ATOMIC: a graph attention network for atopic dermatitis prediction using human gut microbiome

Introduction: Atopic dermatitis (AD) is a chronic inflammatory skin disease driven by complex interactions among genetic, environmental, and microbial factors; however, its etiology remains unclear. Recent studies have reported the role of gut microbiota dysbiosis in AD pathogenesis, leading to increased interest in microbiome-targeted therapeutic strategies such as probiotics and fecal microbiota transplantation. Building on these findings, recent advances in computational modeling have introduced machine learning and deep learning-based approaches to capture the nonlinear relationships between gut microbiota and diseases. However, these models focus on diseases other than AD and often fail to capture complex microbial interactions or incorporate microbial genomic information, thereby offering limited interpretability.

Methods: To address these limitations, we propose ATOMIC, an interpretable graph attention network-based model that incorporates microbial co-expression networks to predict AD. Microbial co-expression networks incorporate microbial genomic information as a node feature, thereby enhancing their ability to capture functionally relevant microbial patterns. To train and test our model, we collected and processed 99 gut microbiome samples from adult patients with AD and healthy controls at Kangwon National University Hospital (KNUH).

Results: ATOMIC outperformed baseline models, achieving an AUROC of 0.810 and an AUPRC of 0.927 for KNUH dataset. Furthermore, ATOMIC identified microbes potentially associated with AD prediction and proposed candidate microbial biomarkers that may inform future therapeutic strategies.

Discussion: By identifying key microbial taxa that contributed to the AD classification through its interpretable attention mechanism, ATOMIC provides a foundation for personalized microbiome-based interventions and biomarker discovery. Additionally, to facilitate future research, we publicly released a gut microbial abundance dataset from KNUH. The source code and processed abundance data are available from ATOMIC GitHub repository at https://www.github.com/KU-MedAI/ATOMIC.

1 Introduction

Atopic dermatitis (AD) is a chronic inflammatory skin disease characterized by intense itching and eczema (1). Affecting over 120 million individuals globally, its prevalence continues to rise (2). Given its chronic and relapsing nature, AD significantly impacts quality of life and is often accompanied by psychological symptoms such as stress and sleep disturbances (3, 4). Multiple factors, including immune-related genetic and environmental factors, skin barrier dysfunction, and microbiota imbalance are known to contribute to its pathogenesis (5–8). However, the precise etiology of AD remains unclear (9). Current treatments, such as dupilumab and topical corticosteroids, can alleviate symptoms but rarely achieve complete remission (10), highlighting the need for a deeper understanding of the disease mechanism.

Among these factors, increasing attention has been directed toward the role of gut microbiota in the pathogenesis of AD (11). The gut microbiota, comprising approximately 10–100 trillion microorganisms, plays a critical role in the maturation of the immune system (12). Through colonization resistance, a competitive process among microbes for nutrients and space, the microbiota maintains homeostasis and suppresses pathogenic organisms (13). Notably, a reduction in the risk of developing AD has been observed in infants with high levels of short-chain fatty acids (SCFA), such as butyrate and propionate (14). Given that SCFAs support epithelial barrier integrity and modulate cytokine production and immune responses (15), their depletion may weaken colonization resistance and contribute to the development of AD.

With growing evidence linking the gut microbiota to AD, interventions such as probiotics and fecal microbiota transplantation (FMT) are being actively explored as novel therapeutic strategies (16). Probiotics, defined as live microorganisms that provide health benefits upon ingestion (17), have been shown to alleviate AD symptoms by restoring gut microbiota balance (18). FMT, which involves the transfer of fecal microbiota from healthy donors to patients (19), has the potential to suppress AD-related allergic responses and improve immune regulation (16, 20). However, the effectiveness of these microbiome-based interventions largely depends on the accurate identification of disease-associated microbial taxa (21). Traditional abundance-based statistical analyses are limited in capturing complex interactions within the microbiome and potentially overlook critical disease-associated signals. To overcome this limitation, machine learning and deep learning-based models are increasingly being applied to learn complex patterns from high-dimensional microbiome data and uncover novel biomarkers relevant to disease prediction and treatment (22).

For example, Pasolli et al. (23) proposed MetAML, which applies random forests (RFs) and support vector machines (SVMs) to predict diseases using 2,424 publicly available microbiome samples. However, it only uses the abundance of the microbiome and does not incorporate microbial genomic information, such as DNA sequences, limiting its biological interpretability. Oh et al. (24) developed DeepMicro, an autoencoder-based deep learning framework that transforms microbial abundance data into low-dimensional representations, followed by classification using RF, SVM, and multilayer perceptron models for disease prediction. While effective in generating low-dimensional representations, this approach may result in information loss and lacks an end-to-end architecture. Liao et al. (25) introduced GDmicro, which constructs graphs in which each node represents an individual sample, and latent features are derived for each node through domain adaptation (DA) (26). These graphs are then used to predict diseases using graph convolutional networks (GCNs) (27). Although the DA enhances generalization across different cohorts, applying the model to new data requires graph reconstruction and retraining. Despite their methodological differences, existing models share common limitations: they often fail to incorporate microbial relationships and lack genomic information, which hinders both prediction performance and biological interpretability. The conventional machine learning methods, such as RF and MLP, treat microbial taxa as independent features and therefore ignore the complex interaction networks inherent in the microbiome. To model these interactions, graph-based approaches have been introduced; however, classical graph models such as GCNs aggregate information from neighboring nodes without considering the relative importance of each neighbor. As a more advanced alternative approach, graph attention networks (GATs) (28) utilize an attention mechanism to dynamically assign importance weights to neighboring nodes, allowing the model to aggregate information reflecting the differential importance between nodes. Such attention-based architectures have shown promise in capturing complex biological relationships and have demonstrated superior performance in chromatin interaction analysis and cell classification for single-cell Hi-C data (29, 30). However, these methods have not yet been applied to the microbiome in AD research. Moreover, overcoming these limitations is complicated by the scarcity of publicly available gut microbiome data. According to a recent study, only 220,017 human gut microbiome datasets are publicly available from the National Center for Biotechnology Information (NCBI) (31), and this number is likely to be substantially smaller when restricted to AD.

To address these limitations, we propose ATOMIC, a graph attention network for ATOpic dermatitis prediction using human gut MICrobiome. ATOMIC is an interpretable deep learning model that incorporates microbial co-expression networks and microbial genomic information to predict AD. Unlike previous models that rely solely on abundance data, ATOMIC integrates the relationships between microbes and their genomic information into a graph-based architecture. By leveraging GATs, the model learns representations that capture microbial relationships and identifies the key microbial contributors to AD prediction. Furthermore, it supports interpretability by highlighting microbe-level importance through attention scores, enabling the discovery of candidate biomarkers relevant to disease prediction and treatment. For development and evaluation, we applied ATOMIC to a new gut microbiome dataset collected from a cohort of adult patients with AD and publicly released this dataset.

2 Materials and methods

2.1 Overview of ATOMIC

An overview of ATOMIC is illustrated in Figure 1. We first constructed a microbial co-expression network, where each node represents a microbe and each edge represents a co-expression relationship. Microbes with zero-count abundance values were removed from each sample to generate sample-specific graphs, which led to variations in graph sizes across samples. The Graph Attention Network v2 (GATv2) (32) layers update the node representations by incorporating information from their neighboring microbes through attention weights. Finally, the graph embedding is obtained through a self-attention readout, and AD prediction is performed through fully connected layers.

Figure 1

Figure 1. Overview of ATOMIC. (A) The microbial co-expression network construction process of ATOMIC. Each node represents a microbe with abundance and genomic vectors as features, while each edge represents a microbial co-expression. The microbial co-expression network is fed into GATv2 layers. (B) The neural network architecture of ATOMIC. Three stacked GATv2 layers with multi-head attention are followed by a self-attention readout to obtain the graph embedding, and fully connected layers use the graph embedding to predict AD. Abbreviations: GATv2, Graph Attention Network v2; COAT, composition-adjusted thresholding; AD, atopic dermatitis.

2.2 Atopic dermatitis sample collection and data preprocessing

We collected gut microbiome data from 99 adult participants recruited through the Department of Dermatology at Kangwon National University Hospital (KNUH) in Chuncheon, Korea. The cohort included 70 patients diagnosed with AD, aged 18–69 years, based on the Hanifin and Rajka criteria, and 29 healthy controls without chronic inflammatory or autoimmune diseases. Participants were excluded if they had used systemic immunosuppressants or steroids, had a history of inflammatory or autoimmune diseases, or were unable to visit within ±1 week of the scheduled date. The study protocol was approved by the Institutional Review Board of KNUH (No. KNUH-2023-08-011-002) and written informed consent was obtained from all participants. Stool samples were collected in sterile containers and stored at –80 °C within 4 hours of collection. All collected samples were stored in ultra-low temperature freezers (–80 °C or below) or liquid nitrogen tanks (–130 °C to –196 °C) in the human biobank. After completion of the analyses, the remaining samples were disposed of according to the disposal protocols of the analysis institution.

