npj Antimicrobials and Resistance

Antimicrobial resistance is an urgent global health threat, with over 2.8 million multidrug-resistant infections killing over 35,000 annually in the US. Machine Learning (ML) has emerged as a potential solution to improve efficiency of antibiotic high-throughput screens (HTS). We report ML-guided high-throughput screening against E. coli. Large-scale Learning-to-Rank models were trained on public and proprietary datasets to maximize phenotypic inhibition and minimize human cell cytotoxicity. We evaluated several pre-plated compound libraries and a set of "cherry-picked", structurally novel compounds. We screened against a hyperpermeable lptD- mutant, followed by hit confirmation, profiling, cytotoxicity counter-screening, and MOA determination. Results demonstrated a doubled hit rate and 3X fewer toxic hits. Additionally, activity improved against both Wild Type E. coli and the lptD- mutant. ML models showed robust predictive power on structurally dissimilar compounds. The combination of large-scale HTS, ML innovation, and both library-wise selection and cherry-picking strategies distinguishes this study in the antibiotic discovery field.

Metagenomic sequencing has greatly expanded our ability to detect and characterize resistance genes across diverse environments. However, translating the metagenomic findings into actionable clinical insights remains challenging. We developed a web-based tool that applies a Multi-Criteria Decision Analysis (MCDA) framework to metagenomic AMR datasets. The tool integrates gene abundance with risk attributes derived from the WHO Bacterial Priority Pathogens List, including mortality, transmissibility, and treatability scores. Users can upload AMR detection results alongside customized scoring matrices to generate risk profiles at the samples, species, and drug class levels. Validation was performed using upper respiratory tract samples from SARS-CoV-2 patients and controls in central India, as well as a publicly available dataset from Tanzania. From 48 SARS-CoV-2 samples, 9 produced a total of 10 records involving Salmonella enterica, Escherichia coli, and Streptococcus pneumoniae across fluoroquinolone, cephalosporin, and macrolide classes (cumulative scores 4.2-146). In contrast, 3 of 48 control samples yielded 3 records, all linked to macrolide resistance in S. pneumoniae (scores 5.46-92.43). Analysis of 17 Tanzanian samples identified 2 records, with Klebsiella pneumoniae (cephalosporin resistance, score 70.2) and S. pneumoniae (macrolide resistance, score 670.02) emerging as priority risks. This framework bridges the gap between raw metagenomic data and clinically relevant risk assessment.

Carbapenem-resistant Gram-negative bacteria (CRGNB) infections remain difficult to manage because treatment decisions must balance heterogeneous patient risk, limited antibiotic options, potential toxicity and emerging resistance. Clinical care in this setting requires not only single-endpoint risk prediction, but also decision-support frameworks that can jointly enable prognosis assessment, result interpretation, and individualized treatment comparison. Here we present Dr.BUG, an interactive clinical AI agent for personalized decision support in CRGNB infection. Dr.BUG integrates stable feature-set selection, multi-task prognostic modelling, interpretability analysis and model-based simulation of antibiotic regimen recommendation into a unified workflow. Using a development cohort, a temporally independent validation cohort, and external cohorts from the MIMIC-IV dataset, we developed and validated models for four clinically relevant tasks: clinical efficacy, survival outcome, polymyxin resistance and treatment duration. Model inputs were derived primarily from routinely available and relatively low-cost clinical variables, supporting translational feasibility. Across the major tasks, selected-feature models matched or exceeded the performance of their full-feature counterparts while using fewer variables, as reflected in 82.0% of optimized-metric comparisons in the development cohort, and remained robust in both temporal and external validation. Dr.BUG further provided both population-level and patient-level interpretability and generated individualized rankings of candidate antibiotic regimens. In the retrospective analysis of non-survivors, clinician review suggested that regimens recommended by Dr.BUG might be associated with higher predicted survival probabilities. These findings support a broader role for clinical AI in complex drug-resistant infections, extending its utility from offline risk prediction to interpretable, deployable, and personalized decision support.

