Clinical vs. chronological skin age: exploring determinants and stratum corneum protein markers of differential skin ageing in 351 healthy women

Abstract

Apparent skin age can be determined by several clinical measurements and may differ from chronological age, hence defining age acceleration/deceleration (Age A/D). Using data from 360 women with dermatological scoring of 21 clinical signs, we defined 3 well-separated co-occurring classes capturing the dryness, the elasticity and the oily nature of the skin. We related the risk of each clinical signs to the stratum corneum levels of 5 pre-selected proteins, we identified specific chronological age-adjusted signatures of each clinical sign. Using variable selection approaches, we identified 6 (of the 21) clinical signs which were jointly predictive of chronological age and used to define the clinical skin age, and subsequently age A/D. Applying univariate and multivariate approaches we found that stratum corneum levels of insulin degrading enzyme (IDE) was protective against (β = − 1.74, p = 3.3 × 10⁻⁶; selection proportion > 90%) accelerated skin ageing. In conclusion, our results support the fact that molecular markers found in the stratum corneum could predict skin ageing acceleration/deceleration.

Introduction

The aspect of facial skin is a major determinant of individual clinical age which may diverge from their chronological age, defining the concept of Age acceleration/deceleration (Age A/D). Several studies have reported that facial skin ageing could negatively impact individual’s self-esteem and overall quality of life^1,2. Among all the factors influencing facial skin ageing, the exposome which refers to the environmental exposures of an individual throughout their life, plays a role as important as inherited characteristics. Numerous studies involving monozygotic twins have quantified the relative role of genetic background and environmental exposures in both the risk and intensity of facial clinical signs in various populations. A study of 67 monozygotic twin pairs in Japan³ showed that within-twin-pair differences in skin ageing as measured by wrinkle score and facial texture increased with chronological age, hence suggesting a role of life course exposure to environmental factors (including smoking and UV filter application) in Age A/D. A more recent study of 388 pairs of monozygotic twins in South Korea⁴ estimated that 43% of the variability of the melanin index could be attributed to genetic variability and the remaining 57% to exposure to environmental factors.

Given the multifactorial aspect of facial skin ageing manifestations, the biological mechanisms involved in exposure-triggered Age A/D of the skin remain mainly unknown. However, a recent study of 122 healthy women has established that skin transcriptomic profiles changed with chronological age and was reflective of the clinical skin appearance⁵, hence opening avenues to further explore the predictive value of biological age in Age A/D. Proteomic analysis of the surface of the skin provides a less invasive and closer to the phenotype alternative to transcriptomics. Biomarkers from the stratum corneum have already been related to skin diseases such as Ichthyosis vulgaris, which is associated with loss of function mutation of the filaggrin gene⁶ and can be easily measured from D-squame samples⁷. Similarly, many molecular events taking place in the stratum corneum have been revealed via proteomics in lesional ad non-lesional skin of patients suffering of Atopic dermatitis^8,9. In the present observational study we use data from 351 healthy women aged 36 to 75 with expert dermatologist scoring of facial features assessing 21 clinical signs and with D-squame measurement of (N = 5) pre-selected protein levels based on prior reported association with chronological age to (i) define homogeneous skin profiles using clustering algorithms and evaluate their relationship with chronological age, (ii) investigate the association between stratum corneum proteins levels and the risk of (groups of) clinical signs, and (iii) to define clinical age as the part of chronological age that could be explained by clinical signs; age A/D as the difference between clinical and chronological age, and explore which proteins are explanatory of age A/D.

Results

Descriptive analyses

Study population

Our study population originally included (N = 375) women, aged 36 to 73, (mean values 55.0 y.o) as part of two independent studies of volunteers from France and Romania. Each participant underwent in-depth clinical assessment by dermatologist experts to score their clinical features and a D-squame sample on the cheek was collected at recruitment for subsequent targeted proteomic profiling. A total of 15 participants were excluded due to missing information for at least one protein. Score plots along the three first Principal Components (PC)—explaining > 80% of the variance in the data—identified two participants with outlying protein profiles along the three first PC (Supplementary Fig. 1). These two participants were excluded from the present analysis leaving (N = 351) participants with (N = 21) clinical signs and (N = 5) proteins assayed. Key characteristics of participants were comparable across study centres, except for participants from Centre 1 who were slightly older than in the two other centres.