To augment the sample size and improve the generalizability of the microbial co-expression network construction, we additionally collected 1,392 samples from patients with AD and healthy controls available from the NCBI database. These publicly available datasets served as external resources for training more robust models by enhancing microbial association inferences. These additional samples were derived from five independent studies conducted in Korea (33), Hong Kong (34), China (35, 36), and Japan (37). An overview of all the collected samples is summarized in Table 1.

Table 1

Table 1. Detailed information about the collected data used in this study.

All sequencing data, including those from KNUH and public repositories, were sequenced using Illumina paired-end sequencing platform targeting 16S rRNA amplicon sequencing (V3-V4) and they were processed using a unified bioinformatics pipeline. The primer used were 314F (CCTACGGGNGGCWGCAG) and 806R (GACTACHVGGGTATCTAATCC). The expected size of the PCR product is 460 ± 10 bp. We performed microbial taxonomic classification and abundance estimation based on the SILVA 138.1 reference database (38) after preprocessing all sequences using QIIME2. However, owing to the low proportion of sequences matching at the species level, we aggregated the abundance of species to obtain genus-level microbial abundance data. Genus-level abundances were then normalized within each sample to a sum of 100.

2.3 Microbial co-expression network construction

We constructed a global microbial co-expression network as an undirected graph $G=(V,E)$, where the nodes represent microbes and the edges indicate statistically significant co-abundance relationships. To define the edges, we computed correlation coefficients between microbial taxa using the composition-adjusted thresholding (COAT) method (39), which infers correlations based on the log ratios between pairs of compositional variables. To construct a robust network, we randomly sampled 80% of the dataset five times and computed COAT correlations in each iteration. Only microbial correlations consistently observed across all five iterations were retained as edges in the final network. The microbial co-expression network was constructed using both the KNUH and public datasets, while model training and evaluation were performed exclusively on the KNUH cohort.

We represent the set of $n$ microbial nodes as $V={\left\{(v_i, s_i)\right\}}_{i=1}^n$ where each node $i$ is characterized by a 64-dimensional abundance vector $v_i\in {ℝ}^{64}$ and a 768-dimensional genomic vector $s_i\in {ℝ}^{768}$. The abundance vector $v_i$ for microbe $i$ was obtained by multiplying the scalar abundance of microbe $i$ by a 64-dimensional learnable vector that was randomly initialized from a uniform distribution. The genomic vector $s_i$ was derived from DNABERT-2 (40), a genome foundation model. Specifically, for each genus-level taxon, we encoded the DNA sequences of all available constituent species and then computed the average of these vectors. Thus, each microbial node $i$ has an initial representation $h_i^{(0)}\in R^{832}$, obtained by concatenating its abundance vector $v_i$ and genomic vector $s_i$. The connections between these nodes are defined by the edge set $E=\left\{e_{ij}|\left|c_{ij}\right|\ge 0.1, i,j=1,2,\dots ,n, i\ne j\right\}$, where $e_{ij}$ represents a co-expression edge between microbial nodes ($v_i$, $s_i$) and ($v_j$, $s_j$), and $c_{ij}$ is their corresponding COAT correlation coefficient.

Although the topology of the co-expression graph (i.e., edge set E) was fixed across all samples, the set of active nodes varied depending on the sample-specific microbial composition. For each sample, we constructed a subgraph by removing nodes corresponding to microbes with zero abundance. As a result, each sample had a distinct graph size, reflecting its individual microbial profile. This strategy enables sample-specific graph representation while preserving a consistent co-expression backbone, facilitating efficient learning of microbial interactions in a biologically meaningful and scalable manner.

2.4 Graph neural network for learning microbial co-expression relationships.

After constructing the microbial co-expression networks, we applied GATv2 to model the relationships among microbes and to predict AD. GATv2 is an improved architecture over the original GAT that addresses the limitation of static attention, where certain key nodes consistently receive high attention weights regardless of the query node. This static ranking reduces the capacity of the model to capture context-dependent interactions. By contrast, GATv2 introduces dynamic attention, allowing attention weights to vary depending on the query node. This enables a more expressive and flexible modeling of complex relational structures in the graph, which is particularly important for capturing sample-specific microbial interactions.

In each GATv2 layer $l$, the importance of a neighboring node $j$ to a target node $i$ is computed using the attention coefficients. GATv2 employs a multi-head attention mechanism in which each attention head k independently learns distinct attention patterns. These attention coefficients are then normalized across all neighbors $j\in N_i$ using the softmax function (41), formally defined as Equation 1:

$\begin{array}{ll}\begin{array}{c}{\alpha }_{ij}^{(l,k)}=\frac{exp\left(a^{(l,k)T}·LeakyReLU\left(W^{(l,k)}·\left[h_i^{(l-1)}\left|\right|h_j^{(l-1)}\right]\right)\right)}{{\sum }_{j^{\prime }\in N_i}exp\left(a^{(l,k)T}·LeakyReLU\left(W^{(l,k)}·\left[h_i^{(l-1)}\left|\right|h_{j^{\prime }}^{(l-1)}\right]\right)\right)}\end{array}& (1)\end{array}$

where $a^{(l,k)}$ denotes the attention vector, $W^{(l,k)}$ denotes the linear transformation matrix at layer $l$ for the $k$-th attention head, and $\left|\right|$ denotes the vector concatenation.

The updated representation of node $i$ in layer $l$ is computed by aggregating its neighbor’s features weighted by the attention coefficients across all $K$ heads (Equation 2):

$\begin{array}{ll}\begin{array}{c}{h}_i^{(l)}=|{|}_{k=1}^K\sum _{j\in N_i}{\alpha }_{ij}^{(l,k)}·W^{(l,k)}{h}_j^{(l-1)}\end{array}& (2)\end{array}$

where $h_i^{(l)}$ denotes the updated representation of node $i$ in layer $l$, $K$ is the number of attention heads; and $W^{(l,k)}$ is the linear transformation matrix corresponding $k$-th attention head in layer $l$.

In the final GATv2 layer $L$, the node embeddings from all the attention heads are averaged (Equation 3):

$\begin{array}{ll}\begin{array}{c}{h}_i^L=\frac{1}{K}\sum _{k=1}^K\sum _{j\in N_i}{\alpha }_{ij}^{(L,k)}·W^{(L,k)}{h}_j^{(L-1)}\end{array}& (3)\end{array}$

The resulting node embeddings $h^L=[h^L{}_1,h^L{}_2,\dots ,h^L{}_n]$, $h^L{}_i\in {ℝ}^F$, where $F$ denotes the dimensionality of the updated node embeddings, are used to compute the graph embedding through an attention-based readout. Each node embedding $h^L{}_i$ is projected to a scalar importance score, ${\widehat{h}}_i$, using a learnable linear transformation matrix $W_p\in {ℝ}^{1\times F}$ (Equation 4):

$\begin{array}{ll}\begin{array}{c}{\widehat{h}}_i=W_p{h}^L{}_i\end{array}& (4)\end{array}$

To incorporate the attention-based node importance into the graph embedding for the prediction of AD, we calculated the attention score$A_i$ for each node i by applying the softmax function as Equation 5:

$\begin{array}{ll}\begin{array}{c}{A}_i=\frac{\exp ({\widehat{h}}_i)}{{\sum }_{j=1}^n\exp ({\widehat{h}}_j)}\end{array}& (5)\end{array}$

The final graph embedding $P$ is then computed as the weighted sum of the node embeddings using the attention score $A_i$. Finally, the graph embedding $P$ is fed into a fully connected network with two hidden layers to predict the AD. In our implementation, we used three stacked GATv2 layers (L = 3), each configured with eight attention heads (k = 8). The output dimensions of each node representation were set to 32 (F = 32). To improve generalization, we applied edge dropout (p = 0.3), Mish activation (42), node dropout (p = 0.3) (43).

