Compounds for the treatment of bacterial infections
SyntheMol-RL addresses the limitations of SyntheMol-MCTS by using reinforcement learning to optimize chemical space exploration and multi-parameter optimization, resulting in compounds with enhanced antibacterial activity and solubility, effectively inhibiting Staphylococcus aureus growth and treating infections.
Patent Information
- Application Number
- PCT/CA2026/050701
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-16
- Filing Date
- 2026-05-15
- Publication Date
- 2026-11-19
AI Technical Summary
Existing generative AI methods for drug discovery, such as SyntheMol-MCTS, struggle with synthesizability and the inability to optimize multiple molecular properties simultaneously, limiting their effectiveness in identifying compounds effective against antibiotic-resistant bacteria like Staphylococcus aureus.
SyntheMol-RL, a reinforcement learning-based model, enhances the chemical space exploration by generalizing across chemically similar building blocks and performs multi-parameter optimization, generating compounds that are easy to synthesize and possess multiple relevant properties like antibacterial activity and aqueous solubility.
SyntheMol-RL effectively identifies compounds with improved predicted properties, generating 79 unique compounds, with 13 showing potent in vitro activity, and one compound, synthecin, demonstrating efficacy in a murine wound infection model against MRSA.
Smart Images

Figure CA2026050701_19112026_PF_FP_ABST
Abstract
Description
COMPOUNDS FOR THE TREATMENT OF BACTERIAL INFECTIONSFIELD
[0001] The present disclosure relates to compounds, compositions and uses thereof for inhibiting growth of a Staphylococcus aureus bacterium and / or treating a Staphylococcus aureus bacterial infection.BACKGROUND
[0002] The rapid spread of antibiotic resistance is a critical challenge facing modern medicine, with global implications for public health. In 2019, approximately 4.95 million deaths were linked to infections caused by drug-resistant bacteria, and this number is expected to rise to 10 million annually by 2050 if the emergence of antimicrobial resistance (AMR) continues to outpace the development of new antibiotics [1], Among the most concerning are the ESKAPE pathogens, six bacterial species notorious for their virulence and resistance to multiple drugs [2], Among these, Staphylococcus aureus stands out as a significant cause of morbidity and mortality in both community and healthcare settings [3,4], In high-income regions, S. aureus is the leading cause of deaths associated with AMR (25.4% of such deaths), and globally, in 2019, over 700,000 deaths associated with AMR were linked to S. aureus infections [1]. Furthermore, of drug-resistant pathogens, methicillin-resistant S. aureus (MRSA) had the greatest increase in attributable burden from 1990 to 2021, doubling from 57,200 to 130,000 attributable deaths over this period. The World Health Organization highlights S. aureus as a high priority for the development of new antibiotics, particularly due to the dwindling number of treatment options available for antibiotic-resistant S. aureus infections [5], Thus, there is an urgent need for novel therapeutic approaches to combat S. aureus and its increasing antibiotic resistance.
[0003] Artificial intelligence (Al) is playing an important role in drug development broadly and in antibiotic discovery in particular [6-9], One Al method commonly employed in drug discovery is property prediction models, which are trained to predict the properties of molecules such as antibacterial activity or aqueous solubility
[0010] , These models can be applied in a virtual screening
[0011] approach to evaluate large chemical libraries to identify compounds that are predicted to be effective. However, since property prediction models must evaluate molecules one-by-one, they do not scale well to the chemical spaces of billions of molecules that are increasingly used in industrial and academic settings. Indeed, in recent years, multiple ultra-large synthetically accessiblechemical databases have been assembled, including Enamine’s REAL Space, WuXi’s GalaXi, and OTAVA’s CHEMriya, all exceeding 1010molecules. Therefore, more recently, generative Al methods have been developed to directly design molecules with promising properties without a slow evaluation process [12-15],
[0004] Generative Al methods for drug discovery are promising, but a major limitation of many of these methods is the synthesizability of the compounds proposed by these models
[0016] , Although many of the ALgenerated molecules are predicted to be effective, they are often synthetically intractable, thus preventing experimental validation and limiting the real-world impact of these models. To address this challenge, SyntheMol
[0017] was previously developed as a generative Al method for drug design that uses a Monte Carlo tree search (MCTS) guided by a trained antibacterial activity property predictor to explore a space of 30 billion easy-to-synthesize compounds. This method enabled the rapid synthesis of 58 SyntheMol-generated compounds and experimentally validate their antibacterial efficacy, with six compounds proving to have high potency against Acinetobacter baumannii and several other ESKAPE species.
[0005] SyntheMol successfully designed compounds with in vitro antibacterial activity, but two important limitations were observed. First, the MCTS algorithm that powers SyntheMol is not the most effective method for exploring chemical space. Indeed, SyntheMol builds compounds by selecting molecular building blocks and combining them with a pre-defined set of chemical reactions to form molecules, and MCTS is responsible for scoring and selecting the most promising building blocks. However, MCTS treats every building block independently and cannot learn patterns of antibacterial activity that are shared across chemically similar building blocks, thereby limiting its search efficiency. Second, SyntheMol can only optimize a single molecular property at a time, even though real-world drug discovery requires identifying compounds with many druglike properties simultaneously. This proved to be a particular challenge in prior work, where only two of six potent antibacterial compounds were sufficiently soluble to be administered for in vivo toxicity experiments.
[0006] The background herein is included solely to explain the context of the disclosure. This is not to be taken as an admission that any of the material referred to was published, known, or part of the common general knowledge as of the priority date.SUMMARY
[0007] In the present disclosure, SyntheMol-RL is introduced. SyntheMol-RL is a significantly enhanced version of SyntheMol that overcomes the two limitations of the original version of SyntheMol (hereafter referred to as SyntheMol-MCTS). First, SyntheMol-RL replaces the MCTS algorithm with a reinforcement learning model that can generalize across chemically similar building blocks to more rapidly and effectively search massive combinatorial chemical spaces for promising, easy-to-synthesize compounds. Second, SyntheMol-RL introduces the ability to perform multi-parameter optimization to generate compounds that simultaneously possess multiple relevant molecular properties for drug discovery applications. Moreover, the chemical space explored by SyntheMol-RL was expanded from 30 billion compounds to 46 billion compounds.
[0008] The application of SyntheMol-RL to design easily synthesizable compounds against the target bacterium 5. aureus with simultaneous optimization for antibacterial activity against S. aureus and aqueous solubility (Fig. 1). Importantly, SyntheMol-RL markedly outperforms SyntheMol-MCTS and a virtual screening approach that uses property prediction models in silico, finding more compounds that are new, diverse, and predicted to be antibacterial. The present disclosure is also directed to the synthesis and experimental testing of 79 compounds designed by two variants of SyntheMol-RL and found two and 11 potent hits, respectively, compared to zero hits for SyntheMol-MCTS and two hits for virtual screening. For example, a de novo generated compound was further investigated, known as synthecin herein, and its ability to fully arrest growth of MRSA in a murine wound infection model was demonstrated. These results demonstrate that SyntheMol-RL is an effective and flexible framework for novel drug design applications.
[0009] In various aspects, the present disclosure provides a compound of Formula (I):R2-C(O)-NR1R7(Formula (I))or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof, wherein:R1is H or C1-C6alkyl;R2is benzyl substituted with up to four R3, wherein each R3is independently selected from hydroxyl, halogen or CF3, or R2is CH2CR4R5R6, wherein R4is H or hydroxyl, R5is hydroxy, halogen or CF3 and R6is benzyl, wherein the benzyl is optionally substituted with one or more substituents independently selected from C1-C6alklyl, hydroxyl, halogen or CF3;R7is C1-C6alkyl, optionally substituted with one or more substituents independently selected from hydroxyl, halogen, benzyl or 5- or 6-membered heterocycle comprising S, N or O, wherein the benzyl or the 5- or 6-membered heterocycle is optionally substituted with one or more substituents independently selected from C1-C6alkyl, hydroxyl, halogen or CF3, or R7is benzyl or thiol, optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
[0010] In various embodiments, R2is benzyl substituted with three R3, wherein each R3is independently selected from hydroxyl or chloro.
[0011] In various embodiments, R2is CH2CR4R5R6, wherein R4is H or hydroxyl, R5is CF3 and R6is benzyl.
[0012] In various embodiments, R7is C1-C4alkyl, optionally substituted with one or more substituents independently selected from fluoro, hydroxyl or benzyl, wherein the benzyl is optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
[0013] In various embodiments, R7is benzyl or thiol, optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
[0014] In various embodiments, R1is H.
[0015] In various embodiments, R6is benzyl optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
[0016] In various embodiments, the compound of Formula (I) is:salt, solvate, and / or prodrug thereof.
[0017] In various embodiments, the compound is:or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof.
[0018] In various embodiments, the prodrug is an ester with a hydroxyl or an amine group of the compound. For example, the ester is a C1-C6alkyl ester.
[0019] In various embodiments, the compounds as disclosed herein are for inhibiting growth of a Staphylococcus aureus bacterium.
[0020] Accordingly, provided is a compound for inhibiting growth of aStaphylococcus aureus bacterium:or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof.
[0021] According to another aspect of the present disclosure, provided is a composition comprising a compound disclosed herein, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, and a pharmaceutically acceptable carrier or vehicle.
[0022] According to a further aspect of the present disclosure, provided is a method for inhibiting growth of a Staphylococcus aureus bacterium, comprising contacting the bacterium with an effective amount of a compound disclosed herein, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof.
[0023] According to another aspect of the present disclosure, provided is the use of a compound disclosed herein, or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof, to inhibit the growth of a Staphylococcus aureus bacterium.
[0024] According to a further aspect of the present disclosure, provided is the use of a compound disclosed herein, or pharmaceutically acceptable salt, solvate and / or prodrug thereof, to treat or prevent a Staphylococcus aureus bacterial infection.
[0025] Other features and advantages of the present disclosure will become apparent from the following detailed description. It should be understood, however, that the detailed description and the specific examples, while indicating embodiments of the disclosure, are given by way of illustration only and the scope of the claims should not be limited by these embodiments, but should be given the broadest interpretation consistent with the description as a whole.DRAWINGS
[0026] Certain embodiments of the disclosure will now be described in greater detail with reference to the attached drawings in which:
[0027] FIGURE 1 shows an overview of the SyntheMol pipeline in exemplary embodiments of the disclosure. The SyntheMol-RL pipeline can be broken into three primary steps. (1) Property prediction models are trained based on a combination of inhouse chemical screening data against 5. aureus as well as aqueous solubility data obtained from the Therapeutic Data Commons. (2) SyntheMol-RL generates molecules guided by these property prediction models. A reinforcement learning (RL) model is used to iteratively select molecular building blocks based on their potential for generating molecules with promising 5. aureus antibacterial activity and aqueous solubility. The generated molecule is then scored using the property prediction models, and these scores are used to update the RL model to improve its generative capacity. (3) A prioritized subset of molecules is selected for synthesis and laboratory testing based on predicted properties as well as novelty and diversity. For example, this SyntheMol-RL pipeline generated a new compound that can effectively treat a MRSA wound infection in a mouse model, showing the utility of SyntheMol-RL for real-world drug discovery applications.
[0028] FIGURE 2 shows property prediction models and chemical spaces in exemplary embodiments of the disclosure, (a) A diverse set of small molecules were screened against S. aureus RN4220 at 50 pM final concentration. The mean normalized growth across two biological replicates is plotted. The hit cutoff is calculated as: μ − 2σ (the mean minus two standard deviations). Points below the cutoff are considered “hits” (n = 1,137), and points above the cutoff are “non-hits” (n = 9,521), totaling 10,658 data points used for training, (b) A t-SNE visualization of chemical spaces represented by ChEMBL antibiotics and the 5. aureus training set in comparison to the chemical spaces explored by SyntheMol-RL. (c) Depictions of the two types of models used in this study. Chemprop-RDKit (top) was consistently used as the property predictor in each experiment. Both architectures were used for the reinforcement learning model, (d) ROC and precision-recall curves for each model in an ensemble of 10 models used as the 5. aureus activity property prediction model. Dark curves represent the average of all 10 models in the ensemble. Area under the curve (AUC) is indicated on the respective plots, (e) Example reactions are shown for the two ultralarge chemical spaces used for SyntheMol-RL: the Enamine REAL Space (top) and the WuXi GalaXi (bottom).
[0029] FIGURE 3 shows the analysis of 5. aureus property predictor model and chemical spaces in exemplary embodiments of the disclosure, (a) Normalized growth of 5. aureus RN4220 in duplicate experiments used as the training data, (b) ROC and (c) precision-recall curves for each 5. aureus activity property prediction model in an ensemble of 10 split by scaffold. Dark curves represent the average across all models. Area under the curve is indicated on respective graphs, (d), Density plots showing the distribution of molecular weight and cLogP across 1) the 5. aureus model training set 2) a collection of known antibiotics from ChEMBL and 3) the REAL and WuXi GalaXi chemical spaces explored by SyntheMol-RL. Not visualized are two molecules with molecular weights exceeding >3,000 g / mol. (e) Same as (d), but focused on the region of the distribution containing the majority of molecules for visual clarity.