BackgroundCatalogues of genetic variants associated with resistance underpin whole-genome sequencing (WGS)-based predictions of drug susceptibility in Mycobacterium tuberculosis, and are essential for molecular diagnostics and surveillance. The current gold standard catalogues are those released by the WHO but the underlying data are not fully released and they are difficult to interpret. Open and reproducible methods would help address these problems, extending the important work already done. MethodsWe have developed an automated method, catomatic, that uses a binomial test to associate informative isolates with resistance or susceptibility, and built a catalogue (catomatic-1) from the same 39,358 samples used to construct the first edition of the WHO catalogue (WHOv1). We performed a sensitivity analysis to optimise statistical and bioinformatic parameters for each drug, and benchmarked catomatic-1 against WHOv1 using an independent Validation Dataset of 14,380 isolates. FindingsBy using simpler statistics, catomatic-1 algorithmically classified 1,329 genetic variants, ranging from five for linezolid to 440 for pyrazinamide. WHOv1 included generalisable rules added by a panel of experts, increasing its predictive coverage, but at the cost of reproducibility. Despite not including such expert rules, catomatic-1 achieves comparable performance for all drugs, with sensitivities for first-line agents above 88% on the independent Validation Dataset. The automated process allowed us to efficiently explore parameter space; for instance, detecting resistant variants with low read support improved the sensitivity for all drugs. InterpretationPerformant resistance catalogues for M. tuberculosis can be built automatically using transparent and reproducible statistical methods. As more data are collected, catalogue content and performance will evolve, highlighting the need for proper versioning, machine/human readability, and open access. This approach demonstrates resistance catalogues used in surveillance and diagnostics can be rapidly and reproducibily updated. FundingThe National Institute for Health and Care Research (NIHR), Engineering and Physics Sciences Research Council (EPSRC) and ORACLE Corporation. Research in contextO_ST_ABSEvidence before this studyC_ST_ABSWe searched PubMed and preprint servers (bioRxiv, medRxiv), and publicly available mutation catalogues for studies linking Mycobacterium tuberculosis genomic variants with drug resistance using whole-genome or targeted sequencing and phenotypic drug-susceptibility testing (pDST). Search terms combined "Mycobacterium tuberculosis", "genome sequencing", "mutation catalogue", "mutation effects", "drug resistance", and individual drug names, with no language or date restriction. We included studies providing paired, clinical genomic and pDST or MIC data, excluding purely in-silico or case-only reports. This work directly builds on methodologies and data published by five prior studies, and makes primary comparisons with the First (WHOv1) and Second (WHOv2) Editions of the WHO Catalogue of mutations in Mycobacterium tuberculosis. Added value of this studyWe developed catomatic, a transparent, reproducible tool for building catalogues of resistance- and susceptibility-associated genetic variants. Trained on the same samples used to build WHOv1 and benchmarked on an independent Validation Dataset, catomatic achieves comparable sensitivity, specificity, and definitive prediction rates to WHOv1 without expert-rule augmentation and despite using simpler statistics. It optimises parameters per drug, produces machine-readable outputs (CSV/JSON), and demonstrates that adjusting read-support thresholds can improve detection of minor resistance subpopulations. Implications of all the available evidenceCatalogues of resistance-associated variants for M. tuberculosis can be rapidly and transparently constructed. Making catalogues available in human/machine-readable formats with uncertainty estimates will improve uptake of WGS for M. tuberculosis surveillance and diagnostics; using a reproducible process permits diagnostic test manufacturers, researchers, clinical and public health laboratories to select the level of statistical support necessitated by their specific use-case, Policymakers should balance the benefits of expert rules against loss of reproducibility. Future work will expand the size of the datasets used, integrate minimum inhibitory concentration data, and establish consensus workflows for routine, transparent catalogue updates.

