Predictive modeling of Pseudomonas syringae virulence on bean using gradient boosted decision trees

Area of Science:

• Bacteriology
• Plant Pathology
• Genomics
• Machine Learning

Background:

• Pseudomonas syringae pathovar designations are assumed to reflect host specificity, but this is rarely tested.
• Understanding host specificity in P. syringae is crucial for managing crop diseases.

Purpose of the Study:

• To assess the virulence of diverse P. syringae isolates on common bean (Phaseolus vulgaris).
• To determine host specificity differences between P. syringae phylogroup 2 (PG2) and phylogroup 3 (PG3) bean isolates.
• To develop a machine learning model for predicting P. syringae virulence on bean.

Main Methods:

• A rapid seed infection assay was used to measure virulence of 121 P. syringae isolates on common bean.
• Gradient boosting machine learning models were trained using whole genome k-mers, type III secreted effector k-mers, and effector/phytotoxin presence/absence data.
• Model predictions were functionally validated on 16 strains.

Main Results:

• Bean-infecting P. syringae isolates were generally more virulent on bean than non-bean isolates.
• PG3 bean isolates were significantly more virulent on bean than PG3 non-bean isolates, indicating higher host specificity.
• Machine learning models using whole genome data achieved high accuracy in predicting virulence (mean absolute error = 0.05).
• 94% of predicted virulence values for 16 strains fell within validated bounds.

Conclusions:

• PG3 P. syringae strains exhibit greater host specificity to common bean compared to PG2 strains.
• P. syringae PG2 strains may have evolved a broader host range, reflecting a different lifestyle.
• Machine learning is a powerful tool for predicting bacterial virulence and understanding host-specific adaptation in plant pathogens.