A generative deep learning approach to de novo antibiotic design

Summary

The antimicrobial resistance crisis necessitates structurally distinct antibiotics. While deep learning approaches can identify antibacterial compounds from existing libraries, structural novelty remains limited. Here, we developed a generative artificial intelligence framework for designing de novo antibiotics through two approaches, a fragment-based method to comprehensively screen >10⁷ chemical fragments in silico against Neisseria gonorrhoeae or Staphylococcus aureus, subsequently expanding promising fragments, and an unconstrained de novo compound generation, each using genetic algorithms and variational autoencoders. Of 24 synthesized compounds, seven demonstrated selective antibacterial activity. Two lead compounds exhibited bactericidal efficacy against multidrug-resistant isolates with distinct mechanisms of action and reduced bacterial burden in vivo in mouse models of N. gonorrhoeae vaginal infection and methicillin-resistant S. aureus skin infection. We further validated structural analogs for both compound classes as antibacterial. Our approach enables the generative deep learning-guided design of de novo antibiotics, providing a platform for mapping uncharted regions of chemical space.

Introduction

Antibiotic-resistant bacterial infections underlie an urgent public health crisis and are associated with approximately five million annual deaths globally¹. Bacterial pathogens such as Neisseria gonorrhoeae and Staphylococcus aureus are categorized as “urgent” and “serious” threats, respectively, by the U.S. Centers for Disease Control and Prevention (CDC) due to the widespread incidence of resistance to existing antibiotics and dearth of new, effective antibiotic treatments². Between 1980 and 2003, only five antibacterial agents were developed by the top 15 pharmaceutical companies³. There is a dire and growing need for structurally novel compounds, especially those with unique mechanisms of action, to overcome antibiotic-resistant infections.

Given the extensive time and resources needed for antibiotic discovery, computational approaches to facilitate the identification of new candidates have emerged. Recently, deep learning approaches have been used to screen millions of compounds in silico for activity against diverse pathogens^4-11. A graph neural network (GNN), for example, is a type of deep learning model that has successfully been used to predict antibacterial activity. GNNs represent chemical structures as mathematical graphs and iteratively update graph node and edge values using “message passing” operations¹². Ultimately, for a given molecule, a GNN produces a single output value between 0 and 1, representing the predicted probability that the molecule possesses a specific property of interest (e.g., antibacterial activity or cytotoxicity).

Although deep learning approaches have substantially increased discovery rates^4-8 and facilitated the discovery of a new structural class of antibiotics⁷, they have largely been applied to existing in silico small molecule libraries, limiting the structural diversity that can be explored. Drug-like chemical space has been theoretically estimated to contain ~10⁶⁰ compounds¹³, but the largest in silico libraries currently contain only ~10¹¹ compounds^14,15. Generative artificial intelligence (AI) approaches can expand beyond these known chemical spaces^16-23 and have recently been applied to both antimicrobial peptide design^24-26 and to the design and optimization of small molecules^27-32. A recent study explored a large chemical space using a Monte Carlo tree search coupled to a GNN and found compounds with antibacterial activity against Acinetobacter baumannii when used in the presence of an outer membrane permeabilizer³³. Developing and experimentally validating generative AI methods that can design structurally novel antibacterial molecules, beyond those enumerated in commercial chemical spaces, would substantially augment the capability to search chemical space for antibiotic candidates.

In recent years, fragment-based drug discovery (FBDD) has emerged as a powerful tool for drug design, offering an efficient framework to screen large fragment libraries against a specific protein target^34-37. In contrast, target-agnostic approaches prioritize phenotypic measurements from the start. By screening against whole-cell activity rather than specific molecular targets, phenotypic approaches can integrate useful structural information across diverse chemotypes, can yield molecules with diverse mechanisms of action, and could reduce the downstream attrition rates commonly associated with target-centric methods³⁸. Despite their potential, phenotypic-guided approaches to both fragment-based and de novo molecule generation remain underexplored, representing a significant opportunity to develop novel chemical entities informed by biological outcomes, rather than predefined structural constraints.

Here, we present a deep learning-based approach to the generative design of antibiotic compounds using fragments as starting points or generate these compounds de novo. In our fragment-based approach, we used GNNs to comprehensively screen >45 million chemical fragments in silico and identified fragments predicted to have selective antibacterial activity against N. gonorrhoeae and S. aureus. To expand the fragments into molecules, we provided them as inputs to two types of generative algorithms – a genetic algorithm based on chemically reasonable mutations (CReM) and a variational autoencoder (VAE). In our de novo approach, we removed the need for a fragment input and allowed the CReM and VAE models to design molecules based on knowledge learned during training. Together, these models generated >36 million previously unenumerated compounds with predicted antibacterial activity. After down-selection, we synthesized 24 compounds and empirically validated seven compounds to be antibacterial. Two of these compounds, NG1 and DN1, displayed high potency and selectivity, with mechanisms of action distinct from those of clinically used antibiotics and efficacy in mouse infection models. Our approach provides a platform for the generative deep learning-guided design of antibiotic candidates, facilitating the discovery of antibacterial compounds and enabling the efficient exploration of vast, uncharted regions of chemical space.

Results

Applying graph neural network models to large fragment spaces

As chemical fragments underlie the biological activity of structural classes of compounds and graph search algorithms can accurately identify substructures associated with selective antibacterial molecules⁸, we reasoned that fragments offer starting points for designing structurally unique antibacterial molecules using generative models. We found that the antibacterial prediction scores of commonly used antibiotics in three major classes are indeed associated with the prediction scores of their corresponding fragments (STAR Methods and Figure S1A). Accordingly, we comprehensively screened large, enumerated chemical fragment spaces using Chemprop models¹² based on graph neural networks (GNN) that predict the antibacterial activity of a given fragment or molecule against N. gonorrhoeae or S. aureus (Figure 1A). These models were previously trained and benchmarked on empirical, binarized growth inhibition data for 38,765 compounds screened against N. gonorrhoeae ATCC 49226 and 39,312 compounds screened against S. aureus RN4220, a methicillin-susceptible strain¹³, at a final concentration of 50 μM^7,8 (Table S1). To further assess our models’ predictive capabilities, we performed additional analyses and observed that the prediction scores for N. gonorrhoeae and S. aureus (1) significantly differed between compounds that were experimentally active and inactive (Figure S1B), (2) were influenced by certain functional groups (Figure S1C) and (3) demonstrated species-specificity (Figure S1D). These analyses indicated that our models could reliably identify antibacterial compounds and fragments as starting points for generative AI algorithms. To avoid selecting human cytotoxic compounds, we counter screened fragments using human cytotoxicity models for hepatocellular carcinoma (HepG2), human primary skeletal muscle (HSkMC), and human lung fibroblast (IMR-90) cells that we had previously trained and benchmarked (Table S1)⁷.

Figure 1. Graph neural networks predict chemical fragments with antibacterial activity against N. gonorrhoeae.

(A) Schematic of the fragment-based generation approach.

(B) t-SNE representation of chemical space, showing that GDB fragment libraries represent chemical space diverse from known antibiotics.

(C) Rank-ordered fragment prediction scores from the N. gonorrhoeae antibacterial activity model.

(D) Computational filters applied to sets of fragments, resulting in final sets of fragments for N. gonorrhoeae.

(E) Rank-ordered plot of fragment (above) and compound (below) prediction scores for available compounds procured for empirical testing at 50 μM.

(F) Chemical structures of fragment F1 and associated active compounds from (E). The fragment is highlighted in the compounds.

(G) MIC and CC₅₀ values for the compounds in panel (F). n=2, * indicates values >32 μg/mL.

See also Figure S1.

To comprehensively search fragment space, we assembled a database of >45 million chemical fragments with chemically meaningful structures (“possible fragments”). The database combines three large libraries: (1) all possible (chemically stable) fragments containing up to 11 atoms of C, N, O, and F from the Generated DataBase-11 (GDB-11) (26,434,571 fragments); (2) all additional possible fragments containing up to 11 atoms of C, N, O, Cl, and S from GDB-13 (1,089,000 fragments); and (3) all fragments in the Enamine REadily AccessibLe (REAL) database, which have improved synthetic accessibility and can vary in the number of atoms per fragment^39-43 (18,338,026 fragments). We then examined the structural novelty spanned by these fragment libraries by visualizing the chemical space they occupied relative to our training library and a manually compiled set of 559 known antibacterial compounds (Data S1). We used a t-distributed stochastic neighbor embedding (t-SNE) plot based on Tanimoto similarity using RDKit fingerprints as a distance metric, a commonly used measure of chemical similarity (Figure 1B). This two-dimensional projection of chemical space, in which structurally-similar fragments or compounds cluster, showed that the fragment libraries substantively expand on the chemical diversity beyond that of known antibiotics (Data S1). Thus, our curation resulted in a consolidated, exhaustive database of 45,858,026 unique fragments spanning a diverse subset of chemical space.

Fragment-based generative design of antibacterial compounds

We virtually screened our fragment library using Chemprop models predicting antibacterial activity against N. gonorrhoeae (Figure 1C) and shortlisted those with prediction scores >0.05 in the GDB databases and, due to better synthetic accessibility, >0.1 in the Enamine REAL database (Figure S1E). This resulted in 3,844,505 fragments (8.4%) predicted to be active against N. gonorrhoeae (Figure 1D). To prioritize structurally novel and selective fragments with no obvious chemical liabilities, we implemented a series of filters on the fragments (Figure 1D and S1E). First, we removed fragments that were predicted to be cytotoxic (cytotoxicity prediction score >0.5) by any of the three human cell cytotoxicity models (HepG2, HSkMC, and IMR-90). Second, we removed fragments containing pan-assay interference substructures (PAINS) or Brenk substructures, which are associated with unfavorable properties such as promiscuous binding, toxicity, chemical reactivity, and metabolic instability^44,45. Third, we required that all remaining fragments be structurally distinct from those in known antibiotics, which we enforced by requiring the Tanimoto similarity to be <0.5 with respect to any compound in our curated set of 559 known antibacterial compounds (STAR Methods, Data S1, and Figure S1E). The number of unique fragments passing these filters was 1,156,945 (Figure 1D).

To validate that our screening approach successfully identifies fragments with antibacterial activity, we searched for full compounds that contained any of our filtered fragments and selected those predicted to be antibacterial (compound prediction score >0.1) and non-cytotoxic (cytotoxicity prediction score <0.5), without PAINS or Brenk substructures, and structurally distinct from known antibiotics (Tanimoto similarity <0.5). Applying these filters to 799,149 compounds from the Broad Institute^6-8,46, we obtained 66 molecules representing 27 fragments for testing against N. gonorrhoeae ATCC 49226 (Figure 1E). We found that 7 of them inhibited the growth of N. gonorrhoeae at 50 μM, resulting in a working true discovery rate of 10.6% (Figure 1E, Table S2, and Data S2). Notably, a single fragment, F1 was represented in two active Broad (BRD) compounds, BRD1 and BRD2 (Figure 1F), and the minimum inhibitory concentrations (MICs) of these compounds ranged from 8 to 16 μg/mL. The compounds also exhibited a range of selectivity values, with half-maximal cytotoxic concentration (CC₅₀) values of 32 to >128 μg/mL when tested against HEK293, HepG2, HSkMC cells, respectively (Figure 1G). The therapeutic indices (TIs; ratio of human cell CC₅₀ to MIC values) were thus between 2 to 8 for the two compounds associated with F1. To identify molecules associated with F1 exhibiting higher therapeutic indices, we searched for analogs containing F1 and found a paucity of compounds in a purchasable compound library comprising >6 million molecules (STAR Methods). Therefore, we reasoned that a generative approach to the design of compounds based on F1 could enable the exploration of much larger chemical spaces and produce compounds with higher potency and selectivity.

We leveraged two generative machine learning algorithms: a genetic algorithm based on CReM, and a fragment-based VAE (F-VAE)^18,21,47 (Figure 2A and 2C). CReM is a computational framework that starts with a compound of interest and generates new molecules by adding, replacing, or deleting atoms and functional groups. When additions are made, they are sampled from up to 1,557,992 distinct structures containing only common atoms (C, N, O, S, P, F, Cl, Br, I, and B) from ChEMBL⁴⁸. In our pipeline, we coupled CReM to Chemprop models for antibacterial activity prediction (hereafter referred to as F-CReM), such that only high-scoring compounds containing F1 (prediction score >0.7) were successively provided as inputs for the next round of molecule generation. After five rounds of selection, where each round resulted in progressively higher-scoring compounds (Figure 2B), we generated 518,203 F1-containing compounds. To obtain an experimentally tractable list of promising molecules, we filtered these compounds based on their predicted antibiotic activity and cytotoxicity, Tanimoto similarity to known antibiotics and calculated synthetic complexity (based on either the synthetic accessibility score [SAscore]⁴⁹, retrosynthetic accessibility score [RAscore]⁵⁰ or R-score from Spaya Iktos⁵¹ (STAR Methods). We further ensured that the compounds of interest did not display PAINS or Brenk alerts. This resulted in 285 F-CReM-generated compounds with predicted activity against N. gonorrhoeae (Figure 2D).

Figure 2. Generative deep learning-guided design of compounds containing fragment F1 with predicted antibacterial activity.

(A) Schematic of the genetic algorithm based on F-CReM and examples of filtered generated compounds, with fragment F1 highlighted.

(B) Distributions of antibacterial activity prediction scores from five rounds of compound generation using F-CReM.

(C) Schematic of the F-VAE algorithm and examples of filtered generated compounds.

(D) Computational filters applied to fragment-based designed compounds, resulting in two final sets of compounds shortlisted for chemical synthesis.

(E) Molecular weights, topological polar surface area (TPSA), calculated log P (clogP), hydrogen bond acceptors, hydrogen bond donors, and retrosynthetic accessibility score (RAscore) values for all filtered compounds generated by F-CReM and F-VAE, compared to the median of known Gram-negative antibiotics.

(F) Structures of two designed and synthesized compounds, NG1 and NG2.

