Prediction and association analyses of skin phenotypes in Japanese females using genetic, environmental, and physical features

Abstract

Background

Skin characteristics show great variation from person to person and are affected by multiple factors, including genetic, environmental, and physical factors, but details of the involvement and contributions of these factors remain unclear.

Objectives

We aimed to characterize genetic, environmental, and physical factors affecting 16 skin features by developing models to predict personal skin characteristics.

Methods

We analyzed the associations of skin phenotypes with genetic, environmental, and physical features in 1472 Japanese females aged 20–80 years. We focused on 16 skin characteristics, including melanin, brightness/lightness, yellowness, pigmented spots, wrinkles, resilience, moisture, barrier function, texture, and sebum amount. As genetic factors, we selected 74 single‐nucleotide polymorphisms of genes related to skin color, vitamin level, hormones, circulation, extracellular matrix (ECM) components and ECM‐degrading enzymes, inflammation, and antioxidants. Histories of ultraviolet (UV) exposure and smoking as environmental factors and age, height, and weight as physical factors were acquired by means of a questionnaire.

Results

A linear association with age was prominent for increase in the area of crow's feet, increase in number of pigmented spots, decrease in forehead sebum, and increase in VISIA wrinkle parameters. Associations were analyzed by constructing linear regression models for skin feature changes and logistic regression models to predict whether subjects show lower or higher skin measurement values in the same age groups. Multiple genetic factors, history of UV exposure and smoking, and body mass index were statistically selected for each skin characteristic. The most important association found for skin spots, such as lentigines and wrinkles, was adolescent sun exposure.

Conclusion

Genetic, environmental, and physical factors associated with interindividual differences of the selected skin features were identified. The developed models should be useful to predict the skin characteristics of individuals and their age‐related changes.

Abbreviations

AUC

area under the receiver operating characteristic curve

BMI

body mass index

ECM

extracellular matrix

GWAS

genome‐wide association study

MAD

median absolute deviation

MMP(s)

matrix metalloproteinase(s)

ROC

receiver operating characteristics

SIA

skin image analyzer

SL

solar lentigo

SNP

single‐nucleotide polymorphism

TEWL

transepidermal water loss

UV

ultraviolet

1. INTRODUCTION

Skin characteristics are influenced in a complex manner by multiple factors. Among them, the effects of genetic factors have been investigated by means of genome‐wide association studies (GWAS), and many of the findings have been collected, organized, and published in the NHGRI‐EBI GWAS Catalog (NIH). However, most research has been disease‐related; for example, GWAS for psoriasis has identified many genetic factors, ¹ , ² and this information has aided drug development. ² Sample numbers in disease‐related GWAS have reached more than 100,000, and this enables even factors with small effects to be identified. ³ In contrast, it is difficult to acquire both skin measurement results and genetic data for large numbers of healthy human subjects. In addition, environmental factors ⁴ , ⁵ , ⁶ , ⁷ and physical factors ⁸ , ⁹ affect individual skin characteristics, which makes it very difficult to accurately clarify the mechanisms and interrelationships of phenotypic changes.

Human genetic diversity in the world has been comprehensively investigated ¹⁰ . In Asia, genetic ancestry is diverse and is strongly correlated with linguistic affiliations and geography. ¹¹ Many GWAS analyses have focused on the people of European ancestry, but few have been done for East Asian populations, especially Japanese. It is well known that skin color differs greatly among races, ¹² and the genes and their single‐nucleotide polymorphisms (SNPs) that determine each skin color have been examined. ¹² , ¹³ , ¹⁴ Furthermore, there are differences in skin color even within a racial group, such as the Japanese, ¹⁵ and these differences are related to SNPs of skin color–related genes. ¹⁵ In addition, the formation of pigmented facial spots in Dutch Europeans has been reported to be associated with SNPs of skin color genes IRF4, MC1R, ASIP, and BNC2, based on the GWAS analysis. ¹⁶ However, in Japanese, those SNPs were not selected monomorphically, and instead the peroxisome proliferator‐activated receptor gamma coactivator 1 beta (PPARGC1B) gene was significantly associated with pigmented spots and freckles. ¹⁷ An examination of 11 selected types of SNPs of melanogenesis genes found that OCA2 rs74653330 and rs1800414 are important determinants of skin color in Japanese. ¹⁵

Measurable skin features, such as skin color, including melanin index, cheek brightness/lightness (L ^*) and cheek yellowness (b ^*), pigmented spots, wrinkles, skin moisture content and barrier function, and mechanical properties, such as resilience, surface texture, sebum level, and so on, are affected by multiple factors, including hormones, such as estrogen ¹⁸ and androgen, ¹⁹ several kinds of vitamins and nutrients such as carotene, ²⁰ skin circulation, ²¹ , ²² inflammation, ²³ antioxidant systems, ²³ and dermal structural components such as extracellular matrix (ECM) proteins. ⁵ , ⁸ However, the effects of SNPs in many individual genes may be small and difficult to identify in the GWAS analysis. ¹⁷ In addition, skin features are also affected by environmental and physical factors. Therefore, it is necessary to comprehensively analyze the contributions of these factors to the skin features. In this study, in order to find even weak effects of SNPs, environmental factors, and physical factors, we used a large number of subjects and performed a multivariate analysis of complementary relationships, taking into account four possible genetic models (recessive, multiplicative, additive, and dominant), as well as the effects of environmental factors such as exposure to ultraviolet (UV) rays in sunshine and smoking behavior (estimated from a questionnaire) and physical factors, such as height, weight, and body mass index (BMI).