Antimicrobial resistance (AMR) is a growing global threat to human health, and rapid methods for characterising emerging antimicrobial resistance genes (ARGs) are needed. Here, we develop a semi-automated workflow using cell-free gene expression (CFE) systems to measure the activity of two ARGs encoded on plasmid DNA that produce rifampicin-inactivating and gentamicin-inactivating enzymes. We validated the use of a small benchtop Myra liquid handling system compared to manual pipetting, with no statistical differences observed. After optimising the pre-incubation time of ARGs and dispensing protocol, expression of aac(3)-IIa increased the half-maximal inhibition concentration (IC50) of gentamicin by over 150-fold, while arr-3 increased the IC50 of rifampicin by approximately 20-fold compared to controls. Future work could extend this platform to characterise novel ARGs identified through genomic surveillance or rapidly profile activity of new or derivative antibiotics. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=87 SRC="FIGDIR/small/720151v1_ufig1.gif" ALT="Figure 1"> View larger version (20K): org.highwire.dtl.DTLVardef@1a61fe3org.highwire.dtl.DTLVardef@1778eadorg.highwire.dtl.DTLVardef@380be4org.highwire.dtl.DTLVardef@194bb63_HPS_FORMAT_FIGEXP M_FIG C_FIG

In vitro methods to characterize drug combinations typically involve phenotypic screenings using checkerboard assays (CBA) or, more recently, DiaMOND. Such approaches rely on the Fractional Inhibitory Concentration Index (FICI), a fixed-time measurement of growth inhibition that, nonetheless, necessitates secondary validation by time-kill assays (TKA). Longitudinal time-kinetics of bacterial killing are considered the gold standard in vitro proxy for antimicrobial activity, but they required increased assay complexity, particularly against the slow growing Mycobacterium tuberculosis. Here, we developed a new methodology named OPTIKA (Optimized Time Kill Assays) that enhances the capacity of traditional TKA by over 1000-fold. This allows for easy and dynamic examination of n-way drug interactions by simultaneously monitoring bactericidal and sterilizing capacities in a longitudinal manner. We then replicated previous DiaMOND studies and performed comparisons using CBA and OPTIKA methodologies. We demonstrate that selection of the efficacy parameters (either routed on bacteriostatic, bactericidal or sterilizing properties) affects the interpretation of in vitro drug interactions and, consequently, its potential translational value. The increased assay throughput provided by OPTIKA offers a novel framework for developing tuberculosis treatment regimens. TeaserOPTIKA is a new methodology that increases time-kill assay performance against Mycobacterium tuberculosis by over 1,000-fold

Treatment of Mycobacterium tuberculosis (Mtb) is challenging and requires administration of at least four different antibiotics. Unfortunately, multi-drug resistant Mtb strains continue to emerge, undermining the effectiveness of current treatment regimens and highlighting the urgent need for new therapeutics. In this study, we evaluated the potential of natural product-derived chlorotonils as anti-Mtb agents. We demonstrate that chlorotonils exhibit nanomolar potency against a range of attenuated and virulent Mtb strains. Mechanistic studies and resistance profiling in Mtb revealed that chlorotonils affect both lipid and energy metabolism. Through systems biology approaches, including the construction of an Mtb CRISPRi library specifically designed for chemical-genomic profiling, we identified MmpR5/MmpL5 as major driver of chlorotonil-resistance in Mtb leading also to cross-resistance with bedaquiline. Our findings highlight chlorotonils as valuable chemical tools to further dissect the role and function of the MmpS5-MmpL5 efflux pump in drug-resistant Mtb.

SynopsisO_ST_ABSBackgroundC_ST_ABSPyrazinamide is one of four first-line antibiotics used to treat tuberculosis, however antibiotic susceptibility testing for pyrazinamide is challenging. Resistance to pyrazinamide is primarily driven by genetic variation in pncA, an enzyme that converts pyrazinamide into its active form. MethodsWe curated a dataset of 664 non-redundant, missense amino acid mutations in pncA with associated high-confidence phenotypes from published studies and then trained three different machine learning models to predict pyrazinamide resistance. All models had access to a range of protein structural-, chemical- and sequence-based features. ResultsThe best model, a gradient-boosted decision tree, achieved a sensitivity of 80.2% and a specificity of 76.9% on the hold-out Test dataset. The clinical performance of the models was then estimated by predicting the binary pyrazinamide resistance phenotype of 4,027 samples harboring 367 unique missense mutations in pncA derived from 24,231 clinical isolates. ConclusionsThis work demonstrates how machine learning can enhance the sensitivity/specificity of pyrazinamide resistance prediction in genetics-based clinical microbiology workflows, highlights novel mutations for future biochemical investigation, and is a proof of concept for using this approach in other drugs.