[0030] FIGURE 4 shows generating compounds with SyntheMol-RL in exemplary embodiments of the disclosure, (a) Venn diagrams showing the number of molecules from each SyntheMol method (RL-Chemprop, RL-MLP, MCTS) or virtual screening method (VS-Chemprop) that pass the respective thresholds for predicted 5.aureus activity and solubility, respectively, (b) Violin plots displaying the distribution of 5. aureus scores across the four methods. The dotted line indicates the threshold used to define “hit” (> 0.5). (c) Violin plots displaying the distribution of predicted log solubility scores across four methods. The dotted line indicates the threshold used to define “hit” (> -4). (d) Line graph representing the number of molecules present at each stage of post-hoc filtering for each version of SyntheMol. (e) Violin plots displaying the distribution of 5. aureus scores across the four methods after applying filters outlined in Methods. Random molecules were randomly selected from the chemical space and serve as a baseline for comparison. The dotted line indicates the threshold used to define “hit” (> 0.5). (f) Violin plots displaying the distribution of predicted log solubility scores across the four methods after applying filters outlined in Methods as well as random molecules. The dotted line indicates the threshold used to define “hit” (> -4). (g) A t-SNE visualization of the training set along with all generated molecules from each of the three SyntheMol models compared to known antibiotics from ChEMBL.
[0031] FIGURE 5 shows the SyntheMol-RL ablation experiments in exemplary embodiments of the disclosure, (a) Distribution of predicted 5. aureus activity and log solubility scores for the final dynamic weighting system (Dyn) and fixed weightings using RL-Chemprop. The sum of weights is equal to one. (b) Number of molecules generated that pass property “hit” thresholds for 5. aureus activity and log solubility (>0.5 and > -4 respectively). Dynamic weighting achieves optimal or near-optimal performance without need for additional fine-tuning, shown by generating the most “hit” compounds, (c) Distribution of predicted 5. aureus activity and log solubility scores for the final dynamic weighting system and fixed weightings using RL-MLP. The sum of weights is equal to one. (d) Number of molecules generated that pass property “hit” thresholds for 5. aureus activity and log solubility (> 0.5 and > -4 respectively), (e) Tanimoto similarities of each rollout compared to previous rollouts for RL-Chemprop set at desired target similarities in range [0.4, 0.8], The model can effectively generate compounds at a user-defined diversity, (f) Tanimoto similarities of each rollout compared to previous rollouts for RL-Chemprop at fixed temperatures and dynamic temperature for reference. Similarity is highly sensitive to changes in temperature, further necessitating the use of dynamic changes to output desired diversity.
[0032] FIGURE 6 shows the comparisons of hits generated by SyntheMol-RL and GFlowNet in exemplary embodiments of the disclosure. SyntheMol-RL and GFlowNet generated “hit” compounds that passed fdters for predicted antibacterial activity (>0.5), log solubility (>-4), novelty, and diversity (Methods). Additionally, GFlowNet molecules were fdtered for synthesizability (Synthetic Accessibility score <4) to ensure a fair comparison to SyntheMol-RL compounds, which are intrinsically synthetically accessible. RL-Chemprop generated 495 hits (total = 10,983), RL-MLP generated 356 hits (total = 9,227), and GFlowNet generated 313 hits (total = 10,304). These hits were compared based on three properties of interest: a, 5. aureus score, b, log solubility score, and c, molecular weight. Representative samples are also visualized for d, RL-Chemprop, e, RL-MLP, and f, GFlowNet.
[0033] FIGURE 7 illustrates that generated molecules show potency in vitro activity in exemplary embodiments of the disclosure, (a) A heat map showing the minimum inhibitory concentrations (MICs) of up to 44 compounds from each generative SyntheMol method (RL-Chemprop, RL-MLP, MCTS), the virtual screening method (VS-Chemprop), or the random molecules. Molecules were tested against 5. aureus RN4220 in LB medium, with compounds tested at concentrations ranging from 128 pg / ml to 2 pg / ml in two-fold serial dilutions. Lighter shades indicate lower MIC values. Experiments were performed in biological duplicate. Annotated squares denote “hit” molecules with MIC ≤ 8 pg / ml. Those denoted with an asterisk (*) indicate molecules identified in the literature, (b) Growth inhibition of S. aureus RN4220 by seven molecules not identified in the literature. Ciprofloxacin is shown as a positive control. Structures of compounds are shown. Experiments were performed using twofold serial dilution series. Experiments were performed in biological duplicate, with each point representing one replicate, (c) Heat maps depicting the MICs of the seven compounds against representative isolates of the ESKAPE pathogens (left) and against a representative sample of S. aureus clinical isolates from the CDC AR Isolate Bank (right). Molecules are presented in the same order as (b) from left to right. Molecules were tested in LB medium, with compound concentrations ranging from 128 pg / ml to 2 pg / ml in two-fold serial dilutions. Lighter shades indicate lower MIC values. Experiments were performed in biological duplicate.
[0034] FIGURE 8 shows synthecin can suppress an S. aureus infection in a mouse wound model in exemplary embodiment of the disclosure, (a) Chemical Release Evaluation on Agar Media (C. R. E. A. M.) assay of vehicle (10% DMSO), MLP-01 (synthecin), RL-CP 08, RL-CP 03, RL-CP 35, RL-CP 12, RL-CP 33, and RL-CP_20 on LB agar. Each molecule was tested at 2% w / v in Glaxal Base. Larger zones of inhibition are expected to correspond with greater efficacy when used as a topical treatment, (b) Killing of S. aureus USA300 in nutrient-replete conditions (LB) at 4x, 2x, lx,!4x, * x, and Ox MIC for 8 h. Experiments were conducted in biological duplicate. Individual replicates with means connected are plotted, (c) Overview of the infection model used to assess the in vivo antibacterial efficacy of synthecin against 5. aureus USA300. In a wound infection model, mice were infected with S. aureus USA300 (-2.15 x 107CFU). At 1-hour post-infection (hpi), mice were treated with vehicle (10% DMSO) (n = 5) or 2% synthecin (n = 5). Mice received treatments 4-, 8-, 12-, and 20-hpi. (d) Bacterial load of 5. aureus USA300 from wound tissue prior to treatment (n = 5) and of vehicle (n = 5) or 2% synthecin-treated (n = 5) mice 24-hpi. Pre-treatment represents the bacterial load at the time of initial treatment (1 -hpi). Black lines represent the geometric mean of the bacterial load for each group. NS means non statistically significant. ***P using unpaired two-sided t-test with Welch’s correction (Pre-treatment vs 2% synthecin, P = 0.1774; Pre-treatment vs Vehicle, P = 0.0005; Vehicle vs 2% synthecin, P = 0.0004). (e) Representative images of the dorsal surface of mice pre-treatment (1-hpi), after 12 h of treatment with vehicle or synthecin, and after 24 h of treatment with vehicle or synthecin. Note the inflammation seen in the vehicle control (arrows) that is absent in the synthecin treated mouse. Note the plaques, which is molecule precipitate, seen in the synthecin treated-mouse (arrow) that is absent in the vehicle-treated mouse.
[0035] FIGURE 9 shows the dorsal surface of mice with an 5. aureus wound infection model in exemplary embodiments of the disclosure. Representative images of the dorsal surface of mice after 4 h, 8 h, and 20 h of treatment with vehicle or synthecin. Note the marked inflammation on the vehicle treated animals.
[0036] FIGURE 10 shows building block 3753222, which SyntheMol-RL used multiple times when constructing hit compounds in exemplary embodiment of the disclosure. Notably, all four compounds shown were validated to have an MIC <8μg / ml in S. aureus RN4220. SyntheMol-RL seemingly employs a strategy of finding a particularly promising building block, then to satisfy the diversity objective it expands on that building block with a diverse set of “secondary” building blocks.DETAILED DESCRIPTIONI, Definitions
[0037] Unless otherwise indicated, the definitions and embodiments described in this and other sections are intended to be applicable to all embodiments and aspects of the present disclosure herein described for which they are suitable as would be understood by a person skilled in the art. It is also to be understood that the terminology used herein is for the purpose of describing particular aspects only and is not intended to be limiting.
[0038] In understanding the scope of the present disclosure, the term “comprising” and its derivatives, as used herein, are intended to be open ended terms that specify the presence of the stated features, elements, components, groups, integers, and / or steps, but do not exclude the presence of other unstated features, elements, components, groups, integers and / or steps. The foregoing also applies to words having similar meanings such as the terms, “including”, “having” and their derivatives. The term “consisting” and its derivatives, as used herein, are intended to be closed terms that specify the presence of the stated features, elements, components, groups, integers, and / or steps, but exclude the presence of other unstated features, elements, components, groups, integers and / or steps. The term “consisting essentially of’, as used herein, is intended to specify the presence of the stated features, elements, components, groups, integers, and / or steps as well as those that do not materially affect the basic and novel characteristic(s) of features, elements, components, groups, integers, and / or steps.
[0039] Terms of degree such as “substantially”, “about” and “approximately” as used herein mean a reasonable amount of deviation of the modified term such that the end result is not significantly changed. These terms of degree should be construed as including a deviation of at least ±5% of the modified term if this deviation would not negate the meaning of the word it modifies. In addition, all ranges given herein include the end of the ranges and also any intermediate range points, whether explicitly stated or not.
[0040] As used in this disclosure, the singular forms “a”, “an” and “the” include plural references unless the content clearly dictates otherwise.
[0041] In embodiments comprising an “additional” or “second” component, the second component as used herein is chemically different from the other components or first component. A “third” component is different from the other, first, and second components, and further enumerated or “additional” components are similarly different.
[0042] The term “and / or” as used herein means that the listed items are present, or used, individually or in combination. In effect, this term means that “at least one of’ or “one or more” of the listed items is used or present.
[0043] The abbreviation, “e.g.” is derived from the Latin exempli gratia and is used herein to indicate a non-limiting example. Thus, the abbreviation “e.g.” is synonymous with the term “for example.” The word “or” is intended to include “and” unless the context clearly indicates otherwise.
[0044] The term “suitable” as used herein means that the selection of the particular compound or conditions would depend on the specific synthetic manipulation to be performed or composition to be prepared, the identity of the molecule(s) to be transformed and / or the specific use for the compound, but the selection would be well within the skill of a person trained in the art.
[0045] The term “cell” as used herein refers to a single cell or a plurality of cells and includes a cell either in a cell culture or in a subject.
[0046] The term “treating” or “treatment” and the like as used herein, and as is well understood in the art, refers to an approach for obtaining beneficial or desired results, including clinical results. Beneficial or desired clinical results include, but are not limited to alleviation or amelioration of one or more symptoms or conditions, arresting development of disease, diminishment of extent of disease, stabilized (i.e. not worsening) state of disease, preventing spread of disease, delay or slowing of disease progression, amelioration or palliation of the disease state, including regression of the disease, diminishment of the reoccurrence of disease, and remission (whether partial or total), whether detectable or undetectable. “Treating” and “treatment” may also refer to prolonging survival as compared to expected survival if not receiving treatment.“Treating” and “treatment” as used herein also include prophylactic treatment. The effect may be prophylactic in terms of completely or partially preventing a disease or symptom thereof and / or may be therapeutic in terms of affecting a partial or complete cure for a disease and / or symptoms of the disease. For example, a subject with an early disease state can be treated to prevent progression, or alternatively a subject in remission can be treated to prevent recurrence. Prophylactic treatment includes preventing the disease or a symptom of a disease from occurring in a subject which may be predisposed to the disease but has not yet been diagnosed as having it (e.g., including diseases that may be associated with or caused by a primary disease).
[0047] Treating may refer to any indicia of success in the treatment or amelioration or prevention of an infection, including any objective or subjective parameter such as abatement; remission; diminishing of symptoms; or making the disease condition more tolerable to the patient; slowing in the rate of degeneration or decline; or making the final point of degeneration less debilitating. The treatment or amelioration of symptoms is based on one or more objective or subjective parameters; including the results of an examination by a physician. Accordingly, the term "treating" includes the administration of the methods of the present disclosure to prevent, delay, alleviate, arrest or inhibit development of the symptoms or conditions associated with diseases.
[0048] The term “subject” as used herein includes all members of the animal kingdom including mammals such as a mouse, a rat, a dog and a human. Thus, the methods and uses of the present disclosure are applicable to both human therapy and veterinary applications.
[0049] The term “pharmaceutically acceptable” as used herein means compatible with the treatment of subjects, for example humans.
[0050] The term “pharmaceutically acceptable carrier” as used herein means a non-toxic solvent, dispersant, excipient, adjuvant or other material which is mixed with the active ingredient in order to permit the formation of a pharmaceutical composition, i.e., a dosage form capable of administration to a subject.
[0051] The term “pharmaceutically acceptable salt” as used herein means either an acid addition salt or a base addition salt which is suitable for, or compatible with the treatment of subjects.
[0052] The term “solvate” as used herein means a compound, or a salt and / or prodrug of a compound, wherein molecules of a suitable solvent are incorporated in the crystal lattice. A suitable solvent is physiologically tolerable at the dosage administered.
[0053] The term “prodrug” as used herein means a compound, or salt and / or solvate of a compound, that, after administration, is converted into an active drug.
[0054] The term “administering” or “administration” and the like as used herein refers to the placement of a compound, one of more compounds or a pharmaceutical composition thereof, as disclosed herein into a cell, either in cell culture or in a subject, by a method or route which results in at least partial delivery to a desired site. The compounds and compositions disclosed herein can be administered by any appropriate route which results in an effective treatment in the subject. Possible routes of administration of the compounds and pharmaceutical compositions disclosed herein include, but are not limited to, topical, intravenous, intraperitoneal, intramuscular, subcutaneous, transdermal, oral, buccal, sublingual, intranasal, or rectal routes of administration, or a combination thereof.
