Introduction

Antimicrobial resistance (AMR) is a severe threat. Currently, an estimated 4.95 million deaths are associated with AMR each year [1], and this number is only predicted to increase. Urgent action is required to prevent further escalation, slow epidemic spread, and preserve our capacity to administer advanced medical therapies relying on surgery or immunosuppression. Surveillance based on whole genome sequencing is playing an increasingly important role in monitoring the emergence and spread of AMR [2]. Taking full advantage of this accumulating data will require automated systems that can transform genome sequences into actionable intelligence, namely, predictions of resistance profiles and identification of new mechanisms of resistance [3–5]. Machine learning (ML) approaches have the potential to fill this niche, though a number of challenges remain in implementing such systems [6].

One underappreciated challenge is in the assumptions made by ML methods themselves. Most classical ML methods assume training data are independent and identically distributed, which is not true of pathogen surveillance samples due to the underlying structure of bacterial populations [7]. During the epidemic spread, successful clones spread rapidly. If this spread is due in part to acquisition of AMR determinants it could lead to an association between the phenotype and phylogenetic markers that do not directly contribute to AMR. These noncausative associations are likely further exacerbated by biased sampling focused on human disease in high-income countries [5,8] leaving large regions of the phylogeny unexplored. The phylogenetic dependence of genes and traits has long been recognized in comparative and evolutionary biology [9], with statistical approaches to correct for this dating at least to the mid-1970’s [10]. In the context of microbial genome-wide association studies (GWAS), these phylogenetic effects have been mitigated in part by the application of mixed-effect models that attempt to correct for sampling biases [11–13]. To date, the implications of biased sampling for ML approaches to AMR prediction have not been systematically investigated, despite the clear risk of confounding [14]. Here, by constructing realistic pathological scenarios, we show that biased sampling combined with population structure can dramatically affect the performance of ML methods for predicting AMR. We investigate the factors underlying the performance of ML methods and provide concrete recommendations for the future evaluation of ML approaches to predicting AMR.

Sampling effects bias machine learning prediction of AMR

To comprehensively evaluate the impact of population structure in predicting AMR, we collected between 3,204 and 7,188 genomes for each of three Gram-negative species and two Gram-positive species representative of current WHO priority pathogens [15], including the gastrointestinal and urinary tract pathogen Escherichia coli, the opportunistic pathogen Klebsiella pneumoniae, the gastrointestinal pathogen Salmonella enterica, the skin commensal and opportunistic pathogen Staphylococcus aureus, and the major agent of community-acquired pneumonia Streptococcus pneumoniae. In total, our dataset includes resistance phenotypes for 27 antibiotics, spanning multiple drug classes and diverse sequence types (S1–S6 Figs, Methods). To limit the effects of sample size and class imbalance, we excluded antibiotic-organism combinations with less than 1,000 genomes or with resistant or susceptible strains exceeding 80% of the data set. The average number of genomes was ~2,700 with ~44% resistant strains, and 80% of the data was in the range of 1,134–3955 genomes, with 25%–71% resistant. A whole genome alignment was constructed for each species using Roary [16], and SNP-sites [17] was used to identify variable positions in the alignment. Both SNPs and the pangenome presence-absence matrix from Roary were used as predictive features. As a representative ML model we used LightGBM [18], a gradient boosted decision tree method representative of the tree ensemble-based methods that generally perform well for predicting AMR [19,20].

We defined discrete clades based on deep divergences between branches observed in the phylogenetic tree for each species constructed from core genome alignments using IQ-Tree (Figs 1A, 1B and S2–S6A) based on deep divergences between branches. These clades align closely with established sequence types, providing biologically meaningful groupings. Using this clade structure, we developed pathological training cases simulating biased sampling. In our first approach, Scheme A (Fig 1C), we selected two clades, and a balanced test set was held-out from one (the test clade). We then devised 4 scenarios, representing biased sampling. Scenarios a and b excluded either resistant or sensitive isolates from training data for the test clade, respectively. As a positive control, we also included scenarios c and d, which excluded sensitive or resistant strains from the nontest clade. Compared to a random split of training and test data, models trained on biased datasets exhibited much lower area under the receiver operating characteristic curve (AUC) scores (Fig 1D). Exclusion of resistant strains led to low precision and high recall of sensitive strains (S7A Fig), while the opposite was observed for exclusion of sensitive strains (S7A Fig), suggesting that the ML models may have been influenced by markers of lineage rather than identifying genuine indicators of AMR. This was not observed in our control scenarios, and, importantly, persisted even when both resistant and sensitive strains were present in the test clade, but at biased proportions (S7BC Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Bacterial population structure and clade-specific sampling affect machine learning predictions of antimicrobial resistance (AMR).