The treatment of tuberculosis (TB), which kills 1.8 million each year, remains difficult, especially with the emergence of multidrug resistant strains of Mycobacterium tuberculosis (Mtb). While there is an urgent need for new drug regimens to treat TB, the process of drug evaluation is slow and inefficient owing to the slow growth rate of the pathogen, the complexity of performing bacteriologic assays in a high-containment facility, and the context-dependent variability in drug sensitivity of the pathogen. Here, we report the development of "DRonA" and "MLSynergy", algorithms to perform rapid drug response assays and predict response of Mtb to novel drug combinations. Using a novel transcriptome signature for cell viability, DRonA accurately detects bacterial killing by diverse mechanisms in broth culture, macrophage infection and patient sputum, providing an efficient, and more sensitive alternative to time- and resource-intensive bacteriologic assays. Further, MLSynergy builds on DRonA to predict novel synergistic and antagonistic multi-drug combinations using transcriptomes of Mtb treated with single drugs. Together DRonA and MLSynergy represent a generalizable framework for rapid monitoring of drug effects in host-relevant contexts and accelerate the discovery of efficacious high-order drug combinations.

Antibiotic resistance poses one of the greatest public health challenges of the 21st century. Yet not all pathogens are equally affected by resistance evolution. Why? Here we examine what underlies variation in antibiotic resistance across human bacterial pathogens and the drugs used to treat them. We document the observed prevalence of antibiotic resistance for pathogen x drug combinations across 57 different human bacterial pathogens and 53 antibiotics from 15 drug classes used to treat them. Using AIC-based model selection we analyze 14 different traits of bacteria and antibiotics that are believed to be important in resistance evolution. Using these data, we identify the traits that best explain observed variation in resistance evolution. Our results show that nosocomial pathogens and indirectly transmitted pathogens are significantly associated with increased prevalence of resistance whereas zoonotic pathogens, specifically those with wild animal reservoirs, are associated with reduced prevalence of resistance. We found partial support for associations between drug resistance and gram classification, human microbiome reservoirs, horizontal gene transfer, and documented human-to human transfer. Global drug use, time since drug discovery, mechanism of drug action, and environmental reservoirs did not emerge as statistically robust predictors of drug resistance in our analyses. To the best of our knowledge this work is the first systematic analysis of resistance across such a wide range of human bacterial pathogens, encompassing the vast majority of common bacterial pathogens. Insights from our study may help guide public health policies and future studies on resistance control.

BackgroundHelicobacter pylori is a significant risk factor for gastric cancer, peptic ulcers, and MALT lymphoma. Rising antibiotic resistance rates complicate treatment strategies. While nucleotide sequence based assays are reliable in predicting clarithromycin and levofloxacin resistance, predicting metronidazole resistance is more challenging due to diverse metabolic pathways contributing to resistance, and high genomic variability. MethodsWe assembled a cohort of 483 H. pylori clinical isolates, combining whole-genome sequencing with phenotypic susceptibility testing. Machine learning models (SVM, XGBoost, FNN) were trained on genomic variants to predict resistance phenotypes. A sliding-window approach and SHAP-based importance scoring were used for feature selection to identify biologically relevant mutations, improving prediction accuracy, particularly for metronidazole resistance. ResultsThe best-performing FNN model improved metronidazole resistance prediction by 16% compared to conventional (non-ML, single polymorphisms) sequence-based detection methods applied to the same strain collection. Feature selection identified 32 feature sets, with 11 sets significantly improving F1-scores over the baseline. Combining 2-4 feature sets revealed 53 synergistic combinations across all models. Validation showed that 87% of these combinations significantly outperformed non-ML molecular testing, with 16 combinations achieving F1-scores above 0.65. ConclusionMachine-learning can significantly improve the performance of sequence-based susceptibility testing for metronidazole in H. pylori. Novel candidate predictive markers identified from whole-genome data offer testable hypotheses about yet unexplored mechanisms of metronidazole resistance. These findings support the potential for ML-based approaches to enable more accurate susceptibility-guided therapies.