[0055] As used herein, the term “effective amount” or “therapeutically effective amount” means an amount of a compound, or one or more compounds, of the disclosure that is effective, at dosages and for periods of time necessary to achieve the desired result.II. Compounds. Compositions. Methods and Uses of the Disclosure
[0056] Herein, SyntheMol-RL, a reinforcement learning-based generative model that can rapidly design synthetically accessible small molecule drug candidates from a massive chemical space of 46 billion compounds was used in the generation of a novel antibiotic agent. SyntheMol-RL generated a candidate antibiotic against 5. aureus by optimizing for both antibacterial activity and aqueous solubility. Results found that SyntheMol-RL generated molecules with improved predicted properties compared to both the previous MCTS version of SyntheMol as well as an Al-based virtual screening baseline. 79 SyntheMol-RL compounds were synthesized that wereunique relative to the training dataset and found that 13 showed potent in vitro activity, of which seven were identified for further study. Furthermore, one hit compound, synthecin, demonstrated efficacy in a murine wound infection model of methicillin-resistant S. aureus (MRSA).
[0057] In one aspect of the invention, a compound of formula (I) is provided:R2-C(O)-NR1R7(Formula (I))or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof, wherein:R1is H or C1-C6alkyl;R2is benzyl substituted with up to four R3, wherein each R3is independently selected from hydroxyl or halogen, or R2is CH2CR4R5R6, wherein R4is H or hydroxyl, R5is hydroxy, halogen or CF3 and R6is benzyl;R7is C1-C6alkyl, optionally substituted with one or more substituents independently selected from hydroxyl, halogen, benzyl or 5- or 6-membered heterocycle comprising S, N or O, wherein the benzyl or the 5- or 6-membered heterocycle is optionally substituted with one or more substituents independently selected from C1-C6alkyl, hydroxyl, halogen or CF3, or R7is benzyl or thiol, optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
[0058] The term “halogen” is used herein to refer to fluorine (F), chlorine (Cl), bromine (Br) and iodine (I).
[0059] The term “Ci-Ce alkyl” refers to alkyl groups having 1-6 carbon atoms including methyl, ethyl, propyl, butyl, pentyl and hexyl groups. Branched alkyl groups are also included, for example, isopropyl, isobutyl, sec-butyl, tert-butyl, isopentyl, neopentyl, tert-pentyl, isohexyl, 1,3-dimethylbutyl, 1,1 -dimethylbutyl and 1,2-dimethylbutyl.
[0060] The term “5- or 6-membered heterocycle comprising S, N or O” refers to a ring-shaped compound with 5 or 6 atoms in the ring, at least one of which is independently S, N or O.
[0061] Also provided in the present disclosure is a compound for inhibiting growth of a Staphylococcus aureus bacterium:or a pharmaceutically acceptable salt, solvate and / or prodrug thereof.
[0062] The compounds of the present disclosure may be suitably formulated in a conventional manner into compositions using one or more carriers. Accordingly, the present disclosure includes a composition comprising one or more compounds of the disclosure and a carrier. The compounds of the disclosure may be suitably formulated into pharmaceutical compositions for administration to subjects in a biologically compatible form suitable for administration in vivo. Accordingly, the present disclosure also includes a composition comprising one or more compounds disclosed herein, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, and a pharmaceutically acceptable carrier or vehicle. Conventional procedures and ingredients for the selection and preparation of suitable compositions are described, for example, in Remington’s Pharmaceutical Sciences (2000 - 20th edition) and in The United States Pharmacopeia: The National Formulary (USP 24 NF19) published in 1999.
[0063] In some embodiments, suitable pharmaceutically acceptable carriers include, but are not limited to, inert solid fillers or diluents and sterile aqueous or organic solutions. Pharmaceutically acceptable carriers are well known to those skilled in the art and include, but are not limited to, aqueous and non-aqueous solutions. Pharmaceutically acceptable carriers can be aqueous or non- aqueous solutions, suspensions and emulsions. Examples of non-aqueous solvents suitable for use in the present disclosure include, but are not limited to, propylene glycol, polyethylene glycol, vegetable oils such as olive oil, and injectable organic esters such as ethyl oleate. Aqueous carriers suitable for use in the present disclosure include, but are not limited to, water, ethanol, alcoholic / aqueous solutions, glycerol, emulsions or suspensions, including saline and buffered media.
[0064] Compounds of the disclosure may also be coupled with soluble polymers as targetable drug carriers. Such polymers can include polyvinylpyrrolidone, pyran copolymer, polyhydroxypropylmethacrylamide-phenol, polyhydroxy-ethylaspartamide-phenol, or polyethyleneoxide-polylysine substituted with palmitoyl residues.Furthermore, compounds of the disclosure may be coupled to a class of biodegradable polymers useful in achieving controlled release of a drug, for example, polylactic acid, polyglycolic acid, copolymers of polylactic and polyglycolic acid, polyepsilon caprolactone, polyhydroxy butyric acid, polyorthoesters, polyacetals, poly dihydropyrans, poly cyanoacrylates and crosslinked or amphipathic block copolymers of hydrogels.
[0065] In some embodiments, depending on the mode of administration, the pharmaceutical composition will comprise from about 0.05 wt% to about 99 wt% or about 0.10 wt% to about 70 wt%, of the active ingredient (one or more compounds of the disclosure), and from about 1 wt% to about 99.95 wt% or about 30 wt% to about 99.90 wt% of one or more pharmaceutically acceptable carriers, all percentages by weight being based on the total composition.
[0066] Salts of the compounds of the disclosure are generally formed by dissolving the neutral compound in an inert organic solvent and adding either the desired acid or base and isolating the resulting salt by either filtration or other known means.
[0067] The formation of solvates of the compounds of the disclosure will vary depending on the compound and the solvate. In general, solvates are formed by dissolving the compound in the appropriate solvent and isolating the solvate by cooling or using an antisolvent. The solvate is typically dried or azeotroped under ambient conditions. The selection of suitable conditions to form a particular solvate can be made by a person skilled in the art. Examples of suitable solvents are ethanol, water and the like. When water is the solvent, the molecule is referred to as a “hydrate”.
[0068] Prodrugs of the compounds of the present disclosure may be, for example, conventional esters formed with available hydroxy, thiol, amino or carboxyl groups. For example, available hydroxy or amino groups may be acylated using an activated acid in the presence of a base, and optionally, in inert solvent (e.g. an acid chloride in pyridine). Some common esters which have been utilized as prodrugs are phenyl esters, aliphatic (C1-C24) esters, acyloxy methyl esters, carbamates and amino acid esters.
[0069] Throughout the processes described herein it is to be understood that, where appropriate, suitable protecting groups will be added to, and subsequently removed from, the various reactants and intermediates in a manner that will be readily understood by one skilled in the art. Conventional procedures for using such protecting groups aswell as examples of suitable protecting groups are described, for example, in “Protective Groups in Organic Synthesis”, T. W. Green, P. G. M. Wuts, Wiley-Interscience, New York, (1999). It is also to be understood that a transformation of a group or substituent into another group or substituent by chemical manipulation can be conducted on any intermediate or final product on the synthetic path toward the final product, in which the possible type of transformation is limited only by inherent incompatibility of other functionalities carried by the molecule at that stage to the conditions or reagents employed in the transformation. Such inherent incompatibilities, and ways to circumvent them by carrying out appropriate transformations and synthetic steps in a suitable order, will be readily understood to one skilled in the art. Examples of transformations are given herein, and it is to be understood that the described transformations are not limited only to the generic groups or substituents for which the transformations are exemplified. References and descriptions of other suitable transformations are given in “Comprehensive Organic Transformations - A Guide to Functional Group Preparations” R. C. Larock, VHC Publishers, Inc. (1989). References and descriptions of other suitable reactions are described in textbooks of organic chemistry, for example, “Advanced Organic Chemistry”, March, 4th ed. McGraw Hill (1992) or, “Organic Synthesis”, Smith, McGraw Hill, (1994). Techniques for purification of intermediates and final products include, for example, straight and reversed phase chromatography on column or rotating plate, recrystallisation, distillation and liquid-liquid or solid-liquid extraction, which will be readily understood by one skilled in the art.
[0070] In another aspect of the disclosure, provided is a method for inhibiting growth of a Staphylococcus aureus bacterium, comprising contacting the bacterium with an effective amount of a compound disclosed herein or a pharmaceutically acceptable salt, solvate and / or prodrug thereof.
[0071] In a further aspect of the disclosure, provided is a method of treating a Staphylococcus aureus bacterial infection comprising administering an effective amount of a compound disclosed herein or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, to a subject in need thereof.
[0072] In some embodiments, the subject is a mammal. In some embodiments, the subject is a human.
[0073] Effective amounts may vary according to factors such as the disease state, age, sex and / or weight of the subject. The amount of a given compound that will correspond to such an amount will vary depending upon various factors, such as the given drug or compound, the pharmaceutical formulation, the route of administration, the type of condition, disease or disorder, the identity of the subject being treated, and the like, but can nevertheless be routinely determined by one skilled in the art. The effective amount is one that following treatment therewith manifests as an improvement in or reduction of any disease symptom.
[0074] In vivo administration of a compound and / or composition of the disclosure can be accomplished by any suitable method and technique presently or prospectively known to those skilled in the art. For example, a compound of the disclosure may be formulated in a physiologically or pharmaceutically acceptable form and administered by any suitable route known in the art including, for example, oral and parenteral routes of administration. As used herein, the term parenteral may include intravenous (such as by injection), intra-arterial, intraperitoneal, subcutaneous, intramuscular, intrathecal transepithelial, nasal, intrapulmonary (for example, by use of an aerosol), rectal and topical (including the use of a patch or other transdermal delivery device) modes of administration. Administration of a compound of the disclosure may comprise a single administration, or a plurality of administrations at continuous or distinct intervals over a selected period of time.
[0075] The pharmaceutical forms suitable for injectable use include sterile aqueous solutions or dispersions and sterile powders for the extemporaneous preparation of sterile injectable solutions or dispersions. In all cases, the form must be sterile and must be fluid to the extent that easy syringability exists.
[0076] A compound of the disclosure may be orally administered, for example, with an inert diluent or with an assimilable edible carrier, or it may be enclosed in hard- or soft-shell gelatin capsules, or it may be compressed into tablets, or it may be incorporated directly with the food of the diet. For oral therapeutic administration, the compound may be incorporated with excipient and used in the form of ingestible tablets, buccal tablets, troches, capsules, caplets, pellets, granules, lozenges, chewing gum, powders, syrups, elixirs, wafers, aqueous solutions and suspensions, and the like. In the case of tablets, carriers that are used include lactose, com starch, sodium citrate and saltsof phosphoric acid. Pharmaceutically acceptable excipients include binding agents (e.g., pregelatinized maize starch, polyvinylpyrrolidone or hydroxypropyl methylcellulose); fillers (e.g., lactose, microcrystalline cellulose or calcium phosphate); lubricants (e.g., magnesium stearate, talc or silica); disintegrants (e.g., potato starch or sodium starch glycolate); or wetting agents (e.g., sodium lauryl sulphate). The tablets may be coated by methods well known in the art. In the case of tablets, capsules, caplets, pellets or granules for oral administration, pH sensitive enteric coatings, such as Eudragits™ designed to control the release of active ingredients are optionally used. Oral dosage forms also include modified release, for example immediate release and timed-release, formulations. Examples of modified-release formulations include, for example, sustained-release (SR), extended-release (ER, XR, or XL), time-release or timed-release, controlled-release (CR), or continuous-release (CR or Contin), employed, for example, in the form of a coated tablet, an osmotic delivery device, a coated capsule, a microencapsulated microsphere, an agglomerated particle, e.g., as of molecular sieving type particles, or, a fine hollow permeable fiber bundle, or chopped hollow permeable fibers, agglomerated or held in a fibrous packet. Timed-release compositions can be formulated, e.g. liposomes or those wherein the active compound is protected with differentially degradable coatings, such as by microencapsulation, multiple coatings, etc. Liposome delivery systems include, for example, small unilamellar vesicles, large unilamellar vesicles and multilamellar vesicles. Liposomes can be formed from a variety of phospholipids, such as cholesterol, stearylamine or phosphatidylcholines. For oral administration in a capsule form, useful carriers or diluents include lactose and dried com starch.
[0077] Liquid preparations for oral administration may take the form of, for example, solutions, syrups or suspensions, or they are suitably presented as a dry product for constitution with water or other suitable vehicle before use. When aqueous suspensions and / or emulsions are administered orally, the compound of the disclosure is suitably suspended or dissolved in an oily phase that is combined with emulsifying and / or suspending agents. If desired, certain sweetening and / or flavoring and / or coloring agents may be added. Such liquid preparations for oral administration may be prepared by conventional means with pharmaceutically acceptable additives such as suspending agents (e.g., sorbitol syrup, methyl cellulose or hydrogenated edible fats); emulsifying agents (e.g., lecithin or acacia); non-aqueous vehicles (e.g., almond oil, oily esters or ethylalcohol); and preservatives (e.g., methyl or propyl p-hydroxybenzoates or sorbic acid). Useful diluents include lactose and high molecular weight polyethylene glycols.
[0078] It is also possible to freeze-dry the compounds of the disclosure and use the lyophilizates obtained, for example, for the preparation of products for injection.
[0079] The compounds of the disclosure may be formulated for parenteral administration by injection, including using conventional catheterization techniques or infusion. Formulations for injection may be presented in unit dosage form, e.g., in ampoules or in multi-dose containers, with an added preservative. The compositions may take such forms as sterile suspensions, solutions or emulsions in oily or aqueous vehicles, and may contain formulating agents such as suspending, stabilizing and / or dispersing agents. In all cases, the form must be sterile and must be fluid to the extent that easy syringability exists. Alternatively, compounds of the disclosure are suitably in a sterile powder form for reconstitution with a suitable vehicle, e.g., sterile pyrogen-free water, before use.
