Genomic and machine learning approaches to predict antimicrobial resistance in Stenotrophomonas maltophilia

Abstract

Stenotrophomonas maltophilia is a multidrug-resistant pathogen, which poses a major challenge to clinical management due to its increasing resistance to common antibiotics, such as levofloxacin (LEV) and trimethoprim-sulfamethoxazole (SXT), and poor clinical response to treatment. There is an urgent need for rapid and reliable antimicrobial susceptibility testing (AST) methods to improve treatment outcomes. This study collected 441 S. maltophilia strains, performed whole-genome sequencing, and used machine learning to identify key resistance determinants for LEV and SXT, constructing predictive models for resistance phenotypes. The 441 S. maltophilia strains we collected show significant genomic diversity and representative lineage distribution. Machine learning identified key resistance markers for LEV and SXT, improving area under the curve values to 92.80% for LEV and 95.44% for SXT. Validation accuracies reached 94.87% for LEV and 96.27% for SXT. Mutations in parC, smeT, and gyrA were strongly associated with LEV resistance. The gene presence of sul1, sul2, and CEQ03_18740, as well as gene mutations in Gsh2, prmA, and gspD, were highly correlated with SXT resistance. These findings suggest that integrating genome-based markers can enhance the prediction of antimicrobial resistance, offering a robust method for clinical application. Genotypic AST can reliably predict resistance phenotypes, providing a promising alternative to traditional AST methods for S. maltophilia infections.

Importance: Stenotrophomonas maltophilia is an emerging multidrug-resistant pathogen, making treatment challenging and requiring more effective diagnostic methods. This study offers a novel approach by integrating whole-genome sequencing with machine learning to identify key resistance markers for levofloxacin and trimethoprim-sulfamethoxazole. The predictive models developed can reliably forecast antimicrobial resistance phenotypes, providing a faster and more accurate alternative to traditional susceptibility testing. This approach not only enhances clinical decision-making but also aids in the timely administration of appropriate therapies. By identifying specific genomic markers associated with resistance, this study lays the foundation for future development of personalized treatment strategies, addressing the growing concern of antibiotic resistance.

Keywords: Stenotrophomonas maltophilia; antimicrobial resistance prediction; machine learning; whole-genome sequencing.

Fig 1

Lineage distribution of the Stenotrophomonas maltophilia complex. Different color ranges on branches represent the 23 lineages. From inside to outside, concentric rings display (i) sources of isolates; (ii) cities of isolation; (iii) origin type (environmental, anthropogenic, or human-associated); (iv) antibiotic-resistant phenotype to LEV; and (v) antibiotic-resistant phenotype to SXT.

Fig 2

Population structure of the Stenotrophomonas maltophilia complex. (A) Geographic origins of the 307 S. maltophilia isolates collected in China, with light and dark blue shades indicating isolate numbers per province. The colored pie charts show the phylogenetic lineage distribution per province. (B) Heatmap representing the clustering of pairwise average nucleotide identity among the 307 isolates, with deeper red indicating higher ANI and lighter blue indicating lower ANI. Left: distribution of phylogenetic lineages of S. maltophilia. (C) Phylogenetic lineage-based presence (red) or absence (yellow) of selected genes among 307 isolates, including DNA gyrase genes gyrA/gyrB, topoisomerase genes parC/parE, RND efflux pump systems smeABC, smeDEF, smeOP, efflux pump regulators smeT and smeU2, plasmid-mediated smqnrR, and dihydropteroate synthase genes sul1 and sul2. Antibiotic susceptibility results for LEV and SXT are represented in varying shades of blue.

Fig 3

Key AMR features screened for LEV and SXT resistance in S. maltophilia from the CARD database or the curated database (CARD plus ORF-based features). (A) LEV resistance features; (B) SXT resistance features. Left: key AMR features identified from the CARD database. Right: resistance markers identified using the ORF-based library method combined with the CARD database. Red indicates detected resistance features, and light pink indicates undetected features. Dark green represents both AST and predicted resistant phenotypes; light green represents both AST and predicted susceptible phenotypes.

Fig 4

Distribution of key AMR features in LEV and SXT resistance prediction models. Bar charts illustrating the distribution of LEV and SXT susceptibility testing values with the number of isolates possessing resistance features. The y-axis shows isolate numbers; the x-axis shows susceptibility values from different methods. The far-left category contains isolates with values less than or equal to the measured value, and the far-right category includes isolates with values greater than or equal to the measured value. Different colored bars represent different key resistance features. Dashed lines separate susceptible phenotypes (left), resistant phenotypes (right), and intermediate (middle). (A) LEV MIC value distribution; (B) LEV KB method distribution; (C) SXT MIC value distribution; and (D) SXT KB method distribution.