Whole-cell screening for Mycobacterium tuberculosis (Mtb) inhibitors is complicated by the pathogens slow growth and biocontainment requirements. Here we present a synthetic biology framework for assaying Mtb drug targets in engineered E. coli. We construct Target Essential Surrogate E. coli (TESEC) in which an essential metabolic enzyme is deleted and replaced with an Mtb-derived functional analog, linking bacterial growth to the activity of the target enzyme. High throughput screening of a TESEC model for Mtb alanine racemase (ALR) revealed benazepril as a targeted inhibitor. In vitro biochemical assays indicated a noncompetitive mechanism unlike that of clinical ALR inhibitors. This is the first report of an antimicrobial activity in an approved Angiotensin Converting Enzyme (ACE) inhibitor and may explain clinical data associating use of ACE inhibitors with reduced Mtb infection risk. We establish the scalability of TESEC for drug discovery by characterizing TESEC strains for four additional targets. SIGNIFICANCE STATEMENTThe challenge of discovering new antibiotics is both scientific and economic. No simple test can determine if a given molecule will be safe and effective in real human patients. Many drug candidates must therefore be advanced for each new antibiotic that reaches the market - a risky and expensive process. In this work we use synthetic biology to engineer the common laboratory model bacterium E. coli as a tool for early stage antibiotic discovery. As a proof of concept we expressed a known tuberculosis drug target and found a novel inhibitor: benazepril. Many other drug targets could be screened similarly using the system that we describe. Because E. coli can be grown safely and cheaply, this approach may help to reduce costs and make drug discovery more accessible.

Tuberculosis remains a global health problem. Making it easier and quicker to identify which antibiotics an infection is likely to be susceptible to will be a key part of the solution. Whilst whole-genome sequencing offers many advantages, the processing of the genetic reads to produce the relevant public health and clinical information is, surprisingly, often the responsibility of the end user which inhibits uptake. Here we characterise how well a freely-available tool we have developed, gnomonicus, predicts the antibiotic resistance profile of a sample (given its variant call file) using our implementation of the second edition of the WHO catalogue of resistance-associated variants (WHOv2). To facilitate this, we have constructed a Diverse Testset of 2,663 publicly-available M. tuberculosis samples which have both genetic and drug susceptibility testing (DST) data. We have chosen to apply the catalogue such that our tool will return a result of (i) Fail if there are insufficient reads at a genetic locus associated with resistance, (ii) Unknown if a genetic variant in a resistance gene not listed in the catalogue is encountered and (iii) Resistant if three or more short-reads support the presence of a resistance-associated variant. The last step increases the sensitivity for all 15 antibiotics but only reaches significance in a few in our testset. Comparing our results to those of TB-Profiler, an existing tool, highlights the different design choices and demonstrates the performance of both tools on our Diverse Testset is comparable. By only considering high confidence DST results we show that gnomonicus, in combination with our translation of WHOv2, achieves sensitivities and specificities in excess of 95% for both isoniazid and rifampicin. Impact StatementWhole genome sequencing clinical samples taken from patients with tuberculosis is a potentially fast and accurate method for determining to which antibiotics the infection will be susceptible. Two barriers need to be overcome; the first, which is knowing which mutations are associated with resistance (or not) to a range of antibiotics is well on the way to be solved thanks to the efforts of the World Health Organization (WHO) who have published extensive catalogues containing lists of such mutations. The second barrier is that the processing of the raw genetic files remains a largely manual process overseen by bioinformaticians. Here we describe gnomonicus, our open-source AMR prediction tool, and report the performance of our translation of the second edition of the WHO catalogue using a carefully designed publicly-available dataset of 2,663 M. tuberculosis samples. We hope that not only will this tool be useful but also that this dataset will be used by other researchers to facilitate comparisons between pipelines, approaches and tools. Data SummaryThe attendant GitHub repository1 allows gnomonicus to be rerun on all 2,663 samples in the Diverse Testset; it therefore includes instructions, the necessary input files (all the variant call files, the version of the WHOv2 catalogue and a link to the H37Rv GenBank file used in this study) and also the output JSON files. The ENA accession numbers for all 2,663 samples (including a bash script to download them) and their corresponding phenotypic drug susceptibility testing results are included with the intention that people can either reproduce our results, or use the same dataset for other analyses. The JSON files output by TB-Profiler have also been added for comparison. The repository contains a series of Juypter notebooks containing Python3 code that allows the user to discover and parse the output JSON files from either tool and save the results as data tables. Other notebooks allow the user to reproduce all the analysis underlying this work, including reproducing the figures and many of the tables.