[0080] A compound of the disclosure may be administered as a topical composition, such as, but not limited to, a solution, gel, cream, lotion, liquid suspension, aerosol, nebulized spray, ointment, drops or patch. In some embodiments, a compound and / or composition of the disclosure is administered topically. In some embodiments, a compound and / or composition of the disclosure is administered as a topical cream.
[0081] Compositions for nasal administration may conveniently be formulated as aerosols, drops, gels and powders. For intranasal administration or administration by inhalation, the compounds of the disclosure are conveniently delivered in the form of a solution, dry powder formulation or suspension from a pump spray container that is squeezed or pumped by the patient or as an aerosol spray presentation from a pressurized container or a nebulizer. Aerosol formulations typically comprise a solution or fine suspension of the active substance in a physiologically acceptable aqueous or nonaqueous solvent and are usually presented in single or multidose quantities in sterile form in a sealed container, which can take the form of a cartridge or refill for use with an atomising device. Alternatively, the sealed container may be a unitary dispensing device such as a single dose nasal inhaler or an aerosol dispenser fitted with a metering valve which is intended for disposal after use. Where the dosage form comprises an aerosol dispenser, it will contain a propellant which can be a compressed gas such as compressedair or an organic propellant such as fluorochlorohydrocarbon. Suitable propellants include but are not limited to dichlorodifluoromethane, trichlorofluoromethane, dichlorotetrafluoroethane, heptafluoroalkanes, carbon dioxide or another suitable gas. In the case of a pressurized aerosol, the dosage unit is suitably determined by providing a valve to deliver a metered amount. The pressurized container or nebulizer may contain a solution or suspension of the active compound. Capsules and cartridges (made, for example, from gelatin) for use in an inhaler or insufflator may be formulated containing a powder mix of a compound of the disclosure and a suitable powder base such as lactose or starch. The aerosol dosage forms can also take the form of a pump-atomizer.
[0082] In another aspect of the disclosure, provided is the use of a compound disclosed herein to inhibit growth of a Staphylococcus aureus bacterium.
[0083] In a further aspect of the disclosure, provided is the use of a compound disclosed herein or a pharmaceutically acceptable salt pharmaceutically acceptable salt, solvate and / or prodrug thereof, to treat or prevent a Staphylococcus aureus bacterial infection.
[0084] It will be understood that any component defined herein as being included may be explicitly excluded by way of proviso or negative limitation, such as any specific compounds or method steps, whether implicitly or explicitly defined herein.EXAMPLES
[0085] The following non-limiting examples are illustrative of the present disclosure:
[0086] Results
[0087] Property Prediction Model Development. SyntheMol-RL is guided by molecular property prediction models that evaluate the properties of generated molecules and thereby provide feedback to improve the generative model’s ability to design molecules with desired properties. Therefore, the first step toward building SyntheMol-RL is to train molecular property prediction models on the properties that this study requires of the molecules. Here, the focus is on two key properties: (1) antibacterial activity, which is defined here as the ability of a compound to effectively inhibit the growth of 5. aureus in vitro, and (2) aqueous solubility, which is necessary for thecompound to be efficiently delivered and tested in vivo. For antibacterial activity, the study performed an assay to build a training set, as described below. For aqueous solubility, the AqSolDB dataset of 9,982 molecules with log solubility values was used as curated by the Therapeutics Data Commons
[0018] , While there are many drug-like properties that could be prioritized at this stage, aqueous solubility was chosen to help circumvent issues of low solubility that were previously observed when SyntheMol-MCTS was applied to antibiotic discovery
[0017] ,
[0088] To create the antibacterial activity training set, 10,716 compounds were physically screened against S. aureus RN4220 at 50 / zM in LB medium (Fig. 2a)
[0019] , Experiments were conducted in biological duplicate, with end-point growth being measured at 600 nm optical density (ODeoo) after 16 hours of incubation at 37°C (Fig.3a). The mean / z was computed and standard deviation cr ODeoo value across the library and used / z — 2 a as a threshold for binarizing the ODeoo values into active (low ODeoo) and inactive (high ODeoo) molecules. After removing duplicate compounds based on SMILES (see Methods), a set of 10,658 unique molecules was obtained with 1,137 active compounds (10.7%) and 9,521 (89.3%) inactive compounds. A t-SNE visualization of the training dataset compounds and 1,007 molecules from ChEMBL20 with known antibacterial activity shows that the active compounds cover both known and new antibiotic chemical space (Fig. 2b).
[0089] After acquiring these two training datasets, two different property prediction model architectures were leveraged for each dataset: (1) Chemprop-RDKit and (2) MLP-RDKit (Fig. 2c). The Chemprop-RDKit architecture consists of the graph neural network (GNN) model ChemproplO augmented with 200 molecular features computed by the cheminformatics package RDKit21. Chemprop-RDKit combines the GNN representation of the molecule with the 200 RDKit features using a multilayer perceptron (MLP). The MLP-RDKit architecture is nearly identical to the MLP contained within the Chemprop-RDKit model, but MLP-RDKit only uses the 200 RDKit features as input and does not have a GNN component. Both model architectures were chosen to experiment with to compare the importance of accuracy and speed; Chemprop-RDKit learns a more comprehensive molecular representation and is thus potentially more accurate, while MLP-RDKit is potentially less accurate but is faster without the GNN component.
[0090] Chemprop-RDKit and MLP-RDKit were trained on both the S. aureus antibacterial activity and aqueous solubility datasets. For each of the four model-dataset combinations, the model was trained using 10-fold cross-validation with splits containing 80% train, 10% validation, and 10% test data. Each of the four model-dataset combinations trained in less than 80 minutes on a machine with 8 CPUs and 1 GPU. The Chemprop-RDKit models slightly outperformed the MLP-RDKit models, although both showed high performance (Fig. 2d, Table 1), including on more challenging splits based on molecular scaffold (Fig. 3b-c).
[0091] Table 1. Property Prediction Model Performance
[0092] Generative Model Development. SyntheMol
[0017] is a generative model that designs easily synthesizable molecules with desirable properties, such as inhibition of bacterial growth in vitro. SyntheMol ensures straightforward synthesis of generated molecules by constraining the generative process to a combinatorial chemical space, where it builds molecules using small molecular building blocks that are readily purchasable and chemical reactions that are well-validated. To generate a molecule, SyntheMol selects two or three building blocks and combines them using a chemical reaction to form a molecule, which is then scored by a property prediction model. This score provides feedback to SyntheMol to guide its generative process toward higher scoring molecules, resulting in a set of generated molecules that are easy to synthesize and likely to possess the desired properties. SyntheMol was previously applied to generate antibiotic candidates against the gram-negative bacterium^, baumannii, and 58 generated compounds were successfully synthesized and six highly potent hits were identified, demonstrating the ability of SyntheMol to design effective antibacterial compounds.
[0093] In its original form, SyntheMol employed a Monte Carlo tree search (MCTS) algorithm [22,23] to guide its selection of building blocks and reactions to design molecules. SyntheMol -MCTS treats the generative process as a tree search, where nodes in the tree are combinations of one or more building blocks. SyntheMol-MCTS computes a value for each node based on statistics about the node that measure both exploration (how many times the node has been selected) and exploitation (the average property score of molecules built using that node). During generation, SyntheMol-MCTS sequentially explores nodes with one, two, or three building blocks, at each step choosing the node with the highest value. When two or three building blocks in a node are compatible with one of the specified chemical reactions, a new node is created with the product of the chemical reaction applied to those building blocks. Once a node with a product molecule is selected, SyntheMol-MCTS updates its statistics for each node involved in generating that molecule based on the property prediction score of the molecule, concluding one rollout. While MCTS is a simple and effective algorithm, it suffers from a key inefficiency by treating each node independently rather than leveraging the chemical similarity between molecular building blocks in different nodes to determine the nodes’ values.
[0094] Here, the new version of SyntheMol overcomes this inefficiency by replacing MCTS with a reinforcement learning (RL) algorithm [24-27] (Fig. 1, see Methods). SyntheMol-RL uses a deep learning model as a value function that determines the value of a node based on the chemical structures of the building block(s) in that node. Specifically, the RL value function learns to predict the expected property score of molecules constructed using the building block(s) in a given node. During each step of generation, SyntheMol-RL computes the value of nodes containing one, two, or three building blocks. SyntheMol-RL then samples a node proportional to the node’s value, which encourages exploration (by not always selecting the highest value node) and exploitation (by sampling higher value nodes more frequently). To simultaneously optimize for multiple properties (in this case, antibacterial activity and aqueous solubility), the value of a node is computed as a weighted combination of multiple RL models - one for each property. SyntheMol-RL automatically adjusts the weights dynamically over time to explore the Pareto frontier of the multiple properties and identify where the greatest density of ‘hit’ molecules exist (those that pass user-defined thresholds for all properties). The RL models are trained during the generative processby taking nodes and molecules from previous iterations and training the RL models to predict the property scores of each molecule from the nodes that created it. Thus, as the RL model generates more molecules and creates more training data, the accuracy of the RL models and the quality of the generated molecules improve.
[0095] For antibiotic generation with SyntheMol-RL, the trained Chemprop-RDKit models were used for antibacterial activity and aqueous solubility as fixed property predictors to score designed molecules since those models were the most accurate. For the RL value function, either a Chemprop-RDKit model architecture or an MLP-RDKit model architecture was used to evaluate tradeoffs between accuracy and speed. In both cases, the RL value models were initialized with the parameters from the trained property prediction models of this study, but then they were further trained during generation on the RL task of predicting the value of nodes (i.e., combinations of building blocks).
[0096] SyntheMol-RL was provided with molecular building blocks and chemical reactions from two easily synthesizable combinatorial chemical spaces: the Enamine REAL Space and the WuXi GalaXi (Fig. 2e, Fig. 3d-e). The Enamine REAL Space
[0028] (2022 version) contains 31 billion molecules that can be produced using -139,000 molecular building blocks and 169 chemical reactions. 13 of the most common reactions were used, which can build 30 billion molecules. The WuXi GalaXi29 (2022 version) contains 16 billion molecules that can be produced with -15,000 building blocks and 36 chemical reactions. Since these two chemical spaces are relatively distinct (Fig.2b), this gives SyntheMol-RL the capability to generate a structurally diverse set of easily synthesizable molecules within a chemical space of 46 billion molecules.
[0097] Generative Model Application. SyntheMol-RL was used to generate antibiotic candidates against X aureus. Two versions of SyntheMol-RL was used for this purpose: (1) RL-Chemprop, which uses a Chemprop-RDKit model architecture as the value function, and (2) RL-MLP, which uses an MLP-RDKit model architecture as the value function. In both cases, the generative model were run for 10,000 rollouts, resulting in the generation of roughly 10,000 unique molecules. To demonstrate the improved efficacy of RL over MCTS, MCTS was also run for 10,000 rollouts. Furthermore, to illustrate the benefit of using generative modeling over virtual screening approaches, the Chemprop-RDKit property prediction models were applied in a virtual screening
[0011] manner (VS-Chemprop) to evaluate the antibacterial activity and aqueous solubility of 21 million randomly sampled molecules (14 million REAL and 7 million GalaXi to match the relative sizes of the chemical spaces). 21 million was chosen since it took VS-Chemprop about 7 days to evaluate that many compounds, which roughly matched the time of the slower generative model, RL-Chemprop, on the slower of the GPUs and CPUs. This gives VS-Chemprop the best possible chance to discover promising compounds compared to the generative models, normalized for time.
[0098] The compounds generated by RL-Chemprop were first evaluated, RL-MLP, and MCTS, as well as those screened by VS-Chemprop, based on their predicted antibacterial activity and aqueous solubility. Defining a hit as any molecule with an antibacterial prediction score > 0.5 and a log solubility > -4 (which constitutes high solubility
[0030] ), RL-Chemprop generated 11.6% hits (1,273 of 10,983) and RL-MLP generated 5.3% hits (493 of 9,228) compared to 3.0% hits for MCTS (347 of 11,630) and 0.006% hits for VS-Chemprop (1,290 of 21,000,000) (Fig. 4a). The distributions of antibacterial prediction scores (Fig. 4b) and log solubility prediction scores (Fig. 4c) show that all methods can similarly identify soluble molecules, but the generative models - and particularly the RL models - excel at designing molecules with high predicted antibacterial activity. This illustrates the benefit of using RL for molecule design. Of note, a set of ablation experiments were performed that elucidate the effect of individual components of the RL model on generation performance (see Methods and Fig. 5).
[0099] To obtain a set of promising compounds for synthesis and testing in vitro, a set of four filters to prioritize compounds were established. First, as above, hits were identified and only kept molecules with an antibacterial prediction score > 0.5 and a log solubility > -4. Second, novel hits were selected by computing the Tversky similarity
[0031] between each hit molecule and all known antibiotics in the training set and the ChEMBL antibiotics set, only keeping molecules with a maximum Tversky similarity of < 0.6. Third, a diverse set of hits were selected by computing the largest set of novel hits such that no two compounds in that set had a Tanimoto similarity
[0032] greater than 0.6 (see Methods). Fourth, to limit the diverse, novel hits to a practical number of compounds to test in the wet lab, the remaining compounds were ranked by antibacterial prediction score and selected the top 150 compounds from each method. This provided 600 molecules: 150 each from RL-Chemprop, RL-MLP, MCTS, and VS-Chemprop. Finally,to provide an unbiased estimate of the baseline antibacterial activity and solubility of compounds in the REAL and GalaXi chemical spaces, 150 random compounds were selected (100 REAL, 50 GalaXi), for a total of 750 molecules for testing (Fig. 4d).
