Unlocking the mechanism of action: a cost-effective flow cytometry approach for accelerating antimicrobial drug development

ABSTRACT Antimicrobial resistance is one of the greatest challenges to global health. While the development of new antimicrobials can combat resistance, low profitability reduces the number of new compounds brought to market. Elucidating the mechanism of action is crucial for developing new antimicrobials. This can become expensive as there are no universally applicable pipelines. Phenotypic heterogeneity of microbial populations resulting from antimicrobial treatment can be captured through flow cytometric fingerprinting. Since antimicrobials are classified into limited groups, the mechanism of action of known compounds can be used for predictive modeling. We demonstrate a cost-effective flow cytometry approach for determining the mechanism of action of new compounds. Cultures of Actinomyces viscosus and Fusobacterium nucleatum were treated with different antimicrobials and measured by flow cytometry. A Gaussian mixture mask was applied over the data to construct phenotypic fingerprints. Fingerprints were used to assess statistical differences between mechanism of action groups and to train random forest classifiers. Classifiers were then used to predict the mechanism of action of cephalothin. Statistical differences were found among the different mechanisms of action groups. Pairwise comparison showed statistical differences for 35 out of 45 pairs for A. viscosus and for 32 out of 45 pairs for F. nucleatum after 3.5 h of treatment. The best-performing random forest classifier yielded a Matthews correlation coefficient of 0.92 and the mechanism of action of cephalothin could be successfully predicted. These findings suggest that flow cytometry can be a cheap and fast alternative for determining the mechanism of action of new antimicrobials. IMPORTANCE In the context of the emerging threat of antimicrobial resistance, the development of novel antimicrobials is a commonly employed strategy to combat resistance. Elucidating the mechanism of action of novel compounds is crucial in this development but can become expensive, as no universally applicable pipelines currently exist. We present a novel flow cytometry-based approach capable of determining the mechanism of action swiftly and cost-effectively. The workflow aims to accelerate drug discovery and could help facilitate a more targeted approach for antimicrobial treatment of patients.