2. METHODS

2.1. Subjects

A total of 1472 healthy Japanese women aged 20–80 years were enrolled in the 12 studies listed in Table S1. Informed consent was obtained from all participants. Each study was separately approved by the Ethics Committee of Shiseido Co., Ltd.

2.2. Skin measurements

The evaluation of skin condition was conducted after all subjects had washed their faces with the same commercially available facial cleanser to remove cosmetics and sebum. Then, each subject rested for 20 min under ambient conditions of 21–23°C and 45% relative humidity before the measurement of their skin color, pigmented spots, wrinkles, moisture, resilience, surface texture, and sebum as summarized in Table S1.

For the skin color measurement of the cheek and upper inner arm, we used a CM‐700d spectrophotometer (Konica Minolta Sensing, Tokyo, Japan) to quantify color in melanin index and several different color spaces. Among them, the L ^* a ^* b ^* color space is widely used: L ^*, a ^*, and b ^* represent the degree of brightness/lightness, redness, and yellowness, respectively. The upper inner arm was chosen as a test site because it is protected from sun exposure. Digital image indexes of wrinkles and pigmented spots at both the right and left halves of the face were acquired using VISIA‐evolution (Canfield Scientific, Inc., Fairfield, NJ). ²⁴ All facial images were captured under the same conditions and were subjected to image analysis. To evaluate individual pigmented spots, we used a skin image analyzer (SIA) to measure the number and mean area of pigmented spots on the cheek. ²⁵ For a three‐dimensional analysis of wrinkles at the corners of the eyes, called crow's feet, we prepared replicas (Silflo, Flexico Developments Ltd., Potters Bar, UK) and examined them using a light‐sectioning method (HEV‐50HS‐20H, Hamano Engineering Co., Kanagawa, Japan). ²⁶ The volume, depth, and area of wrinkles were calculated. Capacitance was measured as an index of moisture content in the stratum corneum of cheek skin with a Corneometer CM 825 (Courage + Khazaka electronic Gmbh, Köln, Germany). The transepidermal water loss (TEWL) of cheek skin was measured using a VapoMeter (Delfin Technologies, Ltd.). Skin resilience or elasticity was determined using a noninvasive, in vivo suction skin elasticity meter, Cutometer MPA 580s (Courage + Khazaka, Cologne, Germany) equipped with a 2‐mm probe, at a negative pressure of 400 mbar. The ratio (R7, Ur/Uf) of immediate retraction (Ur) to final distension (Uf) was calculated and used as an index of skin resilience. ²⁷ Skin surface textures were analyzed with an image analyzer, SkinVision (Shiseido, Yokohama, Japan). The images of the skin surface were divided into pixels, and a binary image of the skin relief was obtained. To assess the homogeneity of the skin furrows, VC1, the coefficient of variation of the number of black dots in the binary image, was measured. Sebum level at the forehead was measured using a Sebumeter (Courage + Khazaka electronic GmbH).

2.3. Questionnaire

The questionnaire regarding skin characteristics was sent in advance and collected on the skin measurement day. In the questionnaire, we asked about age, height, weight, smoking, past UV exposure history, and skin factors. Body‐mass index, BMI (kg/m²), was calculated from height and weight. The UV exposure history was designed to assess exposures to sunlight during the periods from infancy to 14 years of age, 15–19 years of age, 20–24 years of age, 25–29 years of age, and in the 30s, 40s, 50s, 60s and 70s, as well as current behavior. Participants chose from three options as follows: “I took full protective measures against UV rays to avoid sunburn in my daily life” (score 0), “I did not really care about UV exposure in my daily life, but I took protective measures in places with strong sunlight” (score 1), and “I actively tanned without taking any protective measures against UV rays, not only in places with strong sunlight, but also in everyday life” (score 2). Total UV score for each subject was also calculated by summing up the individual scores for past sun‐related behavior. As for smoking, participants chose from three options: “Never smoked at all” (score 0), “Never smoked more than 20 cigarettes per day” (score 1), “Have smoked more than 20 cigarettes per day in the past or present” (score 2).

2.4. Selected SNPs and genotyping

We compiled a list of 74 SNPs (Table S2) in genes related to the degradation and synthesis of ECM proteins, hormones and vitamins, circulation and inflammation, and antioxidant systems, all of which could potentially affect skin condition. These 74 SNPs were extracted from microarray data obtained at Genesis HealthCare company. Genomic DNA was purified from saliva samples collected from each subject by the ethanol precipitation method using PrEP‐it from DNA Genotek or Beckman Agencourt DNAdvance. For genotyping, Axiom MyDesign human genotyping arrays (Affymetrix) or Infinium CoreExome‐24 custom BeadChip (Illumina) with optional SNPs covering the 74 SNPs were used. QC assessment for each sample at both the individual and SNP levels gave call rates over 0.97.

