Machine learning-based discovery of GW3965 as a therapeutic compound against invasive emm92-type group A Streptococcus

Abstract

A multi-drug resistant emm92-type strain of group A Streptococcus (GAS) has emerged as an important causative agent of invasive infections—particularly affecting people who inject drugs—in the United States. To curtail this developing threat, we aimed to identify and repurpose FDA-investigated compounds as antimicrobials. To identify growth-inhibiting compounds, a machine learning-based model was trained on the emm92-iGAS growth response to 2560 bioactive compounds. The model was used to screen a 6111 FDA-evaluated drug library in silico. Of the 9 validated compounds, GW3965 experimentally exhibited a 99% reduction in iGAS survival at an MIC of 6.25 µM. Treatment with GW3965 aided complete wound closure in a human skin equivalent model, and decreased lesion size and reduced bacterial burden in a mouse model of skin and soft tissue infection. Application of a machine learning model expedited the discovery of GW3965 as a therapeutic for iGAS skin and soft tissue infections.

Introduction

Streptococcus pyogenes or group A Streptococcus (GAS) is a gram-positive bacterial pathogen responsible for ~730 million infections per year globally, which range in severity from superficial pharyngitis and pyoderma to invasive infections¹. Presentation of invasive GAS (iGAS) infections includes skin and soft tissue infections (SSTI), necrotizing fasciitis, streptococcal toxic shock syndrome, and bacteremia². The number of iGAS infections has been on the rise in the U.S., doubling in number of cases and deaths since 2009³. The standard treatment for iGAS disease has been combination therapy with a β-lactam and clindamycin⁴. However, per the 2019 CDC antibiotic resistance threats report³ and the WHO’s 2024 priority pathogens list⁵, iGAS resistance to erythromycin and clindamycin is a growing concern^6–9. In the U.S., the MLS_B phenotype of resistance to 3 classes of antibiotics - macrolides, lincosamides and streptogramin B - is primarily attributed to the acquisition of erm (erythromycin resistance methylase) genes, which encode methyltransferase enzymes^6,10,11.

In the absence of an approved GAS vaccine, tracking the predominant and emerging iGAS emm-types is critical to determine shifts in epidemiological trends. Surveillance of iGAS isolates by the CDC’s Active Bacterial Core at 10 sites across the U.S. shows that the emm1, emm12, emm28, and emm3 strains continue to cause the majority of invasive infections^8,12,13. However, CDC and local surveillance reports support a significant and growing presence of MLS_B-, aminoglycoside-, and tetracycline-resistant isolates between 2010-2022, of which the emm92 strain was a predominant cause of infection^9,13–17. The emm92 strain has been majorly associated with the at-risk intravenous drug use (IVDU) population^{6,9,10,13,18,19}. This trend is particularly evident in the rural West Virginia population, where from January 2020–2023, 46% of infections were caused by a multidrug-resistant emm92 strain, among which 62% were associated with a patient history of IVDU^15,16. Therefore, the emm92 strain serves as a model organism for evaluation of compound efficacy against resistant iGAS infections due to recent emergence and carriage of MLS_B, aminoglycoside, and tetracycline resistance.

Our goal here was to identify novel antimicrobials that would be effective against the GAS pathogen that is responsible for invasive skin and soft tissue infections. We developed a machine learning model trained to predict compounds with inhibitory activity against emm92 iGAS, which serves as a model multi-drug resistant strain, through application of growth inhibitory responses to a diverse compound library, similar to as discussed previously^20,21. The trained neural network was then used for in silico screening of 6111 FDA-evaluated small compounds. Investigation of FDA-evaluated compounds provides the advantage of identifying drugs with known safety profiles that could be repurposed as antimicrobials. In this study, a hit compound was determined as any compound with a growth inhibition score of ≥0.415, wherein our model, a score of 1 would be indicative of 80% or greater growth inhibition (pattern 1–1). According to these parameters, our in silico screening identified 200 hit compounds with predicted growth inhibition scores ranging from 0.415 to 0.712. Compounds were grouped by their annotated mechanism of action, then ranked based on their predictive score. The 9 compounds with the highest prediction score per group were selected for validation. Known antibiotics were excluded, as the goal in this study was to identify novel antimicrobial compounds. Through this approach, we identified GW3965 as an antimicrobial for iGAS infections. GW3965 promoted wound closure in a human skin equivalent in vitro model and reduced disease severity in a murine model of iGAS skin and soft tissue infection. This study employed machine learning to expedite antimicrobial discovery to combat the emerging threat of MLS_B-resistant iGAS.

Results

Generation of an inhibitory compound dataset for training of the machine learning model

To identify novel antimicrobials against the emm92-15 iGAS isolate using machine learning, we generated a training dataset of inhibitory and non-inhibitory compounds by screening a library of bioreactive and natural small compounds containing metabolites, medicinal compounds, and antimicrobials, including known antibiotics (Fig. 1a). Compounds were classified as inhibitory if culture turbidity (OD_600nm) was ≤20% of the medium-drug-solvent control (THY broth, 2.5% DMSO), labeled as relative growth. Of the 2560 compounds screened at a final concentration of 50 μM, we identified 22 inhibitory compounds after a 2-h incubation, 255 after 4-h, 274 after 6-h, and 191 after 24-h (Fig. 1b). Most of the inhibitory compounds were functionally categorized as antibacterial, followed by antineoplastic, antifungal, and anti-inflammatory, as well as antidepressants and antihypertensives (Tables S1). Thus, we generated a diverse training dataset containing known antibiotics and novel inhibitory compounds.

Fig. 1. Development of a machine learning model for prediction and identification of inhibitory compounds.

a Workflow for model development and compound prediction. To generate a training dataset, a library of 2650 small compounds was screened against emm92-15 WV-iGAS; growth curves that represent the 4 temporal patterns of: no inhibition, temporary inhibition, inhibition at endpoint, and complete inhibition are shown in the second panel. b Relative bacterial growth in 50 µM of each compound to the DMSO control at indicated timepoints. The OD_600nm measurement of growth at the 2-, 4-, 6-, and 24-h timepoints was used to determine the percentage of growth inhibition relative to the DMSO culture control. The dotted line represents 80% growth inhibition according to turbidity, where a value of 1 represents normal growth. c Patterns of growth inhibition used to train the machine learning model. The AUC variance values are reported as ± the standard deviation (sd) of ROC-AUC across the 20 cross-validation folds. Training supported the model’s capacity to identify motifs associated with antimicrobial activity based on identification of inhibitory structures and substructures. Some images created with BioRender.com.

