Abstract

Introduction

   Asthma has an annual increasing morbidity rate and imposes a heavy
   social burden on public healthcare systems. The aim of this study was
   to use machine learning to identify asthma-specific genes for the
   prediction and diagnosis of asthma.

Methods

   Differentially expressed genes (DEGs) related to asthma were identified
   by examining public sequencing data from the Gene Expression Omnibus,
   coupled with the support vector machine recursive feature elimination
   and least absolute shrinkage and selection operator regression model.
   Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG),
   Gene set enrichment analysis and correlation analyses between gene and
   immune cell levels were performed. An ovalbumin-induced asthma mouse
   model was established, and eukaryotic reference transcriptome
   high-throughput sequencing was performed to identify genes expressed in
   mouse lung tissues.

Results

   Thirteen specific asthma genes were obtained from our dataset analysis
   (LOC100132287, CEACAM5, PRR4, CPA3, POSTN, LYPD2, TCN1, SCGB3A1, NOS2,
   CLCA1, TPSAB1, CST1, and C7orf26). The GO analysis demonstrated that
   DEGs linked to asthma were primarily related to positive regulation of
   guanylate cyclase activity, gpi anchor binding, peptidase activity and
   arginine binding. The renin-angiotensin system, arginine biosynthesis
   and arginine and proline metabolism were the key KEGG pathways of DEGs.
   Additionally, the genes CEACAM5, PRR4, CPA3, POSTN, CLCA1, and CST1
   expression levels were positively associated with plasma cells and
   resting mast cells. The mouse model revealed elevated nos2 and clca1
   expression in the asthmatic mouse group compared with that in normal
   mice, which was consistent with the findings in asthmatic patients.

Discussion

   This study identified new marker genes for the prediction and diagnosis
   of asthma, which can be further validated and applied clinically

   Keywords: asthma, differentially expressed genes, machine learning,
   support vector machine recursive feature elimination (SVM-RFE), least
   absolute shrinkage and selection operator (LASSO) regression model

1. Introduction

   Asthma is a chronic respiratory illness marked by inflammation and
   remodeling of the airways that is triggered by complex genetic
   regulation and environmental exposure to allergens ([44]1). It is
   caused by swelling and increased mucus in the airways (inflammation)
   and narrowing of the airways due to muscle spasm (bronchospasm) leading
   to bronchial overreaction and airway obstruction ([45]2). Asthma
   affects 5%–10% of the population in many developed countries. In China,
   approximately 4.2% of adults over the age of 20 years and older have
   asthma. There are more than 300 million people with asthma worldwide,
   and its prevalence is increasing every year ([46]3). Asthma usually
   occurs in preschool years and causes reducing quality of life and early
   death, which leads to a large public health burden.

   Many patients with asthma use self-care and pharmacological therapies
   to control their symptoms. Recently, monoclonal antibodies have been
   applied to benefit patients ([47]4). However, their efficacy differs
   owing to the diverse forms of asthma, and some individuals do not
   respond to current asthma treatments ([48]5). In recognition of
   inadequacies in the current understanding of asthma mechanisms, our
   research highlights the need of having a thorough grasp of diagnosis
   and immunological variability in asthma.

   Asthma involves multiple inflammatory responses. The initial barrier to
   resistance against microorganisms, gases, and allergens is the
   bronchial epithelial cells, which is also the centre of the
   inflammatory response ([49]6). First, allergens can be eliminated by
   airway epithelial cells via mucus ([50]7). Second, by using pattern
   recognition receptors, airway epithelial cells may identify chemicals
   associated with pathogens or hazards, and release cytokines and
   chemokines (such as IL25, TSLP, CCL5, and CCL22), activate dendritic
   cells, connect innate and adaptive immunity, and trigger local immune
   responses ([51]8, [52]9). Third, they function as antigen-presenting
   cells, which help trigger type II immunological responses by causing
   naive T cells to differentiate into CD4+ T cells ([53]10). Therefore,
   we chose samples of bronchial epithelial cells from public datasets for
   sequencing and analysis in this study.