(G) MICs against N. gonorrhoeae ATCC 49226 and CC₅₀ values of the compounds for four different human cell types. n=2, * indicates values >128 μg/mL.

(H) MICs of compound NG1 against various bacterial species. Results are representative of two biological replicates.

See also Figure S2.

In addition to F-CReM, which relies on computational sampling of chemical modifications, we employed a deep learning-based variational autoencoder for generative design. Specifically, we developed a fragment-based variational autoencoder (F-VAE) trained on all 1,686,695 compounds from ChEMBL (v2019)⁴⁸. The F-VAE architecture (Figure 2C) consists of a graph convolutional network encoder module, in which molecular graphs are encoded as latent vectors, as well as a recurrent graph decoder module, in which latent vectors are converted back to molecular graphs. The decoding procedure starts from a fragment and expands it atom-by-atom until the assembled structure converges into a compound. Compounds are designed by the F-VAE by sampling the latent space generated from an input fragment (e.g., by creating random latent vectors and passing them to the decoder). By applying the F-VAE, we generated 6,937,677 molecules containing fragment F1. We then used our Chemprop models to predict antibacterial activity for all of these molecules. Filtering the molecules similarly to above (STAR Methods, Figure S1F), we obtained 678 F-VAE generated compounds with predicted activity against N. gonorrhoeae (Figure 2D).

To better understand and comparatively assess the chemical matter produced by the two generative models, we evaluated their physicochemical properties and synthesizability (Figure 2E). We found that F-CReM-generated compounds typically possessed lower molecular weights, lower calculated partition coefficients (calculated log P; clogP)⁵², and lower calculated topological polar surface area (TPSA) values, compared to F-VAE-generated compounds. Both models generated synthesizable molecules, with comparable RAscores as computed by a machine learning (ML)-based synthetic feasibility estimator⁵⁰. In general, all compounds generated based on F1 exhibited lower molecular weights, lower number of H-bond acceptors and donors, larger clogP values, and lower TPSA values compared to known Gram-negative antibiotics (Figure 2E). Taken together, our analyses suggest that our generative approach can produce realistic and synthesizable compounds with promising Chemprop prediction scores. We therefore aimed to synthesize and test several of these compounds.

Synthesis and experimental validation of fragment-based designed compounds

Synthesizing deep learning-generated compounds can be challenging due to the possibility of formidable or undefined synthesis routes. Therefore, to obtain a diverse and reasonably-sized set of compounds for synthesis, we prioritized compounds that were structurally dissimilar from each other and had the highest predicted antibacterial scores (>0.7) (Figure 2D, STAR Methods). This resulted in 80 compounds, which were evaluated by commercial chemical synthesis providers. After attempting to synthesize 27 compounds (9 from CReM and 18 from F-VAE), we successfully synthesized two (NG1 and NG2) with high (>95%) purity (Figure 2F). Both molecules were generated by the F-VAE model. We empirically tested the compounds and found that NG1, but not NG2, inhibited the growth of N. gonorrhoeae ATCC 49226 with an MIC of 0.5 μg/mL (Figure 2G). Determining selectivity as before, we found that NG1 had CC₅₀ values of 25 to 128 μg/mL against HEK293, HepG2 and HSkMC cells, resulting in TIs of 50 to 256 (Figure 2G). NG1 is structurally dissimilar to active compounds in the training dataset and possesses favorable physicochemical properties (Table S5). Interestingly, NG1 exhibited narrow-spectrum activity against N. gonorrhoeae and Neisseria meningitidis – the only other pathogenic Neisseria species – but not against Neisseria cinerea and Neisseria mucosa (two human commensal species) or any other Gram-positive or Gram-negative species tested (Figure 2H). NG1 also exhibited potent activity against highly drug-resistant N. gonorrhoeae strains, including the first strain found in the United States that had lost susceptibility not just to ceftriaxone but to all other drugs previously recommended for first-line treatment (Figure 3A, Table S4). These data consistently indicate that NG1 is antibacterial against multidrug-resistant and pathogenic strains of N. gonorrhoeae, suggesting that it might act through a mechanism of action to which resistance has not yet evolved.

Figure 3. Mechanistic study and in vivo efficacy of compound NG1.

(A) MICs against wild-type and antibiotic-resistant strains of N. gonorrhoeae (listed in Table S4). Each data point represents a strain tested in biological duplicate.

(B) Time-kill curves for ATCC 49226 treated with NG1 or azithromycin. Mean ± SD; n=2.

(C) Minimum bactericidal concentration of NG1 in ATCC 49226. Mean ± SD; n=2.

(D) Membrane rigidification of ATCC 49226 treated with NG1 at 64 μg/mL and control compounds, as measured by the Laurdan generalized polarization (GP). Mean ± SD; n=2.

(E) Loss of membrane integrity of ATCC 49226 treated with NG1 and DMSO, as measured by the uptake of the hydrophobic fluorescent probe, 1-N-phenylnaphthylamine (NPN). Mean ± SD; n=2.

(F) Membrane permeabilization of ATCC 49226 treated with NG1 or Triton X-100, as seen by a time-dependent increase in SYTOX Green fluorescence signal normalized to the untreated control. Mean ± SD; n=2.

(G) Volcano plot of PISA results from FA1090 lysate treated with NG1 compared to the vehicle control. The most significantly affected protein, LptA, is shown in purple.

(H) Expression levels of lptA in FA1090 treated with NG1 at 4× MIC, expressed in log₂ counts-per-million (CPM).

(I) Checkerboard study of NG1 and polymyxin B in ARB #0187 revealing a synergistic effect. Data are representative of biological duplicate.

(J) Schematic of the in vivo study of NG1 in the N. gonorrhoeae vaginal infection model using strain ATCC 49226. Treatment with vehicle, NG1, or ceftriaxone was administered intravaginally for five doses.

(K) Bacterial titers in vaginal lavage fluid. Horizontal lines represent median log10 CFU/mL values. Data represent treatment with vehicle control (n=8), NG1 (n=9), or ceftriaxone (n=5). Two-sided Mann-Whitney U-test compared to vehicle: *p ≤ 0.05, **p ≤ 0.01. Horizontal lines represent medians.

(L) Summary of a structure-activity relationship analysis for NG1.

See also Figures S2 and S3.

Mechanism of action, toxicology, and in vivo efficacy of NG1

We investigated the mechanism of action of NG1 by first examining whether the compound was bactericidal. In a time-kill experiment using N. gonorrhoeae ATCC 49226, NG1 exhibited concentration-dependent killing, with efficacy similar to that of azithromycin (Figure 3B), and a minimum bactericidal concentration (MBC) of 1 μg/mL (Figure 3C). In suppressor mutant generation experiments on solid agar, the frequency of resistance against NG1 was 4.3 × 10⁻⁸ at 8× MIC. The spontaneously arising NG1-resistant isolates retained susceptibility to ceftriaxone, azithromycin, and ciprofloxacin, with unchanged MICs relative to those of the parental susceptible strain (Figure S2A). The lack of cross-resistance was supported by checkerboard assays demonstrating that NG1 acted indifferently (neither synergistically nor antagonistically) to ceftriaxone, fosfomycin, and CCCP, indicating that NG1 does not act similarly to other cell wall- and membrane proton motive force (PMF)-targeting antibiotics (Figure S2B). Indeed, NG1 did not alter the PMF, as measured by the potentiometric fluorophore DiSC₃(5) (3,3-dipropylthiadicarbocyanine iodide)^4,7,53, which displays an increase in fluorescence when the membrane potential, ΔΨ, is disrupted and a decrease in fluorescence when the pH gradient, ΔpH, is dissipated⁵³ (Figure S2C).

To investigate the mechanism of action further, we tested whether NG1 treatment alters membrane fluidity using a Laurdan dye assay⁶, where cells with decreased membrane fluidity exhibit increases in Laurdan fluorescence. Treatment of N. gonorrhoeae cells with NG1 resulted in a modest increase in Laurdan fluorescence, suggesting that NG1 may act, in part, by decreasing membrane fluidity (Figure 3D). We hypothesized that this would compromise membrane integrity and tested this hypothesis by measuring the uptake of a hydrophobic fluorescent probe, 1-N-phenylnaphthylamine (NPN), which fails to penetrate intact outer membranes. NG1 treatment resulted in a significant increase in NPN fluorescence, suggesting that the outer membranes of cells were indeed compromised (Figure 3E). NG1-treated N. gonorrhoeae cells also exhibited increased fluorescence of SYTOX Green (a DNA-intercalating dye) that only penetrates cells with compromised membranes (Figure 3F), supporting the suggestion that membrane damage leads to cell death. We then directly examined the morphological changes induced by NG1 treatment by performing cryogenic transmission electron microscopy (cryo-TEM) on N. gonorrhoeae ATCC 49226 cells. We found pronounced changes, where the NG1-treated cells showed reduced area, roundness, and increased elongation (Figure S2D and S2E).

As NG1 might compromise bacterial membranes by acting on membrane-related proteins, we performed a proteome integral solubility alteration (PISA) assay to study possible protein targets on a proteome-wide scale⁵⁴. The assay revealed a striking de-stabilization of the lipooligosaccharide (LOS) export system protein, LptA (p value = 1.9 × 10⁻⁷, log₂ fold change = −2.4) upon treatment of N. gonorrhoeae lysate with NG1 (Figure 3G). Time-resolved RNA sequencing experiments revealed a dose-dependent up-regulation of lptA in response to NG1 treatment, in addition to changes in either upstream or downstream proteins of the LOS biosynthesis pathway (Figure 3H and S2F). Given NG1’s potential effect on LOS, we hypothesized that NG1 would synergize with polymyxin B, an antibiotic which extracellularly binds lipid A⁵⁵, and indeed observed a potent synergistic interaction (fractional inhibitory concentration index <0.5, Figure 3I). Taken together, these experimental results strongly suggest that LptA is the main target of NG1. This mechanism of action is notable, as LptA has been a proposed⁵⁶, but as yet undrugged, antibiotic target in the LOS biosynthesis pathway.

Given that NG1 appears to inhibit LptA, a protein absent in eukaryotic cells⁵⁷, we aimed to further study its translational potential. We first carried out in vitro toxicology studies measuring human red blood cell hemolysis and bacterial mutagenesis. We found that NG1 was neither hemolytic nor mutagenic up to concentrations of 64 μg/mL, the highest tested (Figure S2G and S2H). Building on these observations, we tested the efficacy of NG1 in a mouse model of N. gonorrhoeae vaginal infection. Here, ovariectomized and estradiol-treated mice were intravaginally inoculated with N. gonorrhoeae ATCC 49226. Two hours later, mice were given their first dose of intravaginal NG1 (1%), ceftriaxone (0.1%), or vehicle control, followed by four additional doses within a 24-hour period (Figure 3J). We found that mice treated with NG1 exhibited a significant decrease in vaginal bacterial load (two-sided Mann-Whitney U-test, p=0.0120) of ~3 logs in colony forming units (CFUs) relative to vehicle-treated mice, with efficacy approaching that of ceftriaxone (Figure 3K).

Furthermore, treatment with NG1 was well-tolerated by all mice. Overall, these results support the translational potential of NG1 with topical in vivo efficacy in treating a model of N. gonorrhoeae infection.

Synthesis and analysis of structural analogs of NG1

Building on the high selectivity of NG1, as well as to further test the functional significance of fragment F1, we investigated its structure-activity landscape by synthesizing and testing structural analogs. We synthesized 74 analogs primarily by altering two functional groups, R1 and R2, connected to the pyrrolidine ring and the amine, respectively (Figure S3A). Testing all analogs, we found eight active molecules against N. gonorrhoeae ATCC 49226, with MIC values between 16 to 32 μg/mL (Figure S3B and S3C). We also procured four analogs in the Enamine REAL Space library and one of the four compounds, named NG1 analog, was active against N. gonorrhoeae ATCC 49226 with an MIC of 4 μg/mL. Additionally, its CC₅₀ values against HEK293, HepG2 and HSkMC cells were higher than those of NG1 (Figure S3D). Given its promise for further development, we investigated NG1 analog’s mechanism of action as above and found that it acted similarly to NG1 (Figure S3E, S3F and S3G). From this analysis, we observed that compounds with an R1 group containing a 2,3-dichlorophenyl group were more effective at inhibiting N. gonorrhoeae than those containing a 2,4-dichlorophenyl group, and compounds containing either group were more efficacious than those containing a 2-chlorophenyl or phenyl group alone (Figure 3L). Together, these findings support the antibacterial activity of fragment F1 and its associated molecules, suggesting the possibility that additional optimization will improve the potency and selectivity of this structural class of antibacterial compounds.

Design of compounds with activity against S. aureus

Having taken a fragment-based approach to designing compounds active against N. gonorrhoeae, we asked whether the same approach could produce compounds active against a Gram-positive pathogen, S. aureus. To start, we predicted antibacterial activity against S. aureus using an ensemble of 20 GNN-based Chemprop models^12,46,58. We used this ensemble of S. aureus models to screen our original set of 45,858,026 fragments using the same prediction score thresholds as before. Of these fragments, 432,919 (0.94%) were predicted to be active against S. aureus (Figure 4A), and 259,807 unique fragments remained after filtering for low predicted human cytotoxicity scores, no obvious chemical liabilities, and structural novelty as before. To validate that the identified fragments were associated with compounds with antibacterial activity against S. aureus, we searched our set of 799,149 compounds from the Broad Institute^6-8,46 as well as the Enamine purchasable space, to identify compounds containing these fragments. Applying similar filters as above to these compounds, we found 130 compounds representing 68 unique fragments that we procured for empirical testing against S. aureus RN4220 (Figure 4B). We found that 19 compounds inhibited the growth of S. aureus at 50 μM or below, resulting in a working true discovery rate of 14.6% (Figure 4B, Table S2, and Data S2). Notably, the active fragment with the highest model prediction score, F2, was represented in the second-highest scoring compound, EN1 (Figure 4C), with an MIC of 8 μg/mL against methicillin-susceptible S. aureus (MSSA) RN4220 and methicillin-resistant S. aureus (MRSA) BAA1556. EN1 also exhibited half-maximal inhibitory concentration (CC₅₀) values ranging from 32 to >64 μg/mL when tested against HEK293, HepG2 and HSkMC cells (Figure 4D), indicating favorable selectivity.

