Machine learning models using gut microbiota signatures demonstrated excellent diagnostic performance for bipolar disorder, with a combined compositional and functional random forest model achieving AUC=0.9499 and AUPR=0.9586.
Key Findings
Results
Gut microbial α-diversity was significantly reduced in bipolar disorder patients compared to healthy controls.
16S rRNA sequencing was used to assess microbial diversity and composition
Both α-diversity (within-sample diversity) and β-diversity (between-sample diversity) were altered in BD compared to healthy controls (HC)
Microbial co-occurrence network connectivity was also weakened in BD patients
Results
Among 12 benchmarked classification algorithms, ensemble-based models—particularly the random forest (RF) classifier—achieved the best diagnostic performance for distinguishing BD from HC.
Twelve classification algorithms were systematically benchmarked to discriminate BD from HC
Ensemble-based models outperformed other algorithm types
The random forest classifier was identified as the top-performing individual model
Results
A set of 35 optimal microbial biomarkers identified via RF feature importance ranking demonstrated excellent classification performance for BD diagnosis.
Three feature selection methods were compared: RF feature importance ranking, independent t-tests, and MaAsLin2 analysis
This 35-feature set achieved AUC = 0.9316 and AUPR = 0.9497
Results
Combining functional pathway features with selected microbial biomarkers in an RF model further improved diagnostic performance beyond using microbial composition alone.
PICRUSt2 functional prediction using KEGG and MetaCyc annotations was applied to derive functional features
The combined compositional and functional RF model achieved AUC = 0.9499 and AUPR = 0.9586
This represented an improvement over the composition-only model (AUC = 0.9316, AUPR = 0.9497)
Results
Functional pathway analysis revealed marked alterations in pathways related to neurodegeneration, lipid metabolism, and heme biosynthesis in BD gut microbiota.
Functional prediction was performed using PICRUSt2 with KEGG and MetaCyc annotations
Altered pathways included those related to neurodegeneration, lipid metabolism, and heme biosynthesis
These findings were interpreted as providing mechanistic insights into the microbiota-gut-brain axis
Li H, Jin Y, Ye D, Liu Q, Su X, Zhang H, et al.. (2026). Machine Learning-based Diagnostic Potential of Bipolar Disorder Using Gut Microbiota Signatures.. IET systems biology. https://doi.org/10.1049/syb2.70056