Machine Learning Model Predicts Synergistic Antibiotic Combinations Against E. coli

Antimicrobial resistance is one of the most pressing threats in modern medicine, and Escherichia coli sits at its center. The bacterium, a WHO Bacterial Priority Pathogen, is notorious for resistance to multiple antibiotics, including third-generation cephalosporins. Because developing entirely new antibiotics is resource-intensive, researchers are increasingly turning to combination therapy, where two or more drugs work together to enhance efficacy and make resistance less likely to emerge. The problem is scale: the number of possible drug pairs is enormous, and testing each one experimentally is slow, costly and often infeasible. A new study published in BMC Artificial Intelligence on 3 March 2026 describes a machine learning framework designed to cut through that combinatorial explosion, integrating drug structure information, pathogen response data and disease microenvironment data to predict which antibiotic combinations are likely to act synergistically against E. coli.

The research team, led by Racheal Claire Kyomukama and Ronald Galiwango, assembled three distinct types of data into a single predictive model. Chemical structure data came from DrugBank, where the structures of antibiotics and two metabolic media, glucose and glycerol, were retrieved in SMILES format and converted into MACCS keys, a set of 166 Boolean descriptors marking the presence or absence of particular substructures in each compound. Pathogen response data came from chemogenomic experiments measuring how single-gene knockout strains of E. coli respond to individual drugs, with fitness scores normalized to z-scores and binarized so that genes with strong effects were flagged as resistance- or sensitivity-associated. Disease microenvironment data captured how antibiotic efficacy fluctuates under different metabolic conditions, an often-overlooked factor, since a combination that is synergistic in one nutrient environment can become indifferent or antagonistic in another.

For training and validation, the researchers used a dataset of 392 drug combinations with known interaction scores drawn from the literature, split 80/20 into 314 training and 78 validation combinations. Each combination was described by 4,145 features: 3,979 E. coli genes and 166 substructural profiles. Because only about 20 percent of the combinations were synergistic, the team applied the synthetic minority oversampling technique, SMOTE, to balance the classes while preserving the data structure. They then trained a suite of classical machine learning algorithms in Scikit-learn, including decision tree, random forest, support vector machine and gradient boost classifier, alongside ensemble combinations, evaluating them with accuracy, precision, recall, F1 score, confusion matrices and ROC-AUC under five-fold cross-validation, with hyperparameter tuning performed on the validation set.

For the final test, the team held out an external dataset of 44 drug combinations with known interaction scores that had not been used in training or validation. Among single classifiers, the gradient boost classifier performed best, achieving 72 percent ROC-AUC, 70 percent accuracy, 66 percent precision, 68 percent recall and 67 percent F1 score. Its confusion matrix showed 21 correctly identified antagonistic combinations, 8 correctly identified synergistic combinations, 8 false antagonistic calls and 5 false synergistic calls. An ensemble of decision tree and gradient boost models performed slightly better, with 74 percent ROC-AUC, but the researchers selected the gradient boost classifier as their primary model because it consistently achieved higher recall for the minority synergistic class, their outcome of greatest interest, while remaining stable across cross-validation folds and on the external test set. A one-way ANOVA across multiple runs found that the performance differences among the models were not statistically significant, suggesting the variation was likely due to sampling variability rather than inherent differences in predictive ability.

With the trained model in hand, the researchers turned it loose on a prediction set of 306 antibiotic combinations, 153 for each of the two media, drawn from a study of the phenotypic landscape of a bacterial cell, for which no interaction scores were known. The model predicted 37 synergistic drug combinations across the two metabolic conditions: 17 in glycerol and 20 in glucose. The metabolic context mattered profoundly. The combination of amikacin and tetracycline, for example, was predicted to be synergistic in glucose but antagonistic in glycerol, a striking illustration of how the nutrient environment can flip the direction of a drug interaction. The model also predicted multiple synergistic pairs involving fusidic acid, including combinations with amikacin, tetracycline, oxacillin, rifampicin and nitrofurantoin.