Clinical signs profiling

The prevalence of the 21 clinical signs ranged from 10.5% for dry skin on the forehead to 78.9% for papyraceus aspect on the cheeks in the full population (Supplementary Fig. 1A). The prevalence of all clinical signs, except those relating to the oily nature of the skin, was consistently higher in older age group (p < 0.0001). Data also supported (more modest) differences in the prevalence of clinical signs across centres with a lower prevalence of most clinical signs in centre 1 (Supplementary Fig. 2B).

Hierarchical clustering calibrated to maximise the decrease in within cluster Jaccard’s index identified 3 clusters of (co-occurring) clinical signs (Fig. 1A): cluster 1 relating to the dryness of the skin (dry skin on the cheeks and on the forehead), cluster 2 relating to skin elasticity (elastosis, elasticity on the cheeks), wrinkles (upper-lip, crow’s feet, underneath eyes, forehead, cheek folds), skin pigmentation and brightness (full face dyschromia, pigmentation of the malar area, full face brightness, papyraceus aspect on the cheeks), and cluster 3 relating to the oily nature of the skin (oily by touching on the cheeks and forehead, shiny on the cheeks and forehead, rough/scaly by touching on the cheeks and forehead, full face pore visibility, and erythrosis on the cheeks).

Figure 1 [Images not available. See PDF.]

Descriptive analysis of clinical signs in the study population. (A) Hierarchical clustering of the (N = 21) assayed clinical signs in the full study population. The number of clusters to retain was calibrated using the Jaccard’s index as a measure of cluster similarity. The optimal number of clusters was defined as the one maximising the decrease in within-cluster Jaccard’s index. (B) Results from a series of univariate logistic regressions for the risk of each of the (N = 21) clinical sign as the outcome. For each clinical sign we report the Odds Ratios (OR) and 95% confidence intervals, adjusting for centre.

Univariate logistic models (Fig. 1B) showed increased risk of any of the (N = 11) cluster 2 (wrinkles/elasticity-related) clinical signs with chronological age (OR ranging from 1.03 [1.01, 1.06] for full face brightness to 1.23 [1.18, 1.28] for elastosis on the full face), while we observed a reduced risk of clinical signs from cluster 3 (oily/shiny cluster) except for full face pore visibility, erythrosis on the cheeks, rough/scaly by touching on the cheeks and on the forehead (OR ranging from 0.95 [0.93, 0.98] for oily by touching on the cheeks to 0.94 [0.92, 0.96] for oily by touching on the forehead). The risk of clinical signs from cluster 1 (dryness) was not associated with chronological age.

Proteins profiling

Skin levels of five proteins was measured in all study participants including Insulin Degrading Enzyme (IDE), Human Chitinase 3-like 1 (YKL40), Lipocalin-1 (LCN1), Transglutaminase-3 (TG3), and Filaggrin-2 (FLG2). We observed a moderate-to-low pairwise correlation across proteins ranging from -0.05 for LCN1-TG3 to 0.64 for the TG3-IDE (Fig. 2A). We found that levels of LCN1 and FLG2 increased with chronological age (p < 0.006) and that levels of TG3 (p < 0.004) and IDE (p < 0.00001) decreased with chronological age (Table 1). Univariate logistic regression models for the risk of each clinical signs as a function of the concentration of each protein separately (Supplementary Table 1A) suggested that higher concentration of FLG2 was associated with a lower risk of oily skin (by touching) on the forehead with OR of 0.59 [0.44, 0.79] and that higher concentration of IDE was associated with a lower risk of full face dyschromia with an OR of 0.59 [0.45, 0.77] and of crow’s feet and underneath eyes wrinkle with ORs of 0.57 [0.43, 0.75] and 0.64 [0.48, 0.84], respectively. None of these associations survived adjustment for chronological age (Supplementary Table 1B) but chronological age-adjusted results suggested that (i) higher concentration of TG3 and IDE were associated with higher risk of dry skin on the cheeks and (ii) higher concentration of IDE were associated with lower risk of a shiny aspect of the cheeks.

Figure 2 [Images not available. See PDF.]