[0100] Among these 750 selected compounds, the RL-Chemprop and RL-MLP compounds had higher average antibacterial prediction scores (median of 0.758 and 0.712, respectively) compared to MCTS, VS-Chemprop, and random (median of 0.605, 0.694, and 0.007, respectively) (Fig. 4e) while all of the compounds had similar predicted log solubilities (Fig. 4f). This computational evidence shows that SyntheMol-RL is the most adept at generating likely antibiotic candidates.
[0101] The availability of these 750 compounds were requested (747 unique compounds due to three duplicates between RL-MLP and MCTS) from Enamine (659 unique compounds) and WuXi (88 unique compounds), depending on the chemical space each compound belongs to. Among those compounds, 435 / 659 (66%) of the Enamine compounds and 49 / 88 (56%) of the WuXi compounds were available for potential synthesis. Notably, RL-Chemprop, RL-MLP, and MCTS demonstrated a strong preference for Enamine compounds and selected very few WuXi compounds (3, 1, and 0, respectively), of which none were available for synthesis, so WuXi compounds were only available for VS-Chemprop and random. This result may be because the Enamine REAL Space has roughly lOx the number of building blocks as the WuXi GalaXi, despite having only 3x as many total molecules. Since SyntheMol makes building block-level decisions, it has disproportionally more opportunities to select Enamine building blocks.
[0102] To further narrow down these available compounds to the most promising set of potential antibiotic candidates, the available compounds were ranked by predicted clinical toxicity according to ADMET-AI33,34, a multi-task Chemprop-RDKit model for ADMET property prediction. 50 compounds were then selected with lowest predicted clinical toxicity for each model type (RL-Chemprop, RL-MLP, MCTS, and VS-Chemprop). For the random baseline, 50 available compounds were randomly selected (33 REAL, 17 GalaXi) from the random set.
[0103] Altogether, 250 compounds were submitted (50 compounds for each of five settings, 248 unique compounds total with two duplicates between RL-MLP and MCTS) for synthesis (Fig. 4g). Enamine successfully synthesized 177 / 224 (79%) unique compounds, while WuXi successfully synthesized 17 / 24 (71%) unique compounds.Altogether, 194 unique compounds were obtained for experimental validation (38 RL-Chemprop, 41 RL-MLP, 38 MCTS, 37 VS-Chemprop, and 44 random).
[0104] A final computational evaluation was conducted prior to in vitro testing, which benchmarked SyntheMol-RL against GFlowNet [35,36], a state-of-the-art-generative model for drug design. While the GFlowNet architecture generated molecules with higher antibacterial activity and aqueous solubility scores than the SyntheMol-RL molecules, the GFlowNet compounds are bulky with complex multi-ring structures and would likely suffer from poor whole cell activity and synthesizability (Fig. 6). Additionally, medicinal chemists at Enamine and WuXi offered limited synthesis of these molecules with >30x higher associated costs (see Extended Discussion). This comparison illustrates the advantages of SyntheMol-RL for molecule generation.
[0105] In Vitro Validation of AI-Generated Molecules. Standard growth inhibition assays were used to determine all synthesized molecules' antibacterial activity against S. aureus RN4220, defined using the minimum inhibitory concentration (MIC). Remarkably, 11 / 38 (29%) molecules synthesized from RL-Chemprop and 2 / 41 (5%) from RL-MLP displayed the ability to completely inhibit bacterial growth at MIC ≤ 8 pg / ml (Fig. 7a). For comparison, our control set of 44 randomly selected molecules synthesized from both the REAL space and GalaXi were also tested. A complete absence of any activity up to our limit of detection of 128 pg / ml was observed. This may be evidence that the ultra-large chemical databases are non-conducive for antibiotic discovery screens and that common benchmarks from existing physical chemical libraries (~1% hit rate for a standard HTS screen) may not be applicable.
[0106] For all compounds deemed a “hit” as defined by MIC ≤ 8 pg / ml, an in-depth manual literature search was conducted to ensure new structures from reported antibacterial activity from literature. Note that this manual literature search was performed after initial in vitro testing to enable maximally detailed literature analyses on a small number of potent molecules. Of the 13 identified hits from both SyntheMol-RL models, six compounds were found to contain or have structural overlap with salicylanilides
[0037] , a small structural class with known antibacterial activity against Gram-positive bacteria. Specifically, niclosamide [38,39], a salicylanilide derivative, is currently being investigated for topical treatment of 5. aureus and is currently in Phase II clinical trials. A similar literature search was conducted for the two hit molecules sourcedfrom VS-Chemprop; both compounds were found to be similar to molecules with antibacterial activity in the literature. All molecules that failed the manual check were removed from further investigations, leaving six compounds from RL-Chemprop, one compound from RL-MLP, and no compounds from VS-Chemprop (Fig. 7b).
[0107] The remaining seven compounds were assessed for their respective spectrum activity against a phylogenetically diverse panel of ESKAPE isolates [2,5], a group of highly virulent and antibiotic-resistant Gram-positive and Gram-negative bacteria that often cause challenging hospital-acquired infections. Molecules displayed no significant potency across all the pathogens (Fig. 7c), including the other Grampositive pathogen E. faecium, indicating all SyntheMol-RL compounds are primarily narrow-spectrum against 5. aureus.
[0108] Importantly, however, the laboratory strain S. aureus RN4220 applied until this point in the study may not be representative of the strains commonly responsible for infection. Therefore, growth inhibition assays were conducted against a set of antibiotic-resistant 5. aureus strains
[0040] especially relevant to the clinic. All seven hit compounds were screened against the most prevalent community-associated MRSA strain for skin and soft tissue infections (SSTIs), USA300, as well as a collection of multidrug-resistant vancomycin-intermediate S. aureus (VISA) isolates from the CDC AR Isolate Bank
[0041] , which covers all resistance mechanisms found within the VISA panel (Fig. 7c). For all 7 compounds, no significant loss of antibacterial potency was found, with all compounds retaining a potent MIC ≤ 8 pg / ml. This provides strong evidence that all seven molecules are able to overcome a wide array of prevalent antibiotic resistance determinants, likely owing to their structure relative to clinical antibiotics.
[0109] In Vivo Validation of SyntheMol-RL-Generated Molecules. S. aureus causes most SSTIs in humans, including infected wounds and ulcers, cellulitis, and folliculitis
[0040] , Not only is MRSA prevalent in hospitals globally, but community-associated MRSA (CA-MRSA) is responsible for a significant proportion of SSTIs in Europe (<1% to 32%), Asia (-17%), and the United States (>50%). SSTIs caused by virulent S. aureus (including MRSA) is a risk factor for invasive infections [40,42], leading to conditions such as bacteremia and endocarditis, which when caused by 5. aureus are associated with mortality rates of 15% and 30-40% respectively
[0043] , Due tothe burden MRSA imposes, the most prevalent CA-MRSA isolate in the United States, USA300 was used, as the translatability of the hit molecules continued to be investigated [4].
[0110] Given their potent in vitro activities, whether each of the seven hit molecules generated by SyntheMol-RL may be amenable to formulation to treat MRSA infected wounds was evaluated. To investigate this possibility, the Chemical Release Evaluation on Agar Media (C. R. E. A. M.) assay was developed in this study, in which control (10% DMSO) and test molecules (2% w / v) were formulated in Glaxal Base for in vivo application and subsequently applied to MRSA-inoculated LB agar plates, prior to overnight incubation at 37°C (Fig. 8a). This in vivo mimicking model serves to evaluate the formulation’s efficacy as a topical antibacterial treatment. It is worth noting that all seven compounds were readily soluble in the Glaxal Base solution, indicating that the in silico optimization of solubility along with antibacterial activity during generation helped remove potential barriers at this stage in development. It was observed that MLP-01 resulted in the largest zone of growth inhibition, indicating that it may demonstrate the most promising in vivo activity in a mouse wound infection model. Interestingly, despite similar growth inhibitory activity in liquid media, there is almost a 2-fold difference in area between MLP-01 and Chemprop-03, showcasing the imperfect correlation between liquid MIC values and growth inhibitory effect when formulated into a cream for in vivo applications. Given its potency and success in the C. R. E. A. M assay, characterization the activity of MLP-01 continued to be investigated, which known as synthecin herein due to its origin of discovery. Synthecin was tested to determine whether its activity was bactericidal or bacteriostatic against 5. aureus USA300. Bacteriostatic activity in LB medium was observed at 2x and 4x MIC (Fig. 8b). For reference, linezolid, a commonly prescribed antibiotic for the treatment of 5. aureus skin infections, is also bacteriostatic
[0044] ,
[0111] Based on these encouraging in vitro results, synthecin was tested for its retention of antibacterial efficacy against MRSA USA300 in a mouse wound infection model, a common infection caused by MRSA. To test the in vivo efficacy of synthecin, a wound infection in cyclophosphamide pre-treated C57BL / 6N mice was established using S. aureus USA300 (~2.15 × 107CFU) (Fig. 8c). The S. aureus infection was allowed to establish for 1 h prior to topical treatments at 1, 4, 8, 12, and 20 h post-infection (hpi)with Glaxal Base supplemented with vehicle (10% DMSO) or synthecin (2% w / v). Mice were then euthanized 24-hpi, and the skin was aseptically dissected and plated to quantify S. aureus burden. Vehicle-treated mice carried ~6.39 × 109CFU / g at the experimental endpoint, with wounded tissues exhibiting significant inflammation (Fig. 8d-e, Fig. 9). In contrast, mice treated with synthecin showed a significantly lower bacterial burden, with ~5.14 × 107CFU / g, similar to the bacterial burden observed in pre-treated infection control mice. The synthecin treated tissues also displayed no signs of inflammation (Fig.8e). Collectively, this data demonstrates that synthecin can effectively suppress an MRSA wound infection, aligning with its exceptional in vitro activity in the C. R. E. A. M. assay.
[0112] Discussion
[0113] SyntheMol-RL is an effective generative Al model for designing easily synthesizable small molecule drug candidates. SyntheMol-RL’s use of reinforcement learning enables it to rapidly explore massive combinatorial chemical spaces with tens of billions of molecules for promising compounds that are easily synthesizable by design. Furthermore, the multiparameter optimization abilities of SyntheMol-RL allow it to identify compounds that simultaneously possess multiple drug-like properties - in this case, antibacterial activity and aqueous solubility - which is a necessity in real-world drug development applications. When applied to antibiotic discovery, SyntheMol-RL generated a superior set of antibiotic candidates for 5. aureus according to in silico metrics and in vitro experiments relative to both SyntheMol-MCTS (previous generative model) and VS-Chemprop, a machine learning-based virtual screening method. Indeed, SyntheMol-RL’s utility as a flexible tool for readily translatable drug discovery is showcased by the discovery of synthecin, a structurally novel molecule effective at treating a wound infection caused by MRSA in a mouse model.
[0114] SyntheMol-RL overcomes limitations present in other generative Al models [12,45] for molecular design. The reinforcement learning component of SyntheMol-RL follows a line of research in reinforcement learning algorithms for molecular generation [24-27], with the sampling of diverse but high scoring building blocks within the reinforcement learning algorithm inspired by a similar sampling mechanism in GFlowNets35. However, the SyntheMol-RL algorithm is explicitly designed for synthesizable molecule generation, which requires learning to explore a combinatorial chemical space of tens of billions of molecules constructed from pre-defined molecular building blocks and chemical reactions. In contrast, many reinforcement learning and GFlowNet models either do not incorporate synthesizability at all [27,35] or only include it as a heuristic to be optimized [36,46], resulting in the generation of many synthetically infeasible compounds, as seen both in GFlowNet experiments and in other work
[0047] , While some models have been designed to generate molecules following known synthetic routes [48-56] like SyntheMol-RL, these studies typically synthesize and experimentally validate few - if any - generated compounds, making it challenging to evaluate those models in terms of real-world synthesizability and biological efficacy.
[0115] Interestingly, the molecules generated by SyntheMol-RL tend to form clusters of a particular chemotype. Specifically, in each cluster, SyntheMol-RL selects one shared building block while exploring diverse “second” building blocks (Fig. 10). This behavior naturally arises from the design of SyntheMol-RL, since the reinforcement learning policy balances both finding building blocks that consistently lead to high-scoring molecules (thus, often repeating a promising building block) while also exploring diverse compounds, which is further reinforced by the post-hoc diversity filtering (thus, selecting diverse second building blocks). The resulting clusters of generated molecules that can be selected for laboratory testing are therefore ideal for both maximizing the probability of finding unique hits across distinct clusters while also naturally exploring the structure-activity relationship (SAR) landscape
[0057] within each cluster. When a cluster of compounds includes multiple hits, as with the RL-Chemprop model, the cluster is likely amenable to SAR optimization based on the presence of those hits and any closely related inactive compounds. In contrast, an isolated hit within a cluster would provide little information about the molecular substructures that are responsible for its activity and may suggest challenges with downstream medicinal chemistry optimization.
[0116] The in vitro results highlight the potential of SyntheMol-RL-generated compounds, particularly in addressing the global burden of MRSA. Indeed, MRS A has seen a dramatic rise in mortality over the past three decades, with deaths associated with MRSA increasing from 261,000 in 1990 to 550,000 in 2021, while attributable deaths doubled to 130,000 in the same period [1], Additionally, MRSA accounts for 26.1% of attributable deaths, making it the foremost cause of deaths attributed to antimicrobial resistance in high-income regions. To address this urgent and growing unmet need, novelantibacterial molecules that display promising activity in such resistant S. aureus strains were identified. The molecules may also have narrow-spectrum activity, which is useful to avoid the widespread disruption of host microbiota and decrease the rate of resistance dissemination [58-60], Among the 79 total SyntheMol-RL compounds synthesized, 15 highly potent active compounds were identified against S. aureus RN4220, a remarkable hit rate of 19%. Seven new compounds were identified that retained potent activity against 5. aureus USA300, as well as VISA strains that cover all resistance mechanisms present for 5. aureus in the CDC AR Isolate Bank, further supporting their robust efficacy in targeting burdensome S. aureus infections.