   Machine learning is a collection of computational intelligence
   techniques that allows a computer to learn a task autonomously,
   improving its experience without being explicitly programmed. This
   method can be used to analyze large amounts of data, establish complex
   and nonlinear relationships, and identify patterns and relationships
   between data and interesting outcomes. Zhang et al. identified the
   immune infiltration-related diagnostic genes SLC27A3 and STAU1 using
   machine learning in patients with Chronic Obstructive Pulmonary
   Disease. The area under the receiver operating characteristic (ROC)
   curve (AUC) was up to 0.900 and 0.971, which demonstrated their high
   diagnostic value ([54]11). Potential biomarkers of idiopathic pulmonary
   frailty were previously identified using the least absolute shrinkage
   and selection operator (LASSO) regression model, random forest
   algorithms, and support vector machine recursive feature elimination
   (SVM-RFE). These methods were very helpful in enabling an early
   diagnosis and enhancing prognosis ([55]12). Research to date shows that
   machine learning helps uncover the potential causal mechanisms of
   asthma with good predictive efficacy and can generate new hypotheses
   ([56]13, [57]14).

   Despite its significant impact on public health, the diagnosis of
   asthma mostly depends on symptoms and accompanying testing, which does
   not allow for prediction and treatment. Our study explored
   differentially expressed genes (DEGs) in asthma by analyzing the
   bronchial epithelial cell transcriptome of the [58]GSE63142 and
   [59]GSE158752 databases from the Gene Expression Omnibus (GEO) public
   sequencing data using machine learning approaches. We evaluated the
   diagnostic utility of these genes using the AUC after doing functional,
   pathway, and gene set enrichment analyses of the DEGs. We hope to
   contribute to the genetic diagnosis and prediction of asthma in the
   future.

2. Methods

2.1. Microarray data processing and analysis

   We obtained the [60]GSE63142 and [61]GSE158752 asthma datasets from the
   GEO database of the National Center for Biotechnology Information
   (NCBI) ([62]https://www.ncbi.nlm.nih.gov/geo/). The [63]GSE63142
   dataset ([64]GPL6480 platform) ([65]15) was uploaded in 2014 and
   included transcriptome studies of the bronchial epithelial cells from
   27 healthy participants and 128 patients with asthma. In the
   [66]GSE158752 dataset ([67]GPL18573 platform) ([68]16), 50 bronchial
   epithelial cell samples from patients with asthma and 17 normal samples
   were obtained. R software (version 4.2.2;
   [69]https://www.r-project.org/) and RStudio software (version 4.2.2;
   [70]https://www.r-project.org/) were used to process and analyze the
   datasets.

2.2. Analysis of differentially expressed genes

   All profiles of the gene transcription microarray data were
   pre-processed utilising the “SVA” package (version 3.46.0), which
   included background correction and normalization. Gene symbols were
   annotated based on the annotation data. To execute advanced feature
   selection and visualization, we utilized [71]GSE63142 as a training set
   for the DEGs analysis and used [72]GSE158752 as a test set. We examined
   transcriptome samples from healthy controls and asthma patients to
   support the expression of the crucial genes identified. The
   [73]GSE158752 dataset was used for verifying the core gene differential
   expression. Using the “limma” R package, we determined the DEGs between
   samples from patients with asthma and healthy subjects using a
   conservative threshold (|log2FC| > 1.0, p < 0.05). We utilised the
   “pheatmap” (version 1.0.12) and “ggplot2” packages (version 3.4.2) to
   generate a heatmap and volcano plot, respectively.

2.3. Function, pathway and gene set enrichment analysis and protein–protein
interaction network analysis of DEGs

   Gene Ontology (GO), Kyoto Encyclopedia of Genes Genomes (KEGG) pathway
   enrichment analyses were used to identify the characteristic
   biological, molecular, and cellular attributes and reveal associated
   enriched pathways. Gene Set Enrichment Analysis (GSEA) was employed to
   investigate the role of genes in biology ([74]17). Analyses using KEGG,
   GO, and GSEA were performed using the R package “clusterProfiler”
   (version 3.14.3). Significantly enriched biological processes,
   molecular functions, cellular components, and pathways were chosen
   based on a threshold of p-value < 0.05 and an false discovery rate
   (FDR) < 0.05. The STRING database ([75]https://cn.string-db.org/) was
   used for the protein–protein interaction (PPI) network analysis of DEGs
   related to asthma. It provides uniquely comprehensive coverage,
   integration, and interactions obtained through text mining ([76]18).