Proteins profiling. (A) Heatmap of pairwise Pearson’s correlation coefficients for each of the (N = 5) proteins. (B) Network representation of a series on LASSO-penalised logistic models regressing the concentrations of the five measured proteins against the risk of each (N = 21) clinical signs, separately. Results are presented for models adjusted for centre (B) and further adjusted for chronological age (C). Proteins are represented as blue nodes and clinical signs are coloured according to the (N = 3) clusters of co-occurring clinical signs: wrinkles (in beige), oily nature of the skin (in red) and skin dryness (in green). Proteins are connected to a given clinical sign are those that were stably selected as jointly and complementarily explanatory of the risk of that clinical sign.

Table 1. Association between chronological age and protein concentrations.

	Regression coefficient [95% confidence interval]	p-value
LCN1	0.02 [0.01; 0.04]	5.71e–03
FLG2	0.02 [0.01; 0.03]	2.07e–03
YKL40	0.01 [− 0.01; 0.02]	3.89e–01
TG3	− 0.01 [− 0.02; − 0.00]	3.53e–03
IDE	− 0.03 [− 0.04; − 0.02]	2.76e–09

Regression coefficients (95% confidence intervals) and p-values are calculated from a series of linear regressions with age as predictor and standardised protein concentrations as outcome.

Significant associations after correction for multiple testing using a Bonferroni corrected per test significance level ensuring a family wise error rate below 0.05 are represented in bold.

To account for the correlation across proteins we performed series of (LASSO) penalised regression models in a stability selection framework¹⁰ relating the measured concentrations of all proteins against each clinical sign separately (Fig. 2B,C). We found that conditionally on the skin concentration of all other proteins, IDE was associated with 8 wrinkle/elasticity-related clinical signs (cluster 2) and four clinical signs relating to the oily aspect of the skin (cluster 3, Fig. 2B). Concentrations of FLG2 were associated with three clinical signs relating to the oily aspect of the skin (cluster 3) and two wrinkles/elasticity-related outcomes (cluster 2), concentrations of LCN1 to two clinical signs relating to the oily aspect of the skin (cluster 3) and to one wrinkle/elasticity-related clinical sign (cluster 2). Concentrations of TG3 and YKL40 were exclusively associated with (N = 4 and 2, respectively) clinical signs relating to the oily nature of the skin (cluster 3). Of the 21 clinical signs, those relating to the dryness of the skin (cluster 1) were not associated with any protein along with oily by touching on the cheeks and rough/scaly by touching on the forehead. Most clinical signs (N = 12) were associated with a single protein, two (crow’s feet wrinkle, and oily by touching on the forehead) with (the same) two proteins (FLG2 and IDE), one (shiny on the forehead) with three proteins and all proteins except FLG2 which was jointly associated with erythrosis on the cheeks and with rough/scaly by touching on the cheeks. Most of these associations, especially those involving IDE and those involving crow’s feet and UE wrinkles-related signs, could be explained by chronological age and were not selected in the models adjusted for chronological age (Fig. 2C). Models adjusted for chronological age showed that (i) YKL40, LCN1, TG3 and IDE jointly and independently of chronological age contributed to the risk of rough/scaly by touching on the cheek and of elasticity on the cheeks, (ii) TG3 and LCN1 to the risk of cheeks folds, (iii) LCN1 and FGL2 to the risk of oily by touching on the forehead, and (v) either a single (N = 8) or no protein (N = 9) were contributing beyond chronological age to the risk of the other clinical signs.

In the chronological age-adjusted model we also found that the risk of dry skin on the cheeks was associated with concentrations of IDE.

Clinical age and its determinants

We defined the clinical age as the part of chronological age that could be explained by the clinical profile of an individual. To account for the correlation across clinical signs we adopted a similar LASSO model calibrated via stability to identify the set of clinical signs jointly explanatory of chronological age. Our stability selection LASSO approach identified that five wrinkle/elasticity-related clinical signs (cluster 2) were jointly and positively related to chronological age (including elastosis on the full face, papyraceus on the cheeks, cheek folds, crow’s feet wrinkle and upper lip wrinkles) (Fig. 3A). Conditional on these clinical signs, we also found that shiny on the forehead was negatively associated with chronological age. Altogether the 6 selected clinical signs explained 59% of the variance of chronological age.

Figure 3 [Images not available. See PDF.]