bacterial AMR globally (3).AMR is a consequence of natural evolution as well as the improper use of antimicrobials (4,5).One of the few ways to tackle AMR is through the development of new antimicrobials.Nonetheless, the number of new compounds brought to the market is declining (6).Development is expensive, usually in the range of hundreds of millions of US dollars.At the same time, profitability is low, leading to a limited economic incentive to invest in new compounds (7).One important step in the development of new compounds is the elucidation of the mechanism of action (MOA).Both the European Medicines Agency (EMA) and the U.S. Food and Drug Administration (FDA) state that, if possible, the MOA of a novel antimicrobial should be known (8,9).Even though it is not strictly necessary to elaborate on the MOA to obtain approval, it can help in selecting which compounds to investigate further.When insight into the MOA is lacking, clinical trials are more likely to fail, resulting in higher costs.Addition ally, the information can be used to modify molecules of interest to enhance their pharmacokinetic and pharmacodynamic properties (e.g., lower host toxicity) (10)(11)(12).Usually, elucidation of the MOA is part of the pre-clinical development after hit-to-lead compound selection.Costs to determine the MOA can become significant as there are no universally applicable pipelines to do so and expertise in microbiology, genet ics, chemical biology, genomics, and biophysics is required (13).For example, techni ques often used for molecular target identification are transcriptomics-based pattern recognition, metabolomics, FTIR spectroscopy, and specialized proteomics (14)(15)(16)(17).
In general, the MOA of antimicrobials can be classified into a limited number of groups.The first group targets the cytoplasmic membrane and aims to disrupt it.Examples include chlorhexidine and polymyxins.The second group tampers with cell wall synthesis, such as the β-lactams and glycopeptides.Both compounds belonging to the first and second groups are considered bactericidal.The third group is composed of compounds that interfere with nucleic acid synthesis.Here, it is possible to distin guish different subgroups: compounds that target replication (e.g., metronidazole) or compounds that target transcription (e.g., rifampicin).The fourth group is inhibitors of protein synthesis.Again, subgroups are distinguished: the compounds that target the 50S subunit of the bacterial 70S ribosome (e.g., macrolides, clindamycin, and streptogra mins), the ones targeting the 30S subunit of the bacterial 70S ribosome (e.g., tetracycline and aminoglycosides), and the ones that cause premature termination of translation (e.g., puromycin).The last group consists of compounds that inhibit folic acid metabo lism (e.g., sulfonamides and trimethoprim) (18)(19)(20)(21).
The occurrence of a limited number of MOA classes hints at the use of predictive modeling based on the MOA of proven compounds to assess the MOA of new com pounds (15,22,23).One way to do so is to assess the phenotypic state of microbial cells after exposure to different antimicrobials.Flow cytometry is one such technique in which the phenotypic attributes of microbial cells can be assessed in a high throughput manner (24,25).Additionally, the measurement of a single sample is cheap, usually costing less than 1 euro.The information obtained from flow cytometric measurements can be used to build a phenotypic fingerprint of the microbial population and is suitable for the study of phenotypic heterogeneity between samples (26)(27)(28).The discriminative power of phenotypic fingerprints can be increased using adaptive binning approaches, such as PhenoGMM (29).
In this study, we explored the use of flow cytometric microbial fingerprinting to predict the MOA of antimicrobial compounds.We exposed axenic cultures of Fusobac terium nucleatum and Actinomyces viscosus to a wide range of antimicrobials.These bacteria, one Gram-negative and one Gram-positive, are particularly relevant in oral microbiology, where the use of antimicrobials is common for the treatment of disease (30)(31)(32)(33).Additionally, the oral microbiome can be considered a reservoir for AMR (34,35).We performed flow cytometric measurements on the treated bacteria and used the data to construct phenotypic fingerprints based on Gaussian mixture models.In turn, these fingerprints were used to train random forest classifiers.Furthermore, we investigated if the grouping of MOA classes could be observed after treating a complex microbial community with different antimicrobials.Saliva was used for this purpose, as it was considered a logical extension for the used axenic oral bacteria.Finally, trained classifiers were used to predict the MOA of cephalothin, a compound that was not used for training or validating the classifiers (further referred to as "unseen compound").

MATERIALS AND METHODS
A general overview of the workflow is provided in Fig. 1.

Treatment
After 16 h of growth, bacterial cultures were divided into aliquots and supplemented with one antimicrobial compound in triplicate (final volume: 300 µL).An untreated sample was included as a control in triplicate.Consecutively, samples were incubated under anaerobic conditions (90% N 2 , 10% CO 2 ) at 37°C for 3.5 h or 24 h.Table 1 shows tested antimicrobial compounds with the respective concentrations in which they were supplemented with the bacterial samples.
After incubation with the antimicrobials, samples were diluted 100× in sterile and 0.2 µm filtered phosphate-buffered saline (PBS) (PBS tablet, Sigma-Aldrich, Steinheim, Germany) in order to stop the treatment.Additionally, both an untreated sample and a heat-treated (40 min at 110°C) sample without consecutive incubation were included as controls in triplicate.

Saliva sampling and treatment
Saliva was collected from a healthy individual (male, 28 years) by passive drooling for 5 min in a sterile container.The donor was asked to refrain from brushing teeth, using mouth rinses, and eating and drinking for 2 h prior to the donation.Consecutively, saliva was diluted 10× in sterile and 0.2 µm filtered PBS (PBS tablet, Sigma-Aldrich, Steinheim, Germany).The cell concentration of the 10× diluted saliva was (6.6 ± 1.9) × 10 6 cells/mL, as determined by flow cytometry.The diluted sample was then divided into aliquots that were each supplemented with one antimicrobial compound in triplicate (final volume: 300 µL).The supplemented antimicrobials were amoxicillin, amoxicillin-clavulanic acid, azithromycin, ciprofloxacin, clindamycin, erythromycin, and metronidazole and were supplemented in the same concentration as the axenic cultures (Table 1).Non-supple mented 10× diluted saliva served as a control and was included in triplicate.Next, samples were incubated under aerobic conditions at 37°C for 24 h.After incubation, samples were diluted 100× in sterile and 0.2 µm filtered PBS to stop the treatment.The saliva donor signed an informed consent before the donation.The study was approved by the Ethics Committee of the University of Ghent (B6702022000406).

