npj Antimicrobials and Resistance

The rapid rise of multidrug-resistant (MDR) bacterial infections has severely limited treatment options, particularly for Gram-negative pathogens such as Klebsiella pneumoniae, a leading contributor to pneumonia, bloodstream, urinary tract, and surgical-site infections. One strategy to restore antibiotic efficacy is the use of resistance-mitigating agents (RMAs), compounds that re-sensitize bacteria to existing antibiotics without displaying independent antibacterial activity. Herein, we report the results of a high-throughput screen of a 3,200-compound fragment-based library against an MDR K. pneumoniae isolate in the presence of subinhibitory ciprofloxacin. This screen identified a tetrahydrocarbazole-containing compound, 1, as a ciprofloxacin potentiator. Subsequent structure-activity relationship studies yielded a difluorinated analog, compound 5, which potentiated multiple antibiotic classes in MDR K. pneumoniae, reducing MICs up to [≥]16 fold. Further testing demonstrated synergistic interactions between compound 5 and ciprofloxacin, ceftriaxone, cefoxitin, and tetracycline across four genetically diverse MDR K. pneumoniae strains. These findings suggest that tetrahydrocarbazole-containing compounds constitute a promising new class of RMAs with potential for future development as therapies against MDR K. pneumoniae infections.

Antimicrobial resistance generates a vast, rapidly growing literature, yet no resource offers a comprehensive, evidence-linked repository of AMR findings at scale. We present ResLit, an automated pipeline and public database that mines the AMR literature for resistance genes, mutations, organisms, and mechanisms. From 2 million candidate PubMed records, BioMistral-7B screened abstracts to 356,000 relevant papers; multi-tier retrieval yielded 117,000 full texts, from which Qwen3-30B performed two-step extraction. ResLit contains 3,120 genes and 13,593 mutations, cross-linked to CARD, ResFinder, and NCBI Reference Gene Catalog across four evidence tiers. It further supports community-driven curation of automated outputs and reference databases. Freely available at www.reslit.info.

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.

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.

Klebsiella pneumoniae is a WHO critical-priority pathogen for which strain-specific bacteriophages are being explored as precision antimicrobials, yet rapid phage-host matching remains a major barrier to therapeutic deployment. We constructed a comprehensive interaction atlas comprising 84 taxonomically diverse phages and 101 globally sourced, clinically representative K. pneumoniae strains, including multidrug-resistant isolates. Systematic pairwise profiling produced 8,484 interaction measurements, of which 2,656 (31.3%) scored positive for bacterial clearance. Genus was the dominant phage-side determinant of host range, while capsule K-serotype was the strongest host-side determinant of susceptibility; aggregate defense, prophage, plasmid, and antimicrobial-resistance features contributed comparatively little. A genome-guided machine learning model predicted interactions without curated host annotations (AUROC, 0.882; AUPR, 0.765), outperforming a model based only on phage genus and K-serotype and modestly exceeding a curated genomic baseline. The model recovered capsule- and lipopolysaccharide-biosynthesis genes, canonical receptors and defense-associated features as major predictors using SHAP analysis. Feasibility tests of expert- and model-selected cocktails exposed a translational constraint. Although all formulations suppressed growth in vitro, only the specific cocktail whose phages replicated robustly within the murine gut reduced colonization, suggesting in vivo amplification rather than predicted host range as the limiting factor for therapeutic efficacy. Together with the activity of a model-selected cocktail built for an isolate completely excluded from training, these results provide a species-wide resource for K. pneumoniae phage matching and support a hybrid workflow combining genome-based ranking with targeted phenotypic validation.

