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

ATOMIC achieved superior predictive performance for atopic dermatitis classification compared to baseline models on the KNUH dataset.

• ATOMIC achieved an AUROC of 0.810 and an AUPRC of 0.927 on the KNUH dataset.
• The model was trained and tested on 99 gut microbiome samples from adult patients with AD and healthy controls collected at Kangwon National University Hospital (KNUH).
• ATOMIC outperformed all baseline models tested in the study.

ATOMIC incorporates microbial co-expression networks with microbial genomic information as node features to capture functionally relevant microbial patterns.

• Microbial co-expression networks incorporate microbial genomic information as a node feature.
• This approach enhances the model's ability to capture functionally relevant microbial patterns.
• The graph attention network architecture allows the model to capture complex microbial interactions that simpler models fail to represent.

ATOMIC identified specific microbial taxa as candidate biomarkers potentially associated with atopic dermatitis prediction.

• Key microbial taxa contributing to AD classification were identified through the model's interpretable attention mechanism.
• The identified candidate microbial biomarkers may inform future therapeutic strategies.
• The attention mechanism provided interpretability by highlighting which microbial features were most influential in the classification.

A novel gut microbial abundance dataset from adult AD patients and healthy controls was collected and publicly released to facilitate future research.

• 99 gut microbiome samples were collected from adult patients with AD and healthy controls at Kangwon National University Hospital (KNUH).
• The dataset was publicly released along with the source code via the ATOMIC GitHub repository at https://www.github.com/KU-MedAI/ATOMIC.
• The study population consisted of adult patients, distinguishing it from pediatric-focused prior studies.

Prior machine learning and deep learning models for gut microbiome-disease prediction had limitations including failure to capture complex microbial interactions, lack of microbial genomic information, and limited interpretability.

• Existing computational models focus on diseases other than AD.
• These models often fail to capture complex microbial interactions or incorporate microbial genomic information.
• Prior models offered limited interpretability, motivating the development of ATOMIC's attention-based approach.
• ATOMIC was designed specifically to address these gaps in the context of atopic dermatitis.

ATOMIC's interpretable attention mechanism provides a foundation for personalized microbiome-based interventions and biomarker discovery in atopic dermatitis.

• The attention mechanism identifies key microbial taxa contributing to AD classification.
• The findings may inform microbiome-targeted therapeutic strategies such as probiotics and fecal microbiota transplantation.
• The model supports personalized microbiome-based interventions through its interpretable outputs.