2.5. Quality control

Because the measured characteristic values of skin fluctuate depending on the season and measurement facilities, the values were standardized in each group of 10‐year age increments from the 20s to 70s using the median and median absolute deviation (MAD) values of data collected in 1611, 1710, 1810, and 1811 (year/month; see Table S1), which were measured at Shiseido Global Innovation Center, Yokohama, in October or November (last two numbers 10 and 11, respectively). In the case of the number and area of pigmented spots (SIA), the median and MAD values of all data were used for standardization due to the small sample size. Samples exceeding the mean ± 3 SD in each age range were excluded from the analysis as outliers, and remaining values were brought closer to a normal distribution by Box‐Cox transformation using EnvStat in R. Except for the skin characteristic values, missing values of genotyping data and environmental factors were predicted and complemented by the chained random forest method using missRanger (Wright & Ziegler). To examine the association between skin phenotype and allele for each SNP, we considered four genetic models (dominant, recessive, additive, and multiplicative). ²⁸ SNPs with a minimum genotype frequency of less than 5% in the genetic model were excluded from the analysis. R ² was calculated as the linkage disequilibrium (LD) coefficient to confirm the relevant SNPs. LD analysis was done using SNP‐disease association analysis software SNPAlyze v9, developed by DYNACOM Co., Ltd. (http://www.dynacom.co.jp/e/products/package/snpalyze/).

2.6. Statistical analysis

Elastic net in glmnet (version 2.0‐13) was used to extract variables associated with skin phenotypes from genetic, environmental, and physical features, such as four genetic models of SNPs, UV exposure history, smoking, age, height, weight, and BMI (kg/m²). A Gaussian function was assigned for continuous objective variables and a binomial function for the binomial objective variables. The data set was randomly divided to 80% for learning and 20% for evaluation. Using the training data set, 10‐fold cross validation was performed while tuning the regularization parameter (λ) to obtain the distribution of the average prediction error. In order to avoid overfitting and underfitting, we selected variables in the models that contained 5–10 SNPs in the range from the minimum mean prediction error (λ min) to 1 standard error (λ min + 1 s). Using the selected explanatory variables, a generalized linear model was constructed. When the objective variable was continuous, a linear regression model was constructed, and the R ² value was calculated. In the case of binary variables, a logistic regression model was constructed and the area under the receiver operating characteristics (ROC) curve (AUC) was calculated. Evaluation values were calculated for both training and evaluation data sets. In the process from model selection to evaluation described earlier, repeated cross validation (100 times) was performed using a new random data set for learning and evaluation. The finally obtained model was considered reasonable from the viewpoint of skin biology.

Statistical analyses were performed using R version 3.1.2 (R Foundation for Statistical Computing, http://www.R‐project.org), EnvStats package version 2.3.1, missRanger package version 1.0.4, glmnet package version 2.0‐13, pROC package version 1.10.0, and parallel package version 3.4.4.

3. RESULTS

3.1. Statistical processing and selected factors

As shown in Table 1, we selected 16 skin features, including cheek and upper inner arm melanin index, cheek brightness/lightness (L ^*) and cheek yellowness (b ^*), pigmented spots, wrinkles, water content, TEWL, resilience, surface texture, and sebum level. The 74 SNPs listed in Table S2 were used in the analysis, together with age, height, weight, BMI (kg/m²), smoking, and past UV exposure history. All data of skin features were plotted against age (Figure 1). Age was usually set in 10‐year periods from the 20s to 70s, but in some cases periods with similar distributions, such as 40s—60s for cheek melanin or 50s–70s for cheek brightness/lightness, were grouped together, as shown in Table S3. In this analysis, the relationship between skin features and genotyping data was examined on the assumption of dominant, recessive, additive, or multiplicative behavior for each selected SNP.

TABLE 1.

Characteristics of selected lower, upper, and linear models for skin features #: lower and upper models were established without MC1R, rs2228479

		Lower model			Upper model			Linear model
			AUC			AUC			R 2
		Count	Train	Valid	Count	Train	Valid	Count	Train	Valid
Skin color	Melanin (cheek)	100	0.67	0.75	100	0.69	0.69	100	0.29	0.33
	Melanin (arm)	100	0.70	0.71	100	0.70	0.70	100	0.16	0.15
	Brightness (L*, cheek)	100	0.70	0.71	100	0.70	0.71	100	0.27	0.28
	Yellowness (b*, cheek)	100	0.72	0.75	100	0.73	0.71	100	0.27	0.35
Pigmented spots	Pigmented spots (VISIA)#	99	0.64	0.64	66	0.62	0.60	100	0.27	0.34
	Number of pigmented spots (SIA)	100	0.72	0.76	100	0.75	0.74	100	0.38	0.36
	Mean area of pigmented spots (SIA)	81	0.72	0.70	68	0.72	0.61	100	0.26	0.18
Wrinkle	Wrinkle (VISIA)	62	0.67	0.62	80	0.66	0.64	100	0.60	0.63
	Crow's feet (area)	93	0.70	0.68	80	0.71	0.62	99	0.37	0.40
	Crow's feet (volume)	83	0.69	0.65	76	0.72	0.67	100	0.38	0.46
	Crow's feet (depth)	96	0.68	0.65	73	0.68	0.64	100	0.53	0.55
Moisture	Water content (cheek)	91	0.65	0.73	51	0.60	0.55	89	0.05	0.03
	TEWL (cheek)	49	0.65	0.62	59	0.67	0.60	95	0.10	0.11
Others	Resilience (R7, cheek)	88	0.61	0.61	44	0.63	0.59	100	0.34	0.38
	Texture (cheek)	42	0.70	0.55	88	0.65	0.60	96	0.09	0.04
	Sebum (forehead)	73	0.67	0.64	67	0.67	0.57	89	0.32	0.35

Note: The influence of genetic, environmental, and physical factors on personal differences of skin features were analyzed. To predict whether individuals show lower 25% or higher 25% levels of skin measurement values among subjects in the same age group, a logistic regression model for binary variables was constructed, and the AUC was calculated. For the continuous objective variable, a linear regression model was constructed, and the R ² value was calculated. The data set was randomly divided to 80% for learning and 20% for evaluation and 100 data sets were prepared. Evaluation values were calculated for both training and evaluation data sets and the best models were shown.