We define clades based on core genome phylogeny and simulate training and testing scenarios that highlight challenges in generalizing AMR prediction across genetically distinct groups. (A) An illustration of defining clades for genomes in one species. (B) Legend for panels C and E, R: resistant (triangles), S: susceptible (circles). (C) An example of training scenarios for scheme A based on the clade splitting in A. Models were trained on clade 1 and one of the other clades, such as clades 1 and 2, clades 1 and 3, and clades 1 and 4. Four training scenarios were tested for each antibiotic and species. In each scenario, either resistant or susceptible samples from one of the paired clades were excluded from the training data, while the held-out genomes from the clade that is not clade 1 were used as the test sets. (D) The area under the ROC curve (AUC) scores in each scenario for scheme A for all antibiotics and species. Light gray: train-test split followed scheme A, dark gray: random train-test split. (E) An example of training scenarios for scheme B based on the clade splitting in A. Models were trained on all samples except for resistant or susceptible samples from one clade. The held-out genomes from the same clade were used as the test sets. (F) The AUC scores for each scenario for scheme B for all antibiotics and species. Light gray: train-test split followed scheme B, dark gray: random train-test split. (G) The precision and recall scores for scheme B; a-random and b-random indicate testing on random train-test splits of the pathological unbalanced data. Underlying data for this figure are available in S3 Data.

https://doi.org/10.1371/journal.pbio.3003539.g001

It is often implicitly assumed that simply increasing sample sizes is sufficient to improve ML models. To test this, we developed a second approach, Scheme B, where we additionally included training data from remaining clades, while still excluding resistant or sensitive isolates from the test clade (Fig 1E). This did lead to an improvement in the AUC compared to Scheme A (Fig 1F). However, extreme biases remained in both precision and recall (Fig 1G), indicating that trained models continued to conflate lineage and AMR markers even in the presence of a larger training data set.

AMR predictivity varies across antibiotics and species

To further understand the factors driving model performance, we trained a random forest meta-model to predict the AUC scores of all trained models (S2 Data), including species, training scenario, clade-wise distance, sample size, and drug class as predictors. We then interpreted this meta-model using Shapley additive explanation (SHAP) values [21] (Fig 2A). SHAP values are a game theoretic approach to model interpretation that can be intuitively understood as providing a quantitative estimate of each feature’s contribution to the model prediction, in this case the AUC.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. AMR predictivity varies across antibiotics and species.

(A) SHAP values for the top 15 features from a random forest model trained on AUC scores from both schemes A and B for all five species. Global feature importance is given by the mean absolute SHAP value (left), while the beeswarm plot (right) illustrates feature importance for each guide prediction. (B) The distribution of AUC scores for each species. Red dashed lines indicate the median values. (C) The distribution of AUC scores for four drug classes shown in A. Red dashed lines indicate the median values. Underlying data for this figure are available in S3 Data.

https://doi.org/10.1371/journal.pbio.3003539.g002

This analysis reaffirmed the inferior performance associated with our pathological training scenarios. Additionally, this analysis showed that training on data sourced from Gram-negative species yielded diminished AUC scores (Fig 2B), presumably due to the presence of the outer membrane [22], and that drug class also had large and at times contradictory effects on model performance (Fig 2C). Given the impact of the species on model performance, we asked how the model interpretation differs when training on each species (see S2–S6C Figs). As an example, the effects of fluoroquinolone varied in three species (S2C, S4C, and S5C Figs) and also in the model encompassing all species (Fig 2A). Subsequent interpretation of models trained on each clade for ciprofloxacin (S9 Fig) further underscored the distinct determinants for prediction across clades and species. Among the known genes associated with ciprofloxacin resistance [23], the gene parC prominently featured in most clades of E. coli and S. aureus. In contrast, the gene gyrA was exclusive to a single clade within S. aureus, possibly reflecting different fitness effects or primary target of ciprofloxacin in this clade [24]. In other clades of E. coli and S. enterica, the strongest predictors diverged, encompassing genes such as transcriptional regulator kdgR, mobilization gene mobB, quinone oxidoreductase qorA, and genes encoding hypothetical proteins.

To test if this lack of common features was general, we trained clade-specific models for all drug classes. Surprisingly, minimal feature overlap was identified, both within the top 10 features and when summing the top features based on SHAP values to account for 50% of total SHAP values (S8A and S8B Fig). This scarcity of common predictive genes or SNPs across clades may in part be a result of the small number of strong predictors (S8C Fig), and provides a partial explanation for the strong effects we observed for our pathological training scenarios. Together with the performance patterns observed in S7A Fig, these results support the conclusion that the models conflate lineage‑associated markers with genuine indicators of AMR.