[0117] Moreover, the C. R. E. A. M. assay provided a simple and informative in vitro platform for evaluating the release and antibacterial activity of molecules formulated for topical administration. The correlation between the zone of inhibition observed in this assay and the subsequent in vivo efficacy of synthecin highlights its utility for prioritizing candidates for in vivo validation. In the mouse wound infection model, synthecin treatment markedly suppressed bacterial proliferation in wounds, resulting in a bacterial burden comparable to pre-treated control mice and mitigated tissue inflammation. These findings demonstrate the C. R. E. A. M. assay as a promising in vitro tool for prioritizing topical antibiotics and the utility of synthecin to effectively treat MRSA-infected wounds.
[0118] Methods
[0119] Antibiotic Training Set Curation
[0120] The training set consists of the “Bioactives 2” chemical library, a structurally and functionally diverse 10,716 compound library housed in the Centre for Microbial Chemical Biology at McMaster University. The library was screened in two biological replicates against S. aureus RN4220 for growth inhibitory activity. Cells were grown overnight at 37°C in 3 ml Luria-Bertani (LB) medium and then diluted 1 / 10,000 in fresh LB. 99 pL of cells were added to each well of Costar 96-well flat-bottom plates manually. Each compound was then added (1 pL) to a final screening concentration of 50 pM in a final volume of 100 pL. Plates were then incubated at 37°C with shaking (900 rpm) for 16 hours. Plates were then read at 600 nm using a BioTek Synergy Neo2 plate reader and data were normalized by plate using interquartile mean (IQM) prior to data compiling and hit identification (see below).
[0121] To assign binary activity labels to the screened compounds, the average normalized optical density at 600 nm (OD600) was first calculated for each compound using two biological replicates. The mean and standard deviation was then determined of these average normalized OD600values across all compounds in the dataset. A threshold of μ − 2σ was applied to binarize these values - values below this threshold were labeled as active, while those at or above the threshold were labeled as inactive. Subsequently, the SMILES were canonicalized for each compound using RDKit version 2023.9.1. For data points with identical SMILES and binary activity labels, one data point was retained and the others were discarded, resulting in 10,660 data points. For data points with identical SMILES but conflicting binary activity labels (at least one labeled as active and one as inactive), all such samples were removed to avoid noise in the activity labels. This process yielded a final 5. aureus antibiotic training dataset of 10,658 unique molecules, of which 1,137 (10.7%) are active and 9,521 (89.3%) are inactive.
[0122] ChEMBL Antibiotics Curation. To compile a set of known antibiotics for structural comparison, the ChEMBL database was queried on November 8, 2023, using the search terms "antibiotic" and "antibacterial." The search term "antibiotic" yielded 636 molecules, with 591 of them having SMILES. The term "antibacterial" returned 611 molecules, with 590 having SMILES. The two sets of compounds were combined, excluded molecules with missing SMILES, converted the SMILES to canonical SMILES using RDKit, and then deduplicated the compounds based on the canonical SMILES. This process resulted in a set of 1,007 unique molecules.
[0123] Morgan Fingerprints. Morgan fingerprints were used for molecular similarity and distance calculations. Morgan fingerprints were calculated with a radius of 2 and 2,048 bits using RDKit’s GetMorganFingerprintAsBitVect function.
[0124] Molecular Similarities. Molecular similarities were calculated using either Tanimoto similarity or Tversky similarity. Tanimoto similarity is defined as Ta(X, Y) =where X and Y are Morgan fingerprints of two compounds. It is asymmetric similarity measure with values between 0 and 1, where higher values indicate that compounds that are more similar (i.e., they have more molecular substructures in common). Tversky similarity is calculated as Ta(X, Y) =where X is the Morganfingerprint of a proposed (i.e., generated) compound and Y is the Morgan fingerprint ofa reference compound. It is an asymmetric similarity measure with values between 0 and 1, where higher values indicate that the proposed compound contains a higher proportion of the reference compound’s substructures. The Tversky similarity is used specifically for calculations to determine of a compound may be new, where it is aimed to measure the proportion of known antibiotic functional groups in reference compounds that are contained in proposed compounds while ignoring extraneous substructures in the proposed compounds.
[0125] Molecular Diversity. Diverse sets of generated molecules were selected using a method based on the idea of finding a maximal independent set of molecules. The goal of this method is to identify the largest set of molecules such that no two compounds in the set have a similarity (e.g., Tanimoto similarity) greater than some user-defined threshold, which thereby provides a strict guarantee that the selected set of compounds is diverse according to that similarity threshold.
[0126] More precisely, given a set of molecules A, a molecular similarity metric s, and a similarity threshold t, the study aims to select D = argmaxD A{D | V%, y G D,s(x,y) < t. If this is reformulated as a graph G=(V, E) where V=A (i.e., every molecule is anode) and E = {x,y G A |s(%, y) > t} (i.e., an edge connects every pair of molecules with a similarity greater than t), then selecting the largest diverse subset of molecules D is precisely equivalent to solving the maximum independent set problem on G. A maximum independent set is the largest independent set in a graph, where an independent set is a set of nodes such that no two nodes in the set have an edge connecting them. Since the maximum independent set problem is NP-hard, solving it is intractable, but instead a maximal independent set is calculated as an approximation (i.e., an independent set that cannot be made larger by adding nodes.) Although a maximal independent set is not necessarily the largest independent set in the graph, it is still an independent set, which means it satisfies the diversity requirement of no two molecules exceeding a given similarity threshold.
[0127] To obtain a maximal independent set, a given set of molecules is formulated as a graph with nodes and edges as detailed above, and the maximal independent set function is applied from networkx version 3.2.1. To slightly improve the approximation, the function is run 10 times with different random seeds andtake the largest of the maximal independent sets. This set of molecules then serves as the diverse set.
[0128] t-SNE Visualizations. t-SNE visualizations were generated using scikit-leam’s61 tSNE, applied to the Morgan fingerprints of molecules with Jaccard (Tanimoto) as the distance metric, squared distances, and a principal components analysis (PCA) initialization. For large datasets, a subset of molecules was randomly sampled to accurately represent the dataset in the t-SNE visualization.
[0129] Property Predictor Architectures. Two property prediction model architectures were used in this study: (1) Chemprop-RDKit and (2) MLP-RDKit. Chemprop-RDKit consists of the graph neural network (GNN) Chemprop augmented with molecular features computed by the cheminformatics package RDKit. Chemprop-RDKit takes as input the graph structure of a molecule with atoms as nodes and bonds as edges. Then, the GNN portion of the model applies three message passing steps that aggregate simple features of each atom and bond in a molecule, such as the atom type and bond type, using neural network layers to build vector representations of local neighborhoods of the molecule. After the message passing steps, these local representations are summed to form a single GNN vector representation for the whole molecule. This GNN vector representation, which is 300-dimensional, is then concatenated with a vector of 200 molecular features computed by RDKit to form a 500-dimensional vector. This combined vector is passed through a multilayer perceptron (MLP) with one hidden layer. The activation function on the output layer of the MLP is a sigmoid layer for binary classification properties (e.g., antibacterial activity) and is a linear layer for regression properties (e.g., aqueous solubility). MLP-RDKit has the same architecture as the MLP at the end of the Chemprop-RDKit model, but it does not have a GNN component and therefore only takes as input the vector of 200 RDKit features. Both architectures were implemented using Chemprop version 1.6.1.
[0130] Model Training. Property prediction models were trained using 10-fold cross-validation with data randomly split into 80% train, 10% validation, and 10% test for each fold. The models were all trained on the training data and evaluated on the test data and used the validation data for early stopping. All models were trained for 30 epochs using the Adam optimizer. The antibacterial activity models used a binary cross-entropy loss and were evaluated using the area under the receiver operating characteristic curve(ROC-AUC) and area under the precision recall curve (PRC-AUC). The aqueous solubility models used a mean squared error loss and were evaluated using mean squared error (MSE) and the coefficient of determination (R2).
[0131] Enamine REAL Space. The Enamine REAL Space comprises 31 billion make-on-demand molecules that can be synthesized through a single chemical reaction using a limited number of molecular building blocks as reactants. For this study, the 11 / 2021 version of the REAL reactions was used, which includes 169 chemical reactions, along with the 2022 ql-2 version of the REAL building blocks, consisting of 139,517 molecular building blocks. The 2022 ql-2 version of the REAL Space was downloaded on August 30, 2022, which encompasses 31,507,987,117 molecules that can be produced using the specified building blocks and chemical reactions.
[0132] Following the approach used in the original SyntheMol-MCTS study, only 13 of the most common reactions were used. To prepare the building blocks for the model, RDKit was first used to convert the SDF file of the building blocks to SMILES format. All building blocks were converted successfully. Next, the molecules by SMILES were deduplicated, reducing the number to 139,444 molecules. Then, the RDKit salt remover was applied to eliminate salts from the building blocks to avoid incorrect reaction template matching during generation, and 24 molecules whose salts could not be properly removed were taken out of the data set. This resulted in a set of 137,656 unique molecules (with 139,493 unique building block IDs due to duplicate SMILES). Using the curated reactions and building blocks, 30,330,025,259 molecules can be produced, representing 96.3% of the total REAL Space.
[0133] WuXi GalaXi. The WuXi GalaXi consists of 16 billion make-on-demand molecules. The 12 / 31 / 2022 version of the WuXi GalaXi was used, which includes 36 chemical reactions, 15,488 building blocks, and 16,146,071,436 molecules that can be produced using these building blocks and chemical reactions. To prepare the building blocks for use in the model, building blocks with missing IDs (4 building blocks) and missing SMILES (4 building blocks) were removed and then deduplicated by SMILES, leaving 14,977 unique molecules. Then, the salts were removed from the building blocks using the RDKit salt remover (all salts were correctly removed).
[0134] SyntheMol-MCTS. The original implementation of SyntheMol
[0017] used a Monte Carlo tree search (MCTS) algorithm to generate molecules. Briefly, the MCTSalgorithm works as follows. The input to the algorithm is a chemical synthesis tree T that consists of a set of nodes N in T. Each node N contains Nmol, which is a set of one or more molecular building blocks B in a chemical space C B c C. MCTS defines a value function on nodes, S(N) =. This value function balances exploitingnodes that lead to high scoring molecules via Q(N), selecting nodes containing building blocks with high property prediction scores via P(N), exploring rarely visited nodes via U(N), and selecting diverse building blocks via D(N) (additional details are available in ref.17).
[0135] To generate a molecule, MCTS first computes the value S(N) of every node containing a single building block. The MCTS policy is then to select the node with the highest value. Given this choice, MCTS creates every possible node with this first building block and a second building block that is synthetically compatible with the first building block in some chemical reaction r contained in the set of chemical reactions R defined by the chemical space r 6 R. MCTS then scores each of these nodes with two building blocks and applies its policy by selecting the node with the highest value. Next, MCTS creates every possible node with the two selected building blocks and a third synthetically compatible building block, and it also creates nodes for every possible molecule that can be formed by applying one of the chemical reactions in R to just the two selected building blocks to form a new molecule. MCTS scores all of these nodes -both nodes with three building blocks and nodes with a single molecule formed from the two building blocks - and again follows its policy by choosing the node with the highest score. If that node contains three building blocks instead of a single molecule, then nodes are created using all compatible chemical reactions to combine those three building blocks, and MCTS selects the node with the highest value.
[0136] At this point, MCTS has completed a single rollout and has obtained a single molecule composed of two or three building blocks. This molecule m e C is scored by a weighted combination of L property predictors, MkC ^ R for k G {1,..., L}, using property weights wkfor k G {1,..., L} (where Efc=iwk = land 0 < wk< 1) to obtain the molecule’s overall property score, p(m) =k=1wk * Mk(m). This overall property score is then used to update the Q(N) exploit score of every node N selected during the generation of molecule m. Additionally, the exploration and diversityscores, U(N) and D(N), are updated to reflect the nodes visited during the generation of molecule m. MCTS repeats this process for a fixed number of rollouts.
[0137] SyntheMol-RL. The reinforcement learning (RL) version of SyntheMol employs the same synthesis tree T, chemical space C, chemical reactions R, and property predictors M1,...,MLwith their weights w1,...,wLas MCTS, but it reformulates the method for computing node values and selecting nodes during each rollout. In SyntheMol-RL, RL replaces the MCTS value function S(N) with an RL value function V(N). V(N) is implemented as a deep neural network that takes as input the building blocks in the node, Nmoland outputs a prediction of the expected overall property score of molecules that can be created from the building blocks Nmoiby following the RL policy. The RL policy is to apply the RL value function V(N) to all nodes created at a given step in generation and then sample a node proportional to the values of the nodes with a temperature scaling, i.e., P N) oc e’TM / y The temperature parameter T can be tuned to affect the amount of exploration or exploitation performed by the RL policy; high temperature prefers exploration (more uniform probabilities) while low temperature prefers exploitation (more spiky probabilities).
[0138] The RL value function V(N) is implemented as a weighted combination of models, similar to the molecule property score p(m). Specifically, V N) = wk* Zk(Nmol) where Z1;..., ZLare deep learning models and wltwLare the same property weights used in p(m). They are then trained on the RL value function objective as follows. After each rollout constructs a molecule m, the RL algorithm stores tuples of N, Mi m),for every node N in the trajectory of nodes that were selected to create the molecule m. This builds an RL value function training set of nodes along with the property prediction scores of the final molecules created from those nodes. After every nrl trainrollouts, the RL models Z1;..., ZLare trained for nrl epochsepochs to take as input the building blocks of the node, Nmol, and predict the relevant score of the generated molecule from M1,...,MLusing a mean squared error loss.