2.4. Techniques for machine learning to find diagnostic indicators

   Using the training set, we applied two types of machine learning
   algorithms to predict asthma disease-associated genes: SVM-RFE and
   LASSO regression models. The LASSO regression is a model for variable
   selection and complexity regularisation. We used the “Venn” package to
   intersect the diagnostic markers of asthma patients generated by the
   LASSO regression and the techniques of SVM-RFE. To gauge the
   performance of these models, we deployed the ROC curve and computed the
   AUC. The AUC quantifies the capacity of these two models to
   discriminate between healthy control and asthma samples, with
   predictions based on the chosen features. In summary, we implemented
   the LASSO regression and SVM-RFE techniques in the training set to
   select and utilise the training and test sets to then identify
   important DEGs for asthma diagnosis and evaluate the models'
   performance by employing the ROC curve and AUC values.

2.5. Connection between immune cells and core genes

   We performed immunoinfiltration analysis using the “CIBERSORT.R”
   package in the training cohort, used the “corrplot” packages to analyze
   the differentiated infiltration of immune cells between healthy people
   and patients with asthma, used the “preprocessCore” package to generate
   immunocyte content, and analyzed the correlation of core gene
   expression and immune cells by the “reshape2”, “ggpubr”, and “ggExtra”
   packages. Data with p-values < 0.05 and q-values < 0.05 were retained
   and volplot, barboplot, lollipop pictures were generated.

2.6. Animal experiments

   Female wild-type C57BL/6 mice were obtained from the Beijing Vital
   River Laboratory Animal Technology Co (Beijing, China). Mice were
   provided unlimited access to water and standard food and were raised
   under specific pathogen-free conditions (22°C ± 1°C, 50% ± 5% humidity)
   with a light/dark cycle of 12 h/12 h. One week before use, the mice
   were isolated and acclimated. Ten mice were classified into two groups
   (n = 5 each): sham and ovalbumin (OVA). Mice in the OVA group were
   given an intraperitoneal injection of sensitized OVA (100 μg; Cat #:
   A5503, Sigma-Aldrich, USA) and aluminium hydroxide (1 mg; Cat #:77161,
   Thermo Fisher Scientific, USA) on days 1 and 13, while saline (100 μl)
   was injected into mice in the Sham group. Mice in the OVA group were
   administered 2% OVA in an aerosol form for over 30 min for seven
   consecutive days, while mice in the sham group were administered
   saline. All animals were sacrificed within 24 h of the last
   nebulization. The Ethics Committee of the Wenzhou Medical University
   Laboratory Animal Resource Center (Wenzhou, China) granted consent for
   all studies to be conducted in accordance with the ARRIVE criteria.

2.7. Transcriptome analysis

   Library building for the high-throughput sequencing of the eukaryotic
   reference transcriptome from the examined mouse lung tissue samples was
   performed by LC-Bio Technology Co., Ltd. (Hangzhou, China). An Illumina
   HiSeq X Reagent Kit (Illumina, San Diego, USA) was used as the
   sequencing platform. The expression of each transcript was calculated
   using the fragments per kilobase per million read technique to identify
   the DEGs between samples.

2.8. Statistical analysis

   For statistical analysis and visualisation, R version 4.2.2 was
   employed, while GraphPad Prism 8.2.0 was used to statistically analyze
   the images. Student's t-tests were used to compare group variables.
   Results were shown as the mean ± standard deviation. At least three
   replicates were performed for each experiment. Statistical significance
   was set at p < 0.05.

3. Results

3.1. Machine learning model development and identification of critical genes
for asthma discrimination

   A total of 30,889 genes from 155 bronchial epithelial cell samples
   obtained from 128 patients with asthma and 27 healthy individuals were
   included in the training set from the GEO dataset. The age range of
   subjects was 18–62 years, with an average age of 37 years. We first
   divided the data into two groups to examine the gene expression matrix
   of the training set: healthy controls and patients with asthma. The
   subsequent DEG analysis revealed a notable variation in gene expression
   between the two groups, as demonstrated in the heatmap, highlighting
   the distinct gene expression landscape of patients with asthma
   ([77]Figure 1A). A comparatively small number of DEGs were found in
   this comparison, with a volcano plot displaying 9 upregulated and 4
   downregulated genes with a conserved threshold of |log2FC| > 1.0,
   p < 0.05 ([78]Figure 1B).