Sample preparation
Directly after stopping the treatment, samples were stained with SYBR Green I/propi dium iodide (SGPI) and analyzed using flow cytometry to determine the phenotypic fingerprint and assess membrane integrity (36).SYBR Green I (10,000× concentrate in DMSO) (Invitrogen, Eugene, USA) and propidium iodide (20 mM) (Invitrogen, Eugene, USA) were diluted 100 and 50 times, respectively, in 0.2 µm filtered DMSO (Merck, Darmstadt, Germany).Samples were then stained with 1% vol/vol SGPI and incubated in the dark at 37°C for 20 min.

Sample measurement
Stained samples were measured using an Attune NxT (Invitrogen, Carlsbad, USA) flow cytometer equipped with a blue (488 nm) and red (638 nm) laser.Performance of the instrument was checked using Attune Performance tracking beads (Invitrogen, Eugene, USA).Only the blue laser was used for the excitation of the stains.A 530/30 nm bandpass filter was used for the detection of green fluorescence (BL1) and a 695/40 nm band-pass filter was used for the detection of red fluorescence (BL3).A 488/10 band-pass filter was used for the detection of forward scatter (FSC) and side scatter (SSC).The flow rate was set to 100 µL/min and stop conditions were set to 100 µL of sample analyzed.The threshold (set on green fluorescence, BL1) and PMT-voltages were determined by using control samples: an untreated sample, a heat-killed sample (110°C for 40 min), sterile BHI, and sterile PBS.

Data analysis
Flow cytometry data were imported in R (version 4.2.0) using the flowCore package (version 2.8.0) (37).Data were transformed using the arcsine hyperbolic function (26), and gated manually on the primary fluorescent channels (BL1 and BL3) to remove background (Supplementary 1).Consecutively, data were normalized by dividing each parameter by the maximum observed value for SYBR Green I fluorescence (BL1) over all treatments within one bacterial strain and treatment duration, for each bacterial strain and for each treatment duration.The same was done for the saliva samples.The application of a Gaussian mixture mask to identify clusters within the flow cytometry data were performed using the "PhenoGMM" function of the Phenoflow package (version 1.1.2)(29).To be able to generate the mask, samples were pooled based on their MOA class and consecutively subsampled to an equal number of cells per MOA class.This was done to avoid biased model training toward a specific MOA class.The Gaussian mixture model (GMM) was optimized using the Bayesian informa tion criterion (BIC) (38,39).The GMM leads to a one-dimensional (1D) vector for each sample that represents the number of cells allocated to each cluster in the model.The parameters upon which the model was built were FSC-H, FSC-A, SSC-H, SSC-A, BL1-H, BL1-A, BL3-H, and BL3-A because these are expected to contain the most information (40).Model output was first converted to relative abundances before further analysis.Converted model output was then used to perform both NMDS (vegan package version 2.6-2) (41) and PCoA (vegan package version 2.6-2) (41) in order to show robustness in observations.Hierarchical cluster analysis was done using the "hclust" function of the stats package (version 4.2.2).For the axenic bacterial cultures, statistical differences between MOA classes were determined through distribution-independent analysis of similarities (ANOSIM) (vegan package version 2.6-2) (41).P values for pairwise compari sons were adjusted using the Benjamini and Hochberg method (stats package version 4.2.2).
Random forest classifiers to predict the mechanism of action were trained on PhenoGMM output of the axenic bacterial strains using the caret package (version 6.0-92) (42).The Matthews correlation coefficient (MCC) was used as a performance metric to optimize the classifiers.The MCC was used to account for class imbalance and was calculated using the mltools package (version 0.3.5)(43).The number of trees was set to 500 for all models.All models were trained using three times repeated fivefold cross-validation scheme.The classification performance of the classifiers was assessed using a fivefold nested three times repeated fivefold cross-validation scheme.Ultimately, classifiers were used to predict the mechanism of action of a compound that was not used in the training and testing data (i.e., cephalothin).