Using this dataset, we trained a machine learning (ML) model to establish common structures and substructures associated with growth inhibition, with the goal of developing an ML model that could reliably predict inhibitory activity of a compound after 24 h of incubation. We found our model could reasonably predict inhibitory activity at the 24-h timepoint (AUC = 0.776 ± 0.052), with variance reported as the standard deviation of ROC-AUC. To improve model prediction, we further classified inhibition activity into four distinct profiles by incorporating data for the intermediate time points of 2, 4, and 6 h. Accordingly, compounds were categorized into four different kinetic growth inhibition profiles: (i) no inhibition, (ii) temporary inhibition (2–6 h), (iii) delayed inhibition (inhibitory at 24 h), or (iv) complete inhibition across all timepoints (Fig. 1c). These 4 categories were employed as growth profiles can allude to a compound’s mechanism of action, where complete inhibition is indicative of bactericidal activity, whereas delayed inhibition can occur from possible cell division arrest. With this approach, we found our model could reasonably predict compounds that exhibited complete (AUC = 0.777 ± 0.049) or no inhibition (AUC = 0.729 ± 0.049), except for the more variable temporary inhibition (0.756 ± 0.197) model. Using the multi-class predictions, we also observe a slight improvement in predicting 24-h inhibition activity. Specifically, we aggregate the predicted probabilities of delayed and complete inhibition as the score for inhibition at 24 h, and those of temporary and no inhibition as the score for no inhibition at 24 h. This yields an AUC of 0.783 ± 0.052. Based on this data, we hypothesize that, rather than focusing solely on the inhibition outcome at 24 h, incorporating intermediate inhibition states provided additional supervisory signals to help the model learn how molecular structures relate to dynamic inhibition behaviors.

In silico prediction and validation of montelukast and GW3965 antimicrobial activity

A key challenge in drug development is obtaining FDA approval for clinical trials; thus, we sought to identify FDA-evaluated drugs suitable for repurposing against iGAS infections. Therefore, our trained model was applied for in silico screening of inhibitory compounds from a Drug Repurposing Hub of 6111 small molecules at various stages of FDA investigation for human diseases^22,23 (Fig. 1a). The predicted bacterial growth inhibition of compounds was implemented by application of a 3-layer directed message-passing neural network (MPNN)²⁴. Each compound was represented as a graph with atoms as nodes and bonds as edges. Bond-centered message passing was applied to support ML of the contextualized atom and bond representations that capture local chemical environments. For the in silico screening of compounds with predicted iGAS growth inhibition, the probability of complete growth inhibition (temporal pattern 1–1) was used to establish predictions. Compounds were then ranked on a scale of 0–1, where 0 represented no predicted inhibition and 1 represented complete inhibition. In silico screening predicted 200 compounds with scores ranging from 0.415 to 0.712. After elimination of known antimicrobial drugs, nine representatives with non-overlapping mechanisms of action and scores ranging from 0.565 to 0.7 were selected. The following compounds were chosen for validation and characterization: montelukast sodium salt, GSK461364, AMG510, siponimod, ABT-263, miltefosine, CFTR inhibitor-172, GW3965 HCl, and probucol (Table 1). We determined minimum inhibitory concentrations (MIC) for each predicted compound by measuring bacterial growth, using two-fold drug dilutions down from 50 µM (Table 1). The leukotriene receptor antagonist montelukast sodium salt and the liver X receptor agonist GW3965 HCl had the best drug potency with inhibitory efficacy of 6.25 µM MIC. Miltefosine inhibited growth at a 25 µM concentration while the remaining compounds had an MIC of >50 µM (Fig. S1a–i and Table 1). Taken together, we selected montelukast and GW3965 as the top candidates for further evaluation (Fig. 2a, b).

Table 1.

Compounds predicted in silico

Compound	Score	Known MOA	MIC
Montelukast	0.700	Cysteinyl leukotriene 1(CysLT1) receptor antagonist	6.25 µM
GSK461364	0.636	Kinase inhibitor	>50 µM
AMG-510	0.634	Covalent inhibitor of K-Rasg12C	>50 µM
Sipionimod	0.621	Agonist of sphingosine-1-phosphate receptor 1 (S1P1) and S1P5	>50 µM
ABT-263 (Navitoclax)	0.599	Inhibitor of the Bcl-2 family proteins	>50 µM
Miltefosine	0.597	Inhibits phosphatidylcholine biosynthesis	25 µM
CFTR inhibitor-172	0.586	Selectively blocks the CFTR channel in a voltage-independent manner	>50 µM
GW3965	0.574	Liver X receptor (LXRα and LXRβ) agonist	6.25 µM
Probucol	0.565	Antioxidant that inhibits oxidation of LDL cholesterol	>50 µM

The two of 9 compounds identified that are shown in bold font were investigated further.

Fig. 2. Assessment of montelukast and GW3965 antimicrobial activity.

a Montelukast sodium salt and B) GW3965 hydrochloride from Cayman Chemical. Chemical structures shown in the top panels were adapted from the manufacturer’s website. Overnight cultures of the emm92 strain were adjusted to a starting OD_600nm value of ~0.05, and a concentration range of 1.56–50 µM of (a) montelukast (middle) or (b) GW3965 (middle) was added, followed by a 24- h incubation. Ampicillin at a concentration of 100 ng/mL was included as a positive control for growth inhibition. Statistical analysis for comparison of growth was completed by a Two-Way ANOVA with repeated measures. a, b Enumeration of viable bacteria at the 24-h time point shown in the bottom panel for each compound. Cultures were plated for bacteria grown with a 6.25 µM concentration of each compound tested, as well as a normal untreated growth control (0) and vehicle control (2.5% DMSO) for comparison of CFU (colony forming units) and percent survival. Statistical significance was determined by Unpaired t-tests with Welch’s correction. There was no statistically significant difference for either the CFU/mL or percent survival between the untreated and the DMSO controls for (a) montelukast (p = 0.0877 and p = 0.0636, respectively) or (b) GW3965 (p = 0.2832 and p = 0.3559, respectively). Data is from three independent experiments ± sd. GraphPad Prism software was used for statistical analysis and figure generation.

We next assessed the bacteriostatic or bactericidal effect of montelukast and GW3965 on iGAS survival with a 6.25 µM concentration. There was no significant difference in viable bacteria between the media and vehicle controls. Treatment with montelukast leads to a 1-log reduction in viable bacteria corresponding to a 76% decrease in survival, which is indicative of bacteriostatic activity (Fig. 2a, bottom). In contrast, GW3965 reduced viable bacteria by 3-log with growth inhibition of 99.9% (Fig. 2b, bottom), which is indicative of bactericidal action²⁵. While both montelukast and GW3965 significantly inhibited bacterial growth (p = 0.007 and p < 0.0001, respectively), montelukast is marked with a black box warning for neuropsychiatric events and has mixed safety data^26–28; thus, we focused on GW3965 for further testing.