Figure 1.

   [79]Figure 1
   [80]Open in a new tab

   DEGs of patients with asthma using the [81]GSE63142 datasets. (A)
   Heatmap of the the [82]GSE63142 datasets; gene upregulation is
   indicated by red and gene downregulation is indicated by
   blue,highlighting the differences in gene expression between healthy
   controls and asthma patients. (B) Volcano plot of the [83]GSE63142
   datasets; significant DEGs with a conservative threshold of
   |log2FC| > 1.0, p < 0.05; red represents gene upregulation and blue
   represents gene downregulation. 13 genes were differentially expressed
   between healthy controls and patients with asthma (9 genes upregulated
   and 4 genes downregulated). (C) The LASSO regression analysis to
   identify the most relevant genes for asthma diagnosis based on the
   differential expression analysis results, which identified 8 diagnostic
   core genes. (D) The SVM-RFE menthod using the e1071, kernlab and caret
   package, which identified 10 diagnostic core genes. (E) Venn plot
   depicting the identification of key genes for distinguishing between
   healthy and asthma patients. The intersection of two machine learning
   algorithms—lasso regression and SVM-RFE method—reveals 8 pivotal genes
   (LOC100132287, CEACAM5, PRR4, CPA3, POSTN, LYPD2, TCN1, and SCGB3A1) as
   robust and discriminative features. DEGs, differentially expressed
   genes; LASSO, least absolute shrinkage and selection operator; SVM-RFE,
   support vector machine recursive feature elimination.

   We constructed the LASSO regression model for feature selection, which
   identified 8 diagnostic core genes of asthma. This approach reduced
   unimportant feature coefficients to zero. The SVM-RFE is a learning
   algorithm used in nonlinear classification, which constructs a
   hyperplane in the feature classes with a maximum margin ([84]19,
   [85]20). The SVM-RFE method is used to identify the most critical
   diagnostic markers associated with asthma progression. Using the
   SVM-RFE approach, 10 asthma diagnostic genes were identified. Through
   the convergence of the SVM-RFE approach and LASSO regression, eight
   important genes were found to be discriminative characteristics that
   might be used to separate asthma sufferers from healthy individuals
   [LOC100132287, carcinoembryonic antigen-related cell adhesion molecule
   5 (CEACAM5); proline-rich 4 (PRR4); carboxypeptidase A3 (CPA3);
   periostin (POSTN); LY6/PLAUR domain cintaining 2 (LYPD2);
   transcobalamin 1 (TCN1); and secretoglobin family 3A member 1
   (SCGB3A1)], as shown in the Venn diagram ([86]Figures 1C–E).

3.2. Pathway enrichment analysis

   Considering the significant variations in asthma gene expression
   patterns, we continued with a more comprehensive study focusing on the
   13 DEGs. According to the results of the GO enrichment analysis, 95 GO
   items comprised 19 molecular functions (MF), 64 biological processes
   (BP), and 12 cellular components (CC). Several processes were
   significantly enriched. Notably, these included the response to oxygen
   levels and positive regulation of guanylate cyclase activity for BP,
   collagen-containing extracellular matrix, zymogen granule and anchoring
   membrane component for CC, and gpi anchor binding, peptidase activity,
   arginine binding and intracellular calcium activated chloride channel
   activity regard to MF. These findings underscore the importance of
   protein metabolism and enzyme regulation in the pathogenesis of asthma
   ([87]Figures 2A–C).

Figure 2.

   [88]Figure 2
   [89]Open in a new tab

   Function, pathway of the DEG enrichment analysis. (A–C) GO analysis of
   DEGs. Enrichment analysis of DEGs, demonstrating a strong association
   with response to oxygen levels and positive regulation of guanylate
   cyclase activity in BP, collagen-containing extracellular matrix,
   zymogen granule and anchoring membrane component in CC, and gpi anchor
   binding, peptidase activity, arginine binding and intracellular calcium
   activated chloride channel activity regard to MF. (D) KEGG pathway
   analysis of DEGs. The DEGs were linked to the pancreatic secretion,
   arginine biosynthesis, renin-angiotensin system and arginine and
   proline metabolism using KEGG pathway analysis. (E) Annotated KEGG
   pathway analysis using the GSEA approach. The results highlighted
   significant enrichment of amino sugar and nucleotide sugar metabolism,
   glycan biosynthesis, parkinsons disease, protein export, steroid
   hormone biosynthesis. (F) Interaction network of the proteins regulated
   by the key genes. PPI network of 13 specific asthma genes reflected the
   core connectivity of CPA3 and TPSAB1 in these proteins. DEGs,
   differentially expressed genes; GO, gene ontology; KEGG, Kyoto
   encyclopedia of genes and genomes; PPI, protein–protein interaction;
   GSEA, gene set enrichment analysis.