Nitrofurantoin is commonly prescribed as a first-line treatment for urinary tract infections (UTIs) and is effective against most uropathogens; however, resistance among Klebsiella spp. remains high. In this study, we investigated synergistic drug combinations--defined as pairs of drugs whose combined effect exceeds that of each agent alone--to enhance treatment efficacy against drug-resistant Klebsiella pneumoniae. A high-throughput drug screen identified six and two small molecules that exhibited >50% synergistic activity in tested strains in combination with nitrofurantoin and ciprofloxacin, respectively. We further validated the top three candidates using nitrofurantoin-susceptible and resistant clinical isolates. Notably, the combination of nitrofurantoin and dequalinium demonstrated strong synergistic activity, observed in 68% of susceptible Escherichia coli strains and 94% of resistant Klebsiella pneumoniae strains tested. Combined inhibition of the citric acid cycle by nitrofurantoin and F1-ATPase by dequalinium resulted in a significant reduction in ATP levels compared to either treatment alone. Importantly, decreased ATP levels did not increase persister cell formation, and dequalinium alone reduced persister cell populations more effectively than nitrofurantoin. However, nitrofurantoin is processed into a poorly understood reactive intermediate whose specific metabolic targets have not been fully elucidated. Using metabolomic analyses, we identified aconitase and isocitrate dehydrogenase--key enzymes in the citric acid cycle--as altered in response to nitrofurantoin. Together, these findings demonstrate that dual targeting of bacterial metabolism by nitrofurantoin and dequalinium represents a promising therapeutic strategy for treating drug-resistant Klebsiella spp. in UTIs.

We previously identified a series of heterobenzamides (HBAs) with potent growth inhibitory activity against Mycobacterium tuberculosis in axenic culture. We also provided evidence that these target QcrB, a component of the terminal cytochrome oxidase in the electron transport chain. We expanded our studies to look at the full microbiological profile: key molecules from the series were tested for activity under different conditions and against additional strains. HBA analogs were active against intracellular bacteria where they exhibited bacteriostatic activity. A strain of M. tuberculosis with a mutation in QcrB (T313I) was resistant to HBAs in both axenic culture and inside macrophages. HBAs retained potency against lineages and mono-resistant strains of M. tuberculosis. HBAs had a narrow spectrum of activity, since they were not active against the ESKAPEE pathogens. Combination of the key HBA with bedaquiline was synergistic, as expected for a QcrB inhibitor, but there was no strong synergy with other drugs. Exposure of M. tuberculosis to the key HBA led to ATP depletion and boosted the oxygen consumption rate. This effect was specific to M. tuberculosis, since human THP-1 macrophage-like cells were unaffected by exposure to the HBA. HBA did not induce the production of reactive oxygen species or affect membrane potential but did affect pH homeostasis. Taken together, these data provide further evidence to support the identification of QcrB as the target and indicate that they are suitable for further drug development.

MotivationIdentifying bacterial antimicrobial resistance (AMR) is critical for diagnostics and treatment, but resistance is a complex trait arising from myriad mechanisms spanning multiple molecular scales. Existing computational approaches often function as black boxes and rarely explore cross-species or multi-drug patterns. We developed amR, an integrated R package suite that provides a complete framework from bacterial genome data curation to interpretable AMR predictions, enabling identification of resistance mechanisms across species and drugs. ResultsThe amR R package suite contains three modular packages. amRdata downloads genomes and paired antimicrobial susceptibility testing data from BV-BRC and processes them, constructs pangenomes, and extracts features at gene/protein cluster, protein domain, annotated Clusters of Orthologous Groups and ResFinder AMR-associated features, and structural variant scales; data are stored in memory-efficient formats (Parquet, DuckDB). amRml trains interpretable machine learning models per species-drug combination, calculates feature importance and performance metrics, and provides rich ground for hypothesis generation and mechanism discovery. amRviz provides an interactive Shiny dashboard to explore metadata distributions and model performance across species and drugs, visualize top predictive AMR features, and analyze cross-model patterns across geographic/temporal strata. We apply the suite to Shigella sonnei, achieving a median Matthews Correlation Coefficient of 0.89 across 23 drugs and drug classes. With thousands of genomes, multi-scale features, and interpretable models, amR provides an accessible, comprehensive framework for AMR research. The amR package suite is installable via GitHub (https://github.com/JRaviLab/amR; BSD-3-Clause license).