Diversity of antimicrobial-treated samples
Axenic cultures of A. viscosus and F. nucleatum were treated with antimicrobials for 3.5 h or 24 h and consecutively were measured by flow cytometry.Membrane integrity was assessed through manual gating (Supplementary 1).An increase in cell population with damaged membrane was observed with increased treatment time for most antimicro bials for A. viscosus and for some antimicrobials for F. nucleatum (Supplementary 2).Phenotypic fingerprints were constructed by applying a Gaussian mixture mask to the flow cytometric data (Supplementary 3).Grouping according to MOA class could be observed based on the phenotypic fingerprint using NMDS for F. nucleatum and A. viscosus after 3.5 h and 24 h of treatment (Fig. 2).Clear separate groups for MOA classes "Membrane Disruption" and "Cell Wall Synthesis" were visible for A. viscosus, except for polymyxin B, which was separate from the "Membrane Disruption" class.For F. nucleatum, distinctive groups were only observed for the 'Membrane Disruption' class at both treatment durations and for the "Cell Wall Synthesis" class after 24 h of treat ment.Robustness in observations was demonstrated by PCoA for both bacterial strains (Supplementary 4).Grouping was driven by strain when both strains were considered together, especially for the shorter treatment duration (Supplementary 5).For A. viscosus, cluster analysis revealed separate clusters for the classes "Membrane Disruption, " "Cell Wall Synthesis, " and "Protein Synthesis: 50S Inhibition" after 3.5 h of treatment.After 24 h of treatment, only the class "Cell Wall Synthesis" formed a separate cluster.As with the ordination, polymyxin B was not clustering with its respective MOA class at both treatment times.For F. nucleatum, only a separate cluster for the class "Cell Wall Synthesis" could be observed after 24 h of treatment (Fig. 3).
Significant statistical differences between phenotypic fingerprints of different MOA classes were found for both A. viscosus (R = 0. 5625, P = 0.0001 at 3.5 h; R = 0.5135, P = 0.0001 at 24 h) and F. nucleatum (R = 0.4764, P = 0.0001 at 3.5 h; R = 0.3763, P = 0.0001 at 24 h).For A. viscosus, pairwise comparison between MOA classes showed a significant statistical difference for 35 out of 45 pairs after 3.5 h of treatment with antimicrobials and for 34 out of 45 pairs after 24 h of treatment with antimicrobials (P < 0.05; Supplementary 6).For F. nucleatum, 32 out of 45 pairs after 3.5 h of treatment and 18 out of 45 pairs after 24 h of treatment were significantly different (P < 0.05; Supplementary 6).The "Cell Wall Synthesis" class was significantly different from all other MOA classes for both strains for both treatment durations."Membrane Disruption" was different from all other MOA classes except for "Heat" for both strains at both treatment durations, for "DNA Transcription" for A. viscosus after 24 h of treatment and for "Protein Synthesis -tRNA Interference" for F. nucleatum after 24 h of treatment."DNA Replication" and "DNA Transcription" classes were different from most other MOA classes for both strains after 3.5 h of treatment and for A. viscosus after 24 h of treatment.For F. nucleatum, the different "Protein Synthesis" classes could not be distinguished from each other.However, for A. viscosus "30S Inhibition" and "50S Inhibition, " as well as "50S Inhibition" and "tRNA Interference" were significantly different after 3.5 h of treatment."50S Inhibition" and "tRNA Interference" were not significantly different from each other anymore after 24 h of treatment.The "Folic Acid Metabolism" and "Protein Synthesis -tRNA Interference" classes showed the least significant differences from other MOA classes.Moreover, they did not differ from the control, except for F. nucleatum after 3.5 h of treatment and for the "Protein Synthesis -tRNA Interference" class for A. viscosus after 24 h of treatment.
A saliva sample was treated with different antimicrobials for 24 h and measured using flow cytometry.Phenotypic fingerprints were constructed by applying a Gaussian mixture mask to the flow cytometric data (Supplementary 3).NMDS revealed grouping according to MOA class based on the phenotypic fingerprint of saliva (Fig. 4).Grouping was most prominent for the "Cell Wall Synthesis" class and the "DNA Replication" class.Cluster analysis showed no distinct clusters between MOA classes (Fig. 4).