[0139] Dynamic RL Parameters. The RL temperature parameter T and property weights wltwLhave a large impact on the diversity of generated molecules and their relative property scores, respectively. By default, the temperature is set to T = 0.1 and the property weights are set to w1,...,wL= However, these values may not beoptimal for every molecule design problem. Rather than manually tuning each parameter, a dynamic tuning mechanism
[0062] was designed that automatically adjusts the temperature and property weights based on pre-defined goals.
[0140] The RL temperature aims to balance exploration and exploitation, so the temperature goal was defined explicitly by setting a target similarity A*, which is the desired maximum Tanimoto similarity between each generated molecule and all previously generated molecules. A dynamic tuning method was then defined to adjust the RL temperature during generation to obtain a molecule similarity of A* on average. During generation, the dynamic temperature tuning mechanism maintains a rolling average molecule similarity Aavfl, which is initialized asavg= A*. After each rollout i generates a molecule mi,avgis updated asavg= y *avg+ (1 — y) * max sim(mi, m.) where y = 0.98 is the rolling average weight and sim(mi, m.) is the Tanimoto similarity between the molecule mtgenerated on rollout i and the molecule mj generated on rollout j. Next, the percent difference between the average similarity Aavfland the target similarity A* is computeddesired temperature is computed asnew= T +diff *T- To smooth out changes to the temperature, the temperature is updated via a rolling average with r = y * r + (1 - y) *Tnew- Finally, the temperature is clipped within reasonable limits with T =0.001 and Tmax= 10.
[0141] The property weights aim to balance the importance of the two property objectives (antibacterial activity and aqueous solubility), so the property weight goal is defined as maximizing the number of molecules that simultaneously possess both properties. Specifically, the success thresholds are defined as t1;..., tLfor each property, and it is considered that molecule m is a hit for property k if Mk(m) > tk. In a similar manner to dynamic temperature tuning, dynamic property weight tuning is performed by computing a rolling average. Here, for each property k G {1,..., L}, a rolling average success rate is defined as SaVgwhich is initialized with SaVg= 0. The aim is to maximize all success rates simultaneously, which is hypothesized to occur when the success rates are equal. Therefore, this study aims to adjust the property weights to make the success rates match. After each rollout i generates a molecule mL. then s^vgis updated as SaVg= y *savg + (! — / ) * I[Mk(mi)' > tk]where y = 0.98 as with the dynamic temperatureand I is in the indicator function that is 1 if Mk(mi) > tkand 0 otherwise. The average success rate across the properties was then computed: savg= ^2fc=isavg- Ifsavg = 0, meaning no successful molecules have been generated, then the property weights are unchanged. If, savg> 0 then the relative amount is determined by which each individual success rate,deviates from their mean, savg. by computing skvgdiff=A rolling average update is then performed of the property weights with wk= y * wk+ (1 — y) * (wk— S vgdiff *wfc)- These property weights are normalized to sum to 1 by computing wk=Wk—. Then, the property values are clipped to wk=Sfc=i Wkmax( \wk,'wmiiLiL„iL's) where wmi„ = 0.001 and are thenAt a high level, this ensures that if the success rate for property k is higher (or lower) than the average success rate across the properties, then the corresponding property weight wkwill be decreased (or increased) proportional to the difference between that property’s success rate and the average success rate.
[0142] Generating Molecules with SyntheMol. SyntheMol was applied in this study, both using MCTS and using RL, to generate antibacterial molecules against 5. aureus. For both RL and MCTS, we had k=2 properties: antibacterial activity (binary classification) and aqueous solubility (regression). Since Chemprop-RDKit performed best on both properties, we set the property predictors Mi and M2 to be the trained Chemprop-RDKit models for antibacterial activity and aqueous solubility, respectively. Specifically, Mi and M2 were the ensemble of ten models for each property from the ten folds of cross-validation, and predictions consist of the average prediction across the ten models. MCTS was adapted to use the same dynamic property weighting scheme as RL to set the weights w1and w2, and for both models, we defined the success thresholds as ti=0.5 for the antibacterial model and t2=- 4 for the log solubility model to match the study’s hit definitions for filtering generated compounds.
[0143] For RL, the RL value function models Z and Z2were implemented as either Chemprop-RDKit or MLP-RDKit models. Whenand Z2are Chemprop-RDKit models, the overall architecture is referred to as “RL-Chemprop”, and when Z and Z2are MLP-RDKit models, the overall architecture is referred to as “RL-MLP”. The ZTand Z2models were initialized with the weights ofand M2. Specifically, ZTand Z2wereset be the first of the ten models in the ensembles of and M2since Z and Z2in the implementation are single models, not ensembles.
[0144] Notably, as an RL value function, these models must make predictions not just on single molecules but on combinations of molecules (e.g., multiple molecular building blocks). For Chemprop-RDKit models, the model itself is unmodified. The Chemprop GNN component is applied in its usual form to the multiple molecules by treating them as a single graph with disconnected components, while the RDKit component uses the average of the RDKit features of the individual molecules, thereby preserving the dimensionality of the RDKit feature input. For MLP-RDKit models, the model is modified by replicating the MLP weights of the first layer nmax bbstimes, where max bbs is the maximum number of building blocks that can be combined into a single molecule nmax bbs= 3. For a given combination of molecules, RDKit features are computed, concatenated, and fed as input to this modified MLP. If there are fewer thannmax bbs molecules, the remaining elements of the input vector are set to 0.
[0145] After modification as described above, the Chemprop-RDKit (Zi) and MLP-RDKit (Z2) RL value function models were then trained during generation every nr( train= 10 rollouts for nri epochs= 5 epochs across all the nodes and molecules that had been previously generated. To dynamically tune the RL temperature T, the target similarity was set to A* = 0.6 to match our diversity filter, which uses 0.6 as a threshold to define diverse molecules.
[0146] All three SyntheMol versions were run - RL-Chemprop, RL-MLP, and MCTS - for 10,000 rollouts, generating 10,983; 9,228; and 11,630 molecules, respectively. 8 CPUs were used for all three versions, and 1 additional GPU was used for RL-Chemprop due to the computational burden of running the Chemprop models on millions of building block combinations (RL-MLP and MCTS were faster with CPU only due to the relatively lightweight computations involved). RL-Chemprop took 4-7 days, RL-MLP took 9-12 hours, and MCTS took 4-5 hours, depending on the precise hardware that was used (GPUs were either NVIDIA A40, TITAN_Xp, TITAN_V, or RTX_2080Ti).
[0147] Virtual Screening with Chemprop-RDKit. To perform virtual screening, 21 million molecules were selected by sampling 14 million molecules uniformly at random from the 31 billion REAL molecules and by sampling 7 million moleculesuniformly at random from the 16 billion GalaXi molecule. The final 21 million compounds roughly approximated the ratio between the two chemical spaces. The Chemprop-RDKit models were then applied for antibacterial activity and aqueous solubility to all 21 million molecules using 1 GPU and 8 CPUs, which took roughly 7 days. This is in line with the time taken by the slowest SyntheMol-RL model.
[0148] Random Selection. 150 molecules were randomly as a baseline for synthesis and validation by sampling 100 molecules uniformly at random from the 31 billion REAL molecules and by sampling 50 molecules uniformly at random from the 16 billion GalaXi molecules.
[0149] Toxicity Predictions. Toxicity predictions for molecules were made using ADMET-AI version 1.2.0. ADMET-AI consists of Chemprop-RDKit models trained on 41 absorption, distribution, metabolism, excretion, and toxicity (ADMET) datasets from the Therapeutics Data Commons. ADMET-AI was applied to make predictions for all 41 properties for each molecule and then extracted the clinical toxicity (ClinTox) predictions. The clinical toxicity predictions are values in the range [0, 1] with higher values indicating a higher probability of toxicity.
[0150] Compound Synthesis. Compound synthesis was performed by Enamine as part of their Enamine REAL Space and by WuXi as part of their WuXi GalaXi. Compound purity of at least 85% was verified using liquid chromatography-mass spectrometry (LC-MS) except in cases of poor solubility, compound instability under LC-MS conditions, or non-informative LC-MS. In these cases, proton nuclear magnetic resonance (1H-NMR) was used to assess chemical purity.
[0151] GFlowNet. GFlowNet molecules were generated using a multi-objective GFlowNet model. The implementation of multi-objective GFlowNets was specifically used, which implemented a multi-objective reward including binding to the target sEH, drug-likeness measure QED, molecular weight, and synthetic accessibility heuristic SAScore. This reward function was replaced with a new reward consisting of the antibacterial activity Chemprop-RDKit model ensemble, the aqueous solubility Chemprop-RDKit model ensemble, and the same SAScore heuristic. The GFlowNet model was run with this new reward function while keeping all other parameters at their default values. After generating molecules, a final set of molecules were selected using the same filtering procedure applied to the SyntheMol compounds, but an additional filterof SAScore < 4 was used (i.e., easy synthesizability according to the SAScore heuristic) to increase the likelihood of obtaining synthetically tractable compounds. Furthermore, instead of selecting the top 150 compounds by antibacterial prediction score, only the top 14 were selected since compounds had to be manually reviewed by chemists at Enamine and WuXi for synthetic accessibility.
[0152] SyntheMol-RL Ablations. A series of ablation experiments were performed to determine the effect of each of the components of SyntheMol-RL on the distribution of predicted antibacterial activity and log solubility scores across the generated molecules. Each ablation experiment was run five times with a different random seed each time. All SyntheMol-RL experiments were run with both RL-Chemprop and RL-MLP. Below are the ablation experiments:
[0153] Final: The final version of SyntheMol-RL. The first seed of this experiment is the one with results reported in the main text.
[0154] Fixed Property Weights: Instead of the dynamic property weights for antibacterial activity and aqueous solubility, fixed weights were used. The weights were (0.00 antibacterial, 1.00 solubility), (0.86 antibacterial, 0.14 solubility), (0.88 antibacterial, 0.12 solubility), (0.90 antibacterial, 0.10 solubility), (0.92 antibacterial, 0.08 solubility), (0.94 antibacterial, 0.06 solubility), (0.96 antibacterial, 0.04 solubility), and (1.00 antibacterial, 0.00 solubility). The fixed weights were chosen to balance the relative scale of the antibacterial activity predictions (ranging from 0 to 1) and the log solubility predictions (roughly ranging from -10 to 2).
[0155] Dynamic Temperature Target: Instead of the dynamic temperature similarity target of 0.6, similarity targets of 0.4, 0.5, 0.7, and 0.8 were used.
[0156] Fixed Temperature: Instead of the dynamic temperature with a similarity target, fixed temperatures of 0.01, 0.05, 0.1, 0.5, and 1.0 were used. As a point of comparison, we also ran SyntheMol-MCTS with different explore weights (the MCTS equivalent of the RL temperature parameter). The explore weights were 0.5, 1.0, 5.0, and 50.0 (the default of 10.0 was used in the final model reported in the main text).
[0157] RL Training: Instead of using reinforcement learning to train the RL value models Zi and Z2, these models were kept fixed after initializing them with weights from the trained property predictors Mi and M2. Thus, they know how to score completemolecules but are not trained to score molecular building blocks for their potential to form complete molecules with desired properties (i.e., the RL objective).
[0158] RL Pretrained Weights: Instead of initializing the RL value models Zi andZ2 with weights from the trained property predictors Mi and M2, these models are initialized randomly and trained from scratch using the RL objective.
[0159] Antibacterial Potency Analyses. S. aureus RN4220, S. aureus USA300, S. aureus clinical isolates (VISA panel; CDC AR Isolate Bank), Escherichia coli BW25113, Pseudomonas aeruginosa PAO1, Klebsiella pneumoniae ATCC 43816, Enterococcus faecium ATCC 19434, and Acinetobacter baumannii ATCC 17978 were grown overnight at 37°C in 3 ml of LB medium with shaking. Overnight cultures were then diluted 1:10,000 into fresh LB. Cells were then introduced to two-fold serial dilutions of each compound under investigation in a final volume of 100 pl in Costar 96-well flat-bottom plates. Plates were incubated at 37°C without shaking (E. coli BW25113, P. aeruginosa PAO1, K. pneumoniae ATCC 43816, E. faecium ATCC 19434, and A. baumannii ATCC 17978) or with shaking at 900 rpm (5. aureus RN4220, S. aureus USA300, and 5. aureus clinical isolates) until untreated control cultures reached stationary phase. Plates were then read at 600 nm using a BioTek Synergy Neo2 plate reader.
[0160] C. R. E. A. M. assay. S. aureus USA300 cells were grown overnight in 3 ml LB medium at 37° C with shaking. ~106CFU in 100 pl liquid LB was deposited onto solid LB agar plates. -100 mg of control or SyntheMol-RL-generated molecule in Glaxal Base was deposited onto inoculated agar plates, and plates were incubated for 18 h at 37° C. For chemical preparation, molecules were weighed and solubilized in 10% DMSO and then added to 100 mg Glaxal Base to a final concentration of 2% w / v. The solution was mixed thoroughly to ensure an even distribution of the compound in the Glaxal Base carrier. For the control, the same amount of vehicle (DMSO) was measured and mixed through Glaxal Base.
