Chronic respiratory diseases like COPD, asthma, and lung cancer pose significant global health challenges. COPD is a leading cause of death, while asthma causes substantial disability. Chronic inflammation in the respiratory tract can increase lung cancer risk. Accurate prediction methods are crucial for early intervention and disease burden reduction. The human microbiome, particularly its functional components, plays a critical role in health and disease. Microbial extracellular vesicles (EVs) are nanosized vesicles that transport microbial components throughout the body, influencing health and disease. Their metagenomic analysis reveals differential proportions between disease states and healthy controls. Circulating microbial EVs, accessible via serum, offer a non-invasive diagnostic target for systemic microbiome activity. This research aimed to develop predictive models for COPD, asthma, and lung cancer using AI modeling of serum microbial EV metagenomic data from respiratory patients. The study also tested these models in mice to identify dietary supplements that might reduce high-fat diet (HFD)-associated risks of these respiratory diseases. The study hypothesized that serum microbial EV profiles could serve as accurate and non-invasive biomarkers for these diseases.

While the microbiome's influence on health is increasingly recognized, the relationship between circulating microbial EVs and respiratory diseases remained largely unexplored before this study. Previous research has highlighted the roles of microbial EVs in intercellular communication and transport of various microbial components. Studies have shown both protective and harmful effects of microbial EVs on health, with some inducing pulmonary inflammation while others exhibiting antitumor responses. Metagenomic analysis of bacterial EVs has indicated differences in their composition between diseased and healthy individuals. However, a comprehensive, systemic approach using serum microbial EVs for predicting asthma, COPD, and lung cancer was lacking.

This study enrolled 1825 participants: COPD (n=93), asthma (n=454), lung cancer (n=283), and healthy controls (n=995) from five South Korean medical centers (2017-2020). Ethical approvals and informed consent were obtained. Serum samples were collected and EVs were extracted using a DNeasy Blood & Tissue Kit (QIAGEN). Microbial genomic DNA was extracted, and the V3-V4 hypervariable regions of the 16S rDNA gene were amplified and sequenced using Illumina MiSeq. Taxonomic assignment was performed using MDx-Pro ver. 2 (MD Healthcare, Korea), clustering operational taxonomic units (OTUs) with a 97% similarity threshold against the Silva 132 database. A novel taxonomic hierarchical accumulation method was employed to code the data, weighting imprecisely classified genera based on their higher-level taxonomy and reducing zero-inflation. This involved accumulating taxonomic values from genus to kingdom levels using a specific formula. Five machine-learning (ML) methods were used to develop prediction models: generalized linear model (GLM) with and without feature selection, gradient boosting machine (GBM), artificial neural network (ANN), and an ANN/GBM ensemble. Feature importance was determined using permutation feature importance analysis. Cross-validation was implemented with a 7:3 train-test split for 30 iterations. An in vivo study using 180 female C57BL/6 mice assessed the effects of various dietary supplements on HFD-associated respiratory disease risk. Mice were assigned to 36 dietary groups: NCD, HFD, and HFD plus various supplements (added to drinking water at a 2% ratio). After 4 weeks, serum was collected, and the developed ML models were applied to predict disease risk. Alpha diversity was assessed using phyloseq in R, while beta diversity was assessed using PCA, MDS, and t-SNE.

After data filtering, 1513 serum microbial EV taxa were used as features. The ANN/GBM ensemble model showed the best performance across all three diseases, with mean AUC values of 0.93 (COPD), 0.99 (asthma), and 0.94 (lung cancer). The model's performance was consistent across 30 test iterations, with low standard deviations. Feature importance analysis revealed different key taxa for each disease. Firmicutes were most important in COPD, while Proteobacteria were dominant in asthma and lung cancer. Specific genera, such as *Megamonas* (COPD), *Fimbriimonas* (asthma), and *Helicobacter* and *Curvibacter* (lung cancer) showed high importance. The in vivo study showed that HFD increased asthma and lung cancer risk in mice, with minimal impact on COPD. Several dietary supplements, including mungbean powder, Kakadu plum powder, and lotus root powder, reduced HFD-associated asthma and lung cancer risk. COPD risk was less affected by dietary interventions.

This study provides the first systemic analysis of serum microbial EV metagenomes for predicting asthma, COPD, and lung cancer. The high AUC values of the developed models demonstrate the potential of serum microbial EVs as diagnostic biomarkers. The novel taxonomic hierarchical coding method addressed the challenges of zero-inflation and imprecise taxonomic assignments in microbiome data. The findings highlight the distinct microbial signatures associated with each respiratory disease, suggesting that the serum EV microbiome reflects the systemic impact of these conditions. The in vivo study offers preliminary support for the diagnostic models, showing that dietary interventions can modify HFD-induced respiratory disease risk. However, further research is needed to confirm the specific roles of identified genera in disease pathogenesis and to validate the efficacy of dietary supplements in humans.

Serum microbial EVs show promise as non-invasive biomarkers for diagnosing asthma, COPD, and lung cancer. The high-performing machine-learning models developed in this study offer a novel diagnostic approach. Future research should focus on validating these findings in larger, more diverse cohorts and exploring the mechanistic links between specific microbial taxa and respiratory diseases. Further investigation of the potential of dietary interventions in mitigating disease risk is also warranted.

While the study used a large sample size, the participants were primarily from a single geographic location and may not fully represent the global population. The study's cross-sectional design does not allow for causal inferences. Further validation in independent cohorts with detailed clinical information is needed to confirm the robustness and generalizability of the findings. The mouse model, while useful for initial assessment, may not fully recapitulate human disease complexities.