GW3965 inhibits GAS growth in vitro

A rationale-extraction (substructure) procedure based on Chemprop physicochemical properties was used to identify the substructure of GW3965 predicted to be critical for activity²⁹. A Monte Carlo Tree Search was used to determine the smallest substructure that preserves the target property while achieving a high predicted score^24,29. The halogen-substituted aromatic ring substructure (e.g. difluoromethylbenzene in GW3965) was identified as the structural component with the highest predicted probability of complete inhibition activity (pattern 1–1). In the GW3965 molecule the predicted active substructure has an additional chloride and fluoride group (Fig. 3a). For comparison, we also assessed two functionally and structurally related liver X receptor agonists LXR-623 and T0901317, for growth inhibition due to shared sub-structural features with GW3965; however, neither compound significantly reduced bacterial growth with an up to 50 µM concentration (p = 0.9973, and p = 0.3844, respectively) (Fig. 3b). LXR-623 and GW3965 both have halogen substitution on an aromatic ring, however LXR-623 lacks the methyl substituted fluorine pattern that is present in the predicted GW3965 rationale substructure. For GW3965, the chloro-trimethyl-aromatic ring with an additional two fluoride groups likely improves compound efficacy in combination with the total structure, although experimental validation will be necessary in further studies. Furthermore, when we evaluated the solvent accessible surfaces for lipophilicity, we found that the substructure agreed between the compounds in large parts of the lipophilicity and hydrophobic areas. This suggests that the activity seen with GW3965 is likely due to the differences in hydrogen bonding versus lipophilic properties (Fig. S2). Since antimicrobial resistance may vary between GAS strains, we tested the growth inhibition by GW3965 on two globally predominant iGAS-M types: M1 (emm1) and M3 (emm3). We found that across the iGAS strains tested (e.g., M92, M1, M3), all had an MIC of 6.25 µM for GW3965 (Fig. 3c). We additionally investigated the effect of GW3965 on a normal flora organism, where GW3965 was determined not to have bactericidal activity towards the gram-positive probiotic Lactococcus lactis (Fig. S3).

Fig. 3. Analysis of GW3965 antimicrobial strain effects and synergy.

a GW3965 (GW) chemical structural interpretability and prediction. Using the rationale-extraction procedure of the Chemprop interpret module, a neural network was trained to identify substructures critical for the predicted antimicrobial activity of GW3965 based on the atomic and bond features represented in the chemical structural formulas or SMILES. The active group with predicted complete inhibition for GW3965 (OC( = O)Cc1cccc(OCCCN(CC(c2ccccc2)c2ccccc2)Cc2cccc(c2Cl)C(F)(F)F)c1) was the chloro-trimethyl-aromatic ring. The SMILE of the predicted difluoromethylbenzene ring sub-group is (F[CH:1](F)c1ccc[cH:1][cH:1]1), as shown in by a broken box. b Growth of the emm92 strain with liver X receptor agonists. LXR-623 and T0901317 were tested for antimicrobial activity against the WV-iGAS emm92-15 strain with a 1.56–50 µM concentration range. Differences in growth were statistically assessed by Two-Way ANOVA with repeated measures. For LXR-623 (p = 0.9973) and T0901317 (p = 0.3844), changes in growth with a 50–1.56 µM concentration of either compound were not significant as compared to the untreated control (0 µM), whereas complete inhibition was seen in control media with ampicillin at 100 ng/mL. Structures adapted from APExBio website. c GW3965 activity against globally predominant iGAS emm-types. The antimicrobial activity of GW3965 was assessed for emm1 and emm3 strains at a 50–1.56 µM concentration range. Statistical comparison of the GW3965 growth effect was completed by a Two-Way ANOVA with repeated measures, where the 6.25 µM concentration significantly decreased growth (p < 0.0001) for either strain as compared to the untreated control. d Checkerboard assay for synergy of GW3965 (GW) and ampicillin (Amp). The emm92 strain was grown with combined 2-fold dilutions of GW3965 (12.5–0.78 µM) and ampicillin (0.1–0.0015 µg/mL). Statistical significance was assessed by Unpaired t-test with Welch’s correction at the 24-h endpoint. For all curves, data from three independent experiments are shown with ± sd. GraphPad Prism software was used for figure generation.

Further, combination therapy of a β-lactam with clindamycin has been used for treatment of iGAS infections, where the additive effect of clindamycin is associated with decreased toxin production⁴. Here, we investigated whether GW3965 could act synergistically with ampicillin, as an exemplar β-lactam. Using a checkerboard assay, we did not find that the two compounds acted synergistically against the emm92 strain (Fig. 3d), with a fractional inhibitory concentration (FIC) index (MIC for ampicillin 0.12 µg/mL, GW3965 6.25 µM) of 2, indicative of an indifferent effect.

GW3965 treatment supports wound closure and bacterial clearance in a human skin equivalent model

Human skin equivalents were utilized in an in vitro SSTI model to determine the effect of GW3965 on emm92 infection. Cultured tissues were wounded by punch biopsy of the epidermal layer to mimic a skin break as a portal of entry for iGAS infection³⁰. The WV-iGAS emm92-3 strain was used for infection studies as it has a higher clindamycin MIC and increased virulence factor expression, compared to the emm92-15 strain used for compound library screening^31,32. Following the replacement of transport media to antibiotic-free culture media for 24 hours, the epidermal wound was infected with the emm92-3 strain (3.3 × 10⁶–1.1 × 10⁷ CFUs) (Fig. 4a), as before³³.

Fig. 4. Effect of GW3965 treatment on wound closure in an in vitro human skin equivalent model.

a Experimental layout of infection, treatment, and endpoint parameters. Daily treatment with fresh culture media with or without 12.5 µM of GW3965 started 24 h post-infection with the emm92-3 strain and continued for 5 days. Images created with BioRender. b Images of wound closure in skin equivalents. Wounds were visualized at 24-h post-infection and on day 6. c–h Histopathology of skin equivalents. H&E-stained tissue shown at 200× magnification, scale bar: 200 µm. The boxed insets show representative Gram-stained cocci at a higher magnification (800×). Closed arrowheads indicate the progress of epidermal closure, while the open arrowheads show bacterial microcolony formation. i Effect of GW3965 treatment on wound closure. Wound diameter was measured as the distance between the keratinocyte layer opening at 1.25× magnification for all histology samples (n = 2 biological replicates per treatment), representative images shown. Diameter was determined for 2 technical replicate sections from each of the 2 skin equivalent tissues ± sd. Comparison of lesion diameter on day 1 versus day 6 was assessed by Unpaired t-tests with Welch’s correction. Statistical significance was compared between groups for either day 1 or day 6 using Unpaired t-tests with Welch’s correction. The corresponding p values represent additional comparisons: D1 Control vs D1 GW3965 (p = 0.99, ns), D1 GW3965 vs D1 emm92 (p = 0.0002), D1 emm92 + GW3965 vs D1 emm92 (p = 0.3461, ns), D6 Control vs D6 GW3965 (p = 0.002), D6 GW3965 vs emm92 D6 (p = 0.0172), D6 GW3965 vs D6 emm92 + GW3965 (p = 0.0034), and D1 emm92 vs D6 emm92 (p = 0.2373, ns). GraphPad Prism software was used for figure generation.