2.5 Model implementation

Our dataset of 99 samples was split into a training set of 59 samples, a validation set of 20 samples, and a test set of 20 samples, corresponding to an approximate 60:20:20 ratio. We implemented ATOMIC architecture using PyTorch (44) and optimized the model using the AdamW optimizer (45). The initial learning rate was set to 0.0001, with a batch size of 16, and the learning rate was scheduled to decay by 1% every 10 epochs. Hyperparameters, including the number of layers, number of attention heads, dropout rates, and learning rate, were optimized based on a grid search using a 4-fold cross-validation strategy to minimize validation loss. We trained our model on a computing machine equipped with an Intel Xeon Gold 6230 CPU, 512 GB of memory, and an NVIDIA A100 GPU.

3 Results

3.1 Performance on the KNUH dataset

We evaluated the performance of ATOMIC on the KNUH dataset by comparing it with several baseline models, including deep learning-based methods such as GDmicro, DeepMicro, GCN, and GraphSAGE (46), as well as traditional machine learning models such as MetAML, RFs, and SVMs (Table 2). GDmicro was originally designed with DA to enhance cross-cohort generalization. However, because our model focused on optimizing performance within a single AD cohort, we disabled the DA component of GDmicro to ensure a fair and direct comparison within the same cohort setting.

Table 2

Table 2. Performance comparison of baseline models on the KNUH dataset in terms of AUROC, AUPRC, and F1 score.

As a result, ATOMIC outperformed most baseline models, achieving an area under the receiver operating characteristic curve (AUROC) of 0.752 $\pm 0.069$ (95% CI: 0.666-0.838), area under the precision-recall curve (AUPRC) of 0.894 $\pm 0.027$(95% CI: 0.860-0.928) and an F1 score of 0.784 $\pm 0.108$(95% CI: 0.650-0.918) across five independent runs with different random seeds. Although some performance variability was observed, likely owing to the limited training sample size, the model demonstrated a strong overall predictive capability. By averaging the output probabilities from models trained in five independent runs with different random initializations, the ensemble achieved more stable and accurate predictions. Specifically, the ensemble-based ATOMIC boosted the AUROC by 7.71%, AUPRC by 3.69%, and F1 score by 7.91% compared with the non-ensemble version. Compared with GDmicro without DA, our ensemble-based ATOMIC achieved competitive AUROC performance and showed substantial gains in AUPRC (17.6%) and F1 score (12.8%), indicating greater robustness and discriminative power. Notably, GDmicro without DA exhibits a higher AUPRC variance, whereas our model maintains a more stable performance across runs. Statistical significance was assessed using the Mann–Whitney U test. ATOMIC with the ensemble achieved significantly higher AUROC scores than the RFs (p = 0.004), SVMs (p = 0.002), MetAML (p = 0.004), GCN (p = 0.004), GraphSAGE (p = 0.004), and DeepMicro (p = 0.004), confirming the robustness of our approach for identifying true positives. Some baseline models showed relatively lower AUROC performance, possibly because of their limited capacity to capture nonlinear inter-microbial dependencies and a lack of microbial genomic feature integration. In contrast, ATOMIC benefits from leveraging microbial co-expression patterns, enabling more effective modeling of complex microbial interactions, and ultimately leading to more accurate and stable predictions of AD.

We visualized sample representations using both graph embeddings derived from the self-attention readout and microbial abundance data to examine the clustering patterns between the AD and healthy control samples. Figure 2 shows the t-SNE plots for the KNUH dataset, where panel A corresponds to graph embeddings and panel B corresponds to microbial abundance data. The graph-based representations revealed more distinct cluster boundaries between the AD and healthy control groups compared to the abundance-based features. To quantitatively assess group separability, we calculated silhouette scores for both representations. The graph embeddings yielded a significantly higher silhouette score (0.566, p = 0.001) than the microbial abundance data (0.055, p = 0.119), suggesting superior class separation. These results demonstrate that the graph-based attention mechanism in ATOMIC captures latent structural patterns in the microbiome that are not apparent from abundance alone, thereby contributing to its enhanced classification performance.

Figure 2

Figure 2. t-SNE visualization of the KNUH dataset (orange: AD; blue: healthy control). The graph embeddings and microbial abundance data were compressed using t-SNE into 2D space. The x-axis represents the t-SNE feature 1, and the y-axis represents the t-SNE feature 2. The ellipses represent covariance-based clusters for each class, and each point indicates a sample. (A) Graph embeddings derived from ATOMIC. (B) Genus-level microbial abundance data. Abbreviations: KNUH, Kangwon National University Hospital; AD, atopic dermatitis.

3.2 Ablation study

To assess the contribution of individual components within ATOMIC, we conducted ablation experiments by removing the microbial genomic information and altering the structure of the co-expression network. As shown in Figure 3, excluding genomic information from node representations led to a significant decline in model AUROC performance (p = 0.045, Mann–Whitney U test), highlighting the importance of integrating genomic context alongside abundance data for microbial feature characterization. Furthermore, when the genomic information was removed and the edges of the co-expression graph were randomly shuffled for each sample, the AUROC performance declined even further (p = 0.028). This indicates that the structured microbial relationships encoded in the co-expression network play a key role in enhancing the predictive accuracy. Collectively, these results confirm that both microbial genomic information and biologically meaningful co-expression structures are essential for robust prediction of AD using ATOMIC.

Figure 3

Figure 3. Ablation study evaluating the impact of microbial genomic information and co-expression graph structure on ATOMIC performance. All bars represent the mean (standard deviation) across five independent runs. The Mann–Whitney U test was performed for significance testing, with absolute rank-biserial correlation (|r|) used to indicate effect size. Abbreviations: AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; w/o, without; ns, not significant; *p-value< 0.05.

3.3 Model interpretation with attention scores

A key advantage of ATOMIC is its interpretability, which is enabled by the attention scores derived from the self-attention readout layer. In the graph representation, each node corresponds to a microbial genus, and the attention score $A_i$ for node $i$ quantifies its relative importance in the model’s prediction. Since the attention scores across all nodes in a sample sum to 1, they allow for within-sample comparisons of microbial relevance. Note that these scores are non-directional; a high score indicates predictive importance but does not imply whether the microbe is enriched or depleted in that sample.

To identify microbes most associated with AD predictions, we aggregated attention scores across samples correctly classified as having AD in the test set. As each sample contained a different subset of microbes, a direct comparison of raw attention scores across samples was inappropriate. To address this issue, we applied a centered log-ratio (CLR) transformation, which is a common approach for compositional data normalization. CLR transforms each attention score by dividing it by the geometric mean of all scores in the same sample, enabling inter-sample comparisons. After normalization, we calculated and ranked the average CLR-transformed attention scores for each microbe. Table 3 lists the top 10 genera with the highest CLR-transformed attention scores, along with the supporting literature, all of which have been previously implicated in AD. Because high attention scores reflect predictive contributions, both protective and risk-associated taxa can be prioritized if they provide informative patterns for AD classification.

Table 3

Table 3. Literature-supported summary of the top 10 microbial genera ranked by CLR-transformed attention scores from the trained ATOMIC model.

The highest-ranked genus was Ruminococcus_gnavus_group (CLR score = 4.877), a well-known producer of SCFAs (47). SCFAs promote the differentiation of regulatory T cells (Tregs) (48), support intestinal and epidermal barrier integrity, and exert broad anti-inflammatory effects (49). Other highly ranked genera, including Butyricicoccus, Coprococcus, Lachnoclostridium, and Eubacterium_hallii_group—are also recognized producers of SCFAs, especially butyrate (50–53). Notably, reduced SCFA levels have been consistently observed in children with AD (54), further supporting their protective role. Additionally, Dorea (family Lachnospiraceae), Flavonifractor (family Oscillospiraceae), and Rothia (family Micrococcaceae) have been reported to decrease or negatively correlate with AD severity (55, 56).

In contrast, ATOMIC identified genera associated with an increased risk of AD. Blautia and Streptococcus both received high attention scores and have been shown to be more abundant in individuals with AD than in healthy controls (57, 58). Notably, Streptococcus has been linked to persistent forms of AD rather than transient presentations (58), suggesting a role in chronic inflammation. These findings demonstrate that ATOMIC is not only effective in predicting AD, but also in suggesting biologically relevant microbial patterns. Its attention-based mechanism provides interpretable outputs that align with known disease mechanisms, offering valuable insights into microbial contributions to AD pathogenesis.