   We performed the KEGG pathway analysis to explore the higher levels of
   biological functions associated with asthma. This approach focuses on
   how a collection of genes in the genome is linked to gene products
   (enzymes) to create pathways that are distinct in any particular
   organism ([90]21, [91]22). According to the KEGG pathway analysis, the
   DEGs were primarily linked to the pancreatic secretion, arginine
   biosynthesis, renin-angiotensin system and arginine and proline
   metabolism ([92]Figure 2D). Nitric oxide is a vasodilator with
   anti-inflammatory and bronchodilatory properties. Since arginine
   stimulates the synthesis of nitric oxide, it is possible that changes
   in the arginine metabolome contribute to the pathophysiology of asthma
   ([93]23). KEGG enrichment analysis further emphasized that the DEGs
   were deeply intertwined with the pathogenesis of asthma.

   Based on the findings of the KEGG enrichment study, we annotated KEGG
   using the GSEA approach. The results highlighted significant enrichment
   of amino sugar and nucleotide sugar metabolism, glycan biosynthesis,
   parkinsons disease, protein export, steroid hormone biosynthesis. There
   was significant enrichment of protein export and amino sugar and
   nucleotide sugar metabolism. Thus, protein, amino sugar and nucleotide
   sugar are crucial for the pathophysiology of asthma ([94]Figure 2E).

3.3. PPI network analysis of asthma related DEGs

   To create a PPI network, thirteen primary DEGs associated with asthma
   were imported, examined, and visualized using SVG in the STING
   database. The total gene scores are shown as the number and colour of
   lines connecting the nodes. Fifteen proteins combined with CPA3 and 19
   proteins combined with TPSAB1, reflecting the core connectivity of CPA3
   and TPSAB1 (Tryptase alpha/beta-1) in these proteins ([95]Figure 2F).
   Among these, POSTN, CLCA1 (Calcium-activated chloride channel regulator
   1), CPA3, and TPSAB2 (Tryptase alpha/beta-2) existed some co-expression
   possibility.

3.4. Validation of the external dataset's main genes

   Our findings from the [96]GSE63142 dataset showed 13 genes differing in
   asthma. We further used a one-to-one format for comparison and
   validated with the [97]GSE63142 dataset The genes CEACAM5, PRR4, CPA3,
   POSTN, TCN1, CST1 (Cystatin-SN), CLCA1, TPSAB1 and NOS2 (Nitric oxide
   synthase 2) were highly expressed in patients with asthma
   ([98]Figure 3A). LOC100132287, LYPD2, SCGB3A1 and C7orf26 (Chromosome 7
   open reading frame 26) had low expression in asthma patients
   ([99]Figure 3B). The validity of these gene transcription patterns as
   diagnostic markers for asthma was supported by the consistency of the
   genome data. Further research is required to fully elucidate the
   underlying molecular pathways. LOC100132287 gene did not present in
   [100]GSE158752 dataset. PRR4, TCN1, CST1, CLCA1, and NOS2 were also
   highly expressed in patients with asthma in [101]GSE158752 dataset
   ([102]Supplementary Figure S1A). LYPD2 and C7orf26 were also lowly
   expressed in patients with asthma in [103]GSE158752 dataset
   ([104]Supplementary Figure S1B).

Figure 3.

   [105]Figure 3
   [106]Open in a new tab

   Verification of the varying expression of diagnostic genes. The
   differential expression of the [107]GSE63142 dataset. (A) The genes
   have high levels of expression in patients with asthma. (B) The genes
   have low levels of expression in patients with asthma.