Skin equivalents were imaged on day 1 and day 6 to assess group differences in wound closure and histology. Here, closure of the punch biopsy site, indicative of epidermal regrowth over time, was seen in both untreated (media only) and GW3965-treated non-infected controls, as well as in infected wounds treated with GW3965 (Fig. 4b). In contrast, the punch biopsy wound remained visible by day 6 for the untreated emm92-infected tissues. H&E- and Gram-staining of tissues on day 1 (24-h post-infection) indicated the presence of emm92 microcolonies (e.g., initial colonization) along the wound bed and minimal keratinocyte regrowth, whereas re-epithelialization around the wound perimeter was observed in the non-infected media control specimens (Fig. 4c, d). The remaining tissues were treated daily for 5 days from the bottom chamber using either media supplemented with 12.5 µM GW3965 (two times the MIC) or drug-free culture media. As expected, significant epidermal wound closure was observed in both non-infected controls (e.g., GW3965-treated (p = 0.0007) and non-treated (p = 0.0006)) by day 6 endpoint (Fig. 4e, f). This was in direct contrast to non-GW3965-treated wounds infected with emm92 iGAS, which uniformly demonstrated incomplete wound closure between day 1 and day 6 (p = 0.2373), and delayed keratinocyte re-epithelization with bacterial microcolonies invading deeply into tissues (Fig. 4g, i). Importantly, wounds infected with emm92 iGAS that were treated with GW3965 exhibited complete or near complete epidermal recovery with only minimal bacteria observed superficially in tissues, as compared to the untreated infected group on day 6 (p = 0.0309) and over the experimental time course (p < 0.0001) (Fig. 4h, i). Altogether, our data suggest that treatment with GW3965 in this model: (i) is non-toxic to human cultured cells, (ii) promotes wound closure in infected tissues, and (iii) supports bacterial clearance in the wound bed and dermis.

GW3965 aids infection recovery in a murine SSTI model

We first established the skin and soft tissue infection (SSTI) model using SKH1 mice for the WV-iGAS emm92 strain based on our previous methods employed for emm1, emm3, emm28, and emm41 strains^33–35, and existing data in CD1 mice³⁶. To determine an effective inoculum dose, we subcutaneously injected hairless immunocompetent male SKH1 mice with a 10⁷–10⁹ inoculum range of the emm92-3 isolate. Of doses evaluated, the 10⁹ CFU inoculum was selected for all subsequent studies because it reliably induces lesion development and a substantial up to 10% decrease in the initial body weight (Fig. S4a). On the contrary, emm92 infection did not lead to significant changes in body temperature (Fig. S4b). In an extended 24-day study, we assessed the general infection course for the emm92 strain using measurements of gross skin pathology and body weight change (Fig. S4c–e). Skin lesions developed in mice by 24–48 h post-infection and self-resolved in 55% of mice by day 24 (Fig. S4c). Bacterial growth in extracts obtained from skin lesions on day 24 demonstrated sterility for 81% of mice, independent of whether the wound had remaining ulceration or was fully resolved. Additionally, GAS was not detected in either the blood or spleen. As illustrated by representative images (Fig. S4e), the natural course of gross pathological findings included a progression from visible and palpable dermal abscess formation (day 2) to epidermal ulceration (day 10), and finally to healed, scarred lesions (day 22).

We next assessed the safety of GW3965 by intraperitoneal (i.p.) delivery of 10, 30, or 50 mg/kg doses to non-infected mice (male and female) for 4 consecutive days. Compared to mice that received saline, GW3965 administration exhibited no significant impact on body weight (p = 0.5062) or temperature (p = 0.5039) for any of the concentrations tested (Fig. S5a, b). Existing pharmacokinetic data for GW3965 in C57BL/6 mice suggest that the compound is cleared from plasma and relevant tissues within one day following oral delivery in food³⁷. Here, we determined the GW3965 concentration in SKH1 mouse plasma following i.p. delivery at 2-, 4-, and 24-hours post-injection, confirming clearance by 24 h (Fig. S5c, d).

Based on these results, we evaluated GW3965 treatment efficacy in the SSTI model using 5 daily i.p. injections of 10 mg/kg or 30 mg/kg doses. Additional control groups included infected mice treated i.p. with 50 mg/kg of penicillin, saline, or vehicle (DMSO in saline), as well as a non-infected vehicle-only group. To simulate therapeutic intervention of clinical infection, treatment was initiated one day post-infection by the emm92 strain, with gross pathology and weight change observed daily for 18 days (Fig. 5a). Notably, the reduction in lesion area for the GW3965- (30 mg/kg) and penicillin-treated groups was indistinguishable (p = 0.9578), while penicillin (p = 0.001) and GW3965 (p = 0.0002) both significantly reduced lesion area compared to the vehicle infected control (Fig. 5b). Infection led to an initial 5-10% weight loss across all infected groups, with weight changing differently over time, yet all groups showed similar rates of weight gain beyond 12 days post-infection. Compared to the vehicle-treated mice, penicillin-treated mice regained weight significantly faster (p = 0.0360) over the course of the experiment (Fig. 5c). Although no such difference was observed for the 30 mg/kg GW3965 treatment group, we did note variations in the time needed to regain body weight above the baseline, which occurred on day 6 for penicillin-treated mice, on day 8 for the GW3965 group, and on day 11 for the vehicle control group (Fig. 5c). Corresponding data obtained in parallel experiments from mice treated with 10 mg/kg of GW3965 indicated that a lower concentration was not sufficient to significantly reduce disease parameters (Fig. S6a–f). Altogether, a 30 mg/kg GW3965 dose significantly reduced disease severity in infected mice as measured by reduced lesion size and accelerated weight gain (Fig. 5b, c).

Fig. 5. Efficacy of GW3965 treatment in an in vivo murine SSTI model of iGAS infection.

a Schematic of experimental procedure. Mice were subcutaneously injected with the emm92-3 isolate. At 24-h post-infection, mice received 30 mg/kg of GW3965 or controls daily through intraperitoneal injection for a total of 5 doses. The spleen, lesions, and blood were collected on day 1 (n = 4, 2 male and 2 female), day 6 (n = 6, 3 male and 3 female, per group), or 18 (n = 10, 5 male and 5 female, per group, aside from the vehicle group, where 1 mouse was removed on day 8 due to morbidity) for assessment. Experimental layout generated with BioRender. b Effect of GW3965 treatment on wound size. The lesion area was assessed according to the equation for a spherical ellipsoid, based on the length and width of the lesion. Statistical significance was determined by mixed model ANOVA with Tukey’s multiple comparison. c Changes in mouse weight over the course of infection and treatment. Weight change was calculated relative to the baseline weight before infection. A Mixed-effects model with Tukey’s multiple comparison test was used to determine significance. Bacterial burden in skin lesions. Tissues were homogenized and plated on blood agar, enumerated bacteria from lesions at endpoint on day 1, day 6 (d) and day 18 (e). Unpaired t-test with Welch’s correction was used to determine significance. f Gross pathology of lesions on day 18. Resolution of the ulcerated scab, resulting in complete skin closure was used to dictate a lesion as resolved. g Infection status on day 18. The percentage of mice that had no countable bacteria from homogenized lesions at endpoint was calculated to determine clearance. GraphPad Prism software was used for figure generation, with ± sd shown.