Random forest classification
Random forest classifiers could be successfully trained using the PhenoGMM-generated fingerprints for A. viscosus and F. nucleatum at both treatment durations (Table 2).Classifiers trained using F. nucleatum as a model strain performed better compared to classifiers trained with A. viscosus.The influence of the duration of treatment was inversed between bacterial strains.
Confusion matrices revealed that incorrect predictions were not distributed evenly over all MOA classes (Supplementary 7).For A. viscosus after 3.5 h of treatment, the "DNA Transcription" class was most often confused with the "Protein Synthesis -50S Inhibition" class, the "Folic Acid Metabolism" class was confused most often with the "Control" class, and vice versa, and the "Protein Synthesis -tRNA Interference" class was confused most often with the "Protein Synthesis -30S Inhibition" class.After 24 h of treatment, the "Protein Synthesis -30S Inhibition" class was mostly confused for being the "Protein Synthesis -50S Inhibition" class, and the "Folic Acid Metabolism" class was most often confused with the "Protein Synthesis -30S Inhibition" class.For F. nucleatum, the "Protein Synthesis -50S Inhibition" class was mostly misclassified as the "DNA Transcription" class after 3.5 h of treatment, and the "Protein Synthesis -30S Inhibition" class was mostly misclassified as the "Protein Synthesis -tRNA Interference" class after 24 h of treatment.

Prediction of MOA of an unseen compound
The MOA of the unseen compound cephalothin could be successfully predicted as the "Cell Wall Synthesis" class by most classifiers (Table 4).Only for the classifier trained on F. nucleatum after 24 h of treatment, two of the replicates were predicted incorrectly as the "Membrane Disruption" class.However, for these replicates, the second most likely MOA class was "Cell Wall Synthesis" with a probability of 38.6% (replicate 2) and 38.6% (replicate 3).

DISCUSSION
The number of novel antimicrobials brought to market is constrained by high develop ment costs and low profitability.The elucidation of the MOA can contribute significantly to this cost (13).To overcome this obstacle, we investigated if flow cytometric micro bial fingerprinting could be used as a fast and cheap alternative to predict the MOA of antimicrobials.Our experimental results show that the phenotypic fingerprints of different MOA groups could be distinguished from each other.Moreover, random forest classifiers predicting the MOA could be successfully trained using these fingerprints.Finally, trained random forest classifiers were able to predict the MOA of an unseen compound.
Phenotypic fingerprints of samples treated with antimicrobials cluster according to MOA class, with clustering of MOA classes that target the integrity of the cell being most prominent.This is the result of the use of SGPI to stain bacterial cells, as SGPI is used to assess membrane integrity (36).Furthermore, the MOA class "DNA Replication" that influences the level of DNA was shown to be statistically different from other MOA classes as well.Again, this may be the result of the staining since SYBR Green I (SG) binds to double-stranded DNA (44).Therefore, the use of different stains may increase discriminative power between MOA classes.For example, the use of bio-orthogonal non-canonical amino acid tagging (BONCAT) could help in discrim inating between the different "Protein Synthesis" classes (45).Additionally, increased

