Identity
XGBoost (Extreme Gradient Boosting) is the machine-learning method used in this study to model genome-based metal tolerance from taxonomy, BacDive phenotype features, and metal-resistance gene-cluster counts. 1
Known alias: Extreme Gradient Boosting. 1
No stable external identifier is reported for this method in the source document. 1
Use in the BacDive metal-tolerance study
The full XGBoost model used 17 features across 3,994 species and included taxonomy, phenotype variables, and n_metal_clusters, the number of metal-resistance gene clusters. 1
In 5-fold phylogenetic-blocked cross-validation, the full model achieved R² = 0.633 and RMSE = 0.028, outperforming the gene-count-only model (R² = 0.063, RMSE = 0.045), phenotype-only model (R² = 0.163, RMSE = 0.043), taxonomy-only model (R² = 0.354, RMSE = 0.038), and taxonomy-plus-phenotype model (R² = 0.345, RMSE = 0.038). 1
In the non-cross-validated model comparison, taxonomy explained 35.4% of metal-tolerance variance, phenotype features explained 16.3%, and taxonomy plus phenotype yielded R² = 0.345, a delta R² = -0.009 relative to taxonomy alone. 1
Adding n_metal_clusters increased the full model to R² = 0.633, supporting environmental-resistome’s interpretation that genome-encoded metal-resistance repertoire was the strongest predictor in this analysis. 1
Feature interpretation
SHAP (Shapley additive explanations) feature importance from the full shap-interpreted XGBoost model placed taxonomic class and order codes together with n_metal_clusters among the top predictors. 1
Phenotype features contributed minimally after taxonomy was included, supporting the interpretation that the measured phenotypes were primarily phylogenetic proxies rather than independent predictors of the composite metal-tolerance score. 1
The model comparison does not establish metal-specific mechanisms because the target was a composite genome-based score rather than direct metal-tolerance measurements. 1
Relation to the metal fitness atlas
The Pan-Bacterial Metal Fitness Atlas provides a methodological contrast rather than a new XGBoost application: its pangenome-scale prediction used a simple gene-presence repertoire score, which failed to predict metal fitness. 2 This refines the interpretation of XGBoost’s BacDive result: predictive performance depends on the modeled target and feature design, so the atlas failure does not contradict the full BacDive XGBoost model. 2
Related source
The method is described in the study summary BacDive Phenotype Signatures of Metal Tolerance, alongside the metal-fitness-atlas scores and bacdive phenotype data. 1