Tuberculosis (TB) remains a leading cause of infectious disease mortality, and the continued emergence of drug-resistant Mycobacterium tuberculosis (MTB) strains threatens the effectiveness of standard treatment regimens. Culture-based antibiotic susceptibility testing (AST) remains the clinical reference standard for resistance determination but typically requires six to eight weeks, delaying initiation of optimized therapy for patients with drug-resistant disease. Whole-genome sequencing (WGS)-based approaches provide a rapid alternative for predicting antimicrobial resistance directly from genomic data and are increasingly being incorporated into diagnostic workflows. This survey reviews computational approaches for genomic resistance prediction in MTB, focusing on two major classes of methods: catalog-based tools that identify established resistance-conferring variants, and de novo machine learning approaches that infer resistance from genome-wide sequence features. We examine the strengths and limitations of these approaches with respect to interpretability, scalability, computational requirements, and concordance with phenotypic testing. We further discuss emerging directions in quantitative minimum inhibitory concentration (MIC) prediction, challenges in pyrazinamide susceptibility testing, and the limited availability of resistant isolates for newer and repurposed drugs used in multidrug-resistant TB (MDR-TB) and extensively drug-resistant TB (XDR-TB) treatment regimens. Continued expansion of paired phenotypic and genomic datasets, standardized MIC testing protocols, and rigorous lineage-aware evaluation frameworks will be essential for improving the clinical reliability and global deployment of genomic resistance prediction for tuberculosis diagnostics.

Background: Tuberculosis, especially drug-resistant tuberculosis (DR-TB) including multidrug-resistant (MDR) and extensively drug-resistant (XDR) strains, remains a leading cause of infectious death worldwide. The rapid accumulation of whole-genome sequencing (WGS) data had spurred numerous computational methods for predicting antimicrobial resistance in Mycobacterium tuberculosis. However, heterogeneous datasets, preprocessing pipelines, and evaluation protocols have made fair comparisons impossible and have hindered clinical translation. A critical yet missing resource is a large-scale, unified benchmark to systematically assess and compare existing methods. Methods: We curated an integrated MTB WGS--phenotypic drug susceptibility testing (pDST) dataset from three sources: the CRyPTIC dataset (Comprehensive Resistance Prediction for Tuberculosis: an International Consortium), a published multi-study compilation, and newly curated literature-derived datasets. The final benchmark contains 54,364 paired WGS-pDST records with broad geographic, lineage, and drug coverage. After harmonizing phenotypes and generating standardized variant features, we evaluated seven models (including classical machine learning and deep learning architectures) across 18 drug-level and six clinical resistance category prediction tasks. Results: XGBoost achieved the highest mean drug-level AUPRC (0.674) and F1-score (0.620) and ranked first in AUPRC for 11 of 18 drugs, whereas WDNN achieved the highest mean AUROC. Random forest yielded the highest mean specificity (0.956) and accuracy (0.933), whereas logistic regression achieved the highest mean recall (0.774), highlighting distinct clinical trade-offs. Drug-level difficulty was highly heterogeneous: rifampicin and isoniazid were predicted robustly, whereas bedaquiline, delamanid, linezolid, and clofazimine remained persistently difficult. In clinical resistance category evaluation, RR-TB, MDR-TB, and pan-susceptibility were well predicted, but XDR-TB and other resistance categories constituted major bottlenecks. Conclusions: Under the largest unified benchmark to date, classical machine-learning methods, particularly XGBoost, provided the strongest precision--recall and F1 performance overall, while neural models remained competitive by AUROC. Emerging drugs (bedaquiline, delamanid, linezolid, clofazimine) and XDR cases remain persistently difficult to predict, identifying key bottlenecks for future method development. This benchmark can serve as a community standard for evaluating MTB resistance prediction and the provided evaluation pipeline offers an actionable baseline for regulatory qualification and clinical decision support system validation, accelerating the translation of WGS-based resistance prediction into practice.

