Gut Microbiome Wellness Index 2 enhances health status prediction from gut microbiome taxonomic profiles

The gut microbiome's crucial role in health and disease is increasingly recognized. However, quantifying an individual's gut health remains a challenge. The Gut Microbiome Wellness Index (GMWI), a stool metagenome-based indicator, was previously developed to assess health by predicting the likelihood of disease from gut microbiome composition, regardless of specific disease type. While promising, GMWI had limitations: it showed higher accuracy in classifying healthy samples than non-healthy ones, assigned equal weight to all species, and only used species-level data. This paper introduces GMWI2, an improved version addressing these limitations to improve the accuracy and generalizability of health status prediction from gut microbiome data.

Landmark studies have established strong links between the gut microbiome and various chronic diseases. The original GMWI, derived from 4347 stool shotgun metagenomes, demonstrated a balanced accuracy of 69.7% in predicting disease presence. A validation cohort showed 73.7% accuracy. Despite its success, limitations in classification bias and weighting of species prompted the development of GMWI2.

This study pooled 8069 stool shotgun metagenomes from 54 published studies across diverse global demographics. Taxonomic profiling was performed using MetaPhlAn3. GMWI2 uses a Lasso-penalized logistic regression model, trained on taxonomic profiles spanning all ranks, to predict disease likelihood. The model learned variable feature importances, avoiding a prevalence-based strategy and allowing for different weights for each taxon. The model's performance was assessed through several validation strategies: training set evaluation, cross-validation (leave-one-out and 10-fold), inter-study validation (ISV) where one study was held out at a time, and external validation using six independent datasets. Additionally, GMWI2 was applied to four longitudinal datasets to demonstrate its utility in tracking gut health over time.

GMWI2 achieved a balanced accuracy of 79.9% on the training dataset (8069 samples), significantly outperforming GMWI (69.7%) and traditional α-diversity indices. Cross-validation and ISV maintained high accuracy (around 80% for high-confidence samples and ~75% overall). External validation on six independent datasets showed a balanced accuracy of 72.1%, further confirming GMWI2's robustness. Analysis of longitudinal datasets revealed GMWI2's ability to track changes in gut health following fecal microbiota transplantation (FMT), dietary interventions, antibiotic exposure, and prebiotic treatment, providing insights often not captured by traditional α-diversity metrics. Importantly, GMWI2 performed well in predicting health in a Parkinson's disease cohort not included in the training data. The majority of taxa with positive and negative coefficients showed higher relative abundance in healthy and non-healthy groups, respectively. Increasing the magnitude cutoff improved accuracy but reduced the number of samples included.

GMWI2 represents a significant advancement in gut microbiome-based health prediction. Its high accuracy across diverse populations and validation strategies demonstrates its robustness and generalizability. The ability to track gut health changes over time adds to its utility. GMWI2's performance surpasses previous methods and provides a valuable tool for various applications, including FMT donor selection and monitoring responses to interventions. The ability to incorporate a 'reject option' allows for a more nuanced interpretation of results, accounting for cases where the gut microbiome alone may not fully explain health status.

GMWI2 offers a robust and versatile tool for predicting health status and tracking gut health changes based on gut microbiome taxonomic profiles. Its high accuracy, generalizability, and open-source availability make it a valuable resource for researchers and clinicians. Future work should focus on incorporating additional microbiome features and expanding the diversity of included populations and disease phenotypes to further enhance its predictive power.

GMWI2 scores reflect an association with, not causation of, health status. It doesn't replace direct clinical measurements. Incorporating strain-level details, functional potential, and broader demographic representation would improve the model. While efforts were made to minimize batch effects, some residual effects may remain. The definition of 'healthy' and 'non-healthy' might impact accuracy. Some well-known pathogens did not show negative coefficients, highlighting the complexity of pathogenicity and the limitations of species-level resolution.