Clinical age and its determinants. (A) Clinical age is defined from a stability selection LASSO using the (N = 21) clinical signs as predictors and chronological age as (continuous) outcome. We report the per-clinical selection proportion as defined by the number times that features were included in the model across (N = 1000) sub-samples of the data (top panel). The threshold in selection proportion to define stably selected predictors was calibrated jointly with the penalty parameter and is represented as a horizontal red dotted line. The clinical signs with selection proportion above that threshold are considered those we report as stably selected and their label in presented in bold on the X-axis. We also report the effect size estimated from a recalibrated model fitting a linear regression with stably selected clinical signs as predictors (bottom panel). (B) Receiver Operating Characteristic (ROC) curves for the recalibrated logistic models fitted on (N = 1000) independent 25% testing sets including stably selected proteins. We report the results for the model including centre and levels of (i) IDE (in beige), (ii) LCN1 (in blue), and (iii) both IDE and LCN1 (in red). (C) Median, 5th and 95th percentiles of the AUC from logistic models for dichotomised age acceleration indicator in models sequentially including each protein in decreasing order of their selection proportion.

We defined the Age A/D as the difference between the chronological age and the clinical age. A negative value of this difference indicating an apparent age lower (age deceleration) than the chronological age and positive value indicating accelerated ageing. Univariate analyses relating Age A/D and concentrations of each of the proteins separately showed that higher concentrations of LCN1 were associated with age acceleration, and higher concentration of TG3 (β =  − 1.40, p = 8.7 × 10⁻⁴) and more markedly of IDE (β =  − 1.74, p = 3.3 × 10⁻⁶) were associated with age deceleration (Table 2).

Table 2. Association between age A/D and protein concentrations.

	Regression coefficient [95% confidence interval]	p-value
LCN1	0.63 [0.19; 1.06]	5.21e–03
FLG2	0.87 [0.10; 1.64]	2.78e–02
YKL40	0.34 [− 0.27; 0.94]	2.73e–01
TG3	− 1.40 [− 2.22; − 0.58]	8.67e–04
IDE	− 1.74 [− 2.46; − 1.02]	3.29e−06

Regression coefficients (95% confidence intervals) and p-values are calculated from a series of linear regressions with each of the protein concentrations as predictor and Age A/D (clinical age–chronological age) as outcome.

Significant associations after correction for multiple testing using a Bonferroni corrected per test significance level ensuring a family wise error rate below 0.05 are represented in bold. Models are all adjusted for centre.

Stability selection models only identified concentrations of IDE as a predictor of Age A/D (selection proportion of 0.91). Selection proportion for LCN1 was 0.63 and did not reach the calibrated threshold of 0.89. When recoding Age A/D as a binary variable defined as true for a clinical age higher than the chronological age, both LCN1 and IDE were selected to be jointly predictive of Age A/D status with selection proportion greater than 0.9. ROC analyses based on logistic regression for the binary Age A/D indicator and recalibrated in 1000 independent validation sets each including 25% of the study population indicated that IDE was yielding an AUC of 0.64, LCN1 an AUC of 0.62, and that LCN1 was only modestly improving the model performance over that of IDE (Fig. 3B) and that the addition of any variable beyond these two was not further improving the model (Fig. 3C). Consistently, our models sequentially adding proteins in descending order of their selection proportion showed that IDE improved the AUC of the model from 0.611 for the model only including centre to 0.637 and further to 0.641 while including LCN1. The addition of any other proteins did not improve the performance of the model, hence suggesting an efficient calibration of our model.

Overall, these results suggest that LCN1, TG3, and IDE are three skin proteins associated with accelerated skin ageing. The information brought about by TG3 seems to be explained by IDE and (to a lesser extent) LCN1.

Discussion

In the present analyses we relate the levels of 5 targeted proteins from the stratum corneum to a series of 21 expert-scored clinical signs including skin dryness, pore visibility, wrinkles and oily nature of the skin. Our descriptive analyses showed that the 21 scored clinical signs had strong co-occurrence patterns and clustering analyses identified 3 main clusters relating mainly to (i) the dryness of the skin, (ii) the skin elasticity and wrinkles, and (iii) the oily nature of the skin. Clinical signs from these different classes showed differential associations with chronological age with stronger associations unsurprisingly observed in clinical signs relating to the elasticity of the skin and presence of wrinkles. Clinical signs relating to the oily nature of the skin appeared to be overall inversely associated with chronological age, and dryness of the skin was not found associated with chronological age.

