Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic

Dermatitis

Ana Duarte

and Orlando Belo

Algoritmi R&D Centre / LASI, University of Minho, Campus of Gualtar, 4710-057 Braga, Portugal

Keywords:

Atopic Dermatitis, Machine Learning, Gene Signature, Sex-Speciﬁc Biomarker, Precision Medicine.

Abstract:

The prevalence of atopic dermatitis is signiﬁcantly higher in women than in men. Understanding the dif-

ferences in the manifestation of the disease between males and females can contribute to more tailored and

effective treatments. Our goal in this paper was to discover sex-speciﬁc biomarkers that can be used to dif-

ferentiate between lesional and non-lesional skin in atopic dermatitis patients. Using transcriptomic datasets,

we ﬁrst identiﬁed the genes with the highest expression difference. Subsequently, several feature selection

methods and machine learning models were employed to select the most relevant genes and identify potential

candidates for sex-speciﬁc biomarkers. Based on backward feature elimination, we obtained a male-speciﬁc

signature with 11 genes and a female-speciﬁc signature with 10 genes. Both candidate signatures were prop-

erly evaluated by an ensemble classiﬁer using an independent test. The obtained AUC and accuracy values

for the male signature were 0.839 and 0.7222, respectively, and 0.65 and 0.6667 for the female signature.

Finally, we tested the male signature on female data and the female signature on male data. As expected, the

analysed metrics decreased considerably in these scenarios. These results suggest that we have identiﬁed two

promising sex-speciﬁc gene signatures, and support that sex affects the ability to distinguish lesions in patients

with eczema.

1 INTRODUCTION

Atopic dermatitis (AD) is a chronic and highly com-

plex inﬂammatory skin condition that signiﬁcantly re-

duces the quality of life of patients. In Europe and

the USA, one in ﬁve children and 10% of adults are

diagnosed with AD (Bylund et al., 2020; Laughter

et al., 2021). Probably due to socioeconomic and

environmental changes, including the growing lev-

els of urbanisation and industrialisation, the preva-

lence of AD is increasing worldwide, having partic-

ularly alarming rates in low-income countries (Nut-

ten, 2015; Schuler et al., 2023; Skevaki et al., 2021;

Tsai et al., 2019). The incidence of the disease varies

signiﬁcantly between different geographical regions

and cultures, age, sex, and ethnicity (Mesjasz et al.,

2023; Nutten, 2015; Schuler et al., 2023). The rea-

sons for this heterogeneity are not clearly understood.

They require further investigation and a deeper under-

standing of how the combination of the multiple fac-

tors involved affects the susceptibility, development,

progression, and treatment of the disease.

https://orcid.org/0000-0001-6505-9888

https://orcid.org/0000-0003-2157-8891

Numerous diseases, including AD, present a

broad spectrum of clinical manifestations, unpre-

dictable courses, and variable responses to therapy. In

this regard, the effective management of these com-

plex diseases associated with multiple phenotypes

and endotypes requires a precision medicine-based

strategy. Concretely, AD is among the disorders that

can beneﬁt most from more personalised and targeted

interventions (Muraro et al., 2016). However, in con-

trast to other diseases, precision medicine in AD is

still in its early stages. Despite the progress made in

recent decades, the clinical reality is still not based on

a multifactorial approach tailored to the needs of each

patient (Mesjasz et al., 2023; Muraro et al., 2016).

Currently, the main challenges for the development of

precision medicine in AD are the discovery of new ef-

fective therapies with few side effects, taking into ac-

count the proﬁle of each patient. Prescribing the most

appropriate therapies for each proﬁle presupposes the

identiﬁcation of the pathophysiological mechanisms

associated with the disease (Arkwright and Koplin,

2023; Leung, 2024). Particularly, the identiﬁcation

of biomarkers capable of distinguishing lesional from

non-lesional skin in AD patients may facilitate the

Duarte, A. and Belo, O.

Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic Dermatitis.