3.4 Biomarker discovery from microbe-mediated genes

We further analyzed the potential functions of microbes that with high attention scores captured by ATOMIC. To accomplish this, we conducted microbe function analysis using Microbe-Set Enrichment Analysis (MSEA) (59). We utilized a publicly available microbe-set library linking human genes to genus-level microbes and applied a custom background consisting of microbial genera detected in the KNUH dataset. To analyze only the microbe sets that primarily contributed to AD prediction, we selected microbes with average CLR-normalized attention scores greater than zero, calculated from AD test samples, as input for MSEA. The resulting enriched genes with an adjusted p-value below 0.05 are shown in Figure 4.

Figure 4

Figure 4. Bar plots illustrating enrichment analysis results. The x-axis shows the enriched genes, pathways, and ontologies, and the y-axis shows the log10 adjusted p-value. (A) Top 10 enriched microbe-mediated genes identified by Microbe-Set Enrichment Analysis (MSEA). (B) Top 10 enriched pathway terms from the KEGG 2021 Human database. (C) Top 10 enriched Gene Ontology terms from the GO Biological Process 2025. (D) Top 10 enriched disease ontology terms from GeDiPNet 2023. Abbreviations: KEGG, Kyoto Encyclopedia of Genes and Genomes; GO, Gene Ontology.

Figure 4A shows the top 10 enriched human genes functionally linked to microbes predictive of AD. Many of these genes are key regulators of T helper 2 (Th2)-mediated inflammation, which is a hallmark of AD pathogenesis. For example, tumor necrosis factor (TNF) encodes TNF-α, a key cytokine elevated in the skin of AD patients (60). TNF-α promotes the expression of Th2 cytokines and thymic stromal lymphopoietin (TSLP) (61), both of which aggravate Th2-mediated immune responses by activating dendritic and mast cells (62).

We also identified innate immune receptors such as Toll-like receptor (TLR)-2 and TLR4 among the top-ranked genes. TLR2 promotes Th2 polarization through interleukin (IL)-5 and high-affinity Fc receptor for immunoglobulin E (FcϵRI) upregulation, while TLR4 broadly modulates inflammatory signaling (63, 64). Polymorphisms in both TLR2 and TLR4 have been associated with increased susceptibility to AD (65). These receptors are functionally linked to the microbial genera identified by ATOMIC, such as Butyricicoccus, Coprococcus, and Ruminococcus_gnavus, which produce SCFAs and modulate host immunity through TLR4 activation. SCFAs are known to suppress inflammatory cytokines including TNF-α, IL-6, and IL-8, yet SCFA level are reduced in patients with AD (66). Moreover, Ruminococcus_gnavus secretes glucorhamnan, which stimulates dendritic cells to produce TNF-α through the TLR4 signaling pathway (67), directly linking this microbe to inflammation-relevant host responses.

Several other inflammation-related genes have been identified in addition to canonical immune mediators. Lymphocyte antigen 96 (LY96), which encodes the myeloid differentiation (MD)-2 protein, functions as an essential coreceptor in the TLR4 signaling pathway and has been proposed as a target for genetic modulation in AD (68). TNF receptor-associated factor 6 (TRAF6) is another key immune regulator previously implicated in IL-23/IL-17-mediated responses in psoriasis-like dermatitis (69), and may play a role in AD.

Genes associated with metabolic dysfunction in AD were also identified in AD. Lipoprotein lipase (LPL) is particularly notable, given the alterations in lipid metabolism observed in patients with AD; LPL agonists are currently under investigation as potential therapeutic agents (70–72). Similarly, serpin family E member 1 (SERPINE1) has been reported to be significantly upregulated in AD lesions and may contribute to dermatitis-associated pruritus (73).

Importantly, our analysis identified several genes that have not been previously associated with AD, highlighting potential novel biomarker candidates. For example, von Willebrand factor (VWF), although not directly linked to AD, is known to activate macrophages and stimulate the production of pro-inflammatory mediators such as TNF, IL-6, chemokine ligand (CCL)2, and CCL3 (74). Similarly, fucosyltransferase 2 (FUT2) and serum vimentin (VIM) have been implicated in other chronic inflammatory conditions, such as inflammatory bowel disease, but their roles in AD remain uncharacterized (75, 76). Nevertheless, given the critical role of inflammatory responses in AD pathogenesis (77), these genes warrant further investigation as potential biomarkers or therapeutic targets.

3.5 Function and disease annotation of microbe-mediated genes

To investigate the biological functions of microbe-mediated genes identified by ATOMIC, we conducted a secondary enrichment analysis using Enrichr (78). Specifically, we selected significantly enriched gene sets from the prior MSEA results (adjusted p< 0.05) and analyzed them across pathways, gene ontology, and disease annotation databases. Pathway enrichment analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) 2021 human database revealed several immune-related pathways that are implicated in AD pathogenesis (Figure 4B). Notably, the TLR signaling pathway (rank 1) was highly enriched. TLRs are pattern recognition receptors essential for microbial sensing and innate immune activation (79). TLR agonists have been shown to suppress Th2-mediated responses in cluster of differentiation (CD)4 $+$T cells, offering potential therapeutic benefits in AD (80). The IL-17 signaling pathway (rank 3) also appeared, consistent with its known role in exacerbating AD through the suppression of filaggrin (FLG) expression, which is vital for epidermal barrier integrity (81). Cytokine–cytokine receptor interactions (rank 5) and chemokine signaling pathways (rank 7) were similarly enriched. These pathways include cytokines, such as IL-4, IL-13, IL-31, and TSLP, and chemokines, such as CCL1, CCL13, and CCL17, which have been implicated in barrier disruption, pruritus, and immune cell recruitment in AD (82, 83).

Gene Ontology analysis (GO Biological Process 2025; Figure 4C) further confirmed the enrichment of the immune and inflammatory processes. Terms such as “inflammatory response” (GO:0006954), “positive regulation of cytokine production” (GO:0001819), and “cytokine-mediated signaling pathway” (GO:0019221) were highly ranked, aligning with both literature-based expectations and KEGG pathway results. Disease ontology enrichment using GeDiPNet 2023 (Figure 4D) also highlighted “dermatitis” as the most significantly associated disease term, reinforcing the AD relevance of the gene sets prioritized by ATOMIC. Collectively, these results demonstrate that ATOMIC prioritizes predictive microbial genera and recovers host gene signatures and immune pathways that are highly relevant to AD pathophysiology, supporting its potential utility for both diagnostic and mechanistic insights.

3.6 Microbial composition analysis

To explore the therapeutic implications of ATOMIC’s microbe-level attention scores, we designed an in silico intervention experiment to simulate the effect of removing specific microbial taxa. This analysis aimed to evaluate whether the microbes identified as influential by ATOMIC contributed meaningfully to AD predictions and could serve as potential targets for microbiome-based treatment strategies in AD.

We stratified microbial genera into two groups based on their average CLR-normalized attention scores in the AD-classified test samples: Group A included taxa with positive attention scores, representing microbes that positively contributed to AD predictions, whereas Group B included taxa with negative attention scores, presumed to have minimal influence. To simulate therapeutic removal, we randomly eliminated 30%, 50%, and 70% of the microbes from each group in the AD test samples and measured the resulting changes in the predicted AD probabilities. If Group A microbes are truly predictive and causally linked to disease severity, their removal should reduce the confidence of the model for AD classification. Conversely, the removal of microbes from Group B was expected to have little to no effect. Each removal scenario is repeated 100 times to ensure statistical robustness. The results are summarized in Table 4.

Table 4

Table 4. Effects of alterations in microbial composition on the prediction of AD probability in ATOMIC.

Removal of Group A taxa led to a progressive decrease in the predicted AD probabilities, with reductions ranging from 0.073 to 0.129, depending on the removal fraction. These decreases were statistically significant at 50% (p = 0.015) and 70% (p = 0.011) removal levels, as determined by the Mann–Whitney U test. In contrast, the removal of Group B taxa had minimal impact across all sampling rates.

These findings support the hypothesis that ATOMIC effectively identifies microbial taxa that substantially influence AD prediction. This in silico intervention framework provides a useful approach for prioritizing microbes for further investigation of their functional roles in AD. While this does not imply that removing these microbes would directly confer therapeutic benefits, it suggests that high-attention taxa may represent ecologically or immunologically important members of the AD-associated microbiota.

4 Discussion

This study demonstrated that ATOMIC outperforms baseline models for the prediction of AD and provides interpretability. Its superior performance in classifying AD results from the incorporation of microbial co-expression networks and microbial genomic information, enabling ATOMIC to capture both structural and functional relationships within the microbiome. ATOMIC assigns microbe-specific attention scores, thereby capturing the relative importance of each microbe. Additionally, to promote broader research on the relationship between atopic dermatitis and the gut microbiome, we made the processed abundance data from this study publicly available to the research community.

