Combining machine learning with high-content imaging to infer ciprofloxacin susceptibility in isolates of *Salmonella Typhimurium*

The study addresses the need for faster, more informative antimicrobial susceptibility testing (AST), particularly for Salmonella Typhimurium treated with ciprofloxacin. Conventional AST is slow and growth-based, limiting timely, appropriate therapy. High-content imaging (HCI) can capture single-cell morphological phenotypes and heterogeneity, and prior work has linked morphological changes to antibiotic mechanisms of action. The research question is whether imaging-derived morphological and intensity features, combined with machine learning, can distinguish ciprofloxacin-susceptible from -resistant S. Typhimurium isolates, potentially even without drug exposure, thereby enabling rapid, mechanism-agnostic susceptibility inference.

The paper situates its approach within prior advances: HCI has been used to profile bacterial morphological responses to antibiotics and infer mechanisms of action and to distinguish antimicrobial classes at single-cell resolution. However, links between morphology and AMR phenotypes are less defined, and most rapid AST technologies remain growth-dependent. Few studies have applied machine learning to quantitatively predict antimicrobial MoA or targets from imaging, and none have predicted AMR status without drug exposure. The authors argue that interpretable machine learning coupled with HCI can overcome limitations of dimensionality-reduction-only analyses and black-box models, and may generalize across Gram-negative bacteria that share phenotypic responses to antibiotics like ciprofloxacin.

Design and isolates: Two isogenic laboratory strains (SL1344 and SL1344gyrA) and two clinical S. Typhimurium isolates (D23580, VNS20081) spanning ciprofloxacin MICs (0.015–1.5 µg/ml) were selected; additional 13 clinical isolates plus three of the initial four were later used for external testing (total 16 at 0×MIC-22h). GyrA spontaneous mutants were generated via stepwise selection (nalidixic acid to ciprofloxacin), PCR-confirmed, and whole-genome sequenced. MICs were determined by M.I.C.E./Etest per CLSI. Time-kill and exposure: Cultures were grown in Isosensitest broth and exposed to 0×, 1×, 2×, 4× MIC ciprofloxacin. Time-kill CFU counts were taken at 0, 2, 4, 6, 8, and 24 h (n=3 replicates per isolate). For imaging, samples were collected every 2 h up to 24 h from cultures at each concentration. Imaging and staining: Fixed cells were stained with CSA-Alexa Fluor 647 (membrane), DAPI (DNA), and SYTOX Green (dead/damaged cell permeability). Opera Phenix high-content confocal imaging used a 63× water objective; 40 fields across two planes per well were captured. Two technical wells per condition and 2–3 biological replicates were used. A second experiment imaged 16 isolates at 0×MIC only, at 22 h, with matched culture handling. Image analysis: PerkinElmer Harmony pipeline identified single bacteria and computed 65 morphology, intensity, and texture features per cell. For each well and time point, features were averaged to produce one datapoint and z-score normalized. Data visualization used heatmaps and principal coordinate analysis (PCoA). Machine learning: Random forests (randomForestSRC) with 1,000 trees were trained for: (1) classifying treatment conditions (combinations of time and concentration) on merged data from all four strains; (2) distinguishing SL1344 vs SL1344gyrA at the most discriminative condition (4×MIC-20h); and (3) discriminating resistant vs susceptible isolates at 0×MIC-22h. Feature importance rankings informed selection of top features. A distance ratio metric quantified group separability over time and concentration. Classifiers evaluated on the 16-isolate, 0×MIC-22h dataset included Naïve Bayes, KNN, SVM, random forest, CatBoost, and a neural network, trained on the top features with hyperparameter grid search. Data splitting used random 50/25/25 training/validation/test partitions repeated 1,000 times; metrics reported included accuracy, sensitivity, specificity, precision, F1, and AUC. Model interpretability employed Partial Dependence and ICE plots to relate top features to resistance probability. Genomics and in silico AMR: Illumina sequencing and mapping pipelines identified mutations in gyrA and other loci; ARIBA/CARD and ResFinder annotated AMR determinants. Ethics approvals were obtained for isolate sourcing.