The presence and quantity of viable bacteria were determined for spleen, blood, and skin lesions of animals at the treatment endpoint (day 6) and experiment endpoint (day 18). In this model, systemic dissemination of iGAS to the blood or spleen was not detected at any time point; also, spleen weight did not differ significantly at either timepoint of day 6 (p = 0.4523) or day 18 (p = 0.0943) (Fig. S6g, h). On day 1 (24 h following s.c. infection with 1.8 to 5.2 × 10⁹ CFUs per mouse), recovery of 1 to 5 × 10⁸ CFUs per gram of tissue verified stable infection prior to treatment initiation (Fig. 5d). By day 6, 67% of penicillin-treated mice had complete clearance (i.e. sterility) of their cutaneous wound, while the remaining mice demonstrated an average 2-log-fold reduction in bacterial burden (p = 0.0537). Likewise, mice treated with 30 mg/kg of GW3965 had an average 2-log-fold reduction in bacterial wound burden (p = 0.054). Bacterial burden for the vehicle control group decreased by 1-log (p = 0.07), indicating a natural course of infection clearance. At the 18-day endpoint, gross examination of skin pathology indicated that ulceration persisted in 56% of vehicle-treated mice, whereas an intact and re-epithelialized epidermis was observed in 90% of penicillin treated and 70% of GW3965 treated groups (Fig. 5f). Differences in bacterial CFUs recovered on day 18 did not reach statistical significance between groups (Fig. 5e). In support of gross pathology findings, 70% of the mice treated with either GW3965 or penicillin had cleared iGAS infection altogether, while only 56% of mice naturally recovered from iGAS infection in the vehicle-treated group (Fig. 5g). Taken together, decreased cutaneous lesion size with more rapid reduction and overall clearance of bacterial infection, combined with faster weight gain supports the predicted antimicrobial effect of GW3965 and suggests clinical efficacy in reducing infection severity by accelerating disease recovery.

Discussion

For iGAS infections, an increase in MLS_B resistance rates over the last decade is a concerning threat that challenges the efficacy of standard treatment³. While GAS remains susceptible to β-lactams, mutations in PBP proteins that decrease penicillin susceptibility have recently been reported^38,39. Furthermore, very high bacterial burdens typically encountered in iGAS infections may limit the efficacy of peptidoglycan synthesis inhibitors via the Eagle effect⁴⁰. As a result, combination therapy with clindamycin has become the standard treatment regimen, with demonstrated improvement in clinical outcomes likely due to reduced expression of GAS toxins^41,42. For iGAS infections caused by clindamycin-resistant strains, a shift to β-lactam and linezolid combination therapy has been proposed⁴³. The acquisition of aminoglycoside and tetracycline resistance in iGAS strains^14–17, including the emergent emm92 strain, further reduces therapeutic options. To address the growing antibiotic resistance of iGAS, we sought to identify novel antimicrobials using the time- and cost-saving capabilities associated with application of machine learning and AI^44,45.

Here, we developed a machine learning model trained to identify inhibitory molecular structures based on a screening dataset of the emm92 growth response to a diverse set of chemical compounds. We found that our model could fairly predict inhibitory activity based on a compound’s SMILES. In addition to predicting well whether a compound was inhibitive at specific timepoints, our model had reasonable predictive power to its kinetics. This model can be used to infer a chemical’s bacteriostatic or bactericidal potential, which may provide clues about possible mechanisms of action for further investigation. Within our training dataset, we identified unconventional growth-inhibiting compounds with functions like antidepressants and antipsychotics. Antidepressants have previously been investigated for antimicrobial properties and have demonstrated effects such as efflux pump inhibition, membrane disruption, and gut microbiota disturbance, notably against gram-positive bacteria and fungi^46,47.

Using the trained model to screen FDA-approved and investigated drugs, we identified the liver X receptor (LXR) agonist GW3965 and the leukotriene receptor antagonist montelukast as the two candidates with in vitro antimicrobial activity. Identification of compounds known to promote the host response to infection while providing direct antibacterial effects would be ideal. While we chose to focus on FDA-evaluated drugs like montelukast and GW3965 in this study, there is further potential to screen larger libraries of small compounds for discovery of uncharacterized antibacterial candidates. This work demonstrates that machine learning is capable of circumventing the common hurdles of limited time and resources⁴⁴, providing a decent model for antibiotic discovery that can be expanded upon.

Here, we evaluated whether the experimental LXR agonist GW3965 could be re-purposed as an antimicrobial compound for iGAS. Our data shows that GW3965 has a bactericidal effect in vitro against several globally predominant iGAS strains (i.e., emm1 and emm3) and reduces gram-positive staining for skin equivalents. Previous in vitro studies showed an anti-viral effect of GW3965 on Newcastle disease virus⁴⁸. Further, reports of an antibacterial effect of GW3965 against Staphylococcus aureus, Listeria monocytogenes, Bacillus cereus, and Streptococcus mutans, altogether support broader GW3965 activity against gram-positive pathogens⁴⁹; however, here, GW3965 did not inhibit the growth of another related gram-positive organism Lactococcus lactis. In S. aureus, GW3965 inhibits growth through direct interaction with the FeoB iron transport protein⁴⁹. Since GAS lacks a FeoB ortholog, the mechanism of action for GW3965 against GAS remains to be elucidated. From our AI-assisted prediction, we suspect its activity is attributed to the difluoromethylbenzene aromatic ring of the molecule, but this has yet to be validated experimentally. While this prediction acts as an important stepping stone, alternative medicinal chemistry approaches need to be employed, as the additionally tested LXR-agonist, LXR-623, also harbors a chloro-trimethyl-aromatic ring but did not have antimicrobial activity against the emm92 strain. Standard affinity chromatography methods and computational approaches like protein-compound docking strategies should be pursued for target-based justification of the GW3965 mechanism of action for GAS.

The activity of GW3965 against additional gram-positive organisms has been primarily measured in vitro, aside from the promising results of increased survival in a Caenorhabditis elegans model of S. aureus infection. Therefore, the application of GW3965 in a murine model of iGAS SSTI was novel. In mice, five 30 mg/kg daily doses of GW3965 reduced bacterial burden by an average of 2-log immediately following treatment (day 6). For comparison, an analogous daily treatment with penicillin had cleared emm92 infection in 67% of mice by day 6, although improvement of GW3965 antibacterial efficacy in vivo could be achieved through additional dose optimization, administration schedule, and delivery route studies. Altogether, the in vitro antibacterial activity of GW3965 in selected gram-positive pathogens is a positive trend and supports further exploration to improve its efficacy for active infections.

Like GW3965, montelukast was also identified to have an antimicrobial effect in vitro. Montelukast has previously been reported as effective for treatment of Streptococcus pneumoniae otitis media in a rat model⁵⁰. Other studies have used this drug successfully in combination treatment regimens for Mycoplasma pneumoniae⁵¹, for acute otitis media⁵², and in sepsis infections in children/infants⁵³. Montelukast is the 14th most prescribed asthma/allergy medicine in the U.S. and functions as a leukotriene receptor antagonist, blocking leukotriene D4 action on the CysLT1 receptor within the lungs/airways⁵⁴. Despite that, montelukast has an FDA’s black box warning for neuropsychiatric effects^26,27. Because of this warning, we decided to focus on GW3965 due to its enhanced safety profile and the limited studies in murine infection models.