Of the five proteins measured in the stratum corneum, the concentrations of LCN1 and FLG2 were found to increase with chronological age, while concentrations of TG3 and IDE were found to decrease with chronological age.

Multivariate analyses regressing the five proteins against each clinical signs showed that IDE was associated with many clinical signs, in particular those relating to wrinkles/elasticity of the skin, and to the oily nature of the skin. Most of the former associations could be explained by chronological age, while most of the latter survived adjustment for chronological age. In the chronological age adjusted models, our results suggested that stratum corneum concentrations of IDE were, independently of chronological age, associated with five clinical signs, LCN1 and TG3 to three clinical signs, YKL40 to three clinical signs and FLG2 to two clinical signs. Despite the strong co-occurrence patterns across clinical signs, there was limited overlap in the clinical signs associated with proteins and most clinical signs were explained by a single protein. Only shiny on the forehead and erythrosis on the cheeks were associated with three or more proteins, suggesting possibly more complex biological mechanisms at stake.

In order to account for the co-occurrence of clinical signs we defined clinical age as the part of the chronological age that could be explained by the clinical signs. From this we derived a measure of Age A/D as the difference between the clinical and chronological ages. Stratum corneum concentrations of IDE (negatively) and LCN1 (positively) were found to be jointly associated to overall accelerated skin ageing and may suggest potential determinants of skin ageing or may point to specific pathways involved in the early onset of skin ageing signs.

To the best of our knowledge, our study is the first to relate levels of LCN1 and ageing and our finding of IDE being associated with accelerated ageing is in keeping with previous reporting of insulin clearance and accumulation of advanced glycation end products in the skin¹¹. Decreased concentration of IDE could lead to accelerated global ageing concerns such as the well characterised deregulated nutrient sensing in skin ageing.

To our knowledge even though this study represents the largest available data with detailed clinical signs assessment and stratum corneum proteins measurements, its size remains modest, which could hamper our ability to detect weak and subtle effects. However, the stability selection approaches we used was able to detect biologically plausible markers of facial skin features and ageing. Second, our findings only apply to the population and skin type under investigation here and generalisation of our findings would warrant (i) the validation in (possibly larger) external cohorts, (ii) the extension of these analyses in participants of different skin types, and (iii) the inclusion of additional stratum corneum proteins. Our definition of clinical skin age is based on a limited number of clinical measurements and could be refined by recent developments estimating apparent age from facial images^12,13.

Despite these limitations our results are providing novel evidence linking the stratum corneum concentrations of IDE, LCN1 and TG3 as being related to skin clinical features and ageing. If validated these represent putative determinant of skin aging and could represent targets for future interventions and treatments.

Methods

In vivo clinical study design

A randomized multicentric clinical study was conducted on European volunteers, in France (Paris and Besançon) and in Romania. These two studies were conducted in accordance with the Declaration of Helsinki principles. According to Decision 147 of 4th March 2015 of the Romanian government, ethics committee approval was not required. The Jardé law was voted in France in 2012. That law was only implemented after its application decree had been published, in November 2016. Our study (which took place in 2015) was subject to the previous law of August 9, 2004, relating to Public Health, which did not require submission to an Ethics Committee for a non-interventional cosmetic study.

Each participant provided written informed consent prior to any procedure.

A total of 376 healthy women, divided in 3 centres were included, aged 36 to 73 years old, skin phototype II or III according to Fitzpatrick’s classification. The main inclusion criteria were: non-smoker or smoking less than 5 cigarettes per day, absence of suntan, absence of dermatological disorder affecting the face (i.e. vitiligo, acne, rosacea or melasma), absence of cosmetic or surgical procedures on the face. Inclusion has been done to respect the same stratification according to 10 years age range in every center, i.e. 20 women between 36 and 45 y.o, 40 women between 46 and 55 y.o, 40 between 56 and 65 and 20 between 66 and 75 y.o. We excluded participants reporting current or past use (for 1 week or more over the 8 weeks prior to invitation) of systemic or topical drugs or cosmetics such as antibiotics, anti-inflammatory drug, corticoids, retinoids, alpha hydroxyl-acids, vitamin C, benzoyl peroxide or any anti acne or antiseborrheic products. All eligible volunteers were provided a gentle cleansing product (Lipikar, La Roche Posay) to standardize their face cleansing on the evening prior to the evaluation visit. Volunteers did not wash nor did they apply any product on their face in the morning prior the visit.