Figure 4. Effect of fragment modification on synthesizability and antibacterial activity.

(A) Rank-ordered fragment prediction scores as determined by the S. aureus antibacterial activity model (above) and computational filters applied to the fragments (below).

(B) Rank-ordered fragment (above) and compound (below) prediction scores for available compounds procured for empirical testing at 50 μM.

(C) Structure of fragment F2 and its associated active compound from (B). The fragment is highlighted in EN1.

(D) MIC and CC₅₀ values for EN1. n=2, * indicates values >64 μg/mL.

(E) Structures of fragment F2 and the truncated fragment F2’, with the number of molecules generated for both fragments by the F-CReM and F-VAE models.

(F) Computational filters applied to down-selected compounds using fragment F2’.

(G) RAscore of molecules generated by F-CReM and F-VAE using fragments F2 and F2’.

(H) Retrosynthesis success rates as determined by ASKCOS for molecules generated by F-CReM and F-VAE using fragments F2 and F2’.

(I) MICs of 12 compounds from the Enamine REAL space with fragment F2 and 16 de novo designed and synthesized compounds with the fragment F2’ against S. aureus RN4220.

See also Figure S4.

EN1 contained a unique 1,4-epoxycyclohexane moiety, and a search within Enamine’s 64 billion compound space yielded only 38 structures containing F2, suggesting that it might be a rare and appropriate starting point for generative AI algorithms that expand on its associated chemical space. We applied F-CReM and F-VAE to F2, and these models generated 141,109 and 3,297,292 molecules, respectively (Figure 4E). Filtering the resulting molecules similarly as before (STAR Methods, Figure S1F), we retained the top 210 compounds predicted to be antibacterial against S. aureus. These remaining compounds were down-selected to five compounds by a medicinal chemist to prioritize properties such as synthesizability, stability, permeability, drug-likeness, and diversity; however, none of these compounds were synthetically accessible within typical synthesis timeframes and costs. To circumvent synthesizability obstacles and determine if preserving the entire F2 fragment is essential for our approach, we truncated F2 to a more synthetically-accessible starting point, F2’ (Figure 4E). We used F-CReM and F-VAE to generate 106,557 and 18,164,746 molecules based on F2’ and down-selected 9,684 molecules as before (Figure 4F, STAR Methods). As expected, the molecules associated with F2’ were more likely to be synthesizable than those associated with F2, as measured by two metrics: the RAscore (Figure 4G) and a retrosynthesis planning tool, ASKCOS (Automated System for Knowledge-based Continuous Organic Synthesis) (Figure 4H). Intriguingly, F-VAE-generated molecules were consistently predicted and selected by a medicinal chemist to be more synthesizable than those from F-CReM and exhibited more drug like properties (Figure S4A, Table S3). Confirming these predictions, we were able to synthesize 16 F-VAE-designed molecules within typical synthesis timeframes and costs. We found that only one of these 16 compounds (6.25%) (Figure S4B, Data S2) demonstrated activity against S. aureus RN4220, with a marginal MIC of 128 μg/mL, the highest concentration tested (Figure 4I). By contrast, six of the 12 additional molecules (50%) (Figure S4B, Data S2) procured from the Enamine REAL Space containing the original F2 fragment, displayed activity, exhibiting MICs as low as 16 μg/mL (Figure 4I). Together, these findings indicate that preservation of F2 is important for antibacterial activity and that designing compounds based on this particular fragment appears to result in a tradeoff between antibacterial activity and synthesizability.

Generation, synthesis, and experimental validation of de novo designed compounds

To further investigate the generality of our approach, we asked if our generative models could design antibacterial molecules without the need for specific fragments as starting points (Figure 5A). We modified our models by providing either ammonia, methane, or water as inputs to CReM (Figure 5B), and no fragment as input to junction-tree variational autoencoders (JT-VAE) (Figure 5C). Together, the models generated 29,014,974 molecules. Interestingly, the JT-VAE model produced more drug-like molecules (Figure 5D, Table S3), while compounds produced by the CReM model tended to have high molecular weights and many H-bond acceptors and donors (which could reduce membrane permeability) (Figure 5D, Table S3). CReM compounds were predicted to be highly un-synthesizable by both the RAscore (Figure 5D) and the ASKCOS retrosynthesis prediction tool (Figure S5A, Table S3). We therefore prioritized molecules generated by JT-VAE and filtered them for favorable properties including high predicted antibacterial scores against S. aureus, low predicted cytotoxicity scores, and Tanimoto similarity to known antibiotics (Figure S5B). The down-selection resulted in 4,831 molecules that were manually inspected and shortlisted to 90 compounds based on synthesizability and diversity (Figure 5E). After review by a chemical synthesis vendor, we procured 22 molecules and tested them against MSSA RN4220 and MRSA BAA1556 at a high starting concentration of 64 μg/mL. Remarkably, six molecules (a working true discovery rate of 27.3%; DN1-DN6) showed antibacterial activity and were structurally dissimilar (Figure 5F and 5G, Data S2), suggesting that our generative models can design antibacterial molecules de novo without the requirement of a fragment as a starting point.

Figure 5. De novo generation of antibacterial compounds without a fragment starting point.

(A) Schematic of the de novo generative approach.

(B) Schematic of the de novo genetic algorithm based on CReM and examples of generated compounds.

(C) Schematic of the de novo JT-VAE algorithm and examples of generated compounds.

(D) Molecular properties and RAscores for a random subset of 10,000 compounds generated by the de novo CReM and JT-VAE models, compared to the median of known Gram-positive antibiotics.

(E) Computational filters and summary of experimental results for JT-VAE-designed compounds.

(F) Experimental validation of 22 JT-VAE-designed molecules against S. aureus when tested at 64 μg/mL. n=2.

(G) Structures of the six de novo designed compounds with activity against S. aureus.

See also Figure S5.

To further investigate the six active compounds, we determined their MICs against MSSA RN4220, MRSA BAA1556, and, as a starting point for assessing spectrum, N. gonorrhoeae ATCC 49226. DN1 was the most potent and selective against S. aureus, with an MIC of 4 μg/mL for both MSSA RN4220 and MRSA BAA1556 and a therapeutic index (TI) of 32 for HEK293, HepG2 and HSkMC cells. By contrast, DN4 was the least potent and selective, with an MIC of 64 μg/mL for S. aureus and a TI of 2 (Figure 6A). Four of the six compounds also exhibited MICs ≤ 8 μg/mL for N. gonorrhoeae, despite being selected based on their S. aureus model prediction scores. To further assess their spectrum of activity, we tested each compound against a panel of Gram-positive and Gram-negative species. We found that DN1-DN3 exhibited broad-spectrum activity against Gram-positive bacteria, including Bacillus subtilis and vancomycin-susceptible Enterococcus faecalis. However, none of the six compounds inhibited the growth of other Gram-negative species or Gramneutral Mycobacterium tuberculosis (Figure 6B, Data S2). Interestingly, DN1-DN3 gained activity against E. coli when the outer membrane was permeabilized through genetic (lptD4213) or chemical (polymyxin B nonapeptide cotreatment) means, or when efflux pumps were genetically disrupted (ΔtolC). In combination with polymyxin B nonapeptide, DN1-DN3 also effectively inhibited the growth of wild-type E. coli, A. baumannii and P. aeruginosa, and DN1 gained activity against K. pneumoniae and M. tuberculosis (Data S2). These findings suggest that the efficacy of DN1-DN3 is hindered by outer membranes, which may act to block penetration or increase efflux of these compounds.

Figure 6. Characterization of de novo designed compounds (DN1-DN6).

(A) MICs against S. aureus, and CC₅₀ values of the compounds in four different human cell types. n=2, * indicate values >128 μg/mL.

(B) MICs against various Gram-positive species (top) and Gram-negative bacterial species (middle), or in combination with a sub-MIC concentration of polymyxin B nonapeptide (bottom). n=2. see also Table S4.

(C) DiSC₃(5) fluorescence of S. aureus RN4220 treated with DN1-DN6, valinomycin (positive control), and DMSO (negative control).

(D) Images of B. subtilis 168 cells treated with the active de novo designed compounds at 2× MIC for 2 hours. DAPI in blue and FM4-64 in red.

(E) Area of untreated and compound treated (2× MIC) B. subtilis cells, where n=number of cells quantified and **** p ≤ 0.0001, evaluated using a two-sided Mann-Whitney U test.

(F) Membrane rigidification of RN4220 treated with DN1 at 32 μg/mL and control compounds, as measured by the Laurdan generalized polarization (GP). Mean ± SD; n=2.

(G) Cryo-TEM images of untreated and DN1 treated (4× MIC) S. aureus.

(H) Quantification of S. aureus membrane thickness (untreated, n = 14; DN1, n = 11). **** p ≤ 0.0001, evaluated using a two-sided Mann-Whitney U test.

See also Figure S6.

Although DN1-DN3 exhibited similar spectra of activity to each other, performing membrane PMF assays as above for S. aureus cells treated with each of DN1-DN6 resulted in markedly different DiSC₃(5) fluorescence profiles (Figure 6C), consistent with the observation that these compounds are structurally distinct. DN1 and DN6 treatment resulted in similar dissipation of ΔΨ; however, microscopic imaging of morphological changes in B. subtilis 168 revealed that treatment with DN6 resulted in filamentation and membrane rupture after two hours, in contrast to treatment with DN1-DN5 (Figure 6D, 6E and S6A). Given DN1’s potency, we investigated its potential membrane-active mechanism by testing whether DN1 alters membrane fluidity using a Laurdan dye assay as before⁶. Treatment of S. aureus cells with DN1 resulted in a modest increase in Laurdan fluorescence (Figure 6F), suggesting a decrease in membrane fluidity. Performing cryo-TEM on both MSSA RN4220 and N. gonorrhoeae ATCC 49226 cells treated with compound at 4× MIC, we found pronounced morphological changes in the membranes of both bacteria. In MSSA, we observed thickening of the membrane and increased perimeter (Figure 6G, 6H and S6B), and in N. gonorrhoeae, we observed dislodging of the membrane and increased cell swelling (Figure S6C) – quantified further by increased cell roundness and decreased elongation (Figure S6D). Overall, these results indicate that our de novo design models can generate bona fide antibacterial compounds, and that a subset of structurally distinct compounds exhibit different, membrane-active mechanisms of action.

Resistance and in vivo efficacy of DN1

Building on our previous findings, we aimed to further investigate the translational relevance of DN1. Time-kill experiments indicated that DN1 substantially decreased log-phase MSSA RN4220 CFUs as early as 2 hours post-treatment (Figure 7A), exhibiting faster bactericidal activity than that of vancomycin, a first-line antibiotic for treating S. aureus infections. Additionally, the MBC of DN1 was 8 μg/mL for MSSA RN4220 (Figure 7B). In spontaneous mutant generation experiments on solid agar, no colonies emerged after seven days of incubation at 4× and 8× MIC on solid agar, suggesting a low frequency of resistance (<9.6 × 10⁻⁹ for MSSA RN4220 and <1.0 × 10⁻⁹ for MRSA BAA1556). Similarly, the MIC of DN1 for MRSA BAA1556 did not meaningfully increase after 30 serial passages in LB (lysogeny broth) or LB supplemented with 0.002% Tween. DN1 also inhibited the growth of 10 multidrug-resistant Gram-positive isolates from the CDC-FDA (Food and Drug Administration) Antimicrobial Resistance Isolate Bank (ARB), including isolates from the vancomycin-intermediate S. aureus (VISA), aminoglycoside/tetracycline-resistant (ATR), and tedizolid/linezolid (oxazolidinone)-resistant staphylococci (TLZD) panels, with MICs of 8 μg/mL across all tested isolates (Figure 7C and Table S4). Consistent with its activity against N. gonorrhoeae ATCC 49226, DN1 exhibited MICs at or below 16 μg/mL against a panel of drug-resistant N. gonorrhoeae isolates (Figure 7D).

Figure 7. Mechanistic study and in vivo efficacy of DN1.

(A) Time-kill curves for S. aureus RN4220 treated with DN1 or vancomycin. Mean ± SD; n=2.

(B) Minimum bactericidal concentration of DN1 in RN4220. Mean ± SD; n=2.

(C) MICs against antibiotic-resistant strains of S. aureus (listed in Table S4). Each data point represents a strain tested in biological duplicate.

(D) MICs against wild-type and antibiotic-resistant strains of N. gonorrhoeae (listed in Table S4). Each data point represents a strain tested in biological duplicate.

(E) Schematic of the in vivo study of DN1 in a neutropenic mouse skin infection model using MRSA strain ATCC BAA1556. Treatment with vehicle, DN1, or fusidic acid was administered topically in six doses.

(F) Bacterial titers in skin tissue. Horizontal lines indicate median CFU/mouse values. Data represent treatment with vehicle control (n=6), DN1 (n=6), or fusidic acid (n=6). Two-sided Mann-Whitney U-test compared to vehicle treatment: *p ≤ 0.05, **p ≤ 0.01. Horizontal lines represent medians.

(G) Summary of a structure-activity relationship analysis for DN1.

See also Figures S6 and S7.

As DN1 exhibited a favorable resistance profile and physicochemical properties (Table S5), we aimed to evaluate its potential for therapeutic development. For this, we tested DN1 for efficacy in treating a mouse model of MRSA BAA1556 skin infection. In this model, each mouse was rendered neutropenic and inflicted with a skin wound. After inoculation of ~10⁵ CFU of MRSA, each wound was topically treated with DN1 (1% w/v) six times before all mice were sacrificed, and the wounds were excised for CFU determination 25 hours after infection (Figure 7E). We found that treatment with DN1 significantly decreased the average bacterial load by ten-fold relative to vehicle (two-sided Mann-Whitney U-test, p=0.0043), with efficacy similar to that of fusidic acid, a positive control antibiotic used in the clinical treatment of Staphylococcus infections (Figure 7F). Together, these findings support the translational potential of DN1, highlighting its efficacy in treating a model of topical staphylococcal infection.

