Machine learning and phylogenetic analysis allow for predicting antibiotic resistance in M. tuberculosis

Area of Science:

• Microbiology
• Bioinformatics
• Genomics

Background:

• Antimicrobial resistance (AMR) is a major global health concern requiring accurate prediction of bacterial resistance patterns.
• Machine learning (ML) models analyze AMR data but often overlook bacterial evolutionary relationships, impacting performance.
• Existing genome-wide association studies (GWAS) account for evolution but only identify well-established variants.

Purpose of the Study:

• To develop a novel method that incorporates phylogenetic relationships into ML models for enhanced AMR prediction.
• To improve the accuracy and biological relevance of identifying resistance-associated genetic features.
• To identify novel candidate mutations linked to antimicrobial resistance.

Main Methods:

• Introduction of a phylogeny-related parallelism score (PRPS) to quantify feature correlation with population structure.
• Integration of PRPS with Support Vector Machine (SVM) and random forest models.
• Application of the pipeline to Mycobacterium tuberculosis AMR data from the PATRIC database.

Main Results:

• The PRPS combined with ML models reduced feature numbers while improving predictive performance.
• Known AMR-associated mutations were successfully re-identified.
• New candidate mutations potentially related to resistance were discovered.

Conclusions:

• Phylogenetic relationships significantly enhance ML model performance in AMR prediction.
• The developed pipeline yields more biologically relevant resistance markers.
• This approach aids in discovering novel AMR-associated mutations.