Abbreviations: AUC, area under the receiver operating characteristics curve; SIA, skin image analyzer; TEWL, transepidermal water loss.

FIGURE 1.

Age‐dependent changes of skin features. The figure shows plots of melanin index of the cheek (A) and upper inner arm (B), brightness/lightness (C), yellowness (D), pigmented spots measured by VISIA (E), number (F), and mean area (G) of pigmented spots measured by skin image analyzer (SIA), wrinkles measured by VISIA (H), volume (I), depth (J), and area (K) of crow's feet, water content (L), transepidermal water loss (TEWL) (M), resilience (N), texture (O), and sebum amount (P) against age. All data of skin features were normalized by Box–Cox transformation and data of the lower 25% (blue) and the upper 25% (red) are shown in color for age groups in increments of 5 years.

Linear regression models were constructed using a large number of data sets (n = 1472) and incorporating two cross validations into the model construction. The distribution of R ² values differed among skin features, as shown in Table S3. As shown in Table 1, the best models gave R ² values in repeated cross validation of 0.28–0.63, which represents a moderate to high degree of association according to Cohen, ²⁹ for 11 skin features, except for arm melanin, mean area of pigmented spots, cheek water content, TEWL, and skin surface texture. The highest associations were found for wrinkle (VISIA) and crow's feet depth.

Because the measured values of each skin feature showed a broad interindividual distribution (Figure 1), we next constructed logistic regression models to analyze the influence of genetic, environmental, and physical factors on personal differences of skin features. Such models may predict whether individuals show lower or higher levels of skin measurement values among subjects in the same age group. To build the logistic regression models, the upper 25% and lower 25% of the groups in increments of 5 years were taken for analysis. As shown in Tables S4 and S5, age increments and % cutoffs were optimized to improve prediction performance. The median AUCs for the lower models were 0.51–0.65, and those for the upper models were 0.49–0.66 in the distribution of evaluation values obtained by repeated cross validation. The best models for upper and lower predictions of each skin feature were selected and are listed in Table 1. The AUC values after repeated cross validation were 0.55–0.75 for the upper models and 0.55–0.75 for the lower models. Most results indicate a moderate association (AUC greater than 0.64), except for water content, resilience, and texture, which showed AUC less than 0.56, corresponding to weak association according to Cohen's criteria. ³⁰

3.2. Involvement of SNPs in personal differences of skin features

The repeatedly selected variables in the models constructed 100 times included 53 SNPs that were selected in 75% or more of generated lower, upper, and linear models for each skin feature (Tables 2 and 3).

TABLE 2.

Lists of frequently selected single‐nucleotide polymorphisms (SNPs) for generated lower, upper, and linear models for skin color and pigmented spots—upper: ●, lower: ○, and linear: ▴

		Skin color				Pigmented spots
Genes	SNP rs number	Melanin (cheek)	Melanin (arm)	Brightness (cheek)	Yellowness (cheek)	VISIA	VISIA (‐MC1R)	SIA number	SIA area
OCA2	rs74653330	▴●○	▴●○	▴●○	▴●○	▴●○	▴●○
OCA2	rs1800414	▴○	▴●○	▴●○	▴		●
ASIP	rs6058017	▴		▴○					○
IRF4	rs1540771		●
MC1R	rs2228479					●
Vit A (BCMO1)	rs7501331		●
Vit A (BCMO1)	rs12934922								▴
Vit B6 (NBPF3)	rs4654748					●
Vit D (GC)	rs2282679	●			▴
Vit D (NADSYN1)	rs12785878	▴○		▴○					▴●
Vit D (CYP2R1)	rs10741657		▴					●
Vit D (DHCR7)	rs11234027								▴
Vit E (CYP4F2)	rs2108622					▴	▴		○
Vit E (SCARB1)	rs11057830				●
ESR1	rs2234693		○
ESR2	rs2987983							○
ESR2	rs1256062				○
Adiponectin RFC4—ADIPOQ	rs6810075			●
Adiponectin CDH13	rs12051272							●
LEPR	rs1137101			○
VEGFA	rs2010963		●		●
VEGFA	rs833061					▴○	▴○	▴●○
VEGFC	rs1485766
FBLN5	rs2246416	●		○					▴○
ELN	rs8326				○				○
COL1A1	rs1107946							▴○	○
MMP1	rs1799750							▴○
MMP2	rs2241145								●
MMP2	rs7201					○	○
MMP2	rs2287074	▴			▴●○
ITGA2	rs1126643		○
HAS2	rs2046571					▴	▴	▴●○	▴
HAS3	rs3785079					●	●

Note: From 100 data sets, models of upper, lower, and linear were constructed, and the 33 SNPs were selected as variables in 75% or more of generated upper, lower, and linear models for skin color and pigmented spots.

Abbreviation: SIA, skin image analyzer.

TABLE 3.