GW3965 has previously been evaluated in mouse models of atherosclerosis³⁷, glioblastoma⁵⁵, diabetes⁵⁶, and sepsis⁵⁷ with results supporting positive effects on disease-specific parameters. However, preclinical studies have shown that prolonged use of LXR agonists may promote cardiometabolic disease (elevated triglycerides) and contribute to atherosclerosis^37,58–60. Based on our data, we reason that short-term use of GW3965 may reduce the potential dangers of prolonged increases in plasma triglycerides, while allowing for its antimicrobial applications. In support, 5 doses of GW3965 at a 30 mg/kg concentration in vivo significantly reduced lesion area following treatment completion. The decrease in lesion area was prolonged after treatment was stopped in mice that were followed for 18 days, supporting the idea that even short-term treatment can reduce disease severity.

Interestingly, the most pronounced effect of GW3965 treatment on emm92 infection was enhanced closure of the wound barrier in skin equivalents and a significant reduction in lesion size in mice. GW3965 activates the nuclear liver X receptors LXRα and LXRβ. In humans and mice, LXRα is expressed by metabolically active tissues (i.e. liver, adipose tissue, kidney, and macrophages), whereas LXRβ is expressed in all cell types⁶¹. Activation of LXRs, either through oxysterols or chemical agonists like GW3965, is thought to have an anti-inflammatory effect^62,63. Wherein, LXRa activation promotes cholesterol efflux in the foamy macrophages associated with atherosclerotic lesions^37,64,65.

Literature reports that GW3965 can shift macrophages to a reparative state⁵⁷, while enhancing cellular debris clearance and tissue regeneration of the nervous system⁵⁵. In the context of iGAS infections, macrophages enhance bacterial clearance and promote wound healing. During the wound healing process, macrophages shift to a reparative anti-inflammatory state in the proliferative and remodeling stages⁶⁶. LXR activation has been shown to promote an M2-like phenotype by increasing IL-10, arginase 1, and 2 expression^67,68, although this effect has not been documented in infection and disease models. GW3965 treatment has also been associated with decreased keratinocyte apoptosis in a model of oral lichen planus⁶⁹. Likewise, our results demonstrated that GW3965 promoted wound closure. Altogether, the existing data indicate that activation of LXRs by GW3965 enhances tissue repair, which may be in part due to keratinocyte modification and promoting an M2-like macrophage phenotype in the wound bed

A reduction in inflammatory macrophages seems counterintuitive for overcoming infection at early stages, but we did not observe a decline in overall disease severity for mice treated with GW3965 starting at 24-hours post infection. This may be explained by the multifaceted function of LXR activation on infection responses. In the context of a polymicrobial sepsis model, GW3965 increased the capacity for macrophage phagocytosis⁵⁷. GW3965 was also shown to have protective effects against organ damage in another sepsis study⁷⁰. Further, in a Mycobacterium tuberculosis infection model, LXRα and LXRβ knockout in mice lead to an increase in bacterial burdens. LXR knockout models showed that effective neutrophil responses early in lung infection and dendritic cell regulation were dependent on LXRα responses. In wild-type animals, GW3965 treatment decreased lung burdens 10-fold and increased Th1/Th17 function⁷¹. Comparatively, in iGAS SSTIs, neutrophils are key mediators of the immune response at infection sites, and Th1 activation has been shown to protect against GAS infection in vaccine challenge models. Therefore, the observed positive effect of GW3965 on iGAS disease severity is likely a result of LXR activation across multiple cell types^72,73. Overall, this work contributes to growing data that supports the potential of GW3965 to be used in a non-antimicrobial capacity to improve infection outcomes.

In this study, the emm92 strain was selected for screening use due to its antimicrobial resistance profile and its growing clinical relevance in the United States since 2010, notably in surveyed areas including West Virginia^{9,13–16,19,31,74,75}. Acquired plasmid-borne MLS_B resistance in the emm92 background has been associated with increased expression of cytolytic and immunomodulatory iGAS virulence factors^32,36, further compounding its growing importance. Thus, in addition to identifying novel antimicrobials against iGAS, we aimed to further characterize the emm92 infection model using human skin equivalents in vitro and murine in vivo models. In the in vivo model, disease presented as an abscess that typically ulcerated within 3 days, followed by a slow reduction in lesion size and skin healing over 18-24 days. While our previous studies of emm1 and emm3 in this murine SSTI model led to high mortality overall, the emm92 strain produced a longer non-lethal infection course. In support translationally, West Virginia mortality rates associated with emm92 infection (based on known patient data) were relatively low in the non-IVDU (8%) and IVDU populations (3%), with most infections presenting as SSTIs of the limbs. Therefore, our previous patient data^15,16 and the infection course observed in the mouse model indicate that the emergent emm92 strain specializes in long-term survival in the wound rather than widespread dissemination. Likewise, drug therapies should address this shift by promoting bacterial killing and stimulating wound healing.

Altogether, we have shown GW3965 to be a potentially beneficial therapeutic agent against an emerging antimicrobial strain of iGAS in several models. Our investigation determined that GW3965 (i) demonstrated in vitro bactericidal activity, (ii) produced no apparent short-term adverse effects in an in vivo murine model, and (iii) promoted wound closure in both an in vitro skin organoid model, as well as in a murine in vivo SSTI model, all of which support its potential for use in combination therapy to expedite wound recovery in conjunction with standard treatment. Further studies are needed to optimize GW3965 treatment to its full potential, but it could be added to the arsenal against growing antimicrobial threats.

Methods

Bacterial strains

The WV-iGAS emm92-15 isolate, displaying MLS_B, aminoglycoside and tetracycline resistance, was collected from an antecubital fossa abscess and identified at the J.W. Ruby Memorial Hospital clinical laboratory in Morgantown, West Virginia¹⁵. It was used in screening assays employed to develop the machine learning dataset. The WV-iGAS emm92-3 strain, isolated from a mediastinal abscess identified at the J.W. Ruby Memorial hospital clinical laboratory in Morgantown, West Virginia¹⁵, was transformed with the pSB027 GFP-expressing plasmid containing a chloramphenicol resistance marker, as described previously⁷⁶ to be used in skin equivalents and mouse infection models. The emm92-3 strain is equally resistant to MLS_B, aminoglycoside, and tetracycline antibiotics, and has no difference in growth or MIC when exposed to GW3965 as compared to the emm92-15 strain (Fig. S7). In addition, the emm92-3 strain has a constitutive MLS_B sub-phenotype, which was associated with a higher clindamycin MIC and increased virulence factor expression, making this isolate a better candidate for efficacy tests in our skin equivalent and murine models^15,31,77. Lastly, the predicted compound was also tested for growth inhibition of the MGAS315 (emm3) and MGAS5005 (emm1) isolates representing historically predominant causes of iGAS disease globally. The MGAS315 isolate was obtained in the 1980s from a patient with streptococcal toxic shock syndrome⁷⁸, whereas MGAS5005 was obtained from the cerebrospinal fluid of a patient in 1996⁷⁹. The Lactococcus lactis strain MG1363⁷⁶ was used in this study to represent a probiotic organism (Fig. S2). All iGAS agar and broth cultures were routinely incubated at 37 °C in 5% CO_2. Overnight cultures of the WV-iGAS emm92-15 strain on brain heart infusion (BHI) agar were suspended in Todd-Hewitt broth supplemented with 0.2% yeast extract (THY medium) to an OD_600nm of 0.05 for starting cultures. Growth optimization tests performed in 48-well plates provided reproducible susceptibility outputs for iGAS growth while remaining high-throughput. All studies were completed with three experimental and biological replicates unless stated otherwise.

