Machine learning-based drug susceptibility prediction from Candida genomic data

Area of Science:

• Medical Mycology
• Genomics
• Computational Biology

Background:

• Invasive Candida infections pose a growing clinical challenge due to increasing antifungal resistance.
• Current antifungal susceptibility testing (AFST) methods lack the speed and accuracy needed for routine clinical practice.

Purpose of the Study:

• To evaluate the species distribution and antifungal susceptibility of invasive Candida isolates in China.
• To assess the feasibility of using whole-genome sequencing (WGS) and machine learning (ML) to predict minimum inhibitory concentrations (MICs).

Main Methods:

• Collected 337 invasive Candida isolates from 20 hospitals in China (2022-2023).
• Determined MICs for nine antifungal agents using broth microdilution.
• Performed WGS on prevalent species and utilized genomic 11-mer features to train and optimize Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) models.

Main Results:

• Identified dominant species: C. albicans (n=103), C. tropicalis (n=71), C. parapsilosis (n=67), and C. glabrata (n=63).
• Observed higher azole and echinocandin resistance in non-albicans Candida species, with notable resistance in C. tropicalis (azoles) and C. glabrata (echinocandins).
• The optimized RF model achieved high accuracy (AUC 0.979), with essential agreement >90.1% and categorical agreement >93.2% for MIC prediction.

Conclusions:

• Non-albicans Candida species present a significant clinical challenge due to emerging antifungal resistance.
• WGS combined with ML offers a highly accurate, potentially rapid method for predicting antifungal susceptibility, supporting earlier and more effective antifungal therapy.