In vivo clinical assessments

Dermatologists involved in the multicentric study were trained at the same time altogether by the same expert to be aligned on clinical assessments realized with an Evalux Bench® (Cosderma, Bordeaux, France) table to guarantee standardized conditions of lightening.

Some evaluations, as brightness, pores visibility, dischromia (skin tone heterogeneity) and elastosis, were performed using a 6 grades scale, on full face.

Papyraceus aspect of the skin (texture), elasticity and erythrosis were assessed on the cheek with the 6 grades scale.

Wrinkles (crow’s feet, forehead, underneath eye, cheek folds) and pigmentation (malar area) assessments were done using referential Skin Aging Atlas¹⁴.

Skin type, including oiliness, shininess, dryness and roughness were assessed by touchy on cheek and forehead, using a 4 grades scale.

For analysis, each clinical scoring of 21 clinical signs evaluated was dichotomized in 2 equal-sized groups “0” for absence and “1” for presence of the clinical sign.

D-squames sampling and protein concentration measurements

Stratum corneum layers samples were collected on the cheeks of these women using D-squame D100 (CuDerm Corporation, Dallas, Texas, USA). These were used to measure the stratum corneum levels of 5 prioritised proteins found associated with age in a previous study. These included filaggrin-2 (FLG2), a cytoplasmic protein involved in the cell adhesion process in the cornified cell layers¹⁵; lipocalin-1 (LCN1) an extra-cellular protein playing an antimicrobial role^16,17; chitinase-3-like protein 1 (YKL40) a protein involved in the allergic skin inflammatory processes¹⁸; protein-glutamine gamma-glutamyltransferase E (TGM3), a cytoplasmic protein involved in the formation of the cornified envelope of the skin¹⁹; and Insulin Degrading Enzyme (IDE), a multifunctional enzyme implicated in the degradation of hormones (insulin) and other peptides such as amyloid β²⁰.

D-squames were incubated 2 min in 550 µl of 50 mM Tris, 150 mM NaCl, pH8.0. This enabled the solubilisation of the 5 proteins of interest (LCN1, FLG2, YKL40, TG3, IDE), and subsequently their dosage on microfluidic cartridges using the FREND instrument (NanoEnTek) in a fast sandwich ELISA type assay.

Statistical analyses

We performed a Principal Component Analysis (PCA) of the concentrations of 5 proteins measured in stratum corneum and represented the data along the first 3 PC explaining 83.7% of the variance in the data to identify possibly outlying observations.

Prevalence of the each of the 21 dichotomised clinical signs was calculated in the full study population and separately for (i) four broad age groups: 36–45 years, 46–55 years, 56–65 years and 66 years and over, and (ii) the three recruitment centres. We tested for differences in prevalence by age group or centre using Chi-squared test and reported the corresponding p-value.

To explore the co-occurrence of the 21 clinical signs we used hierarchical clustering with complete linkage using Jaccard’s index as a measure of similarity. The number of clusters was determined as the one providing the largest decrease in the within-cluster Jaccard index.

We subsequently investigated the risk of each of the clinical sign separately using a logistic model for clinical sign status using chronological age as predictor and adjusting for recruitment centre.

The concentrations of the five proteins measured in stratum corneum were normalised and centred to ensure comparability of the effect size estimates and were related (as outcome) to age (as predictor) using a linear model adjusted for recruitment centre.

We subsequently related the protein concentrations to each (binary) clinical signs using two series of logistic regression models without and with adjustment for chronological age. To identify a sparse set of proteins that are jointly and complementarily predictive of each (N = 21) clinical signs, we used a series of least absolute shrinkage and selection operator (LASSO) logistic models in a stability selection framework¹⁰. Briefly, the model was run on 1000 independent 50% subsamples of the population and we derived for a given value of the penalty the per-protein selection proportion across the 1000 models as measure of clinical signs importance. The two hyper-parameters of the model controlling the sparsity (the penalty term) and the stability (the selection proportion above which a protein is considered as stably selected) were calibrated jointly using a likelihood-based stability score. Results for the (N = 21) stability selection logistic LASSO were plotted as a network where edges represent stably selected proteins jointly associated with each clinical sign. This model was subsequently adjusted (via non-penalisation) for chronological age.