DOI: 10.5220/0012890700003838

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 27-35

ISBN: 978-989-758-716-0; ISSN: 2184-3228

understanding of the mechanisms involved in the de-

velopment of lesions. A thorough comprehension of

these mechanisms is essential to intervene with tar-

geted therapies in the critical pathways. Gene ex-

pression proﬁling based on transcriptomic data can

be used to compare lesional and non-lesional skin tis-

sues and identify the key biomarkers involved. Given

the multiplicity of factors that inﬂuence the disease,

the inclusion of supplementary patient characteristics

such as sex, age, or existing comorbidities can lead to

more accurate biomarker detection. Hence, a holis-

tic approach with the creation of more complex pro-

ﬁles can be tremendously valuable for the manage-

ment and control of the disease, as well as for enhanc-

ing its treatment (Muraro et al., 2016).

With this paper, we expect to contribute to the ad-

vancement of precision medicine in AD. Speciﬁcally,

we aim to discover candidate biomarkers for AD by

applying machine learning (ML) algorithms to gene

expression data of two distinct patient proﬁles. ML

and classical statistical methods have demonstrated

great potential in biomedicine, namely in the process-

ing of high-dimensional datasets such as those used

in the identiﬁcation of gene signatures (Karthik and

Sudha, 2018; Liu et al., 2022). We also hypothesise

that sex plays a role in identifying a reliable gene sig-

nature to differentiate lesional from non-lesional skin

in AD patients. For this reason, we speculate that

males and females have different molecular mecha-

nisms involved in the manifestation of AD and, con-

sequently, a separate analysis is required.

The remaining part of this paper is organised as

follows. Section 2 reports some of the literature that

addresses the application of ML techniques to the dis-

covery of gene signatures, and the inﬂuence of sex on

the manifestation of AD and other diseases. Section 3

summarises the methodology followed to identify the

gene signatures and section 4 presents and discusses

the results obtained. Finally, Section 5 concludes with

an overview of the main ﬁndings and possible future

research directions.

2 GENE SIGNATURES FOR AD:

OPPORTUNITIES AND

CHALLENGES

Despite the urgent need to ﬁnd candidate gene signa-

tures that could revolutionise current dermatological

care, very little research has been conducted at the

AD level using ML algorithms. In fact, most research

applying ML to advance precision medicine focuses

on the most fatal diseases, such as cancer. Neverthe-

less, a few ML-based studies have explored the iden-

tiﬁcation of biomarker candidate genes for AD. One

such example is the work developed by Zhong et al.

(Zhong et al., 2021). Based on a bioinformatics ap-

proach and using LASSO, the authors used transcrip-

tomic datasets and identiﬁed GZMB, CXCL1 and

CD274 as potential biomarkers to distinguish AD le-

sions from non-lesions. On the other side, M

obus and

colleagues (M

obus et al., 2022), also based on tran-

scriptomic datasets, observed two distinct endotypes

for AD associated with notable clinical differences,

allowing patients to be stratiﬁed into eosinophil-high

and eosinophil-low groups. Implementing the Boruta

algorithm and a random forest model, the authors

identiﬁed the most relevant genes to predict the clus-

ters to which the patients belong. Both investigations

demonstrate that some key genes are promising can-

didate biomarkers that have a major impact on the di-

agnosis and management of AD. However, these stud-

ies do not explore the different phenotypes of the pa-

tients, such as age, sex, or ethnicity, which are known

to play a preponderant role in the manifestation of the

disease.

In contrast, other researchers have analysed the

discovery of differentiated biomarkers depending on

the phenotypic characteristics of the patients, namely

sex. For example, Moon et al. (Moon et al., 2013)

proposed a procedure to ﬁnd sex-speciﬁc biomark-

ers based on three datasets from patients with acute

myeloid leukaemia, chronic lymphocytic leukaemia,

and cutaneous melanoma. The considered method-

ology consisted of an algorithm based on the impor-

tance of each feature in order to extract the top-ranked

genes for male and female patients. The selected

genes were properly tested and the results obtained

revealed high accuracy values, conﬁrming the valid-

ity of the strategy followed and underlining the rele-

vance of sex-speciﬁc biomarkers for enhancing prog-

nosis prediction. Further experiments have also been

conducted to unveil sex-speciﬁc biomarkers for dif-

ferent diseases. For instance, some papers exploit the

use of ML methods to ﬁnd sex-speciﬁc biomarkers for

Alzheimer’s disease, emphasising the importance of

considering clinical features in addition to genes for

a more thorough and sensitive analysis (Bourquard

et al., 2023; Ji et al., 2022).