With the rise of antimicrobial resistance, urinary tract infections (UTIs) have become increasingly more difficult to treat, prompting renewed interest in bacteriophage (phage) therapy as an alternative or adjunct to antibiotics. UTIs are an attractive target for phage therapy because they generate a high density of actively replicating bacteria that supports phage propagation, and because the urinary tract is readily accessible for administration and monitoring. Yet studies of phage therapy for UTIs report mixed outcomes, including failures to meet clinical and microbiological endpoints. Here we follow the population dynamics of a clinical Escherichia coli UTI strain and two phages, HP3 and ES19, to which the strain appears susceptible by standard testing. Despite this apparent susceptibilty, both phages fail to suppress the strain, with resistance emerging almost immediately. Using the measured mutation rate, our mathematical model shows that traditional resistance cannot account for these dynamics. We instead demonstrate, including by a phage-specific population analysis profile assay we developed, that heteroresistance drives this rapid failure, offering a plausible explanation for treatment failures in UTI phage therapy

In this technical report, we introduce NO_SCPLOWESSOC_SCPLOW-1, a coarse-grained cofolding framework for binding- affinity prediction. NO_SCPLOWESSOC_SCPLOW-1 requires[~] 1 second per prediction on a single GPU. This offers more than one order of magnitude speed-up over the leading open-source baseline, Boltz-2, which significantly expands the regions of chemical space that can be explored during high-throughput virtual screening. Importantly, NO_SCPLOWESSOC_SCPLOW-1 matches or surpasses the accuracy of Boltz-2 over the same benchmarks adopted in their study--which we show reflect in-distribution scenarios--as well as over more challenging out-of-distribution data encompassing the OpenBind affinity benchmark and 25 internal biochemical assays. Notably, NO_SCPLOWESSOC_SCPLOW-1 maintains robust predictive accuracy even on assays with extremely low similarity to the training data. Moreover, we highlight examples where NO_SCPLOWESSOC_SCPLOW-1 demonstrates meaningful selectivity, separating the binding affinities of identical compounds between on-targets and related off-targets. Nonetheless, zero-shot generalization to real- world medicinal chemistry remains an inherently challenging task; consequently, we acknowledge specific assays where the models performance is limited. We open-source NO_SCPLOWESSOC_SCPLOW-1: code and weights are available at https://github.com/recursionpharma/nesso

The rise of antibiotic resistance necessitates the discovery of antibacterial compounds with novel mechanisms of action (MoAs). Recent machine learning approaches have shown promise in antibacterial compound discovery, but often identify derivatives of known antibiotic classes rather than mechanistically novel compounds. Previous approaches applied Tanimoto similarity filters at the end of screening pipelines, but this method has substantial drawbacks: Tanimoto similarity can be misleading in chemical space, and post-hoc filtering does not influence what activity models learn to prioritize. Here, we present a machine learning pipeline that addresses chemical novelty upfront by employing an XGBoost-based MoA classifier to explicitly prioritize compounds predicted to have mechanisms distinct from known antibiotic classes, combined with graph neural networks for antibacterial activity and toxicity prediction. Applied to the Zinc20 database, our approach successfully identified non-toxic antibacterial compounds structurally distinct from known antibiotics. Notably, the majority of these hits exhibited membrane-targeting activity with selectivity for bacterial cells over mammalian cells, suggesting potential for next-generation membrane-active antibiotics. However, we did not identify compounds with novel protein targets. Systematic analysis revealed that this limitation stems from mechanistic bias in training data rather than model architecture. Specifically, our activity model learned to preferentially score compounds similar to specific groups in the training data, thus overrepresenting certain MoA classes including membrane-active compounds. Even substantial model architecture and training data enhancements did not overcome this constraint. Our findings demonstrate that the primary bottleneck for discovering mechanistically novel antibiotics is the scarcity of diverse, mechanistically-annotated training data. This work provides both a methodological framework for mechanism-aware screening and critical insights into data requirements for genuinely novel antibiotic discovery.

