Machine learning with random subspace ensembles identifies antimicrobial resistance determinants from pan-genomes of three pathogens

Abstract

The evolution of antimicrobial resistance (AMR) poses a persistent threat to global public health. Sequencing efforts have already yielded genome sequences for thousands of resistant microbial isolates and require robust computational tools to systematically elucidate the genetic basis for AMR. Here, we present a generalizable machine learning workflow for identifying genetic features driving AMR based on constructing reference strain-agnostic pan-genomes and training random subspace ensembles (RSEs). This workflow was applied to the resistance profiles of 14 antimicrobials across three urgent threat pathogens encompassing 288 Staphylococcus aureus, 456 Pseudomonas aeruginosa, and 1588 Escherichia coli genomes. We find that feature selection by RSE detects known AMR associations more reliably than common statistical tests and previous ensemble approaches, identifying a total of 45 known AMR-conferring genes and alleles across the three organisms, as well as 25 candidate associations backed by domain-level annotations. Furthermore, we find that results from the RSE approach are consistent with existing understanding of fluoroquinolone (FQ) resistance due to mutations in the main drug targets, gyrA and parC, in all three organisms, and suggest the mutational landscape of those genes with respect to FQ resistance is simple. As larger datasets become available, we expect this approach to more reliably predict AMR determinants for a wider range of microbial pathogens.

Fig 1. S. aureus genomes clustered by shared genetic content compared to known subtypes and antibiotic resistance patterns.

(a) Genomes clustered using hierarchical clustering with average linkage, based on pairwise Jaccard distances between the sets of genetic features present in each genome. Clusters extracted from this hierarchy align well with (b) experimentally observed resistance patterns and (c) subtype annotations from PATRIC. Antibiotics shown are ciprofloxacin (CIP), clindamycin (CLI), erythromycin (ERY), gentamicin (GEN), sulfamethoxazole/trimethoprim (SXT), and tetracycline (TET).

Fig 2. Comparison of SVM ensemble approaches and statistical tests for detecting AMR-conferring genes and alleles in S. aureus.

(a) Workflow for SVM ensemble approaches. Beginning with genomes from PATRIC, open reading frames (ORFs) are identified and clustered by coding sequence to identify putative genes and alleles. Each genome is encoded based on the presence or absence of each gene and allele to capture genomic variation in the pan-genome as a sparse binary matrix. Genomes and/or features of this matrix are randomly sampled 500 times and used to train SVMs to predict binary AMR phenotype for a single antibiotic from genotype. Weights for each feature are averaged across all models in the ensemble and used to rank features by association to AMR. (b) Associations between known AMR-conferring genomic features and AMR phenotype, as ranked by Fisher’s Exact test, Cochran-Mantel-Haenszel test, and four different SVM ensemble types (SVM: ensemble by bootstrapping genomes, SVM-RSE: bootstrapping genomes and features; “random subspace ensemble”, SVM-RSE-O: SVM-RSE with oversampling to balance subtypes, SVM-RSE-U: SVM-RSE with undersampling to balance subtypes). Features were ranked either by p-value for statistical tests or by average feature weight for SVM ensembles. Fractional ranking was used for ties. Only features detected by at least one method are shown, colored by rank (green: in top 10, yellow: 11–50, orange: 51–100, gray: >100). Features shown are either genes or individual alleles (denoted as -#).

Fig 3. Predictive performance of SVM-RSE on 16 organism-antibiotic cases.

(a) Distribution of AMR phenotypes for each case. Organisms examined are S. aureus (SA), P. aeruginosa (PA), and E. coli (EC). Antibiotics examined are ciprofloxacin (CIP), clindamycin (CLI), erythromycin (ERY), gentamicin (GEN), tetracycline (TET), sulfamethoxazole/trimethoprim (SXT), amikacin (AMK), ceftazidime (CAZ), levofloxacin (LVX), meropenem (MEM), amoxicillin/clavulanic acid (AMC), imipenem (IPM), and trimethoprim (TMP). (b) SVM-RSE performance metrics from 5-fold cross validation. Performance values shown are averages and standard errors from 5-fold cross validation. The left-most column “log2(R/S)” shows the extent of class imbalance, the log2 of the number of resistant genomes divided by the number of susceptible genomes.

Fig 4. Characterization of mutations in four predicted AMR-conferring alleles in S. aureus.

For each of the predicted AMR-associated genes (a) kdbB, (b) SA_RS10745, (c) oppD and (d) ahpF, the AMR phenotype distributions and locations relative to InterPro structural domains are shown for individual mutations. Mutations in the predicted AMR-associated allele are in orange, while all other mutations observed for that gene are in black (only mutations in at least 5 genomes are shown). For kdbB, the first five annotations in light blue are associated with P-type ATPase. Abbreviations include superfamily (SF), domain superfamily (DSF), nucleoside triphosphate hydrolase (NTH), ATP-binding cassette transporter (ABCt), pyridine nucleotide-diphosphate oxidoreductase (PNDOR), and alkyl hydroperoxide reductase (AHPR), in addition to those used in InterPro annotations.