An interpretable fecal microbiota-based GBM-SHAP machine-learning model for distinguishing ankylosing spondylitis from healthy controls

Abstract

Previous studies have identified associations between ankylosing spondylitis (AS) and the gut microbiota. This retrospective secondary analysis aimed to develop and interpret a fecal microbiota-based machine-learning model for distinguishing AS samples from healthy-control samples using publicly available shotgun metagenomic datasets. Paired-end fecal shotgun metagenomic sequencing data were obtained from PRJNA375935/SRP100575 and PRJEB28545/ERP110757. PRJNA375935/SRP100575 served as the model-development cohort and was divided into training and internal test sets, whereas PRJEB28545/ERP110757 was reserved as an independent external-validation cohort. Reads underwent quality control, host-read removal, taxonomic classification with Kraken2, and species-level abundance estimation with Bracken. Candidate features were screened using a nominal P-value threshold and selected by LASSO logistic regression within the training set. Hyperparameters for six machine-learning models were optimized using five-fold ROC-AUC-based GridSearchCV within the training set. After hyperparameters were fixed, ten-fold cross-validation was used to summarize training-set performance. SHAP values were used to quantify microbial-feature contributions to predictions generated by the final gradient boosting machine (GBM) model. A total of 361 samples were included: 211 from PRJNA375935/SRP100575 (97 AS and 114 healthy controls) and 150 from PRJEB28545/ERP110757 (113 AS and 37 healthy controls). The model-development cohort comprised a training set of 158 samples (69 AS and 89 healthy controls) and an internal test set of 53 samples (28 AS and 25 healthy controls). Alpha-diversity ( P = 0.024) and beta-diversity (PERMANOVA P = 0.001) analyses indicated differences in fecal microbial community structure between AS and healthy-control samples. Exploratory univariate screening identified 1,418 species-level candidate features meeting the nominal P < 0.05 criterion, and LASSO retained 82 microbial features. The GBM model showed the highest observed discrimination, with AUCs of 0.980 in the internal test set and 0.810 in the independent external-validation cohort. A GBM model based on public fecal metagenomic profiles distinguished AS samples from healthy-control samples with encouraging discrimination in an independent external-validation cohort. These findings support further investigation of fecal microbiota-based classification for AS; however, prospective clinical validation, standardized sample processing, and more complete clinical metadata are required before clinical application.