Synthesis and analysis of structural analogs of DN1

As a starting point for further developing and optimizing DN1, we investigated its structure-activity landscape similarly to NG1. We modified the position or type of halogen atom (Cl or F) on the aryl ring and replaced the central pyrrolidine with 4- or 6-membered nitrogen-containing heterocycles. Of the 19 analogs synthesized (Data S2) that include these modifications, we found that five analogs exhibited MICs ranging from 2 to 8 μg/mL when tested against MSSA RN4220, MRSA BAA1556, and a panel of 10 MRSA isolates from the ARB (Figure S7A and S7B). A more potent analog, DN1-164, exhibited MICs ≤4 μg/mL for all antibiotic-resistant S. aureus strains tested and against wild-type N. gonorrhoeae without increases in cytotoxicity against human cells, resulting in TIs >32 (Figure S7C). All the more potent analogs (DN1-154, DN1-164, and DN1-173) increased DiSC₃(5) fluorescence in S. aureus RN4220 cells, suggesting that, similar to DN1, these analogs dissipate the ΔΨ component of the PMF (Figure S7D). Furthermore, the chemical structures of the active analogs suggest that substitutions of the central pyrrolidine are tolerated, the placement of a halogen in the meta position of the aryl ring might improve potency, and the carboxylic acid is needed for activity (Figure 7G). Thus, these findings highlight a broad structure-activity landscape relevant to improving the selectivity of our de novo-generated compound, DN1, and its associated chemical series.

Discussion

The unmet need for novel antibiotics is large and growing. Virtual screening of chemical compound libraries offers a way to leverage the diversity of enumerated chemical space; yet these libraries sample an infinitesimal fraction of drug-like chemical space, which is estimated to contain ~10⁶⁰ compounds. Here, we have presented a deep learning-enabled approach to the generative design of compounds with antibacterial activity. Using GNNs as scoring functions, we first comprehensively screened >45 million chemical fragments to identify selective antibacterial starting points against N. gonorrhoeae or S. aureus. We then developed two models – a genetic algorithm-based model (CReM) and a variational autoencoder (VAE) – to generate compounds based on these fragments. While fragments offer consistent and tractable starting points, these models can also generate molecules de novo, without any fragment input, as we demonstrated. Of the 24 compounds that we designed using these generative approaches and subsequently synthesized, two compounds – NG1 and DN1 – possessed narrow-spectrum activity against pathogenic Neisseria and broad-spectrum activity against Gram-positive bacteria and N. gonorrhoeae, respectively. Both compounds exhibited mechanisms of action distinct from those of often-used antibiotics and were effective in reducing bacterial titers in different animal models of infection. Additionally, both compounds exhibited a structure-activity landscape that can be productively used for further optimization. Together, our results enable the generative design of two unique structural classes of antibacterial compounds and demonstrate the ability of our platform to explore uncharted regions of chemical space.

Here we focused on implementing two molecular generation models: CReM and VAE. Despite CReM’s reliance on rule-based features of bioactive molecules to either grow or introduce modifications and generate structurally diverse molecules, our VAE-based models consistently generated compounds that were more drug-like and synthesizable. Our platform is modular in that other generative models, such as generative adversarial networks^16,59, molecular transformers⁶⁰, flowbased models⁶¹, diffusion-based methods^22,62-64, and fragment linkers⁶⁵, can be readily incorporated. Combining these diverse approaches and continuing to explore fragments that lead to readily-synthesizable molecules will enable new portions of chemical space to be harnessed for successful de novo design efforts.

Our approach also allowed for the models to have different chemical starting points, enabling either constrained (fragment-based) or unconstrained (de novo) design. Using fragments as starting points possesses several unique advantages. First, it allows one to comprehensively search molecular building blocks (e.g., fragments with all possible combinations of up to 11 atoms comprised of C, N, O, and F) to systematically identify and exploit underexplored areas of chemical space that would otherwise be unlikely to be sampled by generative methods. The disparity between the millions of fragments that our GNN models predicted to be antibacterial and the hundreds of fragments that could be found within commercially available libraries suggests that much of antibacterial chemical space indeed remains underexplored. Second, by screening fragments based on target-agnostic phenotypic outcomes, specifically bacterial growth inhibition, our method increases the likelihood of discovering molecules with unique modes of action that potentially circumvent existing antimicrobial resistance mechanisms. Lastly, if chosen carefully, a highly synthesizable starting point can constrain the search space to realistic, drug-like molecules, reduce the risk of generating implausible structures, and facilitate a coherent structure–activity relationship analysis. In contrast to the fragment-based approach, de novo design without predefined chemical starting points enables unrestricted exploration of chemical space and more variability in the properties of the generated compounds. After manual inspection and stringent filtering for structural novelty, low predicted cytotoxicity, and synthesizability, the tested de novo designed compounds produced a high working true discovery rate of 27% and were structurally diverse with drug-like properties. Importantly, compounds from both approaches exhibited antibacterial activity against highly drug-resistant strains of N. gonorrhoeae and Gram-positive bacteria, highlighting the versatility of generative AI models when coupled to antibacterial scoring functions.

Overall, these findings underscore the value of multiple approaches to the generative design of antibiotics and that the choice of a chemical starting point should be guided by strategic considerations including chemical novelty, diversity, and synthesizability. We anticipate that future studies will leverage multi-objective optimization frameworks to identify compounds that simultaneously satisfy multiple criteria, including high antibacterial activity, synthetic accessibility, low toxicity, and favorable drug-like properties such as metabolic stability, low protein binding, and optimal pharmacodynamics.

Limitations of the study

Despite our observations of high working true discovery rates and promising chemical diversity of generated molecules, it is possible for the ML models used here to produce false positive predictions or generate structurally non-diverse molecules. Model outputs should therefore be analyzed on a case-by-case basis, and when such outputs appear promising, the compounds should be empirically tested. Furthermore, the utility of our generative approach heavily relies on the ability to synthesize compounds for validation. As evidenced by the small number of generated compounds that we could synthesize, chemical synthesis remains a challenge when molecules are not designed according to prespecified synthetic routes. Improvements in synthesizability predictors and complex retrosynthesis-based algorithms should enable larger sets of de novo designed compounds to be experimentally tested and validated.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Cell culture.

Human cell lines were obtained from ATCC: HEK293 (CRL-1573), HepG2 (HB-8065), HSkMCs (PCS-950-010) and IMR-90 (CCL-186). HEK293 and HepG2 cells were grown to log phase in high-glucose Dulbecco’s Modified Eagle Medium (DMEM; Corning 10-013-CV) supplemented with 10% fetal bovine serum (FBS; ThermoFisher 16140071) and 1% penicillinstreptomycin (ThermoFisher 15070063). HSkMCs were grown in mesenchymal stem cell basal medium for adipose, umbilical and bone marrow-derived MSCs (ATCC: PCS-500-030) supplemented with ATCC’s primary skeletal muscle growth kit (ATCC: PCS-950-040) and 1% penicillin-streptomycin. IMR-90 cells were grown in in Eagle’s Minimum Essential Medium (EMEM; ATCC 30-2003) supplemented with 10% FBS and 1% penicillin-streptomycin. Cells were cultured at 37°C and 5% CO2 in humidity-controlled incubators and passaged <5-10 times. All cell lines were authenticated by the supplier, ATCC, using STR profiling.

Bacterial cell culture.

The main bacterial strains used in this study include methicillin-susceptible Staphylococcus aureus RN4220 (MSSA DSM 26309; DSMZ German Collection of Microorganisms and Cell Cultures, Leibniz Institute), methicillin-resistant Staphylococcus aureus FPR3757 (MRSA USA300; ATCC BAA-1556), and Neisseria gonorrhoeae (ATCC 49226). Other common strains include Bacillus subtilis 168 ATCC 23857, Escherichia coli BW25113 DSMZ 27469, MG1655 ATCC 700926, JW5503-KanS (ΔtolC832::FRT, CGSC 14206), and RFM795 (lptD4213) (CGSC 14179), Klebsiella pneumoniae ATCC 13883, Pseudomonas aeruginosa PAO1 ATCC 47085, Acinetobacter baumannii ATCC 17978 and Mycobacterium tuberculosis MTB H37Ra, ATCC 25177. Additional bacterial isolates in Table S4 were obtained from the CDC & FDA Antibiotic Resistance (AR) Isolate Bank (Atlanta, Georgia)⁶⁶. All bacterial strains were authenticated by their respective suppliers: ATCC, DSMZ and/or CDC, using next-generation sequencing or Sanger sequencing. Unless specified, LB medium containing 1.5% Difco agar (Becton Dickinson 244520) was used to grow individual colonies. Cells were grown in liquid LB (Becton Dickinson 244620) from single colonies aerobically in 14 mL Falcon tubes using 2-3 mL working volumes without antibiotic selection. The cultures were incubated at 37 °C in a light-insulated, humidity-controlled chamber, shaking at 300 rpm. All Neisseria strains were passaged twice on chocolate agar plates (CAP; Hardy Diagnostics, H25). Several colonies were picked from plates to obtain a desired OD₆₀₀ and grown in specialized Graver Wade media⁶⁷ at 37°C with 5% CO₂ in a humidity-controlled incubator.

Mouse models.

All animal studies were performed at the Wyss Institute at Harvard in accordance with protocol IS00000852-6, approved by the Harvard Medical School Institutional Animal Care and Use Committee and the Committee on Microbiological Safety. All mice were all purchased from Jackson Laboratory and housed in pathogen-free facility maintained at 20-26°C ambient temperature, 40-65% relative humidity, and a 12:12 light-dark cycle. Enrichment devices were included in the animal environments as required and changed bi-weekly. For DN1, female C57BL/6J mice had a 48-hour acclimation period upon arrival to the mouse facility prior to testing the toxicity (healthy; 6-12 weeks old, weight of 22 ± 2 g) and efficacy (rendered neutropenic with cyclophosphamide on Day −4 and Day −1; 19 ± 2 g). For NG1, female BALB/c had a 1 week acclimation period to the mouse facility prior testing the toxicity (healthy; 6-8 weeks old, 20 ± 2 g) and efficacy (healthy, ovariectomized at Jackson Laboratory; 23 ± 3 g). None of the mice had been involved in previous procedures.

METHOD DETAILS

Computational methods

Machine learning models.

Chemprop¹² is a software package that implements directed message-passing neural networks (D-MPNNs). Here, these models are trained on binarized data representing the empirical growth inhibitory activity of compounds. As described previously¹², D-MPNNs, which are a form of graph neural networks (GNNs), take as input a simplified molecular-input line-entry system (SMILES) string of a compound and convert it to a molecular graph representation. Graph convolutional operations are applied to each atom and bond, and information from local substructures of the compound are aggregated. An intermediate vector representation is generated by collapsing the information associated with each bond and atom into a single embedding. This embedding is then concatenated to a list of global biophysical features, such as molecular weight, which are used to potentially improve the predictions made, and the output is then passed to a fully-connected feed-forward neural network. The output of each trained model is a score between 0 and 1, representing the probability that the input molecule is antibacterial (0: does not inhibit bacterial growth and 1: inhibits bacterial growth). The output of an ensemble was taken as the average of all models in the ensemble.

In this work, ensembles of 20 binary classification Chemprop models for two bacterial species, N. gonorrhoeae and S. aureus, were trained on 38,765 and 39,312 compounds (as described in detail in Anahtar et al.⁸ and Wong et al.⁷), respectively. Training datasets containing the experimentally-determined inhibitory activity of each compound are available in Data S1. Hyperparameter optimization was performed, and each ensemble was previously benchmarked^7,8. Chemprop models predicting cytotoxicity against human cells were previously developed and described in Wong et al.⁷; in this work, we use these models to predict the cytotoxicity of a given fragment or compound against human hepatocarcinoma (HepG2), primary skeletal muscle (HSkMC), and lung fibroblast (IMR-90) cells. The output of these models is a score between 0 and 1, describing the probability that the inputted compound is cytotoxic (0: non-toxic and 1: toxic).

SHAP analysis of molecular feature importance

Model setup and background data.

A SHAP (SHapley Additive exPlanations) analysis was conducted to quantify the contribution of 2D RDKit-derived descriptors (200 features) to antibacterial activity predictions. Features were stored in .npz format, and background data comprised the first 100 compounds from the dataset. Data were split into 240 equal-sized chunks to enable embarrassingly parallel execution across a high-performance computing (HPC) cluster.

SHAP feature attribution for antibacterial activity models.

To interpret the contribution of molecular descriptors to model predictions, we applied SHAP analysis to neural network models trained to predict antibacterial activity from RDKit-derived 2D chemical features. Separate SHAP analyses were performed for models trained on S. aureus and N. gonorrhoeae datasets, each using the same set of 200 standardized descriptors as model input. SHAP values were computed using a model-agnostic approach with the shap.Explainer class, where predictions were generated via a custom wrapper around the Chemprop command-line interface. This wrapper accepted NumPy arrays of feature vectors, saved them to temporary files compatible with Chemprop, and invoked model inference using subprocess calls. A background set of 100 representative molecules was used to approximate conditional expectations for feature ablation. To parallelize SHAP computations, the full dataset was divided into 240 chunks and processed independently across a CPU-based HPC cluster. For each chunk, SHAP values were computed and saved as serialized objects. After completion, all chunks were aggregated into a single SHAP explanation object per species. We calculated global feature importance by taking the mean absolute SHAP value across all molecules for each descriptor. The top 10 most influential features were visualized in SHAP summary plots, annotated with short chemical descriptions and colored by feature value. All 200 features were also exported to a CSV file reporting their names, descriptive annotations, and mean SHAP contributions. These results enabled identification of key structural and physicochemical properties associated with antibacterial activity across species-specific models. SHAP analysis was performed in Python 3.10 on CPU nodes on the high performance computing cluster. Predictions were generated using the Chemprop CLI (version 1.6.1). RDKit (version 2024.3.6) was used for descriptor generation. All scripts utilized shap, matplotlib, numpy, and pandas for computation and visualization.