Through a literature-based analysis of high-attention microbes, we identified associations between known AD-related mechanisms, including SCFA production and host immune modulation. Furthermore, enrichment analyses of microbe-mediated genes revealed significant involvement in inflammation-related pathways, such as TLR signaling, IL-17 signaling, and cytokine–cytokine receptor interactions. Interestingly, genes, such as VWF, FUT2, and VIM, which are not directly linked to AD, have been implicated in broader inflammatory conditions, suggesting novel avenues for future research.

Beyond prediction, ATOMIC provides a framework for individualized microbiome interpretation. Given the known inter-individual variability in the gut microbial composition, patients with similar clinical presentations may respond differently to treatment. ATOMIC addresses this challenge by identifying the most influential microbes in each individual, thereby supporting personalized diagnostic and therapeutic strategies. Its attention mechanism highlights taxa most predictive for each individual, offering interpretable insights that can inform clinical decision-making. While conventional probiotics and FMT aim to restore the overall microbial balance, their broad-spectrum and non-specific nature limits their utility for disease-specific interventions (84). These findings highlight the need for next-generation probiotics (NGPs), which involve targeted modulation of disease-associated taxa (84, 85). In this context, ATOMIC’s attention scores offer a data-driven approach to prioritize candidate NGP strains for AD, enabling precision targeting on a per-patient basis, bridging the gap between microbiome analysis and personalized medicine in AD.

Despite these contributions, several limitations should be acknowledged. First, we used genus-level rather than species-level data, potentially obscuring species-specific effects. Second, while our training data consisted solely of adult samples, some supporting literature referenced pediatric studies, introducing possible age-related confounders (86). Nevertheless, specific treatment guidelines for adult-onset AD have not yet been established; consequently, therapeutic approaches developed for pediatric AD are commonly applied in adults (87). The findings of this study may serve as a valuable reference for future studies aimed at developing targeted therapeutic agents for adult-onset AD. Third, the KNUH dataset is relatively small (n = 99), which may have limited the generalizability of our findings. Although cross-validation was employed to reduce overfitting, external validation was essential.

To address this, we applied ATOMIC to an external pediatric AD dataset from Korea (88). This dataset was obtained from the European Nucleotide Archive (accession No. PRJEB41351) and comprised 346 samples, including 234 patients with AD and 112 healthy controls. On this dataset, the ATOMIC ensemble achieved an AUPRC of 0.886 and an F1-score of 0.762, surpassing GDmicro-AUPRC (0.773) and F1 (0.733), although GDmicro yielded a higher AUROC of 0.867 compared to 0.735 for ATOMIC. This result was consistent with the pattern observed in the KNUH dataset. Although the AUROC was relatively low, ATOMIC maintained a higher AUPRC and F1 score, suggesting better precision in identifying true AD cases, which may be advantageous for clinical applications that prioritize specificity. Interestingly, the performance of ATOMIC decreased on the pediatric dataset, despite the larger sample size. This performance gap is likely due to age-dependent differences in the gut microbiome composition (89), which in turn limits the applicability of adult-based microbial co-expression networks to pediatric data.

Future studies will focus on improving the generalizability of the model by incorporating diverse microbiome datasets across age groups, geographic regions, and disease subtypes. In addition, we plan to enhance the scalability of ATOMIC by integrating techniques such as deep domain adaptation and knowledge distillation (90), enabling effective transfer to large-scale and heterogeneous datasets.

5 Conclusion

In this study, we presented ATOMIC, an interpretable graph attention network-based model that integrates microbial co-expression networks with genomic information to predict AD using gut microbiome data. By combining genus-level abundance profiles and DNABERT-derived genomic features as graph node attributes, ATOMIC effectively captures both structural and functional relationships among microbes. Trained on microbiome data from adult patients with AD and healthy controls, ATOMIC outperformed the existing baseline models in predictive performance. Importantly, the attention scores derived from the model enabled the identification of key microbial taxa that contributed to the AD classification. Functional enrichment analysis of microbe-mediated genes revealed host immune pathways and inflammation-related mechanisms associated with these microbes. These findings highlight ATOMIC’s potential as a predictive model and tool for discovering candidate microbial biomarkers and therapeutic targets for AD. However, the small sample size and the use of single-cohort data in this study may limit the model’s generalization performance. Therefore, further validation studies by collecting larger, multi-cohort data are necessary to ensure the reliability and generalizability of the model. To support continued research, we publicly released the processed gut microbiome data. ATOMIC offers a foundation for future efforts toward personalized microbiome-based interventions and biomarker discovery for AD.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: KAP241559 (K-BDS; https://kbds.re.kr/KAP241559.

Ethics statement

The studies involving humans were approved by Institutional Review Board of KNUH (No. KNUH-2023-08-011-002). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

Author contributions

HB: Conceptualization, Methodology, Validation, Visualization, Writing – original draft, Writing – review & editing. JM: Data curation, Resources, Writing – review & editing. SK: Data curation, Investigation, Writing – original draft, Writing – review & editing. WL: Validation, Writing – original draft, Writing – review & editing. DL: Conceptualization, Resources, Writing – review & editing. HE: Conceptualization, Resources, Writing – review & editing. YH: Data curation, Resources, Writing – original draft, Writing – review & editing. MJ: Conceptualization, Methodology, Resources, Supervision, Writing – original draft, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This study was supported by the Korean Ministry of Science and ICT (MSIT) (S0301-25-1001) to DL; MSIT under the Technology Development Program of the Korean Ministry of SMEs and Startups (S3364091) to DL and MJ; the Research Grant from Institute of Medical Sciences, Kangwon National University 2024 (KNUH_2024-02-04) to YH; the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health & Welfare, Republic of Korea (RS-2025-24535069) to YH; the ICT Challenge and Advanced Network of HRD program supervised by the Institute of Information & Communications Technology Planning & Evaluation (IITP) (IITP-2025-RS2022-00156439) to MJ; the Bio&Medical Technology Development Program of the National Research Foundation funded by the MSIT (RS-2024-00441029) to MJ.

Conflict of interest

The authors declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Weidinger S, Beck LA, Bieber T, Kabashima K, and Irvine AD. Atopic dermatitis. Nat Rev Dis Primers. (2018) 4:1. doi: 10.1038/s41572-018-0001-z

PubMed Abstract | Crossref Full Text | Google Scholar

2. Tian J, Zhang D, Yang Y, Huang Y, Wang L, Yao X, et al. Global epidemiology of atopic dermatitis: A comprehensive systematic analysis and modelling study. Br J Dermatol. (2024) 190:55–61. doi: 10.1093/bjd/ljad339

PubMed Abstract | Crossref Full Text | Google Scholar

3. Matthewman J, Schultze A, Strongman H, Bhaskaran K, Roberts A, Denaxas S, et al. Cohort studies on 71 outcomes among people with atopic eczema in uk primary care data. Nat Commun. (2024) 15:9573. doi: 10.1038/s41467-024-54035-1

PubMed Abstract | Crossref Full Text | Google Scholar

4. Silverberg JI, Gelfand JM, Margolis DJ, Boguniewicz M, Fonacier L, Grayson MH, et al. Patient burden and quality of life in atopic dermatitis in us adults: A population-based cross-sectional study. Ann Allergy Asthma Immunol. (2018) 121:340–7. doi: 10.1016/j.anai.2018.07.006

PubMed Abstract | Crossref Full Text | Google Scholar

5. Leung DY and Guttman-Yassky E. Deciphering the complexities of atopic dermatitis: shifting paradigms in treatment approaches. J Allergy Clin Immunol. (2014) 134:769–79. doi: 10.1016/j.jaci.2014.08.008

PubMed Abstract | Crossref Full Text | Google Scholar

6. Flohr C and Mann J. New insights into the epidemiology of childhood atopic dermatitis. Allergy. (2014) 69:3–16. doi: 10.1111/all.12270

PubMed Abstract | Crossref Full Text | Google Scholar

7. Kim BE and Leung DY. Significance of skin barrier dysfunction in atopic dermatitis. Allergy Asthma Immunol Res. (2018) 10:207–15. doi: 10.4168/aair.2018.10.3.207

PubMed Abstract | Crossref Full Text | Google Scholar

8. Mohammad S, Karim MR, Iqbal S, Lee JH, Mathiyalagan R, Kim YJ, et al. Atopic dermatitis: pathophysiology, microbiota, and metabolome–a comprehensive review. Microbiol Res. (2024) 281:127595. doi: 10.1016/j.micres.2023.127595