The New Delhi metallo-{beta}-lactamase (NDM) is a major determinant of carbapenem resistance. This has prompted the development of novel therapeutic strategies to treat NDM producers. Since these therapies impose new selective pressures, their clinical deployment may favor NDM variants carrying escape mutations. To anticipate these events and inform future therapeutic decisions, we explored the evolutionary landscape of NDM-1 under different selective constraints. To this end, we generated a highly diverse library of blaNDM variants and challenged it with zinc starvation, antibiotics or {beta}-lactam/{beta}-lactamase inhibitor combinations that represent novel and emerging therapies. Selection under zinc limitation and cefotaxime identified mutational trajectories that recapitulate clinical NDM evolution, validating the diversity of the library and its predictive nature. Drug-specific selections revealed sharply different evolutionary pathways. Mecillinam selected a narrow evolutionary pathway centered on residues N220 and M67 which enhance productive active-site interactions with this penicillin. Cefepime/taniborbactam selected multiple escape routes, dominated by substitutions at E152 and, secondarily, K211, that impair productive interaction with taniborbactam. In contrast, cefiderocol/xeruborbactam and aztreonam/avibactam failed to select NDM variants conferring an improved resistance phenotype. These results show that NDM evolution is constrained by the chemistry of each therapeutic challenge. Substrate adaptation is possible for mecillinam, inhibitor escape is readily accessible for taniborbactam, whereas aztreonam- and xeruborbactam-based strategies impose high evolutionary barriers on NDM. Mapping drug-specific evolutionary landscapes can help anticipate resistance before clinical deployment and prioritize therapeutic strategies less likely to drive NDM-mediated escape. ImportanceCarbapenem-resistant infections by Enterobacterales are increasingly difficult to treat because bacteria can destroy some of the most powerful antibiotics used in the clinic. This study focuses on Escherichia coli carrying variants of the New Delhi metallo-{beta}-lactamase, an enzyme that enables bacteria to resist carbapenem antibiotics. New therapies are being developed to overcome this problem, but bacteria may evolve again when exposed to these treatments. Here, we tested how New Delhi metallo-{beta}-lactamase can adapt under the evolutionary pressure of last-resort therapies. The results show that not all treatments carry the same evolutionary risk. Some drugs allow the enzyme to adapt through specific mutations, whereas other drug combinations make such escape much harder. This work helps predict which treatments are more likely to remain effective and which may more readily select resistance. This information can guide the design and use of future therapies against resistant bacterial infections.

Antimicrobial resistance is an evolutionary response to antimicrobial exposure that has been extensively studied across some bacteria, including pathogens and model organisms. Yet, for most species the capacity to develop resistance remains unresolved. Here, we used computational methods to assess patterns of streptomycin resistance evolution across the bacterial tree of life. We curated a panel of high-confidence streptomycin resistance mutations, including eight mutations in the rpsL gene and four mutations in the rrs gene. We then used this panel to screen over 20 000 bacterial genomes from diverse clades. We assessed both evolvability, defined by codon-level accessibility to resistance-conferring mutations via single-nucleotide substitutions, and intrinsic resistance, where resistance-associated variants are already present. Our results suggest that most bacterial species can readily acquire rpsL-resistant mutations. Furthermore, we find that approximately 7% of bacterial species intrinsically carry rpsL resistance variants, with a wide taxonomic distribution but notable enrichment within Alphaproteobacteria. Our study provides a global view of the streptomycin resistance mutational landscape and generates testable predictions for future research.

Understanding how molecular structure encodes biological function remains a grand challenge in drug discovery. Here, we present PubCheF-1, a deep learning model that predicts literature-derived biological function directly from chemical structure. PubCheF-1 was trained on a dataset linking molecules to labels derived from the scientific articles in which they appear, a strategy that connects disparate compounds through the language used to describe their functionalities. When tasked with identifying inhibitors of {beta}-lactamases, including enzymes considered largely refractory to inhibition, PubCheF-1 predicted structurally distinct compounds that collectively have activity against all {beta}-lactamase classes. Furthermore, hit compounds directly bind the enzyme active site, restore antibiotic efficacy in multidrug-resistant high-priority pathogens, and demonstrate potent activity in animal infection models. Together, these findings establish that machine learning-based prediction of biological function derived from the language of scientific literature allows the identification of bioactive molecules at high hit rates, thereby accelerating therapeutic discovery.

Ribosomally synthesized host defense peptides (HDPs) are promising candidates for novel antibiotics. However, non-lytic HDPs, which target intracellular machinery, remain underexploited due to limitations including low potency in serum and narrow activity spectra. To enhance their therapeutic profile, we fused non-lytic HDPs generating "chimerophores" with multimodal mechanisms of action (MOAs). Using a high-throughput self-screening platform (Mex), we synthesized and evaluated a combinatorial library of 99,235 variants, identifying over 30,300 active chimerophores, vastly expanding the functional space of chimeric HDPs. Functional screening of 18 chimerophores revealed candidates with potent, broad-spectrum activity displaying serum-tolerance, low cytotoxicity, orthogonal uptake pathways and multimodality, such as simultaneously targeting of ribosomes and DNA. Integrating these distinct mechanisms into a single molecule allowed lead candidate cp9 to suppress the emergence of resistance in Pseudomonas aeruginosa, establishing a scalable platform for the systematic engineering of next-generation multimodal antibiotics.