As far as we know, no previous research has ad-

dressed the identiﬁcation of sex-speciﬁc genes for

AD. However, many studies have suggested that sex

has a signiﬁcant contribution to the prevalence and

severity of the disease. For example, Johansson et al.

examined the distribution and characteristics of AD in

a Swedish population and concluded that the disease

is more common in females among young adults (Jo-

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

hansson et al., 2022). Similarly, Kiiski et al. noted

that in a Finnish cohort, the prevalence of AD was

higher among women aged 30 to 49 years than among

men of the same age (Kiiski et al., 2022). These

results are consistent with other investigations con-

ducted in other countries, such as Italy (Pesce et al.,

2015) or the USA (Silverberg and Haniﬁn, 2013),

where the female sex also appears to be a risk fac-

tor for AD. Therefore, we argue that the process of

searching for suitable candidate gene signatures for

AD requires a sex-separated analysis. Accordingly,

this paper presents a novel contribution to the identi-

ﬁcation of sex-speciﬁc biomarkers for AD based on a

ML approach.

3 MATERIALS AND METHODS

From a broader perspective, the methodology consid-

ered in our proposal can be divided into two major

stages. Each of these stages must be performed in-

dependently for the male and female data. First, two

datasets sharing the same platform were used to iden-

tify the differentially expressed genes (DEGs). Subse-

quently, based on the selected DEGs, an extra dataset

was also taken into account to determine gene sig-

natures for AD. A more detailed description of the

sequential processes considered at each stage is pro-

vided in Figure 1.

3.1 Data Gathering and Exploration

The datasets used to conduct this project were ob-

tained from the Gene Expression Omnibus

(GEO),

a public repository containing experimental gene ex-

pression data. A preliminary search was performed

in order to obtain transcriptomic datasets for AD

with information about the sex of the patients. With

this objective in mind, we selected the datasets

GSE130588, GSE58558, and GSE150797. Since

some academics state that datasets from different

manufacturers should not be combined to avoid po-

tential bias in the data, all these data sources were

generated by the Affymetrix manufacturer (Liu et al.,

2021; Serio, 2023). These datasets contain nor-

malised microarray data from skin samples collected

from AD patients. As each of these datasets was de-

signed with the purpose of assessing the patients’ re-

sponse to treatments, only the samples that referred

to the start of the therapies were considered. Further-

more, we took into account only patients with lesional

(AD-L) or non-lesional (AD-NL) AD – Table 1 indi-

cates the main characteristics of the selected data.

Combining the three datasets, the number of sam-

ples by sex is balanced. In total, 87 samples corre-

spond to male patients and 88 samples to female pa-

tients. Particularly, 51 samples from males refer to

AD-L, while 36 refer to AD-NL. Conversely, 49 sam-

ples from females correspond to AD-L and 39 to AD-

NL.

3.2 Differential Expressed Genes

Since the identiﬁed DEGs may have a signiﬁcant im-

pact on the overall results, we need to plan the dif-

ferential gene expression analysis in order to avoid

the exclusion of important genes. Even if they are

from the same manufacturer, datasets from distinct

platforms are more likely to introduce bias into the

data (Campain and Yang, 2010). For this reason,

we decided to perform the differential gene expres-

sion analysis using exclusively the datasets produced

on the same platform, i.e., datasets GSE130588 and

GSE58558.

Figure 1: Sequential processes for obtaining candidate gene signatures for AD.

https://www.ncbi.nlm.nih.gov/geo/

Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic Dermatitis

Table 1: Summary of the properties of the datasets used.

Samples

Dataset Manufacturer Platform Requirements Sex AD-L AD-NL

GSE130588 Affymetrix GPL570

Time: week 0 Female 22 21

Tissue: LS or NL Male 29 21

Female 6 7

GSE58558 Affymetrix GPL570 Time: day 1

Male 12 10

GSE150797 Affymetrix GPL23159

Treatment: Female 21 11

untreated Male 10 5

The ﬁrst step in determining the DEGs was to di-

vide each of the datasets into two groups, each cor-

responding to a proﬁle (male or female). Each proﬁle

was analysed separately in order to ﬁnd speciﬁc DEGs

for the group of male patients and speciﬁc DEGs for

the group of female patients. All the necessary pro-

cesses were performed in R (version 4.3.1) using the

limma package. One of the ﬁrst data treatments was

to merge the datasets associated with the same proﬁle,

i.e., the male samples in the GSE130588 dataset were

combined with the male samples in the GSE58558

dataset, and the same was done for the female sam-

ples. As a result, we obtained a speciﬁc dataset for

males with 41 AD-L and 31 AD-NL samples and a

dataset for females with 28 AD-L and 28 AD-NL

records. Since the merged data came from different

experiments, we corrected the batch effect using the

“removeBatchEffect” function from the limma pack-

age. Moreover, because we were working with mi-

croarray data, some additional cleansing tasks were

also required for the following encountered situations:

• Multiple probes corresponding to the same

gene – only the probe with the highest average

expression was kept, and the other probes were

discarded. Ties in average counts were resolved

by choosing one of the probes and eliminating the

rest (Miller et al., 2011).

• Probes associated with various genes – these

records were removed (Hu et al., 2023).

• Probes that did not match a speciﬁc gene –

these records were removed (Wang and Yu, 2023).

To ﬁnalise the data treatment, the probe IDs were

converted into their corresponding gene symbols. Fi-

nally, genes with an absolute log

(fold change) ≥ 1

and p

adj

< 0.05 were identiﬁed as DEGs and saved in

text ﬁles.

3.3 Gene Selection Strategy

After identifying the DEGs for each proﬁle, we

prepared the datasets GSE130588, GSE58558, and

GSE150797 for feature selection and ML modelling.

As with the differential expression analysis, this pro-

cess was also performed using R, analysing each pa-

tient proﬁle separately. Each of the three datasets

was therefore split into male and female groups, re-

sulting in a total of six subsets. For each subset, we

checked the existence of multiple probes correspond-

ing to the same gene and addressed the issue with a

similar approach to the one we used to identify the

DEGs. Probe IDs were also transformed into the cor-

responding gene symbols. By importing the text ﬁle

containing the determined DEGs, we ﬁltered the data

in order to have only DEGs. For each proﬁle, the cor-

responding subsets were merged, and the batch effect

was removed. The resulting male and female datasets

were then ready to support the next steps of the work.

The following tasks, including additional data

processing, feature selection, and construction of ML

models, were performed in Python 3.6 using the

scikit-learn library. All these tasks were applied in

parallel to the male and female datasets. For each

dataset, we divided the data into train (80%) and test

(20%), using a stratiﬁed strategy to maintain the pro-

portion of AD-L and AD-NL samples in both sets.

Only the training data were used for feature selection

and for the construction of ML classiﬁers to iden-

tify potential gene signatures. For ML modelling,

we considered a shufﬂed and stratiﬁed 5-fold cross-

validation. The optimal hyperparameters for each al-

gorithm were found using BayesSearchCV with 30

iterations.

Although the identiﬁcation of DEGs helps to re-

duce the dimensionality of the data, the resulting high

number of genes is still a drawback for efﬁcient pro-

cessing by ML methods. Feature selection is a com-

mon strategy used to minimise these gaps. Therefore,

before building ML algorithms, we conducted a fea-

ture selection approach using Boruta, Support Vector

Machine Recursive Feature Elimination (SVM-RFE),

and Least Absolute Shrinkage and Selection Operator

(LASSO). These three methods reduced the number

of genes differently and thus originated three distinct

gene sets. The Random Forest (RF), XGBoost, Ad-

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

Figure 2: Before (left) and after (right) batch effect correction in the male proﬁle.

aBoost, linear Support Vector Machine (SVM), and

Logistic Regression (LR) methods were applied to

each of these sets. For each ML method, we extracted

the importance of each gene and used the Min-Max

method to normalise the obtained value between 0

and 1. Thus, for each of the three gene sets, we ob-

tained the normalised importance values of each gene

for each algorithm, which we then summed to gen-

erate a gene score. The top genes with signiﬁcantly

higher scores were selected as candidate genes for the

creation of a gene signature.

From the selected set of candidate genes, we con-

ceived new ML models using the same ﬁve meth-

ods and constructed a soft-voting classiﬁer that com-

bines the predictions of the models. This ensemble

method was used to evaluate the predictive power of

the models, considering both AUC and accuracy val-

ues. Adopting a backward feature elimination strat-

egy, we discarded the least important gene and cre-

ated new classiﬁers to compare the AUC and accuracy

values with the prior solution. These steps were per-

formed iteratively until the performance metrics were

worse than the values obtained in the previous step.