Implications for AMR prediction

ML-based modeling of AMR can serve multiple purposes. These include at least two major applications: 1) prediction of strain resistance, and 2) identification of causal variants. While these two applications are related, achieving one does not necessarily imply achieving the other. If biased sampling reflects biological reality, i.e., all strains of a particular lineage truly are resistant to an antibiotic, then models that conflate lineage markers with resistance may still provide accurate predictions for surveillance purposes. This can be seen, for instance, in applications that have used phylogenetic placement to predict AMR [25]. However, such models will not provide insights into the underlying molecular mechanisms of resistance, and will fail on strains from previously unseen lineages or resistant strains emerging from a former sensitive lineage. Conversely, identifying causal variants requires disentangling lineage effects from true resistance determinants, a task complicated by the structure and biases present in genomic datasets.

Our results demonstrate that current ML approaches are particularly vulnerable to the confounding effects of biased sampling and population structure. Even when sample sizes are increased, models struggle to generalize beyond the specific clades included in the training set, underscoring the limitations of scaling as a solution to bias. This is supported by previous work that has shown phylogenetic structure is an important feature for predicting AMR [19], and that many AMR prediction tools perform poorly when tested on phylogenetically-constructed benchmark sets [26]. The lack of overlap we found in predictive features across clades further highlights the challenges in identifying universal resistance markers, with most predictions being dependent on the phylogenetic background. While some of these differences likely reflect true clade‑specific biology, they may also be influenced by the reduced proportion of shared core genomic variation between more distantly related clades. Our ML framework, which incorporates both core genome SNPs and accessory gene presence–absence data, partially mitigates this limitation by capturing variation beyond the strict core genome. Nonetheless, this remains an important factor to consider when interpreting cross‑clade predictive patterns. These findings are consistent with previous observations in bacterial GWAS, where lineage effects are a persistent source of confounding and require explicit correction.

In addition to sampling considerations, feature representation is another important factor influencing model performance. Our use of a combination of core genome SNPs and accessory gene presence–absence patterns captures broad genomic differences but does not capture sequence variation within accessory genes. This limitation is particularly relevant for species with large and diverse accessory genomes, such as E. coli and K. pneumoniae, where within‑gene variation may carry additional predictive signals for AMR. To address this concern, we performed a reanalysis of a subset of our Klebsiella benchmark data using unitigs as predictive features, producing results consistent with our SNP and accessory gene features (see Methods, S10 Fig). While this is not a comprehensive exploration of alternative feature encodings, including k-mers and unitigs, it does suggest that feature engineering alone is unlikely to be sufficient to address the underlying problem of nonindependence of samples.

The statistical challenge posed by analysis of samples related by descent from a common ancestor has long been recognized [10], and our results demonstrate that ML is not immune from such concerns. Future efforts to improve ML-based AMR prediction must incorporate strategies to address this challenge. A wide range of methods have been developed in the field of comparative biology [9,27,28] that may serve as inspiration for adapting ML algorithms. One potential approach is the development of lineage-aware ML algorithms that explicitly model the hierarchical structure of bacterial populations, borrowing ideas from mixed-effect models commonly used in bacterial GWAS [11–13]. Another promising direction is the integration of population-genomic frameworks that incorporate evolutionary signals, such as measures of phylogenetic conservation, into feature selection and model interpretation [29]. Additionally, benchmark datasets that include well-curated, globally representative samples are essential for evaluating and comparing new approaches, ensuring models are tested in realistic scenarios that capture the diversity encountered in clinical and environmental surveillance. Our strategy of phylogeny-aware cross-validation, i.e., testing on held-out clades, will be an important tool to demonstrate generalizability of ML predictors of AMR.

Importantly, sampling strategies must also evolve to support these developments. Expanding surveillance efforts to underrepresented regions and settings, particularly in low- and middle-income countries (LMICs), will provide more balanced datasets and reduce biases toward high-income countries and outbreak scenarios [8]. Targeted sampling that prioritizes diversity, both within and across clades, will be critical for developing models capable of generalizing across phylogenetic backgrounds. Beyond sampling, careful experimental design, including the use of balanced test sets and rigorous evaluation metrics such as those employed here, will further ensure that models are robust to confounding effects. Including algorithms based on simple assumptions, such as phylogenetic placement [25], as baseline models for ML development should help in determining if machine learning is providing additional insights or simply recapitulating phylogenetic structure.

In conclusion, addressing the interplay between population structure and AMR prediction will require a multifaceted approach that includes improved sampling, algorithmic innovation, and systematic evaluation of proposed prediction methods. Only by confronting these challenges can we unlock the full potential of machine learning to provide actionable insights into AMR, advancing both our surveillance capabilities and our understanding of resistance mechanisms. These efforts will be critical to combating the global threat of AMR and ensuring the continued efficacy of life-saving antimicrobial therapies.

Methods

Supporting information