Small open reading frames (sORFs) are a potentially rich, yet error-prone, source of antimicrobial-peptide (AMP) candidates: short sequences are readily prioritized by AMP classifiers but may derive from incomplete gene calls. We developed a genome-context-aware discovery workflow that separates AMP-like sequence properties from evidence for a complete, recurrent coding locus. From 649,653 RefSeq assemblies representing 327 clinically relevant bacterial species, species-aware clustering and length filtering yielded 4,442,548 representative 10-100-aa sequences. AmpScanner v2, Macrel and AMPlify identified 585 non-haemolytic records supported by all three models. However, genome-context auditing of 11,918 mapped candidates showed that 529 of 536 mapped consensus candidates were supported exclusively by partial ORFs near contig termini. By contrast, 3,382 candidates had at least one complete non-edge occurrence; 1,069 recurred in [≥]2 assemblies and 251 in [≥]10 assemblies. We therefore assembled a 20-peptide panel through two explicitly labelled routes: sequence/structure-led selection (n=8) and genome-supported selection (n=12). Broth microdilution against Escherichia coli ATCC 25922 and Staphylococcus aureus ATCC 25923 identified low-micromolar activity in both routes. CAND_04141, a recurrent complete non-edge candidate, had the strongest combined profile (MICs of 4 and 2 M, respectively), while CAND_07825 and CAND_04265 were also active at low micromolar concentrations. In plate-count MBC assays, all three advanced peptides achieved [≥]3-log10 reductions at 128 M. These findings show that high classifier agreement is not a substitute for genomic evidence and provide an auditable framework for prioritizing both synthetic AMP-like sequences and candidate genome-encoded peptides.

Aminoglycosides are critical antibiotics with partially elucidated mechanisms of uptake and action in Gram-negative bacteria. Importantly, although energy-dependent uptake across the inner membrane has been well-established, the specific molecular mechanisms involved have not been definitively identified. To deepen understanding of genetic factors influencing susceptibility and resistance to streptomycin, we applied transposon insertion sequencing in Escherichia coli K-12. This approach identified both known and novel genes whose disruption increased susceptibility, including those involved in respiration, protein export, cell division, and uncharacterised functions. Notably, voltage-sensitive membrane dye-based assays revealed that many susceptible mutants did not display inner membrane hyperpolarisation as often assumed. Conversely, disruption of certain genes, such as the inner membrane antimicrobial peptide transporter sbmA, conferred low-level resistance, with sbmA overexpression increasing streptomycin sensitivity, suggesting its role in aminoglycoside uptake. These findings refine the model of aminoglycoside interaction with various pathways and highlight potential targets for adjuvant therapies to combat antimicrobial resistance.

The global spread of antimicrobial-resistant (AMR) bacteria represents one of the greatest threats to modern medicine, making the development of novel therapeutic strategies increasingly urgent. Phage therapy, which exploits bacteriophages (phages, viruses that specifically infect and kill bacteria) has regained attention as a therapeutic approach for multidrug-resistant infections. One critical determinant of treatment outcome is the capacity of phages to sustain bacterial growth suppression; however, no metric has previously existed to directly quantify the duration of effective lytic activity. Here, we propose the Sustainability Index (SusI), a novel metric that quantifies both the duration and extent of phage-mediated bacterial growth suppression, which is restricted to the primary lysis period from lysis initiation to resistance emergence. Evaluation of individual phages and two-phage cocktails against both laboratory and clinical isolates of Escherichia coli demonstrated that SusI provides information independent of the Virulence Index, which primarily reflects bactericidal activity during the initial phase of infection, and serves as a complementary metric to the Suppression Index, which may incorporate behavior beyond primary lysis. Cocktails composed of phages targeting different receptors specificities consistently exhibited higher SusI values, consistent with the notion that multifaceted selective pressure delays resistance emergence. Furthermore, in a mouse model of systemic infection established by intraperitoneal administration, cocktails with higher SusI values demonstrated superior therapeutic efficacy. These results confirm SusI as a practical metric for rational phage cocktail design. As phage therapy advances toward clinical implementation, standardized quantitative metrics such as SusI are expected to facilitate evidence-based selection of therapeutic phages across diverse pathogens and infection conditions. ImportanceThe global spread of antimicrobial-resistant bacteria is making bacterial infections increasingly difficult to treat. Phage therapy, which uses bacteriophages (viruses that specifically infect bacteria), has re-emerged as a therapeutic alternative; however, reliable methods to determine in advance which phages will be therapeutically effective remain limited. Current evaluation metrics are well-suited for quantifying how rapidly phages kill bacteria but were not designed to directly measure how long lytic activity is sustained before resistant bacteria emerge. Here, we developed the Sustainability Index (SusI), a novel metric that specifically quantifies the duration of effective bacterial growth suppression. Evaluation of multiple phages and their combinations (cocktails) against both laboratory and clinical bacterial isolates demonstrated that SusI can distinguish phage combinations that existing metrics fail to differentiate. Moreover, in a mouse model of lethal bacterial infection, higher SusI values correlated with improved therapeutic outcomes. SusI has potential as a practical tool for selecting phages with greater likelihood of therapeutic success.