PubMed Abstract | Crossref Full Text | Google Scholar

9. Brunner PM, Leung DY, and Guttman-Yassky E. Immunologic, microbial, and epithelial interactions in atopic dermatitis. Ann Allergy Asthma Immunol. (2018) 120:34–41. doi: 10.1016/j.anai.2017.09.055

PubMed Abstract | Crossref Full Text | Google Scholar

10. Facheris P, Jeffery J, Del Duca E, and Guttman-Yassky E. The translational revolution in atopic dermatitis: the paradigm shift from pathogenesis to treatment. Cell Mol Immunol. (2023) 20:448–74. doi: 10.1038/s41423-023-00992-4

PubMed Abstract | Crossref Full Text | Google Scholar

11. Craig JM. Atopic dermatitis and the intestinal microbiota in humans and dogs. Vet Med Sci. (2016) 2:95–105. doi: 10.1002/vms3.24

PubMed Abstract | Crossref Full Text | Google Scholar

12. Gill SR, Pop M, DeBoy RT, Eckburg PB, Turnbaugh PJ, Samuel BS, et al. Metagenomic analysis of the human distal gut microbiome. science. (2006) 312:1355–9. doi: 10.1126/science.1124234

PubMed Abstract | Crossref Full Text | Google Scholar

13. Deng L and Wang S. Colonization resistance: the role of gut microbiota in preventing salmonella invasion and infection. Gut Microbes. (2024) 16:2424914. doi: 10.1080/19490976.2024.2424914

PubMed Abstract | Crossref Full Text | Google Scholar

14. Roduit C, Frei R, Ferstl R, Loeliger S, Westermann P, Rhyner C, et al. High levels of butyrate and propionate in early life are associated with protection against atopy. Allergy. (2019) 74:799–809. doi: 10.1111/all.13660

PubMed Abstract | Crossref Full Text | Google Scholar

15. Maslowski KM and Mackay CR. Diet, gut microbiota and immune responses. Nat Immunol. (2011) 12:5–9. doi: 10.1038/ni0111-5

PubMed Abstract | Crossref Full Text | Google Scholar

16. Mashiah J, Karady T, Fliss-Isakov N, Sprecher E, Slodownik D, Artzi O, et al. Clinical efficacy of fecal microbial transplantation treatment in adults with moderate-to-severe atopic dermatitis. Immun Inflammation Dis. (2022) 10:e570. doi: 10.1002/iid3.570

PubMed Abstract | Crossref Full Text | Google Scholar

17. World Health Organization: Food and Agriculture Organization of the United Nations. Probiotics in food health and nutritional properties and guidelines for evaluation. FAO Food Nutr Pap. Rome (Italy): WHO/FAO (2006) 85, 1–50.

Google Scholar

18. Rios-Carlos M, Cervantes-García D, Córdova-Dávalos LE, Bermúdez-Humarán LG, and Salinas E. Unraveling the gut-skin axis in atopic dermatitis: exploiting insights for therapeutic strategies. Gut Microbes. (2024) 16:2430420. doi: 10.1080/19490976.2024.2430420

PubMed Abstract | Crossref Full Text | Google Scholar

19. Grehan MJ, Borody TJ, Leis SM, Campbell J, Mitchell H, and Wettstein A. Durable alteration of the colonic microbiota by the administration of donor fecal flora. J Clin Gastroenterol. (2010) 44:551–61. doi: 10.1097/MCG.0b013e3181e5d06b

PubMed Abstract | Crossref Full Text | Google Scholar

20. Kim J-H, Kim K, and Kim W. Gut microbiota restoration through fecal microbiota transplantation: A new atopic dermatitis therapy. Exp Mol Med. (2021) 53:907–16. doi: 10.1038/s12276-021-00627-6

PubMed Abstract | Crossref Full Text | Google Scholar

21. Fang X, Wang Y, Wei H, and Huang Y. Precision microbiome: A new era of targeted therapy with core probiotics. Research. (2025) 8:0658. doi: 10.34133/research.0658

PubMed Abstract | Crossref Full Text | Google Scholar

22. Curry KD, Nute MG, and Treangen TJ. It takes guts to learn: machine learning techniques for disease detection from the gut microbiome. Emerg Topics Life Sci. (2021) 5:815–27. doi: 10.1042/ETLS20210213

PubMed Abstract | Crossref Full Text | Google Scholar

23. Pasolli E, Truong DT, Malik F, Waldron L, and Segata N. Machine learning meta-analysis of large metagenomic datasets: tools and biological insights. PloS Comput Biol. (2016) 12:e1004977. doi: 10.1371/journal.pcbi.1004977

PubMed Abstract | Crossref Full Text | Google Scholar

24. Oh M and Zhang L. Deepmicro: deep representation learning for disease prediction based on microbiome data. Sci Rep. (2020) 10:6026. doi: 10.1038/s41598-020-63159-5

PubMed Abstract | Crossref Full Text | Google Scholar

25. Liao H, Shang J, and Sun Y. Gdmicro: classifying host disease status with gcn and deep adaptation network based on the human gut microbiome data. Bioinformatics. (2023) 39:btad747. doi: 10.1093/bioinformatics/btad747

PubMed Abstract | Crossref Full Text | Google Scholar

26. Long M, Cao Y, Wang J, and Jordan M. (2015). Learning transferable features with deep adaptation networks, in: International conference on machine learning, ICML (2015). Cambridge, MA, USA: PMLR.

Google Scholar

27. Jiang B, Zhang Z, Lin D, Tang J, and Luo B. Semi-supervised learning with graph learning-convolutional networks. Proc IEEE/CVF Conf Comput Vision Pattern Recogn. (2019). 11313–20. doi: 10.1109/CVPR.2019.01157

Crossref Full Text | Google Scholar

28. Veličković P, Cucurull G, Casanova A, Romero A, Lio P, and Bengio Y. Graph attention networks. arXiv preprint arXiv:171010903. (2017). doi: 10.48550/arXiv.1710.10903

Crossref Full Text | Google Scholar

29. Li L, Li X, and Wu H. A novel deep learning framework with dynamic tokenization for identifying chromatin interactions along with motif importance investigation. Briefings Bioinf. (2025) 26:bbaf289. doi: 10.1093/bib/bbaf289

PubMed Abstract | Crossref Full Text | Google Scholar

30. Zhou X and Wu H. Schiclassifier: A deep learning framework for cell type prediction by fusing multiple feature sets from single-cell hi-C data. Briefings Bioinf. (2024) 26:bbaf009. doi: 10.1093/bib/bbaf009

PubMed Abstract | Crossref Full Text | Google Scholar

31. Abdill RJ, Adamowicz EM, and Blekhman R. Public human microbiome data are dominated by highly developed countries. PloS Biol. (2022) 20:e3001536. doi: 10.1371/journal.pbio.3001536

PubMed Abstract | Crossref Full Text | Google Scholar

32. Brody S, Alon U, and Yahav E. How attentive are graph attention networks? arXiv preprint arXiv:210514491. (2021). doi: 10.48550/arXiv.2105.14491

Crossref Full Text | Google Scholar

33. Ko MM, Shin S, Kim MH, Kang M, Baek M-G, Yi H, et al. Multi-omics analysis of gwakhyangjeonggi-san for gastrointestinal complications in atopic dermatitis: A randomized, double-blinded, placebo-controlled, parallel-group clinical trial. J Ethnopharmacol. (2024) 319:117256. doi: 10.1016/j.jep.2023.117256

PubMed Abstract | Crossref Full Text | Google Scholar

34. Wang Y, Choy CT, Lin Y, Wang L, Hou J, Tsui JCC, et al. Effect of a novel E3 probiotics formula on the gut microbiome in atopic dermatitis patients: A pilot study. Biomedicines. (2022) 10:2904. doi: 10.3390/biomedicines10112904

PubMed Abstract | Crossref Full Text | Google Scholar

35. Wang Y, Hou J, Tsui JC-C, Wang L, Zhou J, Chan UK, et al. Unique gut microbiome signatures among adult patients with moderate to severe atopic dermatitis in southern chinese. Int J Mol Sci. (2023) 24:12856. doi: 10.3390/ijms241612856

PubMed Abstract | Crossref Full Text | Google Scholar

36. Bian G, Gloor GB, Gong A, Jia C, Zhang W, Hu J, et al. The gut microbiota of healthy aged chinese is similar to that of the healthy young. Msphere. (2017) 2:e00327–17. doi: 10.1128/msphere.00327–17