Thus, the genes that were not eliminated by this iter-

ative process became part of our proposed gene sig-

nature. In the end, we obtained one candidate gene

signature for males and another for females.

To test our initial hypothesis that sex must be

taken into account when establishing gene signatures

for AD, we used the independent test to compare the

AUC and accuracy values obtained when the male

signature is applied to the male and female datasets

and vice versa. The additional step of testing the

male signature against the female dataset and vice

versa was necessary to determine whether the identi-

ﬁed genes can be considered sex-speciﬁc candidates.

4 RESULTS

4.1 DEGs Identiﬁcation and Initial

Gene Selection

At the beginning of the differential gene expres-

sion analysis, two distinct groups were identiﬁed af-

ter merging the datasets GSE130588 and GSE58558.

In both proﬁles, the samples belonging to the same

datasets had similar expression values, but these val-

ues were signiﬁcantly different from the expression

values of the samples in the other dataset. Batch ef-

fect correction removed these technical differences.

Figure 2 shows the box plots obtained before and af-

ter removing the batch effect in the male proﬁle. For

females, the batch effect correction led to similar re-

sults. After processing the data, we identiﬁed 188 and

764 DEGs for the male and female datasets, respec-

tively.

Once the DEGs were determined, we proceeded to

the individual treatment of the three datasets, which

involved ﬁltering the genes according to the DEGs

found. As the GSE150797 dataset was generated us-

ing a different platform, some DEGs did not match the

probe IDs. Consequently, after merging the datasets,

the number of DEGs was reduced to 172 in the male

proﬁle and 700 in the female scenario. Moreover, we

also observed that there were larger differences be-

Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic Dermatitis

tween the GSE150797 dataset (GPL23159 platform)

and the datasets from the GPL570 platform before

batch effect correction. After removing the batch ef-

fect, the existing differences were successfully min-

imized. Boruta, SVM-RFE and LASSO yielded the

three gene subsets indicated in Table 2.

Table 2: Number of selected genes by each feature selection

approach.

Proﬁle Boruta SVM-RFE LASSO

Male 12 90 62

Female 23 75 92

4.2 Determination of Sex-Speciﬁc Gene

Signatures

Since the number of DEGs in the male dataset is rel-

atively small, in this case, we decided to include an

additional scenario corresponding to the training and

validation of the ML algorithms without using any

feature selection method. Table 3 lists the top genes

obtained after implementing the ML process and de-

termining the scores. In general, the different fea-

ture selection strategies identiﬁed the top genes con-

sistently for the same proﬁles.

Based on the results, we selected an initial set

of 11 candidate genes for males, and for females,

we considered the 16 most important genes (Ta-

ble 3, highlighted in bold). Our backward feature

elimination strategy led to the discovery of an op-

timal 11-gene signature speciﬁc to males (KIF2C,

AKR1B10, PHYHIP, FOSL1, FPR1, HS3ST3A1,

MX1, KANK4, PPARG, BCL2A1, and KLHDC7B)

and a 10-gene signature speciﬁc to females (CEP126,

FCHSD1, C17orf96, IL18RAP, P2RY10, PTAFR,

ANKFN1, TBX18, P2RY2, and AEN). Interestingly,

none of the genes are common to both signatures.

This may indicate that the molecular pathways in-

volved in the development of AD lesions may dif-

fer between men and women, suggesting that the sex

of the patients should be considered for better dis-

ease management and treatment. The genes KANK4,

PHYHIP and PPARG from the male proﬁle were

downregulated in the lesions, while the rest were

upregulated. In the female proﬁle, only ANKFN1,

CEP126 and TBX18 were downregulated, while all

others were upregulated. There is scientiﬁc evidence

that some of these genes may have a major impact on

AD. For example, the downregulation of the PPARG

gene, identiﬁed in the male signature, may be associ-

ated with inﬂammation, keratinisation, and sebaceous

gland function (Konger et al., 2021). On the other

hand, some studies suggest that P2Y receptors, such

as the P2RY10 and P2RY2 genes found in the female

signature, can be involved in skin inﬂammation (Pas-

tore et al., 2007).

