Integrating respiratory microbiome and host immune response through machine learning for respiratory tract infection diagnosisOriginal paper
Karen Pendergrass is a microbiome researcher specializing in microbiome-targeted interventions (MBTIs). She systematically analyzes scientific literature to identify microbial patterns, develop hypotheses, and validate interventions. As the founder of the Microbiome Signatures Database, she bridges microbiome research with clinical practice. In 2012, based on her own investigative research, she became the first documented case of FMT for Celiac Disease, four years before the first published case study.
Read MoreWhat was studied?
This study built a diagnostic tool for lower respiratory tract infections (LRTIs), which conventional microbiology identifies in only 30-40 percent of cases. Using metatranscriptome sequencing of bronchoalveolar lavage fluid, researchers profiled both the respiratory microbiome and host gene expression. They then trained machine-learning models on combined microbial and host features to separate LRTIs from non-infectious illness.
Who was studied?
The study enrolled 136 patients with suspected LRTIs at Peking University People's Hospital, China, between 2020 and 2021, all providing bronchoalveolar lavage fluid. Of these, 68 were used for model building (41 LRTIs, 27 non-LRTIs), split into training and validation sets. A separate 68 patients (45 LRTIs, 23 non-LRTIs) served as an independent external validation cohort.
What were the most important findings?
Microbial diversity dropped in LRTIs, with commensals falling and opportunistic pathogens rising. Klebsiella pneumoniae showed the most significant increase. Host analysis found 649 differentially expressed genes, 613 upregulated, enriched in immune and infection pathways. Klebsiella pneumoniae correlated strongly with host genes TNFRSF1B, CSF3R, and IL6R. A Random Forest model using 12 combined features reached ROC AUC 0.937 (95 percent CI 0.832-1), beating single-source models. External validation accuracy was 76.5 percent.
What are the greatest implications of this study?
The results suggest that pairing the respiratory microbiome with the host immune response improves diagnosis of respiratory infection beyond pathogen detection alone. Such a model could support clinicians facing ambiguous cases and reduce unnecessary antibiotic use. The sample was relatively small, external specificity was low (39.1 percent), and biological functions were not experimentally validated, so the tool supplements rather than replaces clinical judgment.