3.5. Analysis of diagnostic indicators using ROC curves

   Using the AUC of the ROC curve, we verified the performance of the 13
   asthma-associated genes in the training set and [108]GSE63142 dataset.
   We created an ROC curve drawing of the diagnostic markers in RStudio to
   determine their diagnostic utility. The AUC ranged from 70.8% to 80.4%
   in the training set ([109]Figure 4). The AUC values showed that the
   [110]GSE63142 dataset performed satisfactorily overall, which suggests
   that these thirteen genes contributed significantly to the diagnostic
   utility of disease classification. Greater accuracy in differentiating
   between healthy individuals and asthma sufferers was indicated by the
   higher AUC values. The AUC ranged from 54% to 83% in the test set.
   CLCA1 reached an AUC value of 0.83 and its predictive efficacy was even
   higher than that of the training set ([111]Supplementary Figure S2).

Figure 4.

   [112]Figure 4
   [113]Open in a new tab

   The AUC of diagnostic genes of the [114]GSE63142 dataset. We used the
   ROC curve and calculated AUC to assess the performance of these models.
   The AUC ranged from 70.8% to 80.4%. ROC, receiver operating
   characteristic; AUC, area under the receiver operating characteristic
   curve.

3.6. Differentiation of immune characteristics and the immune cell
correlation analysis

   The proportion of immune cells in each sample was obtained by immune
   cell infiltration analysis. plasma cells, T cell CD4 naive, T cell
   gamma delta, monocytes, macrophages M0, activated dendritic cells and
   Neutrophils were differential between healthy controls and asthma
   patients ([115]Figures 5A,B). The findings of our investigation into
   the relationship between DEGs and immune cells are displayed in
   lollipop charts ([116]Figure 6). The genes CEACAM5, PRR4, CPA3, POSTN,
   LYPD2, CLCA1, and CST1 were significantly associated with plasma cells
   and resting mast cells. The genes LOC100132287, TCN1, and C7orf26 were
   significantly associated with naive B cells. The genes SCGB3A1 were
   significantly associated with plasma cells, neutrophils and active
   memory CD4 T cells. The genes CEACAM5, PRR4, CPA3, POSTN, CLCA1, and
   CST1 had positive correlations with plasma cells and resting mast cells
   (p < 0.05), whereas PRR4 and CPA3 had negatively correlated with
   neutrophils and active mast cells (p < 0.05). The genes CPA3, POSTN,
   CLCA1, CST1, and TPSAB1 were significantly associated with regulatory T
   cells Tregs (p < 0.05).

Figure 5.

   [117]Figure 5
   [118]Open in a new tab

   Immune cell infiltration analysis between asthma and healthy subjects.
   (A) The proportion of immune cells in each sample. (B) Differential
   infiltration of immune cells between healthy controls and asthma
   patients. Plasma cells, T cell CD4 naive, T cell gamma delta,
   monocytes, macrophages M0, activated dendritic cells and neutrophils
   have significant differences between two groups.

Figure 6.

   [119]Figure 6
   [120]Open in a new tab

   Immune cell correlation analysis of diagnostic genes. The genes
   CEACAM5, PRR4, CPA3, POSTN, LYPD2, CLCA1, and CST1 were significantly
   associated with plasma cells and resting mast cells (p < 0.05).

3.7. Genetic validation in an asthmatic mouse model

   To determine gene expression levels, we extracted RNA from mouse lung
   tissues and performed high-throughput eukaryotic sequencing analysis.
   These four murine homologous genes, cpa3, postn, lypd2, and scgb3a1,
   did not differ significantly between the asthma group and the normal
   group. However, these findings indicated that the mRNA expression of
   nos2 and clca1 in the asthma group was higher than that in the normal
   group, which was in line with the transcriptome observations made in
   asthmatic patients in [121]Figures 1A,B ([122]Figure 7).

Figure 7.

   [123]Figure 7
   [124]Open in a new tab

   Animal experiments and sequencing analysis. (A) Schematic of the OVA
   asthma model construction; (B) Gene expression levels in mouse lung
   tissue. Control: normal mice; Asthma: asthma model mice. OVA,
   ovalbumin. Transcript analysis of the gene mRNA expression levels of
   cpa3, lypd2, scgb3a1, nos2, and clca1 in murine lung tissues, showing
   the gene nos2 and clca1 expression trends consistent with the
   transcriptomic data in patients with asthma.