Table 4 presents the AUC and accuracy values

obtained for each candidate signature using the vot-

ing classiﬁer. A closer analysis of the results shows

that the AUC and accuracy values of the male sig-

nature in the independent test are considerably high

when applied to the male data (0.839 and 0.7222, re-

spectively). However, when this signature is tested

with the female data, the AUC (0.575) and accuracy

(0.6111) values deteriorate substantially. Although

the difference is not as marked, the same is true for

the female signature. The AUC and accuracy values

of the female signature when applied to the female

data are 0.650 and 0.6667, respectively. These values

decrease when the female signature is tested on the

male dataset (AUC = 0.552 and accuracy = 0.6111).

These ﬁndings thus demonstrate that considering sex-

speciﬁc biomarkers leads to improved gene signatures

for distinguishing lesions from non-lesions, reinforc-

ing the beneﬁts of a sex-separate analysis to establish

candidate gene signatures for AD.

Table 3: Top genes identiﬁed by the feature selection methods for each proﬁle.

Male proﬁle Female proﬁle

Boruta SVM-RFE LASSO All DEGs Boruta SVM-RFE LASSO

KIF2C KIF2C FOSL1 KIF2C CEP126 FCHSD1 C17orf96

AKR1B10 FOSL1 KIF2C FOSL1 FCHSD1 C17orf96 IL18RAP

PHYHIP FPR1 MX1 PHYHIP C17orf96 IL18RAP MS4A14

FOSL1 MX1 PHYHIP AKR1B10 GNA15 PTAFR STRIP2

FPR1 KANK4 HS3ST3A1 KLHDC7B IL18RAP WIF1 TBX18

HS3ST3A1 PPARG KANK4 FPR1 P2RY10 STRIP2 P2RY2

BCL2A1 HS3ST3A1 PLAG1 AEN

MS4A14 GNA15

ANKFN1 HSD11B1

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

Table 4: AUC and accuracy values of the optimal gene signatures when tested against the male and female datasets.

Male data Female data

AUC accuracy AUC accuracy

train 0.9737 0.8099 0.8543 0.8429

Male signature

test 0.839 0.7222 0.575 0.6111

train 0.76 0.6659 0.975 0.9286

Female signature

test 0.552 0.6111 0.650 0.6667

5 CONCLUSION AND FUTURE

WORK

AD presents an unequal distribution between men and

women, and its incidence is increasing worldwide.

Nevertheless, there are very few studies on this dis-

ease to identify biomarkers through ML techniques.

Speciﬁcally, we have not found any scientiﬁc study

aimed at ﬁnding sex-speciﬁc biomarkers for the dis-

ease. This could be of particular relevance to gain

deeper insights into how AD manifests in men and

women. To ﬁll this gap in the literature, we developed

a ML approach using transcriptomic datasets and in-

tended to identify male and female biomarkers that

distinguish normal from lesional skin in patients with

atopic eczema.

Our research led to the deﬁnition of a male-

speciﬁc gene signature consisting of the KIF2C,

AKR1B10, PHYHIP, FOSL1, FPR1, HS3ST3A1,

MX1, KANK4, PPARG, BCL2A1, and KLHDC7B

genes, and a female-speciﬁc gene signature com-

prising the CEP126, FCHSD1, C17orf96, IL18RAP,

P2RY10, PTAFR, ANKFN1, TBX18, P2RY2, and

AEN genes. For some of the identiﬁed genes, there

is evidence in the literature to support their possible

inﬂuence on the skin. The difference between the

genes of the two signatures could indicate that dif-

ferent mechanisms are involved in the manifestation

of AD in men and women. A better understanding

of these mechanisms could promote the emergence of

targeted treatments and contribute to the development

of precision medicine in AD.

Although the results obtained emphasise the need

to investigate sex-speciﬁc biomarkers for AD, our

study has certain limitations. The main shortcom-

ings are the limited number of samples and the lack

of public databases providing gene expression data

in combination with clinical phenotypes. Therefore,

new studies on this topic and the availability of new

datasets integrating transcriptomic and phenotypic

data are currently a priority. It would also be valuable

for future research to replicate the proposed method-

ology to other diseases, particularly those where it is

suspected that different molecular mechanisms may

be involved depending on the sex. In addition, a thor-

ough investigation of the discovered biomarkers as

well as the associated molecular mechanisms is re-

quired to gain a comprehensive understanding of how

men and women differ in the development of AD le-

sions. Finally, any research of this nature requires

clinical validation.

ACKNOWLEDGEMENTS

This work has been supported by FCT – Fundac¸

para a Ci

encia e Tecnologia within the R&D Units

Project Scope: UIDB/00319/2020, and the PhD grant:

2022.12728.BD.