Small compound library screening

To develop the machine learning model, a diverse library of 2,560 small compounds from Spectrum Library Micro Source Discovery was used to screen for growth inhibition. Compounds in the Spectrum Library Micro Source Discovery screening library were originally dissolved in 100% DMSO and stored at -20°C. DMSO can have an antimicrobial effect⁸⁰, therefore, we confirmed that a 1.25–5% concentration of DMSO did not affect bacterial growth nor viability (Fig. S8a, b). Starting cultures were added to 48-well plates containing each compound at a 50 µM final concentration in 2.5% DMSO. Bacterial growth was measured as the OD_600nm value at 2, 4, 6, and 24-hour timepoints. Compounds were classified as inhibitory at and above 80% growth inhibition compared to the bacterial growth in control wells with the DMSO and saline vehicle and separated into functional groups (Table S1).

Generation of a machine learning model and prediction of inhibitory compound structures

From the library screening dataset, a directed-message passing neural network (MPNN) was used to ascertain the compound structures and substructures associated with inhibition, as described previously²⁰. A directed MPNN was implemented via Chemprop to predict bacterial growth inhibition for compounds²⁴. Each compound is represented as a graph with atoms as nodes and bonds as edges, according to compound SMILES (Simplified Molecular Input Line Entry System)⁸¹. Bond-centered message passing is used to learn contextualized atom and bond representations that capture local chemical environments. The MPNN consists of three message-passing layers with a hidden size of 300. The resulting atom-level representations are aggregated by mean pooling to obtain a compound-level representation, which is further concatenated with 200-dimensional normalized two-dimensional physicochemical descriptors generated using RDKit, yielding a 500-dimensional feature vector⁸¹. This vector is passed to a two-layer feed-forward neural network comprising a linear transformation, ReLU activation, and a final linear layer to produce logits, followed by a softmax to obtain probabilities over four activity classes according to the following binary labels that indicate the effect of compounds on bacterial growth: 0–0 (no inhibition), 1–0 (temporary inhibition: inhibitory within 2–6 h), 0–1 (delayed inhibition: non-inhibitory at early timepoints but inhibitory at 24 h), or 1–1 (complete inhibition across all timepoints). The model is trained as a multiclass classifier using cross-entropy loss. To promote robust generalization and mitigate scaffold-level information leakage, we adopted a scaffold-balanced splitting strategy and perform 20-fold cross-validation, evaluating performance across folds on structurally diverse compounds. To interpret model predictions, we employed a rationale-extraction procedure based on Chemprop²⁹, which operates on top of the trained predictor to identify substructures critical for the predicted activity. For each compound, candidate substructures are explored using Monte Carlo Tree Search (MCTS), where substructures are scored using the predicted probability of complete inhibition from the trained model^24,29. After a fixed number of rollouts, substructures that satisfy predefined constraints on size and predicted score are retained as candidate rationales. The final rationale is defined as the smallest substructure that preserves the target property while achieving a high predicted score.

In silico screening of drug repurposing hub library

We used the trained model to predict inhibitory structures with antimicrobial potential from an FDA-approved and/or evaluated Drug Repurposing Hub of 6111 small molecules in silico. This library was selected for in silico screening to identify compounds with known safety profiles that could be re-purposed as antimicrobials. The output was a predictive score for inhibition between 0 (no inhibition) and 1 (full inhibition), where 1 was indicative of 80% or greater growth inhibition. The 200 hit compounds had predicted inhibitory activity at scores ranging from 0.415–0.712. The virtually screened compounds were filtered by their predicted inhibitory scores and known mechanism of action. Based on the filtration criteria, nine candidates were selected for further investigation to establish minimum inhibitory concentrations (MIC), whereas known antimicrobials were excluded.

In vitro validation of predicted compounds

The top nine predicted compounds (montelukast, GSK461364, AMG510, siponimod, ABT-263, miltefosine, CFTR inhibitor-172, GW3965, and probucol) were purchased from MedChemExpress (Monmouth Junction, NJ) and tested for growth inhibition of the emm92-15 isolate in vitro, essentially as described for the compound library screening assays. Briefly, overnight cultures on BHI agar were used to inoculate THY medium. Starting cultures of iGAS at an OD_600nm of 0.05 and two-fold dilutions of drugs in a concentration range of 1.56–50 µM were combined in wells. Bacterial growth was measured spectrophotometrically at OD_600nm over a 24-h period in triplicate wells.

For the remaining in vitro and in vivo experiments, montelukast sodium salt and GW3965 hydrochloride were purchased from Cayman chemical (Ann Arbor, MI). To determine the effect of each compound on bacterial survival, starting cultures were grown with 6.25 µM of montelukast or GW3965, as before. At the 24-hour timepoint, bacteria from wells treated with GW3965 or montelukast were plated on blood agar to determine CFU/mL. For comparison of viability, untreated and vehicle-treated bacteria were also plated. The percentage of survival was calculated according to the proportion of viable bacteria from the untreated (0 µM) versus montelukast or GW3965-exposed samples. For subsequent GW3965 testing against the MGAS5005, MGAS315, and emm92-3 iGAS strains, a concentration range of 1.56–50 µM was used to assess growth inhibition, as described above.

GW3965 and ampicillin synergy

Checkerboard assays were used to assess synergy between GW3965 and ampicillin. As the standard of care for GAS infections are β-lactam antibiotics, we tested GW3965 with ampicillin to determine the combined effect on emm92-15 isolate growth. For each replicate, ampicillin and GW3965 were serially diluted 2-fold in a checkerboard pattern at concentration ranges of 0.78–12.5 µM for GW3965, and 0.0015–0.1 µg/mL for ampicillin; controls at the established MIC for each compound alone were included. The OD_600nm value was measured over 24 h. The MIC for each compound separately or in combination at the 24-hour timepoint was used to calculate the fractional inhibitory concentration (FIC) index, as discussed previously⁸². The FIC index was calculated as follows for compound A (GW3965) and compound B (ampicillin): FIC index = FIC_A + FIC_B = A/MIC_A + B/MIC_B. FIC-values of <0.5, >4, or 0.5–4 were considered as synergistic, antagonistic, or indifferent effects, respectively.