There remains a clinical need for better approaches to rapid drug susceptibility testing in view of the increasing burden of multidrug resistant tuberculosis. Binary susceptibility phenotypes only capture changes in minimum inhibitory concentration when these cross the critical concentration, even though other changes may be clinically relevant. We developed a machine learning system to predict minimum inhibitory concentration from unassembled whole-genome sequencing data for 13 anti-tuberculosis drugs. We trained, validated and tested the system on 10,859 isolates from the CRyPTIC dataset. Essential agreement rates (predicted MIC within one doubling dilution of observed MIC) were above 92% for first-line drugs, 91% for fluoroquinolones and aminoglycosides, and 90% for new and repurposed drugs, albeit with a significant drop in performance for the very few phenotypically resistant isolates in the latter group. To further validate the model in the absence of external MIC datasets, we predicted MIC and converted values to binary for an external set of 15,239 isolates with binary phenotypes, and compare their performance against a previously validated mutation catalogue, the expected performance of existing molecular assays, and World Health Organization Target Product Profiles. The sensitivity of the model on the external dataset was greater than 90% for all drugs except ethionamide, clofazimine and linezolid. Specificity was greater than 95% for all drugs except ethambutol, ethionamide, bedaquiline, delamanid and clofazimine. The proposed system can provide quantitative susceptibility phenotyping to help guide antimicrobial therapy, although further data collection and validation are required before machine learning can be used clinically for all drugs.

Antimicrobial resistance (AMR) poses a significant public health challenge, particularly in resource-limited settings such as Zimbabwe, where surveillance systems are often underdeveloped. This study aims to characterise AMR patterns at the Gweru provincial hospital (GPH) and evaluate machine learning (ML) models for predicting resistance to enhance surveillance. This retrospective cross-sectional study comprised 4 054 clinical isolates from 874 patient records (2022-2024). Five ML models, namely, support vector machine (SVM), random forest, logistic regression, gradient boosting, and k-nearest neighbors (KNN), were trained and evaluated, focusing on predictive performance for surveillance purposes. Among all evaluated models, SVM achieved the highest accuracy (72.08%), precision (73.25%), recall (79.78%), F1 score (0.76), and AUC-ROC (0.79), indicating it as the most effective model for AMR surveillance in this study. Feature importance analysis revealed that antibiotic class, hospital ward, patient age, and pathogen type were significant predictors of resistance. Notably, resistance was high for tetracycline (72.1%) and nitrofurantoin (75.7%), whereas imipenem (7.7%) showed the lowest resistance rates. Multidrug resistance was high among S. aureus (30%), whereas Shigella spp. and Serratia marcescens showed no multidrug resistance. This study highlights the significant AMR burden in Gweru and demonstrates the potential of ML, particularly SVM, for use in predictive surveillance. These findings support targeted interventions in high-risk hospital wards against specific pathogens, offering a scalable approach to AMR monitoring in resource-limited settings.

We present BIOTIA-DX RESISTANCE (BDXR), our submission to the CAMDA 2026 AMR Challenge. This work extends our CAMDA 2025 submission [1] to a new set of six species-drug pairs and adds k-mer-based feature engineering (both targeted and whole-genome) for pairs where the 2025 gene-presence base model underperforms. BDXR achieved a mean accuracy of 86.1% across the six pairs on the CAMDA 2026 test set, ranking first on four pairs, tied for first on Streptococcus pneumoniae (penicillin), and second on Campylobacter jejuni (nalidixic acid); per-pair test accuracy ranged from 69.9% (C. jejuni, nalidixic acid) to 98.8% (S. pneumoniae, penicillin). We refer the reader to our 2025 preprint [1] for the underlying workflow, dataset curation, and clinical motivation; this preprint focuses on the results and methodological changes that are new in 2026.