Species specificity of Chemprop models.

To examine the species specificity of the different models for N. gonorrhoeae and S. aureus, our original training data that were empirically tested against both species were split into a new training set (80%) and test set (20%). Both N. gonorrhoeae- and S. aureus-specific models we re-trained and the mean predicted scores were compared with the antibacterial activity of the test set. Compounds with high mean scores against one species demonstrated selective activity against that species, and compounds with high (or low) mean scores against both species demonstrated activity (inactivity) against both species.

Fragment and compound libraries.

A fragment library comprising 45,858,026 fragments was assembled from the Generated DataBase (GDB, courtesy of Dr. Jean-Louis Reymond at the University of Bern, Switzerland) and the Readily AccessibLe (REAL) fragment database from Enamine⁶⁸. The GDB libraries enumerate all theoretically-possible fragments up to a prespecified number of atoms following simple chemical stability and synthetic feasibility rules⁶⁹. To compile a set of 11-atom fragments including C, N, O, F, Cl, and S atoms, we combined the entire GDB-11⁴⁰ (fragments up to 11 atoms including C, N, O, and F atoms, N = 26,434,571) and all up-to-11-atom fragments in the GDB-13⁴¹ (which includes C, N, O, Cl, and S atoms, N = 1,089,000). Enamine fragments (version October 5, 2022, N = 18,338,026) were obtained from the Enamine REAL database.

To experimentally assess chemically diverse compounds containing the above fragments, as well as to search for analogs, we assembled a database from two large purchasable libraries: 799,149 compounds from the Broad Institute and 6,138,200 commercially available compounds from MolPort, MayBridge, and Ambinter.

Identification and selection of fragments and compounds associated with antibiotic activity.

All 45,858,026 fragments, containing readily-accessible fragments from the Enamine REAL Fragment database, were scored using the trained N. gonorrhoeae⁸ and S. aureus ensemble⁷ models, as described above. All 6,937,349 compounds were scored using both N. gonorrhoeae and S. aureus ensembles. For the analysis shown in Figure S1A, three fragments representative of known antibiotics were encoded using the following SMILES strings: nitrofuran: ‘O=[N+](O)c1ccco1’; fluoroquinolone: ‘O=C(O)c2c[nH]c1ccc(F)cc1c2=O’; and carbapenem: ‘O=C(O)C1=CCC2CC(=O)N12’. Compounds containing these moieties were identified in a manually-compiled set of 559 known antibiotics and antiseptics⁶ (Data S1). Compounds containing the fragments were identified using RDKit’s HasSubstructMatch() function, and the known antibacterial compounds were subsequently scored using either the N. gonorrhoeae or S. aureus ensembles. We found that the prediction scores for compounds correlated with the prediction scores of the fragment alone (Figure S1A), supporting the performance of an algorithm that identifies fragments with high predictive scores for the discovery and design of selective antibacterial compounds.

As detailed in the main text, our approach to fragment selection and compound testing involved the following steps. For each of N. gonorrhoeae and S. aureus, fragments with antibiotic activity prediction scores >0.05 or >0.1 were retained in the GDB and REAL libraries, respectively. Of these fragments, those with cytotoxicity prediction scores <0.5 across all three human cell types (HepG2, HSkMC, and IMR90) were retained. Fragments were then tested for the presence of PAINS or Brenk substructures using RDKit’s FilterCatalogParams.FilterCatalogs.PAINS and FilterCatalogParams.FilterCatalogs.BRENK built-in structure lists, respectively. To focus on structurally novel fragments, only fragments with maximum Tanimoto similarity <0.5 with respect to the aforementioned set of 559 known antibiotics were retained. Here, Tanimoto similarity was calculated using Morgan fingerprints, with radius equal to 2 and number of bits equal to 2048. The remaining fragments were matched to compounds using RDKit’s HasSubstructMatch() function. Matched compounds were considered if they had an antibacterial activity prediction score >0.1, and matched compounds were filtered in the same way for cytotoxicity and structural novelty as described for fragments above (GitHub repository, see Code availability).

De novo design of compounds with generative models.

CReM: chemically reasonable mutations framework.

For the genetic algorithm based on F-CReM described in this work, a modified version of CReM²¹ was implemented. For each run of the algorithm, the SMILES string of a fragment and molecule containing the fragment was provided as input; the former was provided in order to ensure that all generated compounds possessed the fragment using RDKit’s HasSubstructMatch() function, and the latter was provided as a seed. For de novo design of molecules based on CReM, the SMILES string of either ammonia, methane, or water was provided as input.

By default, CReM provides two methods: ‘grow’ and ‘mutate’ (as implemented in the CReM Python package). Ranges for parameter combinations were provided as additional input. Parameter combinations for grow algorithms included max_atom (4, 6, 8, and 10 were used), min_atom (0 was used), and radius (2 and 3 were used), while the parameter combinations for mutate algorithms included the max_size (4, 6, 8, and 10 were used), min_size (0 was used), radius (2 and 3 were used), min_inc (−2 was used), and max_inc (2 was used). All possible combinations of these parameters were considered using a full grid search. For a given parameter set, the algorithm proceeded as follows:

1. The original molecule was used to generate de novo molecules with either the grow or mutate function from CReM, and molecules that did not contain the original fragment were discarded.

2. Compounds containing PAINS and Brenk substructures were excluded.

3. Chemprop scores for the resulting molecules were calculated using either the S. aureus or N. gonorrhoeae ensemble.

4. If the user-defined scoring method was set to ‘regular score’, then the compounds were ranked according to the Chemprop models (chempropsco). If not, modified scores that incorporate additional criteria were calculated. The additional criteria included the following variables: SAScore (predicted synthesizability score), tansim (Tanimoto similarity to known antibiotics), hepg2 (predicted toxicity score for HepG2 cells), and prim (predicted toxicity score for HSkMC cells). To improve speed, prediction scores from the IMR-90 cytotoxicity ensemble were not considered as additional variables here, but all generated compounds were eventually filtered based on cytotoxicity prediction scores from all three cytotoxicity ensembles (as described in Down-selection of generated compounds for synthesis and testing below). Modified scores ($adj\_score$) were calculated based on the original score (chempropsco) and the additional variables according to the formula:

$$adj\_score=(2.0\ast chempropsco)\text{-}((sascore∕10.0)+tansim+hepg2+prim).$$

1. Finally, to seed the next iteration of the algorithm, we calculated the number of compounds generated and scored in steps (i)-(iv) above (N_gen) and compared it to the sum of two prespecified numbers, N_top and N_rand. If max_atom_range ≥ 8, then N_top = 2 and N_rand = 1; else N_top = 5 and N_rand = 5. If N_gen ≤ N_top + N_rand, then all N_gen compounds were used to seed the next iteration. Otherwise, the N_top highest-scoring compounds among the N_gen generated compounds, as well as N_rand other randomly chosen compounds among the N_gen generated compounds were used to seed the next iteration.

Fragment-based Variational Autoencoder (F-VAE).

The F-VAE described in the main text was inspired by fragment-based drug discovery strategies often used by medicinal chemists. The algorithm starts with fragments associated with specific properties of interest (e.g., antibacterial activity) and expands these fragments into molecules. In particular, our rationale-conditioned molecular graph generator is a variational autoencoder which completes a full molecule, $G$, given a fragment, $S$. Since each fragment ($S$) may be realized into many different molecules, we introduce a latent variable, $z$, to generate diverse outputs:

$$P(G\mid S)={\int }_{z}P(G\mid S,z)P(z)dz,$$

where $P(z)$ is the prior distribution. The F-VAE model consists of a graph encoder and a graph decoder. In this case, the encoder is a message passing neural network (MPNN) which learns the approximate posterior for variational inference. In the graphical representation of each molecule, each atom or bond is represented by a one-hot encoding of its atom or bond type. The encoder then uses three message-passing layers, followed by an average pooling operation, to transform the initial graph representation into a 20-dimensional latent vector, $z_G$.

The decoder then generates a molecule (a molecular graph) according to its breadth-first order. In each step, the model generates a new atom and all its connecting edges. During generation, we maintained a queue that contained frontier nodes in the graph with neighbors to be generated. Let $G_t$ be the partial graph generated by step $t$. To ensure that the full molecule, $G$, contains $S$ as a subgraph, we set the initial state of $G_0=S$ and put all the leaf atoms (atoms with degree = 1 in the graph) in the queue. In each generation step, the decoder first runs an MPNN over the current graph to compute an atom representation, $h_v$, for each atom, $v$. Suppose that the first atom in the queue is $v$. The decoder then expands the current molecule using three decision steps:

1. Predict whether there will be a new atom attached to $v$. The probability of this expansion step is predicted based on the latent vector, $z_G$, and atom representation of $v$: $p_v=\text{sigmoid}(\text{MLP}({\mathrm{h}}_{\mathrm{v}},{\mathrm{z}}_{\mathrm{G}}))$.

2. If $p_v>0.5$, discard $v$ and move on to the next node in the queue. Otherwise, create a new atom, $u$, predict its atom type, and append it to the queue.

3. Predict the bond type between $u$ and other leaf nodes in the queue. Since atoms are generated in breadth-first order, there are no bonds between $u$ and atoms not in the queue. To fully capture edge dependencies, we predict the bonds between $u$ and atoms in the queue sequentially, and we update the representation of $u$ when new bonds are added to the molecule.

The F-VAE was pre-trained on 1,686,695 molecules from ChEMBL (version 22) to enable the model to generate realistic molecules. Each training example was a pair, ($S$, $G$), where $S$ is a random connected subgraph of a molecule $G$ with up to 15 atoms. We trained the generative model to maximize the likelihood of the ground truth molecule $G$ given fragment $S$, using the following hyperparameters: hidden_size = 400, batch_size = 16, MPNN depth = 3, learning rate = 1e-3, optimizer = Adam, and epoch = 20.

Junction Tree Variational Autoencoder (JT-VAE).

The JT-VAE described in the main text is a deep generative model designed to generate a molecule without a starting fragment. It represents a molecule at two different levels: an atom graph that represents how atoms are connected to each other and a junction tree that captures how chemical motifs are connected to each other. The model consists of an encoder that maps a molecular graph and a junction tree to a latent representation and a decoder that reconstructs a molecule by sequentially assembling chemical substructures based on the learned junction tree structure. The encoder of JT-VAE is a hierarchical MPNN with three layers.

1. Motif layer: This layer represents how the motifs are coarsely connected in the graph. This layer provides essential information for the motif prediction in the decoding process.

2. Attachment layer: This layer encodes the connectivity between motifs at a fine-grained level, highlighting the attachment points between two chemical motifs.

3. Atom layer: The atom layer is a molecular graph representing how its atoms are connected, where each node represents an atom, and each edge represents a bond.

The decoder generates a molecule motif by motif according to their depth-first order, using the information encoded by the latent representation. In each generation step, the decoder expands the current molecule based on two predictions:

1. Motif Prediction: The model predicts the next motif to be attached to the current graph.

2. Attachment Prediction: Next, the model needs to predict the attachment configuration between the newly selected motif and the current graph.

The above three predictions together give an autoregressive factorization of the distribution over the next motif and its attachment. We trained JT-VAE on 1,686,695 drug-like compounds from the ChEMBL database, the same training set as F-VAE. We trained the generative model to maximize the likelihood of the molecules in the training set. The model hyperparameters are as follows: hidden_size = 400, batch_size = 16, MPNN depth = 3, learning rate = 1e-3, optimizer = Adam, and epoch = 20. By sampling from JT-VAE latent space 40 million times, we generated 28,534,490 unique molecules and ranked them based on their predicted antibacterial properties.

Down-selection of generated compounds for synthesis and testing.

All compounds were down-selected based on (i) de-duplication of SMILES, (ii) predicted antibacterial score, (iii) maximum Tanimoto similarity with respect to the set of known antibiotics as well as all active antibacterial compounds in the respective training set, (iv) synthesizability score (SAscore⁴⁹ or RAscore⁵⁰), and (v) predicted HepG2 and/or HSkMC cytotoxicity score. Both SAscore and RAscore were used only to sample different synthetic accessibility scoring approaches. For fragments F1, F2 and F2’, F-CReM compounds were down selected using predicted antibacterial score >0.7, SAScore <3, Tanimoto similarity <0.5, and HepG2 and HSkMC cytotoxicity score <0.2. Compounds that contained a β-lactam motif, as defined by those returning True using HasSubstructMatch() with the molecule described by 'O=C1CCN1', were additionally filtered out to preserve structural novelty. For fragments F1, F2 and F2’, F-VAE generated-compounds were filtered using predicted antibacterial score >0.3, RAScore >0.8, Tanimoto similarity <0.4, and HepG2 cytotoxicity score <0.2.

Visualization with t-SNE.

t-SNE plots were generated using openTSNE’s TSNE() function with perplexity 45, Euclidean distance as the metric, and PCA initialization. Molecules were represented as RdKit fingerprints with min_path=1, max_path=7, n_bits=2048.

ASKCOS retrosynthesis planning tool.

The Monte Carlo tree search retrosynthesis model in ASKCOS⁷⁰ was used to generate retrosynthetic pathways to each compound. The retrosynthetic search was run on nodes with four CPUs with an expansion time of 60 seconds, a branching factor of 25, and a maximum depth of 6. All other search parameters were set to ASKCOS defaults. We consider the search successful if ASKCOS identified at least one retrosynthetic path from the target compound to purchasable compounds.

Experimental methods

Chemical compounds.