PubMed Abstract | Crossref Full Text | Google Scholar

37. Odamaki T, Kato K, Sugahara H, Hashikura N, Takahashi S, Xiao J-Z, et al. Age-related changes in gut microbiota composition from newborn to centenarian: A cross-sectional study. BMC Microbiol. (2016) 16:1–12. doi: 10.1186/s12866-016-0708-5

PubMed Abstract | Crossref Full Text | Google Scholar

38. Quast C, Pruesse E, Yilmaz P, Gerken J, Schweer T, Yarza P, et al. The silva ribosomal rna gene database project: improved data processing and web-based tools. Nucleic Acids Res. (2012) 41:D590–D6. doi: 10.1093/nar/gks1219

PubMed Abstract | Crossref Full Text | Google Scholar

39. Cao Y, Lin W, and Li H. Large covariance estimation for compositional data via composition-adjusted thresholding. J Am Stat Assoc. (2019) 114:759–72. doi: 10.1080/01621459.2018.1442340

Crossref Full Text | Google Scholar

40. Zhou Z, Ji Y, Li W, Dutta P, Davuluri R, and Liu H. Dnabert-2: efficient foundation model and benchmark for multi-species genome. arXiv preprint arXiv:230615006. (2023). doi: 10.48550/arXiv.2306.15006

Crossref Full Text | Google Scholar

41. Goodfellow I. Deep learning. Cambridge, MA, USA: MIT press (2016).

Google Scholar

42. Misra D. Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:190808681. (2019). doi: 10.48550/arXiv.1908.08681

Crossref Full Text | Google Scholar

43. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, and Salakhutdinov R. Dropout: A simple way to prevent neural networks from overfitting. J Mach Learn Res. (2014) 15:1929–58. doi: 10.5555/2627435.2670313

Crossref Full Text | Google Scholar

44. Paszke A. Pytorch: an imperative style, high-performance deep learning library. arXiv preprint arXiv:191201703. (2019). 10.48550/arXiv.1912.01703

Crossref Full Text | Google Scholar

45. Loshchilov I and Hutter F. Decoupled weight decay regularization. arXiv preprint arXiv:171105101. (2017). doi: 10.48550/arXiv.1711.05101

Crossref Full Text | Google Scholar

46. Hamilton W, Ying Z, and Leskovec J. Inductive representation learning on large graphs. Adv Neural Inf Process Syst. (2017) 30, 1024–34. doi: 10.48550/arXiv.1706.02216

Crossref Full Text | Google Scholar

47. Crost EH, Coletto E, Bell A, and Juge N. Ruminococcus gnavus: friend or foe for human health. FEMS Microbiol Rev. (2023) 47:fuad014. doi: 10.1093/femsre/fuad014

PubMed Abstract | Crossref Full Text | Google Scholar

48. Ahn JR, Lee SH, Kim B, Nam MH, Ahn YK, Park YM, et al. Ruminococcus gnavus ameliorates atopic dermatitis by enhancing treg cell and metabolites in balb/C mice. Pediatr Allergy Immunol. (2022) 33:e13678. doi: 10.1111/pai.13678

PubMed Abstract | Crossref Full Text | Google Scholar

49. Ou L-S, Goleva E, Hall C, and Leung DY. T regulatory cells in atopic dermatitis and subversion of their activity by superantigens. J Allergy Clin Immunol. (2004) 113:756–63. doi: 10.1016/j.jaci.2004.01.772

PubMed Abstract | Crossref Full Text | Google Scholar

50. Stiemsma LT, Arrieta M-C, Dimitriu PA, Cheng J, Thorson L, Lefebvre DL, et al. Shifts in lachnospira and clostridium sp. In the 3-month stool microbiome are associated with preschool age asthma. Clin Sci. (2016) 130:2199–207. doi: 10.1042/CS20160349

PubMed Abstract | Crossref Full Text | Google Scholar

51. Lee S-Y, Lee E, Park YM, and Hong S-J. Microbiome in the gut-skin axis in atopic dermatitis. Allergy Asthma Immunol Res. (2018) 10:354–62. doi: 10.4168/aair.2018.10.4.354

PubMed Abstract | Crossref Full Text | Google Scholar

52. Shrode RL, Knobbe JE, Cady N, Yadav M, Hoang J, Cherwin C, et al. Breast cancer patients from the midwest region of the United States have reduced levels of short-chain fatty acid-producing gut bacteria. Sci Rep. (2023) 13:526. doi: 10.1038/s41598-023-27436-3

PubMed Abstract | Crossref Full Text | Google Scholar

53. Shetty SA, Zuffa S, Bui TPN, Aalvink S, Smidt H, and De Vos WM. Reclassification of eubacterium hallii as anaerobutyricum hallii gen. Nov., comb. Nov., and description of anaerobutyricum soehngenii sp. Nov., a butyrate and propionate-producing bacterium from infant faeces. Int J System Evolut Microbiol. (2018) 68:3741–6. doi: 10.1099/ijsem.0.003041

PubMed Abstract | Crossref Full Text | Google Scholar

54. Trompette A, Pernot J, Perdijk O, Alqahtani RAA, Santo Domingo J, Camacho-Muñoz D, et al. Gut-derived short-chain fatty acids modulate skin barrier integrity by promoting keratinocyte metabolism and differentiation. Mucosal Immunol. (2022) 15:908–26. doi: 10.1038/s41385-022-00524-9

PubMed Abstract | Crossref Full Text | Google Scholar

55. Yang J, McDowell A, Seo H, Kim S, Min TK, Jee Y-K, et al. Diagnostic models for atopic dermatitis based on serum microbial extracellular vesicle metagenomic analysis: A pilot study. Allergy Asthma Immunol Res. (2020) 12:792. doi: 10.4168/aair.2020.12.5.792

PubMed Abstract | Crossref Full Text | Google Scholar

56. Li W and Li A. Exploring the causal relationship between gut microbiota and atopic dermatitis: A mendelian randomization study. Medicine. (2024) 103:e40193. doi: 10.1097/MD.0000000000040193

PubMed Abstract | Crossref Full Text | Google Scholar

57. Zhou Z, Yang J, Liu Q, Gao J, and Ji W. Patho-immunological mechanisms of atopic dermatitis: the role of the three major human microbiomes. Scand J Immunol. (2024) 100:e13403. doi: 10.1111/sji.13403

PubMed Abstract | Crossref Full Text | Google Scholar

58. Park YM, Lee S-Y, Kang M-J, Kim B-S, Lee M-J, Jung SS, et al. Imbalance of gut streptococcus, clostridium, and akkermansia determines the natural course of atopic dermatitis in infant. Allergy Asthma Immunol Res. (2019) 12:322–37. doi: 10.4168/aair.2020.12.2.322

PubMed Abstract | Crossref Full Text | Google Scholar

59. Kou Y, Xu X, Zhu Z, Dai L, and Tan Y. Microbe-set enrichment analysis facilitates functional interpretation of microbiome profiling data. Sci Rep. (2020) 10:21466. doi: 10.1038/s41598-020-78511-y

PubMed Abstract | Crossref Full Text | Google Scholar

60. Park HH, Lee S, Yu Y, Yoo SM, Baek SY, Jung N, et al. Tgf-β Secreted by human umbilical cord blood-derived mesenchymal stem cells ameliorates atopic dermatitis by inhibiting secretion of tnf-α and ige. Stem Cells. (2020) 38:904–16. doi: 10.1002/stem.3183

PubMed Abstract | Crossref Full Text | Google Scholar

61. Danso MO, Van Drongelen V, Mulder A, Van Esch J, Scott H, Van Smeden J, et al. Tnf-α and th2 cytokines induce atopic dermatitis–like features on epidermal differentiation proteins and stratum corneum lipids in human skin equivalents. J Invest Dermatol. (2014) 134:1941–50. doi: 10.1038/jid.2014.83

PubMed Abstract | Crossref Full Text | Google Scholar

62. Luo J, Zhu Z, Zhai Y, Zeng J, Li L, Wang D, et al. The role of tslp in atopic dermatitis: from pathogenetic molecule to therapeutical target. Mediators Inflammation. (2023) 2023:7697699. doi: 10.1155/2023/7697699

PubMed Abstract | Crossref Full Text | Google Scholar

63. Yu Y, Zhang Y, Zhang J, Dou X, Yang H, Shao Y, et al. Impaired toll-like receptor 2-mediated th1 and th17/22 cytokines secretion in human peripheral blood mononuclear cells from patients with atopic dermatitis. J Trans Med. (2015) 13:1–10. doi: 10.1186/s12967-015-0744-1