Lists of frequently selected SNPs for generated lower, upper and linear models for skin features such as wrinkle, moisture, resilience, texture and sebum. upper ●, lower ○, linear▴

			Wrinkle			Moisture		Others
Genes	SNP rs number	VISIA	Crow's feet (Area)	Crow's feet (Volume)	Crow's feet (Depth)	Water content (cheek)	TEWL (cheek)	Resilience (cheek) (R7)	Texture (cheek)	Sebum (forehead)
OCA2	rs1800414				●	○	●
ASIP	rs6058017						●			○
IRF4	rs1540771		▴●○	▴●○	○
MC1R	rs2228479				○
Vit A (TTRB4GALT6)	rs1667255					▴			●
Vit B2 (MTHFR)	rs1801133				▴○	▴
Vit.B12 (MS4A3)	rs2298585		▴	▴○					○
Vit.B12 (FUT6)	rs3760776							○
Vit.B12 (FUT6)	rs1047781									○
Vit.B12 (ASS1P10PRELID2)	rs10515552							●
Vit D (CYP2R1)	rs10741657		●
Vit D (DHCR7)	rs11234027							▴
Vit D (NADSYN1)	rs3829251							▴		●
Vit E (CYP4F2)	rs2108622			●	●	●		○
Vit E (SCARB1)	rs11057830		○	▴
Vit E (ZNF259)	rs964184					▴●○	○
ESR2	rs1256062								▴	▴
Adiponectin	rs1501299							●
Adiponectin RFC4 ‐ ADIPOQ	rs6810075						○
Adiponectin ADIPOQ	rs182052				○
LEP	rs7799039	○					○
LEPR	rs1137101				●
VEGFA	rs2010963								○
VEGFC	rs1485766					○
FBLN5	rs2246416	●				●				●
COL1A1	rs1107946	●			●		▴●
MMP1	rs1799750	●
MMP2	rs2241145								▴●
MMP2	rs7201		●	●
MMP2	rs1030868							▴
MMP2	rs2287074							●
MMP2	rs2287076						▴
HAS3	rs3785079		●	●					▴
HAS3	rs2232228					○		●	●
TNFa	rs1799724		○
TNFR2	rs1061622			○			○
PLAU(uPA)	rs2227564				●
PLAU(uPA)	rs4065								●
GPX1	rs1050450							○
SOD2	rs4880									○

From 100 data sets, models of upper (●), lower (○) and linear (▴) were constructed and the 40 SNPs were selected as variables in 75 % or more of generated upper, lower and linear models for skin features such as wrinkle, moisture, resilience, texture and sebum.

In the case of skin color, two SNPs of OCA2, rs74653330 and rs1800414, are common in East Asians ³¹ and play an important role in determining melanin level at the cheek and inner upper arm skin, cheek brightness/lightness, and yellowness, in accordance with previous findings. ¹⁵ Selected SNPs included SNPs related to vitamin D (rs2282679 for cheek melanin and yellowness, rs12785878 for melanin and brightness/lightness, and rs10741657 for arm melanin) and vitamin E (rs11057830), as well as hormone‐related SNPs (ESR1 rs2234693 for arm melanin, ESR2 rs1256062 for yellowness, adiponectin rs6810075 for brightness/lightness, and leptin receptor rs1137101 for brightness/lightness). VEGFA rs2010963 was listed for both arm melanin and yellowness. SNPs related to ECM also appeared, including FBLN5 rs2246416 for cheek melanin and brightness/lightness, MMP2 rs2287074 for cheek melanin and yellowness, and ITGA2, integrin α2, rs1126643 for arm melanin.

In pigmented spots measured by VISIA, which detects color differences of spots using an unpublished algorithm, OCA2 rs74653330, MC1R rs2228479, vitamin B6‐related SNP rs4654748, vitamin E rs2108622, VEGFA rs833061, MMP2 rs7201, HAS2 rs3785079, and HAS3 rs2232228 were necessary for model construction. As an SNP of MC1R, rs2228479, is involved in the formation of pigmented spots, ³² the construction of upper and lower prediction models of pigmented spots analyzed by VISIA was carried out with or without rs2228479, and compared. The loss of rs2228479 reduced the constructed upper models from 90 to 66, although the lower models were almost unchanged at 98 and 99, suggesting that MC1R rs2228479 may be involved in the formation of pigmented spots. OCA2 rs1800414 may replace MC1R rs2228479. Other SNPs, such as OCA2 rs74653330, VEGFA rs833061, MMP2 rs7201, and HAS3 rs2232228, appeared in common. In pigmented spots measured by SIA, which evaluates numbers and areas of pigmented spots separately, selected SNPs in the models were different from those in the case of evaluation by VISIA and included ECM‐related SNPs, such as FBLN5 rs2246416, ELN rs8326, COL1A1 rs1107946, MMP2 rs2241145, MMP1 rs1799750, and HAS2 rs2046571. In the model using the mean area of pigmented spots, more ECM‐related SNPs appeared. In models using the number of pigmented spots measured by SIA, selected SNPs were independent of color‐related SNPs, and instead hormone‐related SNPs, such as ESR2 rs2987983 and CDH13 (adiponectin) rs12051272, were selected. Interestingly, VEGFA rs833061 was common in both SIA and VISIA.

In the case of half‐face shallow wrinkles measured by VISIA, LEP rs7799039 was selected in the lower model, and FBLN5 rs2246416, COL1A1 rs1107946, and MMP1 rs1799750 were selected in the upper model, whereas no predominant SNP was present in linear models. In the case of crow's feet wrinkles, area, volume, and depth were separately analyzed, and IRF4 rs1540771, Vit B12 (MS4A3) rs2298585, Vit E (SCARB1) rs11057830, MMP2 rs7201, and HAS3 rs3785079 were commonly selected for area and volume. As for depth, OCA2 rs1800414, MC1R rs2228479, Vit B2 (MTHFR) rs1801133, ADIPOQ rs182052, COL1A1 rs1107946, and PLAU rs2227564 were selected.