Motivation: Antimicrobial resistance (AMR) has been identified as a top global public health threat. Accurate AMR phenotype prediction from whole-genome sequencing data is an essential tool for accelerating clinical decision-making and mitigating resistance spread. Although many previous works have explored the use of tree-based machine learning (ML) models to predict resistance, the field lacks a systematic evaluation of the training pipeline across a variety of pathogenic species and antibiotics. Results: Using nine clinically relevant species-antibiotic combinations from the NCBI antimicrobial susceptibility testing database, we present a detailed analysis of the ML pipeline and identify key factors affecting model performance and evaluation. We begin by relabelling all isolates using current CLSI minimum inhibitory concentration breakpoints to resolve inconsistencies and increase available data, resulting in up to a 19% label swap and 56% data enlargement per species-antibiotic combination. We identify several key training parameters including k-mer length, which can increase classification F1 scores by over 20 points compared to commonly used k-values, feature matrix truncation, which can induce polynomial time reductions with limited performance reduction, and ML model class. By comparing 5-fold cross-validation with evaluation on an unseen clinical dataset, we show that random cross-validation splits--often criticized as overly optimistic--can act as a strong proxy for downstream clinical performance, yielding closer F1 scores than phylogeny-aware splits in all cases. We finally present an interpretability study which shows that over 95% of k-mers used by our models are associated with identifiable genomic features. Our results highlight the importance of feature design, evaluation protocol, and biological analysis in genomic AMR prediction, and support tree-based models as a robust and interpretable method.

Collateral sensitivity (CS) - resistance to one antibiotic inducing hypersensitivity to another - offers an evolutionary trap for multidrug-resistant pathogens. A strongly connected component (SCC) in the directed CS graph is a closed cycle in which every drug is reachable from every other. Prior evidence for CS SCCs is exclusively in vitro. We mined 104,337 susceptibility records from BV-BRC spanning four WHO critical-priority pathogens (Klebsiella pneumoniae, Escherichia coli, Staphylococcus aureus, Pseudomonas aeruginosa; 18,821 isolates), using Fisher's exact tests with Benjamini-Hochberg FDR correction, Tarjan's algorithm, and permutation testing (n = 1,000). Two species yielded qualifying SCCs. In K. pneumoniae (4,286 isolates), a 3-node SCC - imipenem, meropenem, tetracycline - was detected (empirical p = 0.001); both carbapenem-tetracycline edges are bidirectional (OR = 1.81-1.82, q < 0.002, n > 850 per edge). In E. coli (6,720 isolates), a bidirectional 2-node SCC links colistin and cefotaxime (OR = 10.13, 95% CI 2.82-46.12, q = 0.042, n = 87; permutation p = 0.008); with a fragility index of 1, we report it as a hypothesis, not an established effect size. The carbapenem signal is tetracycline-specific: tigecycline shows co-resistance (OR < 0.35), as its distinct RamA/AcrAB-TolC mechanism predicts. ORs of 2.2-2.7 persisted across independent year bands (2009-2014). S. aureus returned no qualifying SCC, but that null is power-limited: only 8% of testable pairs could detect the K. pneumoniae effect size. Prior clinical analyses characterised pairwise and three-way collateral effects; to our knowledge these are the first closed CS cycles identified in clinical surveillance data, motivating experimental follow-up.