Compounds with high purity (>90%) were procured either from the Broad Institute Center for the Development of Therapeutics (CDoT) or from commercial chemical vendors including BIONET-Key Organics Ltd.(Cornwall, UK), ChemBridge (San Diego, CA), ChemDiv (San Diego, CA), Maybridge (Altrincham, UK), MedChemExpress (Monmouth Junction, NJ), TargetMol (Boston, MA), Vitas-M (Hong Kong, China), and Enamine (Kyiv, Ukraine). Stock solutions and serial dilutions of all compounds were freshly prepared in dimethyl sulfoxide (DMSO; MilliporeSigma D5879), unless stated otherwise. Known antibiotics were obtained as follows: vancomycin (Fisher Scientific AAJ6279003), valinomycin (Thermo Fisher, V1644), triclosan (MilliporeSigma, 72779), azithromycin (Cayman Chemical, 15004), ceftriaxone (sodium salt hydrate, Cayman Chemical, 18866), fosfomycin sodium (MilliporeSigma, 34089), CCCP (MedChemExpress, HY-100941), and fusidic acid (Millipore Sigma F0881), all dissolved in DMSO. Kanamycin sulfate (MilliporeSigma, 60615) was dissolved in ultrapure MilliQ-water and ciprofloxacin powder (MilliporeSigma 17850) was dissolved in dilute acid (0.1 M HCl). De novo generated compounds, and their respective analogs were synthesized and procured from Enamine (Kyiv, Ukraine). Analogs of compounds containing fragment F1 (NG1 analogs) were designed and synthesized by CC4CARB (Chemistry Center for Combating Antibiotic-Resistant Bacteria), an NIAID-led (National Institute of Allergy and Infectious Diseases) partnership with RTI (Research Triangle Institute).

MIC, MBC, and bacterial growth inhibition assays.

For S. aureus, a bacterial suspension of ~10⁵ CFU/mL was obtained either by performing a 1:10,000 dilution of an overnight culture, picked from a single colony, or a 1:500 dilution of an OD₆₀₀ 0.08 suspension in fresh LB (Becton Dickinson 244620). Cells were seeded in a 96-well plate, with 99 μL of bacteria and one μL of two-fold serially diluted compound in DMSO. Plates were sealed with breathable membranes (Millipore Sigma Z763624) and incubated for 18-24 hours at 37°C with 5% CO₂. The MIC was determined as the minimum concentration for which OD₆₀₀ < 0.1, as measured using a SpectraMax M3 plate reader. For initial screening experiments, active compounds were determined as those for which OD₆₀₀ < 0.15. For calculating the MBC (maximum bactericidal concentration), overnight treated cells were removed from incubation and serially diluted 10-fold in room temperature LB. Five μL was spotted on LB agar plates and allowed to dry at room temperature before stationary incubation at 37 °C overnight for 16 to 24 h. CFUs were determined by manual counting, and all measurements are based on counts containing at least six colonies.

For N. gonorrhoeae, MICs were determined via broth microdilution when screening compounds, via agar dilution when confirming values for a given candidate small molecule (e.g., NG1), or via ETEST when testing a standard-of-care antibiotic, per CLSI M100 and M07 guidelines. Prior to MIC testing, frozen stocks were passaged twice on chocolate agar plates (CAP; Hardy Diagnostics, H25). N. gonorrhoeae broth microdilution was performed by first preparing the bacterial inoculum by picking individual colonies from an overnight CAP, suspending in PBS to OD₆₀₀ 0.08, and diluting the suspension 1:200 in Graver Wade media⁶⁷. Within 15 minutes of inoculum preparation, each well of a 96- or 384-well plate was inoculated with bacteria and compound (serially diluted in DMSO) such that the final DMSO concentration in each well was ≤1% and bacterial concentration was ~5 × 10⁵ CFU/mL. Plates were incubated at 36-37°C with 5% CO₂ for 20-24 hours. The MIC was determined as the concentration of compound resulting in complete inhibition of growth both visually and as measured by PrestoBlue HS Cell Viability Reagent (Invitrogen) after 1-2 hours of incubation. N. gonorrhoeae agar dilution was performed for candidate small molecules by first preparing agar dilution plates following CLSI M07 guidelines. Briefly, serial dilutions of 100× compound stock solutions were made in DMSO and added to molten agar, made of gonococcal (GC) medium base (BD Difco 228950) with 1% IsovitaleX Enrichment (BD 211876), which had been equilibrated to 45-50°C in a water bath. The agar and compound solution were mixed thoroughly and poured into a 6-well plate and allowed to solidify at room temperature. A DMSO-only growth control was included with every dilution series. Plates were used immediately or stored in sealed plastic bags at 4°C for up to a week and allowed to equilibrate to room temperature before use. The bacterial inoculum was prepared by making a 1:10 dilution of a 0.5 McFarland standardized inoculum of each bacterial strain in sterile saline and inoculating 2 μL of the suspension onto each marked location of a plate. The inoculated plates were allowed to dry and incubated at 36-37°C with 5% CO₂ for 16-20 hours while inverted. MICs were read on a dark surface, with growth on the growth control plate required for validity. Neisseria gonorrhoeae ETESTs (bioMérieux) were performed as described previously⁷¹. Briefly, a sterile swab was soaked in a 0.5 McFarland standard bacterial suspension, excess fluid was removed, and the swab was used to evenly cover the entire surface of a plate of GC agar base with 1% IsoVitaleX. The plate was allowed to completely dry before placing an ETEST with sterile forceps, incubating for 18-24 hours at 37°C with 5% CO₂, and reading the MIC as the lowest antibiotic concentration that inhibited growth.

For M. tuberculosis, 100 μL of exponentially growing bacteria was seeded at a density of 5 × 10⁴ per well in 7H9 supplemented with ADS (albumin dextrose saline), incubated with drug for 5 days at 37°C, then incubated for another 24 hours with 15 μL of 0.02% resazurin (w/v), and fluorescence was read with a SpectraMax M3 plate reader (Ex = 530 nm, Em = 590 nm).

For other species not previously specified, an OD₆₀₀ 0.08 suspension was made in PBS and diluted 1:500 in fresh LB or Haemophilus Test Medium Broth (Remel) for Haemophilus. Each well of a 96-well plate received 99 μL of the bacterial suspension and 1 μL of compound. Plates were incubated at 37°C in ambient conditions, except for Haemophilus (5% CO₂), for 18-24 hours and MICs were read visually and confirmed by measuring OD₆₀₀ values. All assays were performed in biological duplicates.

Bacterial time-kill assays.

For MSSA RN4220, cells were diluted 1:10,000 from an overnight culture into fresh LB and plated into 96-well flat-bottom clear plates using 99 μL working volumes. Plates were then sealed with breathable membranes, and cells were grown to early exponential phase, OD₆₀₀ ~ 0.01 (corresponding to ~10⁶ CFU/mL) in a 37°C incubator with shaking at 300 rpm. One μL of compound in two-fold serial dilutions in DMSO was then added to each well to the final concentrations indicated, and bacterial cell cultures were sealed and re-incubated at 37°C with shaking at 300 rpm. At indicated times, cells were removed from incubation, serially diluted in room-temperature LB, and spotted on LB agar (Becton Dickinson 244520) in rectangular plates. Plated cells on LB agar were allowed to dry at room temperature before stationary incubation at 37°C overnight (18-24 hours). CFUs counts were manually determined. For N. gonorrhoeae, time-kill assays were performed as previously described⁷². Specifically, a 0.5 McFarland suspension was prepared in sterile PBS using individual colonies picked from chocolate agar plates that had been grown for 18-20 hours at 37°C in a humid 5% CO₂-enriched atmosphere. The suspension was diluted 1:500 in pre-warmed (37°C) GW media and 90 μL was added to each well in a round-bottom 96-well plate. The plate was pre-incubated for 4 hours with shaking at 150 rpm in a 35°C, 5% CO2-enriched incubator. At time 0, 10 μL of PBS (growth control) or antimicrobial (to achieve a final concentration of 0.5×, 1×, 2×, or 4× MIC, where the MIC of NG1 was 0.5 μg/mL and azithromycin MIC was 0.25 μg/mL) was added to each well of pre-incubated bacteria in duplicate, with a separate row for each time point. At indicated times, the corresponding row of cells was removed from incubation, serially diluted in PBS, and spotted on GC agar in rectangular plates. After drying at room temperature, plates were incubated overnight (18-24 hours) without shaking and CFU counts were manually determined.

Cytotoxicity assay and CC₅₀ determination.

A resazurin-based assay, which quantifies the number of live cells in a sample, was used to monitor human cell viability in the presence of a compound. For CC₅₀ determination, 99 μL cells were plated into 96-well clear flat-bottom black tissue-culture-treated plates (Corning 3603) at a density of 10⁴ cells/well and incubated at 37°C with 5% CO₂. Twenty-four hours after plating, each well received one μL of two-fold serially-diluted test compound. Additional wells were treated with 1 μL of DMSO as a negative control and Triton X-100 (0.1% final concentration) as a positive control. Cells treated with the compound were re-incubated for 24 hours, after which 0.15 mM resazurin (Millipore Sigma R7017) was added to each well. After an additional 24 hours of incubation, the fluorescence was read at excitation/emission at 550/590 nm using a SpectraMax M3 plate reader. CC₅₀ values were calculated by normalizing the fluorescence values based on the positive and negative controls and performing a nonlinear fit with the [Inhibitor] vs. response – Variable slope function in GraphPad Prism (v.10.1.0). HepG2 and HEK293 CC₅₀ values for BRD1 and BRD2 were determined as above, except in 384-well plates seeded with 4500 viable cells/well, where cells were treated with 1 μL (for HepG2) or 500 nL (for HEK293) of compound that were transferred using an acoustic dispenser (Labcyte Echo 555). All experiments were performed in biological duplicate.

Spontaneous mutant generation and frequency of resistance experiments.

N. gonorrhoeae: Isolated colonies of N. gonorrhoeae ATCC 49226 were picked from an overnight plate to make a heavy suspension in PBS. The suspension was serially diluted and plated on chocolate agar to determine the initial inoculum, and 500 μL of the suspension was added to each GC agar plate containing 0, 1×, 2×, 4×, or 8× MIC of NG1. The suspension was spread and allowed to dry before stationary incubation at 37°C in a 5% CO₂ incubator for 3 days. Colonies that emerged on each plate were individually picked and the elevated NG1 MIC was confirmed using both agar dilution and broth microdilution methods in parallel with the parental strain for comparison. The frequency of resistance was calculated as the total number of colonies counted divided by the total number of bacteria inoculated on each plate. S. aureus: MSSA RN4220 was picked from single colonies and grown overnight in 2mL of fresh LB. OD₆₀₀ was measured and the suspension was serially diluted and plated on solid agar (with no compound) to determine the initial inoculum. One mL of the same overnight culture (~10⁹ CFU) was aliquoted and centrifuged at 3700 × g for 5 min. The cell pellet was resuspended to a final volume of 500 μL in fresh LB, then pipetted onto the surface of LB agar plates containing 1×, 2×, 4× or 8× agar MIC of DN1 (4 μg/mL). Cells were then spread using a bent, sterile inoculating loop, and plates were dried and inverted before stationary overnight incubation at 37°C. The next day, plates were removed from incubation, and colonies that grew on each plate were counted to calculate the frequency of resistance as above.

DiSC₃(5) fluorescence.

Individual colonies of S. aureus RN4220 were picked and grown in 2 mL liquid LB overnight at 37°C with shaking at 300 rpm. Cells were diluted 1:100 from the overnight cultures into liquid LB and grown to mid-log phase, OD₆₀₀ ~0.5, at 37°C with shaking at 300 rpm. For N. gonorrhoeae, multiple colonies from the overnight CAP were resuspended in prewarmed Graver Wade media to achieve OD₆₀₀ ~0.5. DiSC₃(5) (Invitrogen D306) was dissolved in DMSO and added to liquid cultures at a final concentration of 1 μM. After additional incubation in the presence of DiSC₃(5) for 1 to 2 h, cells were plated in 200 μL working volumes in black, opaque flat-bottom 96-well plates, after which fluorescence was measured every 30 seconds for 5 minutes at an excitation/emission of 622/670 nm using a SpectraMax M3 plate reader. Cells were either untreated or treated with DMSO (1%), DN1, DN1 analogs, and NG1 at a final concentration of 10 μg/mL (10× MIC). Other control antibiotics were also tested at 10× MIC. Fluorescence was measured immediately following treatment according to the same specifications as above.

Laurdan membrane fluidity assay.

An OD₆₀₀ of 0.5 cell suspension of either N. gonorrhoeae ATCC 49226 in Graver Wade media or S. aureus RN4220 in LB was prepared. As previously described⁶, a solution of 1 mM Laurdan (Sigma-Aldrich 40227) was prepared in 100% DMF (Sigma-Aldrich PHR1553) and stored in the dark at −20°C. Then, 100 μL of the 1 mM Laurdan stock was added to the cell suspension and incubated for 10 minutes at 37°C with shaking at 300 rpm while covered with aluminum foil. The cells were then centrifuged at 4,000 rpm for 5 minutes and washed three times in 10 mL of PBS with 0.2% (w/v) glucose. Cells were then resuspended in 10 mL of the same solution and distributed into a single 96-well opaque flat-bottom plate (Costar Black Polystyrene Plate, 266) with 100 μL per well and 198 μL in column 1. Two μL of NG1, NG1 analog or DN1, starting at 128 μg/mL, was added to column 1, and serially diluted across columns 2-12 to obtain a final total volume of 100 μL in each well. To the control wells, Tween-20 (Sigma-Aldrich 655204), a known membrane fluidizer (Ipsen et al., 2022) starting at a concentration of 0.5% was added. Untreated or DMSO treated (1%) cells served as the negative control and triclosan (starting at 62.5 μg/mL) and azithromycin (starting at 250 μg/mL) were included as positive or neutral antibiotic controls. A kinetic read was taken every 45 seconds at 37°C using a SpectraMax M3 plate reader, exciting the plate at 330 nm and taking two emission readings at 460 nm and 500 nm. The Laurdan generalized polarization (GP) was calculated with the formula: $\text{GP}=(\mathrm{I}\_460\text{-}\mathrm{I}\_500)∕(\mathrm{I}\_460+\mathrm{I}\_500)$, where $\mathrm{I}\_460$ indicates the fluorescence intensity at 460 nm and $\mathrm{I}\_500$ indicates the fluorescence intensity at 500 nm. All fluidity measurements were performed in biological duplicate.