In the cases of moisture, water content, and TEWL, the selected SNPs were different, although OCA2 rs1800414 and Vit E (ZNF259) rs964184 were commonly selected. As for water content, Vit A (TTRB4GALT6) rs1667255, Vit B2 (MTHFR) rs1801133, Vit E (CYP4F2) rs2108622, VEGFC rs1485766, FBLN5 rs2246416, and HAS3 rs2232228 were selected, whereas for TEWL, ASIP rs6058017, adiponectin (RFC4‐ADIPOQ) rs6810075, LEP rs7799039, COL1A1 rs1107946, MMP2 rs2287076, and TNFR2 rs1061622 were selected.

In the case of resilience (R7) measured by the cutometer, vitamin‐related SNPs, such as Vit B12 (TMEM215‐ASS1P12) rs12377462 and (FUT6) rs3760776, Vit D (DHCR7) rs11234027 and (NADSYN1) rs3829251, Vit E (CYP4F2) rs2108622, and ECM‐related SNPs, such as MMP2 rs1030868 and rs2287074, and HAS3 rs2232228, were selected. For skin surface texture, Vit A (TTRB4GALT6) rs1667255, Vit B12 (MS4A3) rs2298585, ESR2 rs1256062, VEGFC rs1485766, MMP2 rs2241145, HAS3 rs2232228, and PLAU rs4065 were selected. In the case of sebum, ASIP rs6058017, Vit B12 (FUT6) rs1047781, Vit D (NADSYN1) rs3829251, ESR2 rs1256062, FBLN5 rs2246416, and SOD2 rs4880 were picked up.

3.3. Environmental and physical factors influencing skin features

The variables of environmental factors, past UV exposure history and smoking, and physical factors, height, weight, BMI, and age, were selected in 75% or more of generated lower, upper, or linear models for each skin feature, as shown in Table 4.

TABLE 4.

Frequently selected variables of environmental factors, past ultraviolet (UV) exposure history and smoking, and physical factors in generated lower, upper, or linear models for skin feature—upper: ●, lower: ○, linear: ▴

		Age						UV exposure history
		20s	30s	40s	50s	60s	70s	Infancy–14 years	15–19 years	20–24 years	25–29 years	30s	40s	50s	Total	Present	Weight	Height	BMI	Smoking
Skin color	Melanin (cheek)	▴	▴	▴	▴	▴				●○▴		●○▴			▴	●○	●▴		●○
	Melanin (arm)							▴		●▴		○				○▴	●○▴		●○▴
	Brightness (cheek)	▴	▴		▴	▴	▴			▴	●▴	●○▴	▴		▴	●○	●○▴		●○▴
	Yellowness (cheek)	▴	▴		▴		▴			●○▴		●○▴			▴	●	●○▴		●○▴	●
Pigmented spots	VISIA (‐MC1R)	▴	▴			▴	▴		▴			▴	▴	▴			○		○▴
	VISIA	▴	▴			▴	▴		▴			▴	▴	▴			●○		○▴
	SIA number	▴	▴		▴	▴	▴		▴		▴	▴		▴
	SIA area	▴	▴			▴	▴				▴				▴
Wrinkles	VISIA	▴	▴		▴	▴	▴		▴	○▴	▴	▴		▴			●○▴	●	▴
	Crow's feet (area)	▴	▴		▴	▴	▴		▴					●▴			○
	Crow's feet (volume)	▴	▴		▴	▴	▴							●▴
	Crow's feet (depth)	▴	▴			▴	▴		▴					○▴					●▴
Moisture	Water content (cheek)	▴	▴			▴	▴						○	▴
	TEWL (cheek)	▴	▴	▴										▴
Others	Resilience (cheek) (R7)	▴	▴		▴	▴	▴		○▴	▴		▴	▴	▴				▴	○▴
	Texture (cheek)	▴	▴	▴	▴													▴		●▴
	Sebum (forehead)	▴	▴	▴										▴					●

Note: From 100 data sets, models of upper, lower, and linear were constructed, and the variables of environmental factors, past UV exposure history and smoking, and physical factors, height, weight, BMI, and age, were selected in 75% or more of generated upper, lower, and linear models for each skin feature.

Abbreviations: BMI, body mass index; SIA, skin image analyzer; TEWL, transepidermal water loss.

Age variables were selected in almost all of the linear models of skin features except for the melanin index of upper inner arm, which was not dependent on age (Figure 1). For some skin features, the age variable was not selected for certain age periods, such as the 40s.

As described in Methods, participants chose from three options for UV exposure history. The UV exposure history influenced the 12 skin measurement values age‐dependently, and the ratios among the three options were quite similar in all cases ((Figure 2)). For skin color, melanin index of both cheek and arm were affected by present UV exposure and exposure during 20–39 years of age, as well as by cumulative UV exposure. Similar effects of UV exposure were observed during 20–49 years of age for brightness/lightness and during 20–24 years of age and in the 30s for yellowness. In contrast to skin color, pigmented spots were influenced by UV exposure only in the linear model. Earlier UV exposure (during 15–19 years of age) affected pigmented spots measured by VISIA and spot number measured by SIA. The area of pigmented spots measured by SIA was not affected by UV exposure, except during 25–29 years of age. Wrinkles measured by VISIA were affected by UV exposure during 15–39 years of age and in the 50s, whereas in crow's feet, wrinkles were affected by UV exposures during 15–19 years of age and in the 50s, but not during 20–39 years of age. As for moisture, UV exposure during the 50s was selected in the linear model, and that during the 40s was selected in the lower model. Skin resilience was affected by UV exposure during 15–59 years of age, except during 25–29 years of age in the linear model, whereas in the lower model, it was selected only during 15–19 years of age. Sebum was influenced by UV exposure during the 50s.