- Growth and morphology under ciprofloxacin: Without drug, all four strains showed similar growth; at 1× MIC, growth slowed variably and rebounded by 24 h. HCI revealed pronounced, time-dependent filamentation (length increase) peaking around 6 h at 1×MIC (mean 9.51 ± 0.20 µm), then decreasing. Strain-specific differences were evident (e.g., D23580 elongated more; SL1344gyrA showed sustained elongation at 24 h). - Feature structure by time and concentration: Heatmaps and PCoA showed clustering primarily by drug concentration (0–1× vs 2–4×) and exposure time, indicating these factors drive morphological variation. A random forest trained to classify treatment conditions achieved an out-of-bag error of 0.25 and identified top features dominated by SYTOX Green (SG) intensity/texture, plus morphology (cell area/length). - SL1344 vs SL1344gyrA discrimination: Maximal separation occurred at 4×MIC-20 h. At this condition, PCoA showed distinct clusters with non-overlapping 95% confidence ellipses, and a random forest achieved OOB error 0 (100% accuracy). Ten key discriminative features (including width-to-length ratio, CSA and SG symmetry/profile features, DAPI radial mean) differed significantly between strains. - Susceptible vs resistant discrimination without drug: The greatest separation of resistant from susceptible isolates occurred at 0×MIC-22 h. PCoA at this condition showed clear segregation. A random forest identified the ten most important features (primarily SG and DAPI intensity/texture, plus CSA profiles), and heatmaps revealed consistent clustering of resistant vs susceptible isolates. - Minimal feature set and classifier performance: Using 16 isolates at 0×MIC-22 h and stepwise feature inclusion, adding many features reduced performance; only five features sufficed for robust classification. With the five most important features, across 1,000 random splits, the neural network outperformed other classifiers with mean (±SD) on test sets: accuracy 0.87 ± 0.08, sensitivity 0.87 ± 0.11, specificity 0.89 ± 0.12, precision 0.90 ± 0.10, F1 0.87 ± 0.08, AUC 0.91 ± 0.07. - Feature-response interpretation: Neural network PDP/ICE analyses suggested resistant isolates exhibit lower SG intensity (indicating altered membrane permeability) associated with higher resistance probability; lower DAPI intensity correlated with lower resistance probability. Higher SG/CSA profile values associated with lower resistance probability, while higher SG radial relative deviation associated with higher resistance probability. Overall, intrinsic morphological and permeability-related features can predict ciprofloxacin susceptibility without exposure.

The findings demonstrate that high-content imaging captures informative, time- and concentration-dependent morphological phenotypes in S. Typhimurium that relate to ciprofloxacin susceptibility. Machine learning models, especially a compact neural network with only five imaging features, can accurately classify isolates as resistant or susceptible without drug exposure or prior knowledge of mechanisms. This addresses a key gap in rapid AST by moving beyond growth-based assays and dimensionality-only analyses toward quantitative, interpretable prediction at single-cell resolution. The approach suggests that QRDR gyrA mutations and other resistance determinants manifest as intrinsic morphological and permeability changes detectable by imaging. Given that many Gram-negative rods exhibit similar morphological responses to antibiotics, the framework may generalize to other pathogens and antimicrobials, enabling family-level imaging databases and broader diagnostic applications. The authors emphasize interpretability through feature importance and PDP/ICE to mitigate black-box concerns and position HCI+ML as a tool for AMR detection, PK/PD studies, drug MoA inference, and potentially vaccine/antibody phenotyping.

This work integrates high-content imaging with machine learning to infer ciprofloxacin susceptibility of S. Typhimurium isolates, even without antibiotic exposure. Key contributions include: (1) establishing that morphological and intensity features, particularly SG/DAPI/CSA-derived metrics, encode susceptibility information; (2) demonstrating robust classification using as few as five features, with a neural network achieving mean test AUC ~0.91; and (3) providing an interpretable analysis pipeline and generalized workflow for broader application. Future research should expand to other species and antibiotic classes, refine image segmentation with deep learning for improved single-cell resolution, validate performance in complex clinical matrices and polymicrobial samples, integrate genomic and imaging data to elucidate genotype–phenotype links, and translate the minimal feature set into practical, cost-effective diagnostic devices for rapid AST.

- Image analysis and segmentation relied on a proprietary pipeline with limited parameter flexibility; improved, specialized algorithms (e.g., U-net/autoencoders) could enhance single-cell segmentation and feature extraction. - Averaging features per well and sampling at discrete time points prevented longitudinal tracking of individual cells, reducing single-cell temporal resolution. - Experiments used pure cultures in controlled laboratory media, limiting immediate clinical applicability to complex samples or polymicrobial infections. - Focus was restricted to S. Typhimurium and ciprofloxacin; generalization to other bug–drug pairs requires further validation and potential method refinement. - Contributions of efflux pumps and porins to resistance phenotypes were not dissected. - Current HCI instrumentation and workflows may be complex and costly for routine clinical labs; simplified assays based on the minimal feature set are needed.