NPN assay.

Ten mL of an OD₆₀₀ of 1 cell suspension of N. gonorrhoeae ATCC 49226 was prepared in Graver Wade media. As previously described⁶, the cell suspension was spun at 4,000 rpm, 4°C for 15 minutes, washed twice in 5 mM HEPES buffer (Sigma-Aldrich SRE0065) with 20 mM glucose, and resuspended in an equal volume of the 20 μM NPN (Sigma-Aldrich 104043) in HEPES buffer. A 96-well opaque flat-bottom plate (Costar Black Polystyrene Plate, 266) was prepared with 100 μL of cells in 20 μM NPN in HEPES buffer added to columns 2-12. In column 1, 198 μL of 20 μM NPN in HEPES buffer was added, followed by 2 μL of NG1 or NG1 analog (starting concentration of 128 μg/mL). One hundred μL from column 1 was serially diluted across columns 2-12 to obtain a final total volume of 100 μL in each well. DMSO (1%) treated cells served as the negative control. Plates were then incubated at room temperature for 1 hour with no shaking and the fluorescence at 355/420 nm (excitation/emission) was read using a SpectraMax M3 plate reader with readings taken from the top. All measurements were normalized to corresponding values from the untreated control and performed in biological duplicates.

SYTOX Green assay.

Ten mL of an OD₆₀₀ of 0.5 cell suspension of N. gonorrhoeae ATCC 49226 was prepared in Graver Wade media. Ninety eight μL of cell suspension was added to a 96-well opaque flat-bottom plate (Costar Black Polystyrene Plate, 266), with 198 μL of cells to column 1. 2 μL of NG1 or NG1 analog (starting concentration of 128 μg/mL) was added to column 1 and 100 μL from column 1 was serially diluted across columns 2-12 to obtain a final total volume of 100 μL in each well. 1 μL of SYTOX Green Nucleic Acid Stain (5mM, S7020, Invitrogen, Carlsbad, CA), a DNA intercalating dye, was then added to each well at a final concentration of 5 μM. DMSO (1%) treated cells served as the negative control, and Triton X-100 served as positive control, resulting in fully compromised membranes. At selected time points (0, 2 and 4 hours post-treatment), fluorescence at 504/523 nm (excitation/emission) was read with a SpectraMax M3 plate reader. All measurements were normalized to the untreated control and performed in biological duplicate.

Proteomics-based target identification with PISA.

Sample preparation.

Colonies from an overnight plate culture of N. gonorrhoeae ATCC 49226 were resuspended in 650 mL of GC medium (Proteose Peptone No. 3, corn starch, dipotassium phosphate, monopotassium phosphate and sodium chloride; pH adjusted at 7.2 and autoclaved). The medium was supplemented with IsoVitalex 1% v/V final (BD 211876). The bacteria were cultured at 37 °C with 5% CO2 and grown to an OD₆₀₀ of approximately 0.3. Both bacterial cultures and lysates were treated in four replicates with 10 μM of each antibiotics for 20 min at 37°C.

The protocols for preparing bacterial cultures and lysates were adapted from ⁷³. For each condition, 20 mL of bacterial culture was used (4 conditions × 4 replicates × 2 antibiotics). The bacteria cells were washed with 37°C PBS. To prepare the native cell lysate, pelleted bacteria were resuspended in 2 mL lysis buffer (PBS + Halt protease inhibitor) in Eppendorf tubes. The bacteria were subjected to five freeze-thaw cycles in liquid nitrogen to extract proteins. Then the samples were centrifuged at 7000g for 5 min at RT to remove debris. The supernatant was then collected and treated with the antibiotic or DMSO for 20 min at 37°C. Each replicate was then aliquoted into 12-wells in a 96-well PCR plate (40 μL per well). The samples were then exposed to a temperature gradient from 46-70°C in the Biometra T-GRADIENT thermocycler for 3 min followed by an additional 3 min at RT and then snap frozen. The 12 samples for each replicate were then pooled into a single tube, transferred to polypropylene ultracentrifuge tubes and centrifuged at 100,000g for 20 min at 4°C. The supernatant was collected for proteomics analysis.

Briefly, as previously published^74-76, 30 μg of each sample was reduced with 10 mM DTT, alkylated with 50 mM IAA in the dark and subjected to methanol chloroform precipitation. Samples were resuspended in 20 mM EPPS buffer, 8 M urea, pH 8.0, and then diluted down to 1.6 M urea and digested with trypsin (1:60 trypsin:protein). Each digest was labeled using TMTpro 16-plex reagents (Thermo Fischer), and the samples were pooled and cleaned using a Sep-Pack C18 column (Waters).

NanoLC-MS/MS analysis.

The pooled sample was fractionated off-line by capillary reversed phase chromatography at pH 10 into 12 fractions and dried. NanoLC-MS/MS analyses were performed on an Orbitrap Exploris 480 mass spectrometer (Thermo Fisher Scientific). The instrument was equipped with an EASY ElectroSpray source and connected online to an Ultimate 3000 nanoflow UPLC system. The samples were pre-concentrated and desalted online using a PepMap C18 nano-trap column (length - 2 cm; inner diameter - 75 μm; particle size - 3 μm; pore size - 100 Å; Thermo Fisher Scientific) with a flow rate of 3 μL/min for 5 min. Peptide separation was performed on an EASY-Spray C18 reversed-phase nano-LC column (Acclaim PepMap RSLC; length - 50 cm; inner diameter - 2 μm; particle size - 2 μm; pore size – 100 Å; Thermo Scientific) at 55°C and a flow rate of 300 nL/min. Peptides were separated using a binary solvent system consisting of 0.1% (v/v) FA, 2% (v/v) ACN (solvent A) and 98% ACN (v/v), 0.1% (v/v) FA (solvent B). They were eluted with a gradient of 3–26% B in 97 min, and 26–95% B in 9 min. Subsequently, the analytical column was washed with 95% B for 5 min before re-equilibration with 3% B. The mass spectrometer was operated in a data-dependent acquisition mode. A survey mass spectrum (from m/z 375 to 1500) was acquired in the Orbitrap analyzer at a nominal resolution of 120,000. The automatic gain control (AGC) target was set as 100% standard, with the maximum injection time of 50 ms. The most abundant ions in charge states 2⁺ to 7⁺ were isolated in a 3-second cycle, fragmented using HCD MS/MS with 33% normalized collision energy, and detected in the Orbitrap analyzer at a nominal mass resolution of 50,000. The AGC target for MS/MS was set as 250% standard with a maximum injection time of 120 ms, whereas dynamic exclusion was set to 45 seconds with a 10-ppm mass window.

Data processing.

Proteome Discoverer 3.2 software (Thermo Scientific) was utilized for the database search and quantification against the UniProt N. gonorrhoeae (UP000000535; 2106 entries) protein database, downloaded on 28 February 2025. Cysteine carbamidomethylation was set as a fixed modification, along with TMT-related modifications, methionine oxidation, deamidation of arginine and asparagine as variable modifications. Enzyme specificity was defined as trypsin with a maximum of two missed cleavages. A 1% false discovery rate was employed as a filter at both the protein and peptide levels. Contaminants were removed, and proteins with missing values were eliminated. The quantified abundance of each protein in each sample (labeled with a different TMT) was normalized to the total intensity of all proteins in that sample. For each protein, the average normalized protein abundance in the antibiotic-treated replicates was divided by the average normalized abundance of that protein in the vehicle-treated replicates. The average ratio across replicates of each compound compared to the vehicle control was calculated, and the Log2 values of these ratios were determined. A two-tailed student’s t-test was employed to calculate the p-value.

Checkerboard assays.

The broth microdilution MICs of NG1, ceftriaxone, fosfomycin, CCCP, and PMB were determined to be 0.5, 0.008, 31.25, and 1 μg/mL, respectively, for ATCC 49226. The broth dilution MICs for NG1 and polymyxin B were found to be 8 and 187.5 μg/mL, respectively, for CDC-FDA ARB #0187. Two-fold serial dilutions of each compound were made to achieve 10× to 0.156× MIC final concentrations. The bacterial inoculum was made with a 1:200 dilution of an N. gonorrhoeae OD₆₀₀ 0.08 suspension in Graver Wade media. Each well of a 96-well round-bottom clear plate (Corning 3799) received 98 μL of the bacterial suspension and one μL each of two-fold serially diluted compound at 100× final concentration, keeping the final DMSO concentration ≤1%. The plates were incubated overnight at 37°C with 5% CO₂ in a humidity-controlled incubator without shaking. After overnight incubation, OD₆₀₀ was measured and then 10 μL of PrestoBlue^™ Cell Viability Reagent (Invitrogen, A13261) was added to each well to obtain quantitative cell viability measurements. Four hours post-incubation at 37°C, fluorescence was read at an excitation/emission of 550/590 nm using a SpectraMax M3 plate reader. MICs were determined as the minimum inhibitory concentration that leads to at least 50% growth inhibition (compared to the untreated control), and the fractional inhibitory concentration index ($\text{FICi}$) for drug combinations was determined. $\text{FICi}$ was calculated as follows:

$$FICi=\frac{MI{C}_{AB}}{MI{C}_A}+\frac{MI{C}_{BA}}{MI{C}_B}$$

where ${\text{MIC}}_{\mathrm{A}}$ and ${\text{MIC}}_{\mathrm{B}}$ are the MIC of each antibiotic when administered individually. ${\text{MIC}}_{\text{AB}}$ is the MIC of antibiotic A in combination with antibiotic B, and ${\text{MIC}}_{\text{BA}}$ is the MIC of antibiotic B in combination with antibiotic A. FICi of <0.5 indicates synergy, values between 0.5-4 indicate indifference, and >4 indicates antagonism.

RNAseq.

N. gonorrhoeae ATCC 700825 (FA1090) was grown to mid-log phase in Graver Wade medium (OD ~0.3). Compounds were added at 4× final MIC, and an equivalent volume of DMSO was added to control cultures. Following treatment, cultures were returned to the incubator and 1 mL aliquots were collected at specified time points: 0, 15, 30, and 60 minutes. Each condition was tested in triplicate. All samples were immediately centrifuged after collection at 15,000 × g for 1 minute to pellet cells. Supernatants were removed and pellets were resuspended in pre-warmed TRIzol reagent (65 °C; ThermoFisher, Cat# 15596026), vortexed, and placed on wet ice for 5 minutes. To all samples, chloroform (Sigma-Aldrich, Cat# C2432) was added, mixed, and centrifuged at 12,000 × g for 15 minutes at 4 °C. The aqueous layer was mixed with 100% ethanol, incubated at room temperature for 5 minutes, and the RNA was purified using the Direct-zol RNA Miniprep Plus Kit (Zymo Research, Cat# R2070) according to the manufacturer’s protocol. RNA was eluted in 100 μL nuclease-free water following a 5-minute incubation. RNA concentration was measured using a Qubit 4 Fluorometer (ThermoFisher) per manufacturer guidelines.

Library preparation was carried out using the Illumina^® Stranded Total RNA Prep, Ligation with Ribo-Zero Plus kit with 10 bp unique dual indices. Prepared libraries were sequenced on a NovaSeq X Plus platform (Illumina), generating paired-end 150 bp reads. FASTQ files were generated using Illumina’s BCL Convert (v4.2.4). Reads were aligned to the N. gonorrhoeae FA1090 reference genome (NCBI: GCF_000006845.1). Gene-level annotations were used to quantify expression levels. Gene read counts were analyzed using the edgeR package in R (verson 4.4.2). Differential expression was calculated using a generalized linear model (GLM) approach. Genes with a false discovery rate (FDR) < 0.05 were considered significantly differentially expressed. KEGG pathway analysis was performed using gene set enrichment analysis in the clusterProfiler R package (v4.10.0).

Bacterial cytological profiling

Bacterial strain and culture conditions.

All cytological profiling experiments were performed using Bacillus subtilis strain 168 in biological duplicate. Cultures were grown overnight in tryptic soy broth (TSB) at 37°C with shaking at 220 rpm. Overnight cultures were diluted from an OD₆₀₀ of 0.7 into fresh TSB and seeded into poly-D-lysine–coated black-wall, clear-bottom 96-well plates (Millipore Sigma, CLS3842) for imaging.

Drug treatment and fixation.

Cells were incubated at 37°C with orbital shaking (50 rpm) for 30 minutes post-seeding before treatment with either vehicle control or antibiotics. Drug exposure was carried out for 1.5 hours. Following treatment, cells were fixed with 4% paraformaldehyde (PFA) in phosphate-buffered saline (PBS) for 15 minutes at room temperature and washed twice with PBS prior to staining.

Staining and imaging.

Fixed cells were stained using FM4-64 (membrane; Thermo Fisher, T13320) and DAPI (nuclei; Thermo Fisher, D1306) according to the manufacturer’s instructions. Imaging was performed using an Opera Phenix 2 high-content confocal microscope (PerkinElmer) equipped with a 63× water-immersion objective. Confocal z-stacks were acquired with five planes spaced 1 μm apart. Brightfield images were acquired in parallel.

Image analysis and feature extraction.

Raw images were processed using Harmony (PerkinElmer) and Signal Image Artist. Bacterial segmentation and morphometric feature extraction were performed using Omnipose. Objects were filtered to remove background and artifacts using built-in heuristics, and single cell areas (in μm²) were extracted. Feature distributions were aggregated across fields of view and exported for downstream analysis. Summary statistics were computed using custom Python scripts. Group comparisons were evaluated using nonparametric Mann-Whitney U tests.

Cryo-TEM analysis of bacterial morphology

Sampling.