PubMed Abstract | Crossref Full Text | Google Scholar

64. Kim H-J, Kim H, Lee J-H, and Hwangbo C. Toll-like receptor 4 (Tlr4): new insight immune and aging. Immun Ageing. (2023) 20:67. doi: 10.1186/s12979-023-00383-3

PubMed Abstract | Crossref Full Text | Google Scholar

65. Zhang Y, Wang H-C, Feng C, and Yan M. Analysis of the association of polymorphisms rs5743708 in tlr2 and rs4986790 in tlr4 with atopic dermatitis risk. Immunol Invest. (2019) 48:169–80. doi: 10.1080/08820139.2018.1508228

PubMed Abstract | Crossref Full Text | Google Scholar

66. Xiao X, Hu X, Yao J, Cao W, Zou Z, Wang L, et al. The role of short-chain fatty acids in inflammatory skin diseases. Front Microbiol. (2023) 13:1083432. doi: 10.3389/fmicb.2022.1083432

PubMed Abstract | Crossref Full Text | Google Scholar

67. Henke MT, Kenny DJ, Cassilly CD, Vlamakis H, Xavier RJ, and Clardy J. Ruminococcus gnavus, a member of the human gut microbiome associated with crohn’s disease, produces an inflammatory polysaccharide. Proc Natl Acad Sci. (2019) 116:12672–7. doi: 10.1073/pnas.1904099116

PubMed Abstract | Crossref Full Text | Google Scholar

68. Liao E-C, Chang C-Y, Wu C-C, Wang G-J, and Tsai J-J. Association of single nucleotide polymorphisms in the md-2 gene promoter region with der P 2 allergy. Allergy Asthma Immunol Res. (2015) 7:249–55. doi: 10.4168/aair.2015.7.3.249

PubMed Abstract | Crossref Full Text | Google Scholar

69. Matsumoto R, Dainichi T, Tsuchiya S, Nomura T, Kitoh A, Hayden MS, et al. Epithelial traf6 drives il-17–mediated psoriatic inflammation. JCI Insight. (2018) 3:e121175. doi: 10.1172/jci.insight.121175

PubMed Abstract | Crossref Full Text | Google Scholar

70. Wu SA, Kersten S, and Qi L. Lipoprotein lipase and its regulators: an unfolding story. Trends Endocrinol Metab. (2021) 32:48–61. doi: 10.1016/j.tem.2020.11.005

PubMed Abstract | Crossref Full Text | Google Scholar

71. Chu H, Jung YW, Kim SM, Bae BG, Bae JM, Kim LK, et al. Exploring the interplay: obesity, lipid metabolism, and clinical manifestations in atopic dermatitis. Ann Dermatol. (2024) 36:310. doi: 10.5021/ad.23.151

PubMed Abstract | Crossref Full Text | Google Scholar

72. Xu G, Chu M, Shen S, Miao H, Bai Y, Liu X, et al. Causal effects of circulating lipids and lipid-lowering drugs on the risk of atopic dermatitis: A mendelian randomization study. Arch Dermatol Res. (2024) 317:154. doi: 10.1007/s00403-024-03635-4

PubMed Abstract | Crossref Full Text | Google Scholar

73. Larkin C, Chen W, Szabó IL, Shan C, Dajnoki Z, Szegedi A, et al. Novel insights into the trpv3-mediated itch in atopic dermatitis. J Allergy Clin Immunol. (2021) 147:1110–4.e5. doi: 10.1016/j.jaci.2020.09.028

PubMed Abstract | Crossref Full Text | Google Scholar

74. Drakeford C, Aguila S, Roche F, Hokamp K, Fazavana J, Cervantes MP, et al. Von willebrand factor links primary hemostasis to innate immunity. Nat Commun. (2022) 13:6320. doi: 10.1038/s41467-022-33796-7

PubMed Abstract | Crossref Full Text | Google Scholar

75. Hu M, Zhang X, Li J, Chen L, He X, and Sui T. Fucosyltransferase 2: A genetic risk factor for intestinal diseases. Front Microbiol. (2022) 13:940196. doi: 10.3389/fmicb.2022.940196

PubMed Abstract | Crossref Full Text | Google Scholar

76. Su L, Pan P, Yan P, Long Y, Zhou X, Wang X, et al. Role of vimentin in modulating immune cell apoptosis and inflammatory responses in sepsis. Sci Rep. (2019) 9:5747. doi: 10.1038/s41598-019-42287-7

PubMed Abstract | Crossref Full Text | Google Scholar

77. Bieber T. Atopic dermatitis: an expanding therapeutic pipeline for a complex disease. Nat Rev Drug Discov. (2022) 21:21–40. doi: 10.1038/s41573-021-00266-6

PubMed Abstract | Crossref Full Text | Google Scholar

78. Kuleshov MV, Jones MR, Rouillard AD, Fernandez NF, Duan Q, Wang Z, et al. Enrichr: A comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Res. (2016) 44:W90–W7. doi: 10.1093/nar/gkw377

PubMed Abstract | Crossref Full Text | Google Scholar

79. Medzhitov R. Toll-like receptors and innate immunity. Nat Rev Immunol. (2001) 1:135–45. doi: 10.1038/35100529

PubMed Abstract | Crossref Full Text | Google Scholar

80. Zuany-Amorim C, Hastewell J, and Walker C. Toll-like receptors as potential therapeutic targets for multiple diseases. Nat Rev Drug Discov. (2002) 1:797–807. doi: 10.1038/nrd914

PubMed Abstract | Crossref Full Text | Google Scholar

81. Kim J, Kim BE, and Leung DY. Pathophysiology of atopic dermatitis: clinical implications. Allergy Asthma Proc. (2019). 40:84–92. doi: 10.2500/aap.2019.40.4202

PubMed Abstract | Crossref Full Text | Google Scholar

82. Yamamura Y, Nakashima C, and Otsuka A. Interplay of cytokines in the pathophysiology of atopic dermatitis: insights from murin models and human. Front Med. (2024) 11:1342176. doi: 10.3389/fmed.2024.1342176

PubMed Abstract | Crossref Full Text | Google Scholar

83. Gros E, Bussmann C, Bieber T, Förster I, and Novak N. Expression of chemokines and chemokine receptors in lesional and nonlesional upper skin of patients with atopic dermatitis. J Allergy Clin Immunol. (2009) 124:753–60.e1. doi: 10.1016/j.jaci.2009.07.004

PubMed Abstract | Crossref Full Text | Google Scholar

84. Singh TP and Natraj BH. Next-generation probiotics: A promising approach towards designing personalized medicine. Crit Rev Microbiol. (2021) 47:479–98. doi: 10.1080/1040841X.2021.1902940

PubMed Abstract | Crossref Full Text | Google Scholar

85. Jan T, Negi R, Sharma B, Kumar S, Singh S, Rai AK, et al. Next generation probiotics for human health: an emerging perspective. Heliyon. (2024). 10:e35980. doi: 10.1016/j.heliyon.2024.e35980

PubMed Abstract | Crossref Full Text | Google Scholar

86. Ramírez-Marín HA and Silverberg JI. Differences between pediatric and adult atopic dermatitis. Pediatr Dermatol. (2022) 39:345–53. doi: 10.1111/pde.14971

PubMed Abstract | Crossref Full Text | Google Scholar

87. Vakharia PP and Silverberg JI. Adult-onset atopic dermatitis: characteristics and management. Am J Clin Dermatol. (2019) 20:771–9. doi: 10.1007/s40257-019-00453-7

PubMed Abstract | Crossref Full Text | Google Scholar

88. Lee M-J, Park YM, Kim B, Tae IH, Kim N-E, Pranata M, et al. Disordered development of gut microbiome interferes with the establishment of the gut ecosystem during early childhood with atopic dermatitis. Gut Microbes. (2022) 14:2068366. doi: 10.1080/19490976.2022.2068366

PubMed Abstract | Crossref Full Text | Google Scholar

89. Radjabzadeh D, Boer CG, Beth SA, van der Wal P, Kiefte-De Jong JC, Jansen MA, et al. Diversity, compositional and functional differences between gut microbiota of children and adults. Sci Rep. (2020) 10:1040. doi: 10.1038/s41598-020-57734-z

PubMed Abstract | Crossref Full Text | Google Scholar

90. Hinton G, Vinyals O, and Dean J. Distilling the knowledge in a neural network. arXiv preprint arXiv:150302531. (2015). doi: 10.48550/arXiv.1503.02531

Crossref Full Text | Google Scholar