MCC (3.5 h treatment) MCC (24 h treatment)
A. viscosus 0.79 0.75 F. nucleatum 0.89 0.92 a Classifiers were trained using the phenotypic fingerprints generated with PhenoGMM.The MCC was used as performance metric to optimize the model to account for class imbalance.An MCC of "0" indicates random guessing; an MCC of "1" indicates perfect classification.
discriminative power could be achieved by combining multiple phenotypic fingerprints acquired with different staining techniques.
A lower number of statistically significant differences between MOA classes were found with increased treatment time for both bacterial strains (Supplementary 6).For F. nucleatum, it has to be noted that seven pairwise comparisons that were significantly different after 3.5 h of treatment but not significant after 24 h, were comparisons involving the untreated control.Figure 2D shows a large difference between the untreated control that did not undergo incubation and the control that did, leading to high variability in this MOA group.This high within variability could explain the reduced statistical differences for these comparisons.For all other cases where statistical differences were not significant after 24 h anymore, we argue that it is the result of phenotypes of bacterial cells becoming more similar with increased treatment time and that these phenotypes are related to cellular death.For example, for both A. viscosus and F. nucleatum, MOA classes "DNA Transcription" and "DNA Replication" were not significantly different anymore after 24 h of treatment.Sigeti et al. found that for Bacteroides fragilis, metronidazole inhibited DNA synthesis after 20 min.It also inhibi ted RNA synthesis, but only after 60 min.Accordingly, secondary inhibition of DNA transcription occurs after inhibition of DNA replication (46).Moreover, in E. coli, inhibition of either RNA synthesis or protein synthesis resulted in the prevention of DNA replication (47).Furthermore, analysis of membrane integrity revealed an increase in cells with membrane damage with longer treatment time for most antimicrobials for A. viscosus.With F. nucleatum, this was only observed for some antimicrobials.However, the control showed a high relative abundance of cells stained with propidium iodide (i.e., damaged) after 3.5 h of treatment (Supplementary 2).Propidium iodide has been demonstrated to be able to stain cells in the exponential growth phase (48) and F. nucleatum has been  shown to reach the stationary phase only after 48 h (49).This could justify why increased membrane damage was only observed for a few antimicrobials with this strain.Notably, polymyxin B was not grouped with its MOA class (Membrane Disruption) for A. viscosus, which can be explained by polymyxin B not being active against Grampositive bacteria (50).The phenomenon implies that an adapted pipeline to the one used in this experimental setup can be used for antimicrobial susceptibility testing.An antimicrobial with a known MOA that does not group with its respective MOA class implies that the compound is not active against the tested bacterium.This can be used in a fast assay where a range of compounds is screened to select the most effective antimicrobial.Moreover, it can be used to assess if a novel antimicrobial is effective against either Gram-positive or Gram-negative bacteria, or both.
A drawback of the presented workflow is that the difference in phenotypic fingerprints is mostly strain specific (Supplementary 5).Therefore, if novel compounds are assessed, it should be done on the same microorganism each time.To account for eventualities, the assessment should ideally be performed on multiple microorganisms.Notwithstanding, we found that grouping according to MOA could also be observed for a human saliva sample (Fig. 4A).The observation shows that the workflow can be robust for complex communities and suggests its potential for the personalization of antimicrobial treatment for patients.Hence, the selection of the most suitable treatment for the patient could be based on the response of the microbial community of the patient.
The performance of random forest classifiers was better for F. nucleatum, even though more statistically different comparisons between MOA classes were found for A. viscosus.As PhenoGMM looks for clusters of cells with similar properties inside the flow cytometry data, bins are set according to subpopulations found in the data (29,38).We hypothesize that the resulting bins lead to distinctive features for each MOA class and that the random forest classification algorithm can select features that are unique for each MOA class.This is different from the statistical analysis, where only the similarity between MOA classes is compared to the similarity within each MOA class.Further investigation is needed to shed light on which factors lead to better classification performance and why differences are found between the used microorganisms.
When it came to predicting the MOA of an unseen compound, the shorter incubation time led to the best classification.Again, we argue that this may be the result of the phenotype of cells becoming more alike with increased treatment time because of cellular death.Remarkably, the classifier with the best performance (F.nucleatum after 24 h of treatment) led to the worst prediction.Further investigation is needed, as this could be caused by multiple reasons.For example, the estimation of model performance could be biased, even though repeated k-fold cross-validation is considered to be robust toward performance estimation (51).Also, feature selection within the random forest model could be biased due to eventualities (e.g., a signal created by the interaction of the compounds with the medium) (52).Alternatively, the (time-dependent) effect of the unseen compound simply resulted in a different phenotype of the bacterium as compared to the ones used for training.
The performance of the classifiers could be increased by expanding the database of phenotypic fingerprints used for training (53).Besides, the use of alternative fingerprinting methods may increase performance as well.Furthermore, the use of different classification algorithms, such as Kernel support vector machines, neural networks, or gradient-boosting decision trees, is worth investigating.
To our knowledge, only a few methods are able to reveal the MOA in a fast way.Bacterial cytological profiling is a fluorescence light microscopy-based technique that can be used to elucidate the MOA by pinpointing the molecular target of antimicrobials.The experimental procedure can be completed in a matter of hours.However, data analysis can be tedious as it involves image analysis (54,55).Another fast technique is the use of AmpC reporter assays, but it can only be used to screen for inhibitors of cell wall synthesis (56).Our proposed workflow offers some clear advantages over other techniques.It is fast, cheap, and can be expanded to cover most MOA classes by including alternative staining methods.Finally, data analysis can be automated, leading to a reduced time for results.
To conclude, we found that the phenotypic fingerprints of bacteria treated with antimicrobials group according to the MOA of the antimicrobials.We found that this grouping could be observed both at the level of a single bacterial strain and at the level of a patient sample (i.e., saliva).Additionally, we showed that flow cytometry could be successfully used for the prediction of the MOA of novel antimicrobials and that a shorter treatment duration is favorable for this purpose.Thus, we provide proof of concept for a cheap and fast alternative for the elucidation of the MOA in the development of novel antimicrobials.