Human skin equivalent GAS infection model

Wounded full-thickness EpiDerm-FT (MatTek, Boston, MA) in vitro cultured human skin equivalents were used to test GW3965 efficacy. Tissue discs with an 8 mm diameter are provided in transwells, with a 3-mm wound made in the center by punch biopsy of the keratinocyte layer. Wounds were inoculated with 10 µL of the GFP-expressing log-phase emm92-3 strain on Day 0, resulting in infection with 3.3 × 10⁶–1.1 × 10⁷ CFU/per tissue. To confirm effective wound establishment by GAS, infected tissues were harvested for histopathology at 24 h post-infection (p.i.) (Day 1, no GW3965 treatment), or after 6 days to assess baseline wound healing. Tissues were incubated in a humid environment at 37 °C in 5% CO₂ with daily-fresh media; with duplicate tissues per condition. Beginning on Day 1, tissues were treated with 12.5 µM of GW3965 delivered daily in culture media placed in the bottom chamber of the transwells. At end points, half of each tissue was fixed with 10% formalin for histopathological evaluation. Histology slides were scanned at 40x magnification on a Huron Technologies TissueScope LE and viewed in a blinded fashion. Scanned images were annotated using SlideVault software (Huron Digital Pathology, Ontario, Canada).

Animal use and welfare

All experiments were performed in 5-week-old male and female, immunocompetent, outbred, hairless mice (strain Crl:SKH1-hrBR) (Charles River, Wilmington, MA). Animal experiments were conducted in compliance with the regulations and standards under the Animal Welfare Act, the Public Health Service Policy on Humane Care and Use of Laboratory Animals, and the Guide for the Care and Use of Laboratory Animals. The protocol 1602000144_R2 ethics approval was completed by the West Virginia University Institutional Animal Care and Use Committee (IACUC). At endpoint, animals were humanely euthanized under deep anesthesia (reagents: Isoflurane 1-5% chamber vol/min) with the methods of exsanguination followed by cervical dislocation.

Skin and soft tissue (SSTI) mouse model of iGAS infection

A range of inocula was initially tested to establish an effective infection dose with the emm92 iGAS bacteria in the murine SSTI model. Mice were anesthetized via isoflurane inhalation (3% in oxygen) and injected subcutaneously (s.c.) in the right flank with 10⁷, 10⁸, or 10⁹ CFU of emm92-3 iGAS, as described previously³⁴. Gross pathophysiology of infection was measured based on whether a lesion developed, changes in mouse weight, and effect on external body temperature for the first 4 days and then every 2 days for the remainder of the 11 days of experiment.

An inoculum size of ~1 × 10⁹ CFU/per mouse was established as optimal based on consistent production of skin pathology with resolution over the length of 24-day experiments. The percentage of weight change over time and lesion size were measured daily unless otherwise noted. Abscess dimensions (length [L] and width [W]) for each mouse were recorded daily to calculate area [A = p(L/2) × (W/2)] and volume [V = 4/3p(L/2)² × (W/2)] with equations for a spherical ellipsoid³⁴.

GW3965 toxicity and plasma pharmacokinetics

Groups of 2 mice were intraperitoneally (i.p.) injected with 10, 30, or 50 mg/kg of GW3965 dissolved in 5% DMSO to assess toxicity in the SKH1 mouse model. GW3965 was administered daily for 4 days, and mice were observed for the effect of treatment on weight and temperature. Changes in weight were calculated for each individual mouse based on the percentage change as compared to baseline weight before treatment began (Day 0).

To determine plasma drug concentration, blood was collected by submandibular puncture from untreated mice in tubes coated with 0.5 M EDTA (pH 7.4). 50 mM EDTA was added to prevent coagulation, and plasma was isolated by centrifugation at 1400 g for 10 min. For standard curve preparation, 40 µL of blank plasma from untreated mice was spiked with 100-900 ng/mL of GW3965, followed by precipitation with 400 µL of acetonitrile. Detection of GW3965 by LC-MS was performed on the AB Sciex Exion LC model and the AB Sciex QTrap5500 mass spectrometer (AB Sciex, Framingham, Massachusetts) using a Phenomenex Luna Omega 1.6 µm C18 100 Å, 50 × 2.1 mm column (Torrance, California), as described⁸³. The transition for GW3965 is 583.9 > 181.2 m/z (Q1 > Q3). Analyst 1.7.2 software was used for data acquisition of LC-MS readouts. A calibration curve was used to calculate the concentration recovered from the plasma standards using MultiQuant3.0.3 software (AB Sciex, Framingham, Massachusetts). For GW3965 pharmacokinetics in mouse blood during infection treatment, mice were bled at 2, 4, and 24 h after i.p injection of the compound. Blood was collected and processed as above, and GW3965 concentration in plasma was measured via the described LC-MS method.

In vivo GW3965 efficacy

Mice were injected s.c. with 1.8–5.2 × 10⁹ CFU of the emm92 strain, as above. Starting on day 1 (24 hr p.i.), a 10-mg/kg ( ~ 16.2 µM) or 30-mg/kg ( ~ 48.6 µM) dose of GW3965 was administered i.p. for 5 days. Control groups of mice received a 50-mg/kg dose of Penicillin G in saline, saline alone, or vehicle (DMSO) using the same schedule. Mice were observed for the effect of treatment on the gross pathophysiology of infection for either 6- or 18-day endpoints. Data was collected as described above. Blood and organ/tissue (e.g., spleen and skin) samples were collected for enumeration of bacterial burden at the indicated times. To assess bacterial burden, the spleens and remaining halves of skin samples were homogenized and serially diluted for plating on blood agar and BHI agar containing 8 µg/mL of erythromycin. Whole blood was plated on blood agar for detection of bacteriemia.

Statistical analysis

GraphPad Prism software was used for statistical analysis. For in vitro growth assays and skin equivalent experiments statistical significance was determined by either One-Way ANOVA or Unpaired t-tests with Welch’s correction unless otherwise noted. Data from mouse experiments was assessed for differences between treatment groups over the course of infection using a Two-Way ANOVA with Tukey’s multiple comparison or a mixed model ANOVA with Tukey’s multiple comparison, as was necessary according to missing values. The error bars in each graph show the standard deviation of values.

Supplementary information

Supplementary Material ^{(1.5MB, pdf)}

Author contributions

L.M.P., M.E.G., W.S., P.R.L., and S.L. conceived the experimental design and performed analyses of the results. P.R.L. analyzed histopathology samples. W.S. performed machine learning model development, training, and artificial intelligence assisted in silico screening. S.B. and W. G. aided in the separation of predicted compounds according to their chemical properties and modeling of structure properties. L.M.P., M.E.G., K.Z., W.S., and S.J.C. collected experimental data. L.M.P., M.E.G., P.R.L., and S.L. drafted the manuscript, which was reviewed and accepted by all authors.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s44259-026-00225-4.