BackgroundFluoroquinolones, while clinically indispensable, carry underappreciated cardiovascular risks, particularly QT prolongation and life-threatening arrhythmias. Emerging evidence suggests geographic and genetic variations in susceptibility, yet Middle Eastern populations remain underrepresented in global pharmacovigilance datasets. ObjectiveThis study investigates the prescribing trends and awareness of fluoroquinolone-related adverse effects among healthcare providers in the UAE using a multimodal combination of artificial intelligence (AI) integrating pharmacovigilance data, environmental exposure mapping, predictive ECG analytics and natural language (NLP) of electronic health records (EHRs) MethodsWe conducted a retrospective cohort study (2018-2023) combining structured ADR reports from UAE MOHAP, WHO-VigiAccess, FAERS, and EMA with unstructured clinical narratives. A hybrid NLP pipeline (BioBERT-based NER, sentiment analysis, and relationship extraction) identified unreported risk patterns. Machine learning (Random Forest, SVM, BioBERT-NLP) stratified high-risk cases, validated against MIMIC-IV ECG waveforms. Geospatial modeling correlated wastewater fluoroquinolone levels with regional arrhythmia incidence. ResultsAmong 1,522 adjudicated ADRs, moxifloxacin demonstrated the strongest cardiotoxicity signal (OR=1.45, 95% CI 1.2-1.8, *p*<0.001), with AI-ECG models detecting subclinical torsades de pointes at 96% sensitivity (AUC 0.97). NLP revealed significant ECG monitoring disparities in Northern Emirates (under documentation rate: 43%). Environmental analyses identified a dose-dependent relationship between moxifloxacin water contamination and arrhythmia hospitalizations (+22% in high-exposure regions, *p*=0.01). Molecular dynamics simulations implicated C7 substituent modification as a viable strategy to reduce hERG channel binding. ConclusionWe integrated multi-omics analysis with pharmacovigilance mining to stratify cardiotoxic risk among fluoroquinolone users in the UAE bridging pharmacovigilance, environmental epidemiology, and structural pharmacology. Our framework enables precision monitoring through AI-ECG integration, policy interventions targeting high-risk prescribing, and drug redesigning to mitigate hERG liability.

The evolution of {beta}-lactamase proteins is shaped by the need to maintain enzymatic activity against previously prevalent {beta}-lactam antibiotics while expanding substrate range against new classes of antibiotics. Using saturation mutagenesis and sequence-barcoding-based quantification, we comprehensively mapped the response of the fitness landscape of TEM-1 {beta}-lactamase, which evolved against penicillin-class antibiotics, to mutational perturbations against six diverse {beta}-lactam antibiotics. This systematic panel of antibiotic substrates, including representatives from penicillin, cephalosporin, and monobactam classes, allowed us to classify resistance mutations into two categories. Generalist mutations conferred resistance to multiple antibiotics and were consistently restricted to three positions critical for substrate recognition and catalytic function R164, G238, and E240. These substitutions produced broad spectrum resistance through mechanisms such as expansion of the active site and improved substrate accommodation. In contrast, specialist mutations conferred resistance to only a single antibiotic and exhibited much wider positional diversity. Ceftazidime selection yielded the greatest number of distinct specialist mutations, which were frequently found in flexible or peripheral regions including the omega loop. One especially unexpected finding was the identification of the E166P variant. E166 is a catalytic residue required for deacylation during hydrolysis, and substitutions at this site are generally assumed to abolish function. However, E166P conferred a significant increase in ceftazidime resistance despite eliminating activity against penicillins. Molecular dynamics simulations and mutational analysis revealed that the E166P mutant employs an alternative catalytic mechanism, involving residue N132, rather than the canonical pathway. Together, our findings reveal, at the molecular level, how specialist mutations open up a wide range of diverse and idiosyncratic solutions at the expense of generalizability. These insights may inform strategic design of antibiotic administration protocols to systematically lower pathogenic evolvability.