g.mol−1

Extrapolation can be conducted for: LogD > 6.3 and LogD < 9.3

Extrapolation can be conducted for: LogD > -5.3 and LogD < -3.3

Chemical having zwitterionic status at tested pH can lead to uncertainty on calculated LogD

If the MW is higher than 741 or the LogD higher than 9.3 and lower than − 5.3, no calculation is performed using this model and an assumption can be made that the level of skin absorption is lower than 1% of the applied dose.

Validation of the Global model using a dataset containing multiple chemical classes

This model was validated by Grégoire et al. (2009a) and was re-evaluated by us by expanding the number of data points to include a set of experimental 289 data points corresponding to 76 chemicals (of which 6 were salicylates). Inter-laboratory studies have shown in vitro experimental variation of about a factor 2 despite a standardized protocol (Wargniez et al. 2017). Thus, a difference between predicted and measured values lower than a factor 2 can be considered to be within the variability of the experimental data. For the majority of chemicals and formulations, the model tends to overestimate the skin absorption. Only a very small fraction of the data, 1.4% (4 data points) were underpredicted by more than a factor of 3. No data point was underpredicted by more than a factor of 4.

Local model

Training set

The training set for the Local model included seven data points associated with six salicylates. There are two experimental values for homosalate which both comply with the SCCS guidance for in vitro skin absorption; therefore, both values were considered. This training dataset comprised values available at the start of the study (the remaining data for 9 substances were only available after the model had been developed). The experimental values for the % applied dose absorbed were: 9.01% for benzyl salicylate, 1.82% for ethylhexyl salicylate, 4.31% for isoamyl salicylate, 20.2% for methyl salicylate, 3.29% for phenethyl salicylate and 0.97% and 3.83% for homosalate (see Supplementary Table 1). The correlation between the LogP values for these substances and the % applied dose absorbed (Fig. 2, R2 = 0.817) resulted in Eq. 12, which was subsequently used for all substances. LogP was used for the correlation rather than the LogD since most of the salicylates are non-ionized (the exceptions being acetyl salicylic acid and capryloyl salicylic acid); therefore, the correction of ionization was not necessary.

$$Local \, Model = Log \left( \right) = 2.016 - 0.2848 \times LogP$$

(12)

Fig. 2

Correlation between skin absorption data and LogP of salicylates. (n = 7 data points)

It is acknowledged that skin absorption is also a function of MW (Bunge and Cleek 1995); however, the correlation of absorption with MW was not as robust as that with LogP, since five of the six substances had MW values within a small range (208 to 262, listed in Supplementary Table 3). Therefore, to avoid over-fit the data and provide a simple model, we used only LogP as the input parameter.

Applicability domain of the Local model

The applicability domain was defined based on the LogP values of the substances used in the training set that ranged from 2.6 to 6.3. There were several substances which were outside the applicability domain of the Local model, with LogP values either lower than 2.6 or higher than 6.3. For substances with a LogP < 2.6, a default value of 50% was applied (according to the SCCS note of guidance (SCCS 2023)) and for substances with a LogP > 6.3, a conservative default value of 1.7% absorption was taken (i.e. the value obtained for a LogP of 6.3 using Eq. 1):

$$Log \left( \right) = 2.016 - 0.2848 \times 6.3 = 0.222$$

Uncertainty assessment of the Global and Local models

The uncertainty of the models is related to the precision of the physicochemical parameters used, particularly LogP. Experimental precision of LogP is usually around 0.3 (according to the OECD 107 Guideline (OECD 1995)). Moreover, even in well conducted in vitro skin absorption studies it was considered that the variability could be equal to 100% (SD = mean). Thus, it was assumed that the model could theoretically not be more precise than a factor 2.

For both models, the same LogP values were used. These were measured values when available and predicted values from the EpiSuite software (EpiSuite Calc KOWIN v1.67) when they were not available. There were measured LogP data for 15 substances available. The correlation between measured LogP and LogP values calculated with EpiSuite is very good (r2 = 0.97, LogPexp = 0.083 + 0.988 log Kow) for these substances, indicating that predictions of skin absorption were not significantly impacted by the source of the LogP within this chemical space.

Metabolism factor (MF) values

The MF values were taken from Najjar et al. (2024). These were generated using the High-Throughput PharmacoKinetics (HTPK) model from Simulations Plus (https://www.simulations-plus.com/), described by Liu et al. (2020) and (Naga et al. 2022). This software estimated the t1/2 in human plasma using input data including predicted values for plasma protein binding and either predicted intrinsic clearance (CLint) or measured CLint calculated from data from incubations with human liver S9 (Najjar et al. 2024). The t1/2 in humans was used to calculate the MF i.e., % of the conversion of the substance in 24 h, according to Eq. 1.

$$Metabolism\,Factor \left( \% \right) = \left( }}} \right)}} } \right) \times 100$$

(13)

where t is the time passed since exposure of the chemical; 24 h (1 day) and in vivo t1/2 is the half-life of the substance in human plasma.

Measured values for Fup were not available for most of the substances; therefore, values were derived using in silico-only input and in silico Fup combined with measured CLint, in vitro using human liver S9 (Najjar et al. 2024).