FIGURE 2.

Age‐dependent distribution of past ultraviolet (UV) exposure. UV exposure history was assessed from a questionnaire covering sun‐related behavior in infancy–14 years, 15–19 years, 20–24 years, 25–29 years, and in the 30s, 40s, 50s, 60s, and 70s. Participants chose from three options, scoring 0, 1, and 2 for protection levels, as described in Section 2. The ratios among the three options at different ages were quite similar among the 12 skin measurements.

Weight and BMI affected skin color and wrinkles measured by VISIA, and height was selected in linear models of resilience and texture and in the upper model of VISIA wrinkles. Crow's feet wrinkles were affected by weight in the lower model of wrinkle area and by BMI in the linear and upper models of wrinkle depth. Smoking was selected in the upper model of yellowness and in the linear and upper models of skin texture.

4. DISCUSSION

In this work, we developed three types of prediction models for 16 skin features, including cheek and upper inner arm melanin index, cheek brightness/lightness (L ^*) and cheek yellowness (b ^*), pigmented spots, wrinkles, water content, TEWL, resilience, surface texture, and sebum level, based on multiple genetic, environmental, and physical factors. As such factors are expected to affect each skin feature in a complex manner, we aimed to explore candidate weak but plausible relationships by using a large number of subjects and conducting multivariate analyses, taking account of four possible genetic models.

OCA2 SNPs, rs74653330 (A481T) and rs1800414 (H615R), were selected as determinants of upper and lower levels of cheek melanin, upper inner arm melanin, cheek brightness/lightness, cheek yellowness, and pigmented spots measured by VISIA. OCA2 is known to regulate melanosome pH, ³³ and the OCA2 rs74653330 (A481T) retains about 70% function in melanogenesis. ³⁴ OCA2 rs74653330 (A481T) and rs1800414 (H615R) are associated with skin melanin at the upper inner arm in Japanese females ¹⁵ and play an important role in susceptibility to tanning in the Japanese population. ³⁵ In this study, OCA2 rs74653330 (A481T) was more influential for the determination of skin melanin index, brightness/lightness, and yellowness than rs1800414 (H615R), in accordance with reported findings. ¹⁵ , ³⁵ Furthermore, it has been reported that rs74653330 (A481T) and rs1800414 (H615R) are characteristics of northeastern Asian and eastern Asian populations, respectively. ³¹ The allele frequency for s74653330 (A481T) was 0.089, which is similar to that (0.082) in Japanese. ¹⁵ In addition, rs1800414 (H615R) was shown to affect skin color in Canadian people of East Asian ancestry. ³⁶ Thus, OCA2 rs74653330 (A481T) and rs1800414 (H615R) may significantly influence the skin color of the Japanese population.

ASIP (Agouti signaling protein) is known to antagonize αMSH and inhibit melanocyte activation. ³⁷ The G allele of ASIP rs6058017 is known to be associated with darker human pigmentation. ³⁸ In this study, rs6058017 affected cheek melanin, brightness/lightness and mean pigmented area measured by SIA, whereas the G allele of rs6058017 was associated with lower cheek brightness/lightness and lower mean area, which may suggest that darker skin shows reduced brightness/lightness and reduces the risk for larger pigmented spots. Further experiments are necessary to clarify the mechanism involved.

IRF4 (interferon regulatory factor 4) gene product is a member of the interferon regulatory factor family of transcription factors, ³⁹ which are involved in the regulation of gene expression in response to interferon and other cytokines, and IRF4 has been reported to enhance tyrosinase expression in collaboration with MITF. ⁴⁰ The A allele of IRF4 rs1540771 is associated with skin sensitivity to sun and freckles. ⁴¹ In this study, the A allele of rs1540771 had a negative effect in the upper model of melanin index of the upper inner arm, suggesting that the A allele is not the major allele in persons with higher melanin levels. The A allele of rs1540771 was also selected in models of the average area, depth, and volume of crow's feet wrinkles, suggesting that it may be associated with increased wrinkle formation at the outer corners of the eyes. This would be consistent with reports that IRF variants are associated with skin aging. ¹⁶ Persons with this allele may benefit especially from sun protection in the crow's feet areas.

For pigmented spots and skin color, VEGFA rs833061 and rs2010963 were selected, respectively. VEGFA induces proliferation and migration of vascular endothelial cells and is essential for both physiological and pathological angiogenesis. ⁴² Vasculature is reported to be associated with increased pigmentation in hyperpigmentary disorders such as solar lentigo (SL) and melasma. ²¹ , ⁴³ Blood flow and volume also affect skin color. ⁴⁴ Thus, VEGFA SNPs may affect vascular condition and consequently skin color and formation of pigmented spots.