Using a similar stability LASSO regression for the 21 clinical signs against chronological age, we defined the dermatological age as the part of clinical age that could be explained a sparse set of clinical signs. Difference between the dermatological age and chronological age defined the skin Age A/D, which was subsequently related to the 5 proteins measures using univariate linear and stability selection LASSO models. A binary skin Age A/D indicator indicating if the participant had a higher apparent than chronological age was finally regressed against (i) each of the proteins separately using logistic regression and (ii) all the 5 proteins measured using a stability selection logistic LASSO to identify which proteins were jointly predictive of accelerated skin ageing. Stability selection LASSO regressions were fitted in a training set of 50% of the participants. A series of logistic models with stably selected variables as predictors were fitted against binary Age A/D indicator in N = 1000 recalibration sets with 25% of un-seen participants. We then constructed Receiver Operating Characteristics (ROC) curves in the (N = 1000) remaining 25% of participants (test set) and reported pointwise median, 5th and 95th percentiles of the True and False Positive Rates and Area Under the Curve (AUC).

Statistical analyses were performed using R version 4.2.2²¹.

Author contributions

AF, SN, LA and NC designed the study and analytical plan. AF and NC supervised the study. LA and NC acquired funding. AF and SM conducted the data analyses. AF, SN, SM and NC drafted the manuscript. SN, VP and LA provided insights in the results interpretation and revised the manuscript. All authors have read an approved the submitted version of the manuscript.

Data availability

Data generated or analyzed during could be accessed upon request to the corresponding author, subject to approval from the access committee.

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-024-65083-4.

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Evangelista, M; Mota, S; Almeida, IF; Pereira, MG. Usage patterns and self-esteem of female consumers of antiaging cosmetic products. Cosmetics; 2022; 9, 49. [DOI: https://dx.doi.org/10.3390/cosmetics9030049]

2. Voegeli, R et al. Cross-cultural perception of female facial appearance: A multi-ethnic and multi-centre study. PLoS ONE; 2021; 16, e0245998.[COI: 1:CAS:528:DC%2BB3MXis1Wgsr4%3D] [DOI: https://dx.doi.org/10.1371/journal.pone.0245998] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33481957][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7822532]

3. Ichibori, R et al. Objective assessment of facial skin aging and the associated environmental factors in Japanese monozygotic twins. J. Cosmet. Dermatol.; 2014; 13, pp. 158-163. [DOI: https://dx.doi.org/10.1111/jocd.12081] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24910280][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4141746]

4. Suh, YJ et al. Genetic and environmental influences on general skin traits: Healthy twins and families in Korea. Twin Res. Hum. Genet.; 2017; 20, pp. 36-42. [DOI: https://dx.doi.org/10.1017/thg.2016.86] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27846924]

5. Xu, J et al. Novel gene expression profile of women with intrinsic skin youthfulness by whole transcriptome sequencing. PLoS ONE; 2016; 11, e0165913. [DOI: https://dx.doi.org/10.1371/journal.pone.0165913] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27829007][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5102383]

6. Thyssen, JP; Godoy-Gijon, E; Elias, PM. Ichthyosis vulgaris: The filaggrin mutation disease. Br. J. Dermatol.; 2013; 168, pp. 1155-1166.[COI: 1:CAS:528:DC%2BC3sXptFemtLo%3D] [DOI: https://dx.doi.org/10.1111/bjd.12219] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23301728][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6317863]

7. Rice, RH et al. Distinguishing ichthyoses by protein profiling. PLoS ONE; 2013; 8, e75355.2013PLoSO..875355R[COI: 1:CAS:528:DC%2BC3sXhs1aksLjP] [DOI: https://dx.doi.org/10.1371/journal.pone.0075355] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24130705][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3793978]

8. Andersson, AM et al. Assessment of biomarkers in pediatric atopic dermatitis by tape strips and skin biopsies. Allergy; 2022; 77, pp. 1499-1509.[COI: 1:CAS:528:DC%2BB3MXis1KjsbvJ] [DOI: https://dx.doi.org/10.1111/all.15153] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34695223]