Interpretability analysis using the Local Interpretable Model-agnostic Explanations framework, or LIME, revealed which features drove the predictions. Of the 4,145 features, only 12.4 percent, or 514, influenced the model’s predictive ability, and the overwhelming majority of those were genes rather than chemical descriptors. Chemogenomic features accounted for roughly 92 percent of the model’s predictive power, with drug structural features contributing the remaining 8 percent. Among the top ten genes, six were predictors of synergy, and the most predictive gene of all was oxyR, a transcription factor that, when activated by hydrogen peroxide, switches on E. coli’s oxidative stress response. The authors suggest a plausible mechanism: several bactericidal antibiotics, including aminoglycosides and fluoroquinolones, induce oxidative stress either directly or indirectly through metabolic disruption, so drugs that simultaneously elevate oxidative stress and impair stress-mitigation pathways can overwhelm the bacterium’s adaptive capacity and kill synergistically.

Pathway enrichment analysis of the top predictive genes, performed with ShinyGO against the KEGG database with Benjamini-Hochberg false discovery rate correction, pointed to amino acid biosynthesis as a highly represented pathway, alongside novobiocin biosynthesis, phenylalanine, tyrosine and tryptophan biosynthesis, and histidine metabolism. The dominance of chemogenomic signals suggests that E. coli’s internal stress-response pathways, rather than drug chemistry alone, largely determine whether combinations act synergistically. The authors note, however, that structural descriptors remain essential for distinguishing chemical classes, target engagement properties and membrane permeability characteristics that influence how drugs interact with the pathogen. The interpretability analysis also indicated that antagonism is not merely the absence of synergy but an active physiological response: induction of chaperones such as dnaJ or stringent response regulators such as sspA may enhance protein folding, damage repair or metabolic adaptation, buffering the combined drug effect and stabilizing bacterial survival under multi-drug stress.

The framework builds on, but differs from, earlier computational approaches. Prior models such as CoSynE, which leverages drug structural fingerprints, INDIGO, which incorporates chemogenomic fitness responses to individual antibiotics, and MAGENTA, which extends chemogenomic modeling to account for the disease microenvironment, each exploited a single data modality. No previous work, according to the authors, had combined all three data types in a unified predictive framework. There is external support for at least some of the model’s calls: fusidic acid combined with amikacin has previously been shown to exhibit strong synergy and clinical effectiveness against Staphylococcus aureus infections, lending credibility to the predictions, although that evidence comes from a different organism.

The study carries important limitations that temper any immediate clinical expectations. High-quality chemogenomic and microenvironment-resolved response data remain limited for most pathogens, which is precisely why E. coli, one of the few organisms with comprehensive multimodal datasets, was chosen. The team restricted the analysis to glucose and glycerol because those were the only media with high-quality, complete chemogenomic datasets, even though eight media were available in the underlying source study. The researchers deliberately used traditional machine learning rather than deep learning, reasoning that the small number of labeled combinations and the high dimensionality of the features would cause deep networks to overfit. Most critically, the predictions were not validated in vivo: drug interactions can differ substantially in complex physiological environments, and metabolic conditions in vitro do not fully reflect host tissues.

The authors emphasize that the 37 predicted synergistic combinations represent a candidate pool for systematic experimental evaluation, not ready-made treatments, and that cross-species validation and experimental testing are needed to support clinical translation. Because the methodology is species-agnostic and many pathogens share orthologous stress-response genes with E. coli, the modular framework could in principle be adapted to other clinically important bacteria once analogous datasets become available. Future work, the team says, should focus on expanding multimodal datasets, refining the model architecture, performing systematic experimental validation under metabolically diverse conditions, and validating the framework across other bacterial species. If those steps succeed, the approach could offer a scalable way to prioritize the most promising drug pairs from a search space far too large to explore at the bench, accelerating the discovery of combination therapies at a moment when resistant infections are outpacing new antibiotic development.

Subject of Research: Machine learning prediction of synergistic antibiotic drug combinations for E. coli using drug structure, pathogen response and microenvironment data

Article Title: Discovery of synergistic drug combinations for E. coli from drug information, pathogen response and disease microenvironment data

Article References: Kyomukama, R. C., Ainembabazi, M., Mutegeki, H., Nabisubi, P., Kakembo, F. E., Nankoma, G. L., Nabukeera, K. C., Walakira, A., & Galiwango, R. (2026). Discovery of synergistic drug combinations for E. coli from drug information, pathogen response and disease microenvironment data. BMC Artificial Intelligence, 2(1), Article 6. https://doi.org/10.1186/s44398-026-00023-y

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00023-y

Keywords: antimicrobial resistance, E. coli, machine learning, drug combinations, synergy prediction, chemogenomics, oxyR, gradient boosting, metabolic microenvironment, antibiotics, computational biology, SMOTE