FIG 1
FIG1 Overview of the experimental workflow.Bacterial cultures of A. viscosus and F. nucleatum were treated with a wide range of antimicrobials for either 3.5 h or 24 h.Consecutively, treated cultures were measured using flow cytometry and data were used to construct PhenoGMM-based phenotypic fingerprints.In turn, these fingerprints were used to train random forest classifiers.Last, the trained classifiers were employed to predict the MOA of the unseen compound cephalothin.

FIG 2
FIG 2 NMDS of flow cytometric fingerprints of A. viscosus (A and B) and F. nucleatum (C and D) after 3.5 h (A and C) and 24 h (B and D) of treatment with antimicrobials.Phenotypic fingerprints were generated using PhenoGMM (Phenoflow package).Compounds with the same MOA are grouped by color.Ellipses were drawn at the 95% confidence level."Heat" indicates the heat-treated control; "Control" indicates the untreated sample that underwent incubation with the antimicrobial treated samples; and "Unincubated Control" indicates the untreated sample that did not undergo consecutive incubation.

FIG 3
FIG 3 Dendrograms of flow cytometric fingerprints of A. viscosus (A and B) and F. nucleatum (C and D) after 3.5 h (A and C) and 24 h (B and D) of treatment with antimicrobials.Compounds with the same MOA are grouped by color.

FIG 4
FIG 4 Diversity analysis of phenotypic fingerprints of a saliva sample treated with different antimicrobials for 24 h.Compounds with the same MOA are grouped by color."Control" indicates to the untreated sample that underwent incubation with the antimicrobial treated samples.(A) NMDS of flow cytometric fingerprints.Ellipses were drawn at the 95% confidence level.(B) Dendrogram of flow cytometric fingerprints.

TABLE 1
Antimicrobial compounds with their respective final concentration and supplier, grouped by mechanism of action

TABLE 2
Matthews correlation coefficients (MCCs) of trained random forest classifiers for A. viscosus and F. nucleatum after 3.5 h and 24 h of treatment with the antimicrobials a

TABLE 3
Performance of each MOA class for the different random forest classifiers trained using the phenotypic fingerprints of antimicrobial-treated cultures of A. viscosus and F. nucleatum

TABLE 4
Predicted MOA classes for the unseen compound cephalothin for random forest models based on the PhenoGMM fingerprint a a Predictions were made for each individual replicate.