REFERENCES

Arkwright, P. D. and Koplin, J. J. (2023). Challenging best

practice of atopic dermatitis. Journal of Allergy and

Clinical Immunology: In Practice, 11:1391–1393.

Bourquard, T., Lee, K., Al-Ramahi, I., Pham, M., Shapiro,

D., Lagisetty, Y., Soleimani, S., Mota, S., Wilhelm,

K., Samieinasab, M., Kim, Y. W., Huh, E., Asmussen,

J., Katsonis, P., Botas, J., and Lichtarge, O. (2023).

Functional variants identify sex-speciﬁc genes and

pathways in alzheimer’s disease. Nature Communi-

cations, 14.

Bylund, S., Kobyletzki, L. B. V., Svalstedt, M., and

Ake

Svensson (2020). Prevalence and incidence of atopic

dermatitis: A systematic review. Acta Dermato-

Venereologica, 100:320–329.

Campain, A. and Yang, Y. H. (2010). Comparison study

of microarray meta-analysis methods. BMC Bioinfor-

matics, 11.

Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic Dermatitis

Hu, Y., Chen, X., Mei, X., Luo, Z., Wu, H., Zhang, H.,

Zeng, Q., Ren, H., and Xu, D. (2023). Identiﬁcation of

diagnostic immune-related gene biomarkers for pre-

dicting heart failure after acute myocardial infarction.

Open Medicine, 18.

Ji, W., An, K., Wang, C., and Wang, S. (2022). Bioinformat-

ics analysis of diagnostic biomarkers for alzheimer’s

disease in peripheral blood based on sex differences

and support vector machine algorithm. Hereditas,

159.

Johansson, E. K., Bergstr

om, A., Kull, I., Mel

en, E., Jons-

son, M., Lundin, S., Wahlgren, C. F., and Ballardini,

N. (2022). Prevalence and characteristics of atopic

dermatitis among young adult females and males - re-

port from the swedish population-based study bamse.

Journal of the European Academy of Dermatology

and Venereology, 36:698–704.

Karthik, S. and Sudha, M. (2018). A survey on ma-

chine learning approaches in gene expression classiﬁ-

cation in modelling computational diagnostic system

for complex diseases. International Journal of En-

gineering and Advanced Technology (IJEAT), 8:182–

191.

Kiiski, V., Salava, A., Susitaival, P., Barnhill, S., Remitz,

A., and Heliovaara, M. (2022). Atopic dermatitis in

adults: a population-based study in ﬁnland. Interna-

tional Journal of Dermatology, 61:324–330.

Konger, R. L., Derr-Yellin, E., Zimmers, T. A., Katona, T.,

Xuei, X., Liu, Y., Zhou, H.-M., Simpson, E. R., and

Turner, M. J. (2021). Epidermal pparγ is a key home-

ostatic regulator of cutaneous inﬂammation and bar-

rier function in mouse skin. International Journal of

Molecular Sciences, 22.

Laughter, M. R., Maymone, M. B., Mashayekhi, S., Arents,

B. W., Karimkhani, C., Langan, S. M., Dellavalle,

R. P., and Flohr, C. (2021). The global burden of

atopic dermatitis: lessons from the global burden of

disease study 1990–2017. British Journal of Derma-

tology, 184:304–309.

Leung, D. Y. (2024). Evolving atopic dermatitis toward

precision medicine. Annals of allergy, asthma & im-

munology, 132:107–108.

Liu, J., Liu, L., Antwi, P. A., Luo, Y., and Liang, F. (2022).

Identiﬁcation and validation of the diagnostic charac-

teristic genes of ovarian cancer by bioinformatics and

machine learning. Frontiers in Genetics, 13.

Liu, L., Wang, T., Huang, D., and Song, D. (2021). Compre-

hensive analysis of differentially expressed genes in

clinically diagnosed irreversible pulpitis by multiplat-

form data integration using a robust rank aggregation

approach. Journal of Endodontics, 47:1365–1375.

Mesjasz, A., Kołkowski, K., Wollenberg, A., and Trzeciak,

M. (2023). How to understand personalized medicine

in atopic dermatitis nowadays? International Journal

of Molecular Sciences, 24.

Miller, J. A., Cai, C., Langfelder, P., Geschwind, D. H.,

Kurian, S. M., Salomon, D. R., and Horvath, S.