Neisseria gonorrhoeae (strain FA1090) and Staphylococcus aureus (RN4220) was grown overnight in stationary cultures. Cells were harvested by transferring the culture to a 5 mL tube and centrifuging at 2000 rpm for 1 minutes. The supernatant was discarded, and the cell pellets were washed twice by resuspending in 3 mL of phosphate-buffered saline (PBS; Sigma-Aldrich) and centrifuging under the same conditions. After the final wash, the cells were resuspended in 99 μL of LB and 1 μL of the drug (NG1 or DN1 at 4× MIC). The samples were incubated at 37°C for 2 hours. The final cell suspension was centrifuged again, adjusted to an optical density at 600 nm (OD₆₀₀) of 0.3 and fixed in 2.5% glutaraldehyde at 4 °C for 1 hour. Following fixation, the cells were washed three times with deionized water. 3 μL of the cell suspension was applied to glow-discharged carbon-coated grids. The grids were vitrified in liquid ethane using a Gatan Cryoplunge system and the imaging was carried out using a JEOL 2100F transmission electron microscope.

Image processing and segmentation.

Raw cryo-TEM images of S. aureus and N. gonorrhoeae were manually segmented to delineate the cell boundary and, when visible, the cytosolic region. Segmented binary masks of cryo-TEM images were analyzed using a custom Python pipeline built with OpenCV, scikit-image, and SciPy. For each object, region properties including area, perimeter, major and minor axis lengths were extracted. Elongation was defined as the ratio of major to minor axis length, and roundness as its inverse. Objects with an area less than 50 pixels were excluded. Image-specific scale factors (μm/pixel), retrieved from a CSV calibration file, were applied to convert all measurements to physical units. When inner membrane masks were available, membrane thickness was estimated by computing the Euclidean distance transform between outer and inner membrane masks. The average distance was doubled to approximate membrane thickness.

Per-cell measurements were aggregated across images and saved as .csv files. Summary statistics (mean ± standard deviation) were computed for each group. Group comparisons (e.g., DN1-treated vs. untreated) were evaluated using two-sided Mann-Whitney U tests (ranksum), appropriate for non-normally distributed data.

Toxicity studies.

Hemolysis and modified Ames genotoxicity studies were performed as described previously⁷. To test the systemic toxicity of DN1, healthy female C57BL/6J mice were given a single 100 mg/kg dose of DN1, which three out of three mice tolerated. An additional three mice were given four doses of 50 mg/kg every 6 hours, which they all also tolerated. To test the systemic toxicity of NG1, healthy female BALB/c mice were given increasing doses of NG1, up to 100 mg/kg. Both compounds were formulated in a 10% compound in DMSO:45% PEG300:45% water solution and administered as a 200 μL intraperitoneal injection. Mice were observed for at least 24 hours for typical signs of toxicity, including impaired movement, lethality, and irritation. Results were representative of at least three mice per dose of each compound.

Mouse S. aureus topical wound infection model.

Female C57BL/6J mice were given over a week to acclimate and then rendered neutropenic with cyclophosphamide (Cytoxan) on Day −4 (150 mg/kg, I.P.) and Day −1 (100 mg/kg, I.P.). On Day 0, a fresh suspension of S. aureus BAA1556 was prepared in tryptic soy broth and titered via serial dilution and plating. Mice were given buprenorphine for anesthesia and kept sedated under isoflurane vapors (3%) during the infection procedure. A ~1.5 cm² patch of skin was prepared on each mouse’s dorsal surface by shaving the fur, sterilizing the underlying skin with iodine and ethanol swabs thrice, allowing the skin to dry completely, and abrading the skin until visibly damaged (reddening and glistening) but not bleeding. Five μL of the S. aureus suspension, corresponding to an inoculum of 1.5 × 10⁵ CFU, was pipetted onto the skin to initiate the bacterial infection. Treatments were administered at 1, 4, 8, 16, 21, and 24 hours post-infection by pipetting 40 μL of formulation topically onto the infected skin and allowing to dry. Treatments groups consisted of n = 6 mice receiving DN1 (1% final concentration, prepared as described above), n = 6 mice receiving vehicle control (DMSO:PEG300:water at 10%:45%:45%), and n = 6 mice receiving the fusidic acid positive control (0.25% final concentration in DMSO:PEG300:water at 10%:45%:45%) At 25 hours post-infection (~1 hour following the last topical treatment), all mice were euthanized by CO₂ asphyxiation, and wounds were wiped with an alcohol pad, excised, weighed, and homogenized in 2 mL of sterile PBS using a Polytron PT10-35 with a 12 mm aggregate that was cleaned with ethanol and water between samples. Homogenized wounds were serially diluted and plated onto BHI agar to determine bacterial titers (CFU/g tissue).

Mouse N. gonorrhoeae vaginal infection model.

Female ovariectomized BALB/cJ mice were given over a week to acclimate prior to handling. As described previously⁷⁷, two days prior to infection (Day -2), vaginal lavage was performed for estrous staging and monitoring of the vaginal microbiota via culture on MacConkey and Brain Heart Infusion agar plates. To increase susceptibility to infection, each mouse received a dose of 17B-estradiol (0.23 mg, I.P.) on Day -2 and Day 0. To reduce the overgrowth of commensal bacteria that occurs with estradiol treatment, mice were given streptomycin (1.2 mg, I.P.) and vancomycin (0.6 mg, I.P.) (1 dose on Day -2, two doses at least 5 hours apart on Day -1, and one dose Day 0) as well as trimethoprim (0.4 g/L) in the drinking water (refreshed on Day 0). On Day 0, the inoculum was prepared by collecting isolated colonies of N. gonorrhoeae ATCC 49226 from an overnight chocolate agar plate in sterile PBS to achieve an OD₆₀₀ 0.2. Within one hour of preparation, the vagina was first rinsed with 30 μL of 50 mM HEPES (pH 7.4) and then 20 μL of the bacterial suspension (~3 × 10⁶ CFU per mouse) was pipetted intravaginally while the mouse was held by the tail with paws grasping the wire cage for at least 1 minute. Serial dilutions of the bacterial suspension were plated onto chocolate agar to determine the initial inoculum. Treatments were administered at 2, 6, 10, 18, and 24 hours after infection by pipetting 20 μL of test compound NG1 (1% final concentration in 10% DMSO, 45% PEG300, 45% water), ceftriaxone (0.1% w/v in water), or vehicle control (10% DMSO, 45% PEG300, 45% water) intravaginally; mice were suspended by the tail for 1 minute before being released into the cage. At 24 hours after infection, mice were euthanized by CO₂ asphyxiation and vaginal lavage was performed using 50 μL of Graver Wade media with 0.05% saponin. N. gonorrhoeae burden was determined by plating 30 μL of neat lavage and 10-fold serial dilutions in PBS onto Thayer Martin Agar Improved (Thermo Fisher); colonies were counted after 18-24 hours of incubation at 37°C with 5% CO₂.

QUANTIFICATION AND STATISTICAL ANALYSIS

All statistical analyses were performed using Python scripts and/or GraphPad Prism (version 10.1.0). Details of the statistical tests used, the exact value of n, what n represents, as well as measures of central tendency (mean or median) and dispersion (e.g., standard deviation) are provided in the Figure legends and STAR Methods section. For each analysis, the choice of statistical test (e.g., student’s t-test or non-parametric tests, such as the Mann-Whitney U or Kolmogorov-Smirnov test) was based on the data type and experimental design. For mouse experiments, sample sizes were determined by power calculations in GraphPad Prism to detect a 1-2 log-fold change in effect size and mouse availability. Group allocation was determined by cage, where each cage contained 3-5 mice that all received the same treatment. No data were excluded from the analyses in the study. The data were assumed to be non-normal; thus, a two-sided Mann-Whitney test was performed using GraphPad Prism. Statistical details can be found in the Figure legends.

Supplementary Material

Data S1

Data S1. Training datasets for deep learning and list of 559 known antibiotics, related to Figures 1 and 4.

Data S2

Data S2. All tested compounds (“actives” and “inactives”), related to Figures 1, 2, 3, 4, 5 and 7

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and Virus Strains
Staphylococcus aureus RN4220 (MSSA)	DSMZ-German Collection of Microorganisms and Cell Cultures GmbH, Leibniz Institute	DSM 26309
Staphylococcus aureus FPR3757 (MRSA)	ATCC - American Type Culture Collection	BAA-1556
Neisseria gonorrhoeae F-18	ATCC	49226
Neisseria gonorrhoeae FA1090	ATCC	700825
Bacillus subtilis 168	ATCC	23857
Escherichia coli BW25113	DSMZ	DSM 27469
Escherichia coli K-12 MG1655	ATCC	700926
Escherichia coli JW5503-KanS (ΔtolC832::FRT)	EC GRC – E coli Genetic Resource Center	CGSC: 14206
Escherichia coli RFM795 (lptD4213)	EC GRC – E coli Genetic Resource Center	CGSC: 14179
Klebsiella pneumoniae NCTC 9633	ATCC	13883
Pseudomonas aeruginosa PAO1	ATCC	47085
Acinetobacter baumannii 5377	ATCC	17978
Mycobacterium tuberculosis H37Ra	ATCC	25177
CDC & FDA Antibiotic Resistance Bank (ARB) isolates	CDC & FDA AR Isolate Bank	https://wwwn.cdc.gov/arisolatebank/ See Table S4 for unique identifiers
Chemicals, Peptides, and Recombinant Proteins
De novo generated compounds and analogs	This study; synthesized by Enamine Ltd. (Kyiv, Ukraine)	See Data S2
NG1 analogs	CC4CARB (Chemistry Center for Combating Antibiotic-Resistant Bacteria)	See Data S2
Vancomycin	Fisher Scientific	AAJ6279003
Valinomycin	Thermo Fisher	V1644
Triclosan	MilliporeSigma	72779
Azithromycin	Cayman Chemical	15004
Ceftriaxone	Cayman Chemical	18866
Fosfomycin sodium	MilliporeSigma	34089
CCCP	MedChemExpress	HY-100941
Fusidic acid	MilliporeSigma	F0881
Ciprofloxacin	MilliporeSigma	17850
Kanamycin sulfate	MilliporeSigma	60615
DMSO	MilliporeSigma	D5879
PrestoBlue Cell Viability Reagent	Invitrogen	A13261
Resazurin	MilliporeSigma	R7017
Triton X-100	MilliporeSigma	T8787
DiSC3(5)	Invitrogen	D306
Laurdan	Sigma-Aldrich	40227
NPN (N-phenyl-1-naphthylamine)	Sigma-Aldrich	104043
SYTOX Green	Invitrogen	S7020
TRIzol Reagent	Thermo Fisher	15596026
Chloroform	Sigma-Aldrich	C2432
FM4-64	Thermo Fisher	T13320
DAPI	Thermo Fisher	D1306
Experimental models: Cell Lines
HEK293	ATCC	CRL-1573
HepG2	ATCC	HB-8065
HSkMC	ATCC	PCS-950-010
IMR90	ATCC	CCL-186
Experimental Models: Organisms/Strains
Female C57BL/6J mice	Jackson Laboratory	000664
Female BALB/cJ	Jackson Laboratory	000651
Female ovariectomized BALB/cJ mice	Jackson Laboratory	000651
Software and Algorithms
Chemprop	Yang et al., 2019	https://github.com/chemprop/chemprop
RDKit	RDKit	https://www.rdkit.org
SHAP	Lundberg & Lee, 2017	https://github.com/slundberg/shap
openTSNE	Policar et al., 2024	https://opentsne.readthedocs.io
CReM	Polishchuk, 2020	https://github.com/DrrDom/chemically-reasonable-mutations
F-VAE	Jin et al. 2020	https://github.com/wengong-jin/multiobj-rationale
JT-VAE	Jin et al. 2018	https://github.com/wengong-jin/hgraph2graph
ASKCOS	Tu et al., 2025	https://askcos.mit.edu
GraphPad Prism v10.1.0	GraphPad Software	https://www.graphpad.com
Harmony & Signal Image Artist	PerkinElmer	N/A
Omnipose	Cutler et al., 2022	https://github.com/kevinjohncutler/omnipose
Proteome Discoverer	Thermo Fisher	v3.2
edgeR v4.4.2	Bioconductor	https://bioconductor.org/packages/release/bioc/html/edgeR.html
clusterProfiler v4.10.0	Bioconductor	https://www.bioconductor.org/packages//2.13/bioc/html/clusterProfiler.html
Deposited Data
Code platform for reproducing all analyses in this work	This study	https://github.com/aartikrish/de-novo-antibiotics
Data generated from the computational pipelines (fragment based and de novo design)	This study	Zenodo repository https://doi.org/10.5281/zenodo.15191826
Data generated from experimental chemical screening	Wong et al. 2024; Anahtar et al. n.d.	Zenodo repository https://doi.org/10.5281/zenodo.15191826
RNA sequencing data	This study	Zenodo repository https://doi.org/10.5281/zenodo.15191826
Proteomics data	This study	PRIDE ID: PXD063107
Other
SpectraMax M3 Multi-Mode Microplate Reader	Molecular Devices	N/A
JEM-2100F Field Emission Electron Microscope for Cryo-TEM	JEOL	N/A
Gatan Cryoplunge	Gatan	N/A
Opera Phenix 2 Confocal Microscope for High-Content Screening	PerkinElmer	N/A
Biometra T-GRADIENT thermocycler	Analytik Jena	8464070201
Orbitrap Exploris 480 Mass Spectrometer	Thermo Fisher Scientific	BRE725539
NovaSeq X Plus	Illumina	N/A
Labcyte Echo 555 Acoustic Dispenser	Beckman Coulter	N/A
Qubit 4 Fluorometer	Thermo Fisher	N/A
Direct-zol RNA Miniprep Plus Kit	Zymo Research	R2070
Sep-Pak C18 Column	Waters	N/A
PepMap C18 trap and RSLC nano-LC columns	Thermo Fisher	N/A