4. Discussion

   Asthma, an prevailing pulmonary malaise, afflicts many individuals
   across the world ([125]24). Wheezing, shortness of breath, coughing,
   and tightness in the chest are some of the symptoms that are caused by
   inflammation and constriction of the airways ([126]25). Currently, we
   know several pathogenetic mechanisms contribute to the development and
   progression of asthma: inflammation, airway hyperresponsiveness, airway
   remodelling, immunological factors, genetic predisposition,
   environmental factors, and neural control ([127]26). Many
   pharmaceutical strategies and self-management methods have been
   developed. However, some patients with asthma cannot control their
   symptoms with available treatment and management strategies, which
   highlights the unmet medical needs of these patients, an incomplete
   understanding of asthma pathogenesis, and the need for continued
   exploration of the topic, such as by targeting the specific genes that
   play cardinal roles in asthma evolution and progression. In this
   context, machine learning and other advanced techniques can be used to
   identify undiscovered genes critical to asthma and further explore
   asthma pathogenesis.

   To address knowledge gaps physiology and pathology of asthma, we used
   two machine learning models to determine the main genes responsible for
   the initiation and progression of asthma. Preliminary exploration
   revealed marked disparities in gene expression between patients with
   asthma patients vs. that in healthy controls. We utilized KEGG, GO,
   GSEA, and PPI network analyses to further confirm the correlation of
   differential gene expression and asthma. Merging the LASSO regression
   with the SVM-RFE method enabled a model that not only showed a
   consensus in identifying key asthma genes but also compensated for
   performance deficiencies of the individual models in terms of accuracy
   and prediction. This synthetic approach identified eight pivotal
   asthma-related genes spanning the training set, test cohorts, and
   experimental animal specimens.

   We identified 8 key genes (LOC100132287, CEACAM5, PRR4, CPA3, POSTN,
   LYPD2, TCN1, and SCGB3A1) associated with asthma by combining the LASSO
   regression model and the SVM-RFE method. Higher AUC values indicated a
   high diagnostic value. The association of the five genes CEACAM5,
   POSTN, TCN1, SCGB3A1, and CPA3 with asthma has been extensively
   identified and validated, and these are the most highly upregulated
   genes in patients with asthma ([128]27–[129]31), and CEACAM5 is
   associated with resting mast cells and eosinophils ([130]32). The genes
   POSTN, TCN1, and CPA3 are associated with the type 2 inflammatory
   response ([131]26, [132]33, [133]34). The gene SCGB3A1 is highly
   expressed in sputum columnar cells in patients with severe asthma and
   associated with non-neutrophilic airway inflammation ([134]35). This is
   consistent with our findings.

   Current reports indicate that PRR4 may impact the efficiency of the
   submucosal glands, leading to pathological changes in the respiratory
   tract ([135]36). The role of the gene PRR4 in asthma require further
   verification. To date, no study has demonstrated an association between
   asthma and LYPD2. Some studies have shown that LYPD2 is predicted to be
   a GPI-anchored Ly6 protein, enriches in non-classical monocytes
   ([136]37, [137]38). However, the gene LYPD2 had high AUC values,
   indicating that it has good predictive efficacy, which requires further
   exploration of the relationship with asthma.

   The current study had several limitations. The foundation of our
   investigation was the computational analysis of gene expression
   samples, which provided estimated results and reflected our reliance on
   computational data. Not all core genes were validated in asthmatic
   mice, partly because some genes are only significantly expressed in
   subjects with severe asthma, relatively small sample size of the mouse
   experiments and not all genes are homologous in humans and mice. In
   future studies, we will continue to examine the association between the
   newly discovered diagnostic genes and asthma. Meanwhile, we are
   continuing to confirm the efficiency of these genes in the diagnosis of
   asthma in clinical applications.

   In conclusion, our study validated several key genes potentially
   associated with asthma (CEACAM5, PRR4, CPA3, POSTN, TCN1, and SCGB3A1)
   and identified new asthma genetic marker, such as LYPD2. We propose
   that the combination of the detection of these genes and patient
   symptoms can lead to the prediction and diagnosis of asthma. We hope
   that by detecting the proteins expressed by these genes, we can
   determine the therapeutic effect of asthma and achieve clinical
   management.

Acknowledgments