(2011). Strategies for aggregating gene expression

data: The collapserows r function. BMC Bioinformat-

ics, 12.

Moon, H., Lopez, K. L., Lin, G. I., and Chen, J. J. (2013).

Sex-speciﬁc genomic biomarkers for individualized

treatment of life-threatening diseases. Disease Mark-

ers, 35:661–667.

Muraro, A., Lemanske, R. F., Hellings, P. W., Akdis, C. A.,

Bieber, T., Casale, T. B., Jutel, M., Ong, P. Y., Poulsen,

L. K., Schmid-Grendelmeier, P., Simon, H. U., Seys,

S. F., and Agache, I. (2016). Precision medicine in

patients with allergic diseases: Airway diseases and

atopic dermatitis - practall document of the european

academy of allergy and clinical immunology and the

american academy of allergy, asthma & immunol-

ogy. Journal of Allergy and Clinical Immunology,

137:1347–1358.

obus, L., Rodriguez, E., Harder, I., Boraczynski, N.,

Szymczak, S., H

ubenthal, M., St

olzl, D., Gerdes, S.,

Kleinheinz, A., Abraham, S., Heratizadeh, A., Han-

drick, C., Haufe, E., Werfel, T., Schmitt, J., and Wei-

dinger, S. (2022). Blood transcriptome proﬁling iden-

tiﬁes 2 candidate endotypes of atopic dermatitis. Jour-

nal of Allergy and Clinical Immunology, 150:385–

395.

Nutten, S. (2015). Atopic dermatitis: Global epidemiology

and risk factors. Annals of Nutrition and Metabolism,

66:8–16.

Pastore, S., Mascia, F., Gulinelli, S., Forchap, S., Dat-

tilo, C., Adinolﬁ, E., Girolomoni, G., Virgilio, F. D.,

and Ferrari, D. (2007). Stimulation of purinergic re-

ceptors modulates chemokine expression in human

keratinocytes. Journal of Investigative Dermatology,

127:660–667.

Pesce, G., Marcon, A., Carosso, A., Antonicelli, L., Caz-

zoletti, L., Ferrari, M., Fois, A. G., Marchetti, P.,

Olivieri, M., Pirina, P., Pocetta, G., Tassinari, R.,

Verlato, G., Villani, S., and Marco, R. D. (2015).

Adult eczema in italy: prevalence and associations

with environmental factors. Journal of the European

Academy of Dermatology and Venereology, 29:1180–

1187.

Schuler, C. F., Billi, A. C., Maverakis, E., Tsoi, L. C., and

Gudjonsson, J. E. (2023). Novel insights into atopic

dermatitis. Journal of Allergy and Clinical Immunol-

ogy, 151:1145–1154.

Serio, P. (2023). Gene expression microarray merging.

https://rpubs.com/Karksus/1013177.

Silverberg, J. I. and Haniﬁn, J. M. (2013). Adult eczema

prevalence and associations with asthma and other

health and demographic factors: A us population-

based study. Journal of Allergy and Clinical Immunol-

ogy, 132:1132–1138.

Skevaki, C., Ngocho, J. S., Amour, C., Schmid-

Grendelmeier, P., Mmbaga, B. T., and Renz, H.

(2021). Epidemiology and management of asthma and

atopic dermatitis in sub-saharan africa. Journal of Al-

lergy and Clinical Immunology, 148:1378–1386.

Tsai, T.-F., Rajagopalan, M., Chu, C.-Y., Encarnacion, L.,

Gerber, R. A., Santos-Estrella, P., Llamado, L. J. Q.,

and Tallman, A. M. (2019). Burden of atopic dermati-

tis in asia. Journal of Dermatology, 46:825–834.

Wang, X. and Yu, G. (2023). Drug discovery in canine py-

ometra disease identiﬁed by text mining and microar-

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

ray data analysis. BioMed Research International,

2023.

Zhong, Y., Qin, K., Li, L., Liu, H., Xie, Z., and Zeng, K.

(2021). Identiﬁcation of immunological biomarkers of

atopic dermatitis by integrated analysis to determine

molecular targets for diagnosis and therapy. Interna-

tional Journal of General Medicine, 14:8193–8209.

Machine Learning Unravels Sex-Speciﬁc Biomarkers for Atopic Dermatitis