ECM‐related SNPs, such as FBLN5 rs2246416, ELN rs8326, COL1A1 rs1107946, MMP‐1 rs1799750, 5 SNPs of MMP‐2, ITGA2 rs1126643, HAS2 rs2046571, and HAS3 rs3785079 and rs2232228, were selected in models of wrinkles, resilience, and texture, for which dermal structure composed of ECM is important. ⁴⁵ They were also chosen in models related to skin color and pigmented spots. This is consistent with reports that fibroblasts in dermal ECM, consisting of collagen, elastin, and mucopolysaccharides, control melanocyte function by secreting melanogenic paracrine mediators, ⁴⁶ and that fibroblast function is controlled by tissue‐specific macromolecular ECM structures. Enzymes such as MMPs‐1 and −2 degrade the ECM proteins in basement membrane (BM) and the dermis ⁴⁷ , ⁴⁸ and affect its functions, such as supporting melanocytes. BM damage has been observed in human SL. ⁴⁹ Thus, ECM‐related SNPs may be involved in susceptibility to both wrinkle formation and skin pigmentation.

Moreover, COL1A1 rs110796, and MMP‐1 rs1799750 were selected in the upper model of face wrinkles measured by VISIA. MMP‐1 has collagenolytic activity after activation in the dermis. ⁴⁷ MMP‐1 rs1799750 is known to increase the expression of MMP‐1 and is indicative of susceptibility to skin aging. ⁵⁰ It has also been reported that the G allele of COL1A1 rs1107946 enhances the production of α1 chain of type I collagen, affects the balance with α2 chain, changes the quality of the collagen chain, and promotes keloid formation. ⁵¹ The hardness of dermal collagen fibers may affect facial recovery and flexibility after skin deformation due to facial expressions such as laughing, which may promote transient wrinkle formation. Altogether, these results suggest that changes of collagen decomposition and repair may influence wrinkle formation, but further work will be needed to establish how this occurs.

Age ranges were selected in linear models of almost all skin features as expected, because skin features change age‐dependently. Unexpectedly, age in the 40s was not selected in many linear models, which may mean that skin features show intermediate values in the 40s. On the other hand, the melanin index of the inner upper arm, which is usually protected from sun exposure, was not affected by aging (Table 4), which is consistent with the finding that the melanin index of inner upper arm during 20–70 years of age resembles that of sun‐protected buttock skin. ⁵² Interestingly, UV exposure history up to 14 years of age was selected in the linear model of inner upper arm melanin index, which is consistent with a report that the melanin index of the inner upper arm during the teens is higher than that during the 20s–70s ⁵² and may suggest greater sun exposure in sun‐protected flexural sites in young persons.

UV exposure in the 20s, 30s, and 40s was selected in the upper, lower, and linear models of skin color factors, that is, cheek melanin index, brightness/lightness, and yellowness, whereas UV exposure was selected mainly in the linear model in the case of pigmented spots. Moreover, UV exposure from 15 to 19 years of age was selected only in the linear model of pigmented spots, but not in linear models of cheek melanin index, brightness/lightness, and yellowness. SL is the most common type of pigmented spots on the face. ⁵³ UV exposure is known to induce melanin synthesis in skin melanocytes ⁵³ and is essential for the development of SL, which appears only on sun‐exposed skin. ⁵³ , ⁵⁴ SL is usually flat with a well‐defined border and is different from suntan, in which skin is evenly darkened after sun exposure and reverts to the normal color when protected from sun exposure. UV is known to induce DNA damage that accumulates age‐dependently due to decreased repair function. ⁵⁵ These results may suggest that early switching of skin cells to form pigmented spots might occur as a result of UV exposure during the teenage years and earlier. Thus, controlling sun exposure during youth might be a good strategy for preventing the later formation of pigmented spots.

BMI was selected in the upper‐25%, lower‐25%, and linear models of cheek and arm melanin indexes, brightness/lightness, and yellowness. As body weight but not height was selected in these models, weight affected skin color. The higher the BMI, the lower the values of cheek melanin indexes, brightness/lightness, and yellowness. Usually melanin index is correlated positively with yellowness but negatively with brightness, due to the reduction of light penetration into the dermis owing to intradermal light scattering. Indeed, OCA2 rs74653330 (A481T) negatively affected melanin indexes and yellowness but positively affected brightness/lightness, which is consistent with the idea that skin melanin plays an important role in heightening yellowness and lowering brightness/lightness. BMI may affect brightness/lightness and yellowness independently of skin melanin content. Brightness is known to be affected by the dermal collagen structure, ⁵⁶ and in overweight people, the density of dermal collagen is decreased. ⁵⁷ Moreover, enlarged adipocytes are known to negatively control the function of dermal fibroblasts by decreasing collagen expression and increasing collagen‐degrading enzyme expression. ⁵⁸ Therefore, higher BMI may be related to reduced brightness arising from lower light scattering in the dermis due to deterioration of collagen fibers.

In conclusion, we have examined the association of 16 skin features with multiple genetic, environmental, and physical factors by using multivariate analysis, and factors associated with interindividual differences of the selected skin features were successfully identified. The developed models are expected to be useful to predict the skin characteristics of individuals and their age‐related changes.

CONFLICT OF INTEREST

This research was supported by internal budget from MIRAI Technology Institute in Shiseido Co., Ltd. SA. KK, MO, TY, and CY are employed by Shiseido Co., Ltd. CI and IM are employed by DYNACOM Co., Ltd.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Amano S, Yoshikawa T, Ito C, et al. Prediction and association analyses of skin phenotypes in Japanese females using genetic, environmental, and physical features. Skin Res Technol. 2023;29:1–13. 10.1111/srt.13231

DATA AVAILABILITY STATEMENT

The data of this study are not publicly available for privacy reasons.