9. Murata, S; Kaneko, S; Morita, E. Interleukin-8 levels in the stratumcorneum as a biomarker for monitoring therapeutic effect in atopic dermatitis patients. Int. Arch. Allergy Immunol.; 2021; 182, pp. 592-606.[COI: 1:CAS:528:DC%2BB3MXhsVeqsrjP] [DOI: https://dx.doi.org/10.1159/000512965] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33486487]

10. Bodinier, B; Filippi, S; Nost, TH; Chiquet, J; Chadeau-Hyam, M. Automated calibration for stability selection in penalised regression and graphical models: A multi-OMICs network application exploring the molecular response to tobacco smoking. J. R. Stat. Soc. C Appl. Stat.; 2023; 72, 1375. [DOI: https://dx.doi.org/10.1093/jrsssc/qlad058]

11. Choi, L-S; Ahmed, K; Kim, Y-S; Yim, J-E. Skin accumulation of advanced glycation end products and cardiovascular risk in Korean patients with type 2 diabetes mellitus. Heliyon; 2022; 8, e09571.[COI: 1:CAS:528:DC%2BB38XitFGrt7nK] [DOI: https://dx.doi.org/10.1016/j.heliyon.2022.e09571] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35711980][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9192809]

12. Xia, X et al. Three-dimensional facial-image analysis to predict heterogeneity of the human ageing rate and the impact of lifestyle. Nat. Metab.; 2020; 2, pp. 946-957.[COI: 1:CAS:528:DC%2BB3cXhslOgtb3I] [DOI: https://dx.doi.org/10.1038/s42255-020-00270-x] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32895578]

13. Bobrov, E et al. PhotoAgeClock: Deep learning algorithms for development of non-invasive visual biomarkers of aging. Aging (Albany NY); 2018; 10, pp. 3249-3259. [DOI: https://dx.doi.org/10.18632/aging.101629] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30414596]

14. de Bazin, R; Doublet, E. Skin Aging Atlas Vol 1: Caucasian Type; 2007; Med’com:

15. Pendaries, V et al. In a three-dimensional reconstructed human epidermis filaggrin-2 is essential for proper cornification. Cell Death Dis.; 2015; 6, e1656.[COI: 1:CAS:528:DC%2BC2MXivVSisro%3D] [DOI: https://dx.doi.org/10.1038/cddis.2015.29] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25695608][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4669814]

16. Fluckinger, M; Haas, H; Merschak, P; Glasgow, BJ; Redl, B. Human tear lipocalin exhibits antimicrobial activity by scavenging microbial siderophores. Antimicrob. Agents Chemother.; 2004; 48, pp. 3367-3372.[COI: 1:CAS:528:DC%2BD2cXnsVehsro%3D] [DOI: https://dx.doi.org/10.1128/AAC.48.9.3367-3372.2004] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15328098][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC514737]

17. Mallbris, L et al. Neutrophil gelatinase-associated lipocalin is a marker for dysregulated keratinocyte differentiation in human skin. Exp. Dermatol.; 2002; 11, pp. 584-591.[COI: 1:CAS:528:DC%2BD3sXlsVyitg%3D%3D] [DOI: https://dx.doi.org/10.1034/j.1600-0625.2002.110611.x] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/12473066]

18. Kwak, EJ et al. Chitinase 3-like 1 drives allergic skin inflammation via Th2 immunity and M2 macrophage activation. Clin. Exp. Allergy; 2019; 49, pp. 1464-1474.[COI: 1:CAS:528:DC%2BC1MXit1Gqs73J] [DOI: https://dx.doi.org/10.1111/cea.13478] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31397016]

19. de Koning, HD et al. Expression profile of cornified envelope structural proteins and keratinocyte differentiation-regulating proteins during skin barrier repair. Br. J. Dermatol.; 2012; 166, pp. 1245-1254. [DOI: https://dx.doi.org/10.1111/j.1365-2133.2012.10885.x] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22329734]

20. Tundo, GR et al. Multiple functions of insulin-degrading enzyme: A metabolic crosslight?. Crit. Rev. Biochem. Mol. Biol.; 2017; 52, pp. 554-582.[COI: 1:CAS:528:DC%2BC2sXhtVChtLnN] [DOI: https://dx.doi.org/10.1080/10409238.2017.1337707] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28635330]

21. R Core Team. R: A Language and Environment for Statistical Computing (2022).