[0161] Bacterial Cell Killing. S. aureus USA300 cells were grown overnight in 3 ml LB medium at 37° C with shaking and diluted 1:10,000 into fresh LB. In 96-well flat-bottom plates, cells were grown to the required density in a final volume of 100 pl at 37°C with shaking at 900 rpm, at which time compound was added at the indicated concentration and cultures were incubated for the required duration. Cells were thenpelleted in 96-well flat-bottom plates by centrifugation at 4,000 x g for 15 min at 4 °C and washed in ice-cold sterile PBS. After washing, cells were ten-fold serially diluted in PBS and plated on solid LB agar medium.
[0162] Mouse Infection Model. Mouse model experiments were conducted according to the guidelines set by the Canadian Council on Animal Care, using protocols approved by the Animal Review Ethics Board and McMaster University under Animal Use Protocol no. 22-04-10. Six- to eight-week-old female C57BL / 6N mice were purchased from Charles River Laboratories (#027, QC, CAN). Animals were housed in a specific pathogen-free barrier facility under Containment Level 2 conditions and maintained on a 12 h light: 12 h dark cycle, which was temperature-controlled (21°C) at 30-50% humidity. No animals were excluded from the analysis, and blinding was considered unnecessary.
[0163] Mice were pre-treated with 150 mg / kg (day T-4) and 100 mg / kg (day T-1) of cyclophosphamide to render mice neutropenic. On day T-0, mice were anesthetized using isoflurane and administered buprenorphine-sustained release as an analgesic at 0.6mg / ml subcutaneously. A 2 cm2abrasion on the dorsal surface of the mouse was inflicted through tape-stripping to the basal layer of the epidermis using approximately 30-35 pieces of autoclave tape. Mice were immediately infected with -2.15 x 107CFU 5. aureus USA300 directly pipetted onto the wound bed. The infection was left to establish for 1 h before the first treatment with Glaxal Base supplemented with vehicle (10% DMSO) or synthecin (2% wt / vol). Mice (n = 5) were treated 1 h, 4 h, 8 h, 12 h, and 20 h post-infection with 150 mg Glaxal Base with synthecin (treatment) or DMSO (control). Mice were sacrificed at experimental endpoint (24 hpi) and wound tissue was aseptically collected, homogenized in PBS and plated on solid LB agar medium to quantify bacterial load. For chemical preparation, synthecin was weighed and solubilized in 10% DMSO and then added to 150 mg Glaxal Base to a final concentration of 2% w / v. The solution was mixed thoroughly to ensure an even distribution of the compound in the carrier. For control groups, the same amount of vehicle (DMSO) was measured and mixed through Glaxal Base.
[0164] While the present disclosure has been described with reference to examples, it is to be understood that the scope of the claims should not be limited by theembodiments set forth in the examples, but should be given the broadest interpretation consistent with the description as a whole.
[0165] All publications, patents and patent applications are herein incorporated by reference in their entirety to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated by reference in its entirety. Where a term in the present disclosure is found to be defined differently in a document incorporated herein by reference, the definition provided herein is to serve as the definition for the term.CITATIONS1. Murray, C. J. L. et al. Global burden of bacterial antimicrobial resistance in 2019:a systematic analysis. The Lancet 399, 629-655 (2022).2. Rice, L. B. Federal Funding for the Study of Antimicrobial Resistance in Nosocomial Pathogens: No ESKAPE. J. Infect. Dis. 197, 1079-1081 (2008).3. Enright, M. C. et al. The evolutionary history of methicillin-resistant Staphylococcus aureus (MRSA). Proc. Natl. Acad. Sci. 99, 7687-7692 (2002).4. Miller, L. S. & Cho, J. S. Immunity against Staphylococcus aureus cutaneous infections. Nat. Rev. Immunol. 11, 505-518 (2011).5. Tacconelli, E. et al. Discovery, research, and development of new antibiotics: the WHO priority list of antibiotic-resistant bacteria and tuberculosis. Lancet Infect. Dis. 18, 318-327 (2018).6. Stokes, J. M. et al. A Deep Learning Approach to Antibiotic Discovery. Cell 180,688-702.el3 (2020).7. Wong, F. et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature 626, 177-185 (2024).8. Wan, F., Torres, M. D. T., Peng, J. & de la Fuente-Nunez, C. Deep-leaming- enabled antibiotic discovery through molecular de-extinction. Nat. Biomed. Eng.8, 854-871 (2024).9. Melo, M. C. R., Maasch, J. R. M. A. & de la Fuente-Nunez, C. Accelerating antibiotic discovery through artificial intelligence. Commun. Biol. 4, 1050 (2021).10. Yang, K. et al. Analyzing Learned Molecular Representations for Property Prediction. J. Chem. Inf. Model. 59, 3370-3388 (2019).Kitchen, D. B., Decomez, H., Furr, J. R. & Bajorath, J. Docking and scoring in virtual screening for drug discovery: methods and applications. Nat. Rev. Drug Discov. 3, 935-949 (2004).Bilodeau, C., Jin, W., Jaakkola, T., Barzilay, R. & Jensen, K. F. Generative models for molecular discovery: Recent advances and challenges. WIREs Comput. Mol. Set. 12, el608 (2022).Gangwal, A. & Lavecchia, A. Unleashing the power of generative Al in drug discovery. Drug Discov. Today 29, 103992 (2024).Bian, Y. & Xie, X.-Q. Generative chemistry: drug discovery with deep learning generative models. J. Mol. Model. 27, 71 (2021).Zeng, X. et al. Deep generative molecular design reshapes drug discovery. Cell Rep. Med. 3, 100794 (2022).Gao, W. & Coley, C. W. The Synthesizability of Molecules Proposed by Generative Models. J. Chem. Inf. Model. 60, 5714-5723 (2020).Swanson, K. et al. Generative Al for designing and validating easily synthesizable and structurally novel antibiotics. Nat. Mach. Intell. 6, 338-353 (2024).Huang, K. etal. Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development, in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) (2021).Kreiswirth, B. N. et al. The toxic shock syndrome exotoxin structural gene is not detectably transmitted by a prophage. Nature 305, 709-712 (1983).Zdrazil, B. et al. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Res. 52, D1180-D1192 (2024).RDKit. https: / / www.rdkit.org / .Kocsis, L. & Szepesvari, C. Bandit Based Monte-Carlo Planning, in Machine Learning: ECML 2006 (eds. Fürnkranz, J., Scheffer, T. & Spiliopoulou, M.) 282-293 (Springer, Berlin, Heidelberg, 2006). doi:10.1007 / 11871842_29.Coulom, R. Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search, in Computers and Games (eds. van den Herik, H. J., Ciancarini, P. & Donkers, H. H. L. M. (Jeroen)) 72-83 (Springer, Berlin, Heidelberg, 2007). doi:10.1007 / 978-3-540-75538-8_7.Zhou, Z., Kearnes, S., Li, L., Zare, R. N. & Riley, P. Optimization of Molecules via Deep Reinforcement Learning. Sci. Rep. 9, 10752 (2019).Ståhl, N., Falkman, G., Karlsson, A., Mathiason, G. & Bostrom, J. Deep Reinforcement Learning for Multiparameter Optimization in de novo Drug Design. J. Chem. Inf. Model. 59, 3166-3176 (2019).Popova, M., Isayev, O. & Tropsha, A. Deep reinforcement learning for de novo drug design. Sci. Adv. 4, eaap7885 (2018).Olivecrona, M., Blaschke, T., Engkvist, O. & Chen, H. Molecular de-novo design through deep reinforcement learning. J. Cheminformatics 9, 48 (2017).Grygorenko, O. O. et al. Generating Multibillion Chemical Space of Readily Accessible Screening Compounds. iScience 23, 101681 (2020).Chemical Spaces • Ultra-Large Compound Collections by BioSolvelT.BioSolvelT https: / / www.biosolveit.de / chemical-spaces / .Klug, D. M. et al. Open Source Antibiotics: Simple Diarylimidazoles Are Potent against Methicillin-Resistant Staphylococcus aureus. ACS Infect. Dis. 9, 2423-2435 (2023).Tversky, A. Features of similarity. Psychol. Rev. 84, 327-352 (1977).Maggiora, G., Vogt, M., Stumpfe, D. & Bajorath, J. Molecular Similarity in Medicinal Chemistry. J. Med. Chem. 57, 3186-3204 (2014).Swanson, K. et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics 40, btae416 (2024). Mukherjee, S. et al. ADMET-AI Enables Interpretable Predictions of Drug-Induced Cardiotoxicity. Circulation 151, 285-287 (2025).Bengio, E., Jain, M., Korablyov, M., Precup, D. & Bengio, Y. Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation, in Advances in Neural Information Processing Systems vol. 34 27381–27394 (Curran Associates, Inc., 2021).Jain, M. et al. Multi-Objective GFlowNets. in Proceedings of the 40th International Conference on Machine Learning (eds. Krause, A. et al.) vol. 202 14631-14653 (PMLR, 2023).Macielag, M. J. et al. Substituted Salicylanilides as Inhibitors of Two-Component Regulatory Systems in Bacteria. J. Med. Chem. 41, 2939-2945 (1998).Rajamuthiah, R. et al. Repurposing Salicylanilide Anthelmintic Drugs to Combat Drug Resistant Staphylococcus aureus. PLOS ONE 10, e0124595 (2015).Weiss, A. etal. Topical niclosamide (ATx201) reduces Staphylococcus aureus colonization and increases Shannon diversity of the skin microbiome in atopic dermatitis patients in a randomized, double-blind, placebo-controlled Phase 2 trial. Clin. Transl. Med. 12, e790 (2022).Lee, A. S. et al. Methicillin-resistant Staphylococcus aureus. Nat. Rev. Dis.Primer 4, 1-23 (2018).Lutgring, J. D. et al. FDA-CDC Antimicrobial Resistance Isolate Bank: a Publicly Available Resource To Support Research, Development, and Regulatory Requirements. J. Clin. Microbiol. 56, 10.1128 / jcm.01415-17 (2018).Moran, G. J. et al. Methicillin-Resistant S. aureus Infections among Patients in the Emergency Department. N. Engl. J. Med. 355, 666-674 (2006).Siddiqui, A. H. & Koirala, J. Methicillin-Resistant Staphylococcus aureus, in StatPearls (StatPearls Publishing, Treasure Island (FL), 2025).Clemett, D. & Markham, A. Linezolid. Drugs 59, 815-827 (2000).Du, Y. et al. Machine learning- aided generative molecular design. Nat. Mach.Intell. 6, 589-604 (2024).Liu, C.-H. et al. RetroGNN: Approximating Retrosynthesis by Graph Neural Networks for De Novo Drug Design. Preprint at https: / / doi.org / 10.48550 / arXiv.2011.13042 (2020).Gao, W. & Coley, C. W. The Synthesizability of Molecules Proposed by Generative Models. J. Chem. Inf. Model. 60, 5714-5723 (2020).Cretu, M., Harris, C., Roy, J., Bengio, E. & Lio, P. SynFlowNet: TowardsMolecule Design with Guaranteed Synthesis Pathways, in ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design (2024). Bradshaw, J., Paige, B., Kusner, M. J., Segler, M. H. S. & Hernandez-Lobato, J.M. A Model to Search for Synthesizable Molecules, in Proceedings of the 33rd International Conference on Neural Information Processing Systems (Curran Associates Inc., Red Hook, NY, USA, 2019).Gottipati, S. K. et al. Learning to Navigate the Synthetically Accessible Chemical Space Using Reinforcement Learning, in Proceedings of the 37th International Conference on Machine Learning (JMLR.org, 2020).Horwood, J. & Noutahi, E. Molecular Design in Synthetically Accessible Chemical Space via Deep Reinforcement Learning. ACS Omega 5, 32984-32994 (2020).Button, A., Merk, D., Hiss, J. A. & Schneider, G. Automated de novo molecular design by hybrid machine intelligence and rule-driven chemical synthesis. Nat. Mach. Intell. 1, 307-315 (2019).Gao, W., Mercado, R. & Coley, C. W. Amortized Tree Generation for Bottom-up Synthesis Planning and Synthesizable Molecular Design, in International Conference on Learning Representations (2022).Klarich, K., Goldman, B., Kramer, T., Riley, P. & Walters, W. P. Thompson Sampling— An Efficient Method for Searching Ultralarge Synthesis on Demand Databases. J. Chem. Inf. Model. 64, 1158-1171 (2024).Gao, W., Luo, S. & Coley, C. W. Generative Artificial Intelligence for Navigating Synthesizable Chemical Space. Preprint at https: / / doi.org / 10.48550 / arXiv.2410.03494 (2024).Koziarski, M. et al. RGFN: Synthesizable Molecular Generation Using GFlowNets. Adv. Neural Inf. Process. Syst. 37, 46908-46955 (2024).Tong, W., Welsh, W. J., Shi, L., Fang, H. & Perkins, R. Structure-activity relationship approaches and applications. Environ. Toxicol. Chem. 22, 1680-1695 (2003).J. Melander, R., V. Zurawski, D. & Melander, C. Narrow-spectrum antibacterial agents. MedChemComm 9, 12-21 (2018).Rea, M. C. et al. Effect of broad- and narrow-spectrum antimicrobials on Clostridium difficile and microbial diversity in a model of the distal colon. Proc. Natl. Acad. Sci. 108, 4639-4644 (2011).Jernberg, C., Löfmark, S., Edlund, C. & Jansson, J. K. Long-term ecological impacts of antibiotic administration on the human intestinal microbiota. ISME J.1, 56-66 (2007).Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. J Mach Learn Res 12, 2825-2830 (2011).Hong, Z.-W. et al. Diversity-driven exploration strategy for deep reinforcement learning, in Proceedings of the 32nd International Conference on Neural Information Processing Systems 10510-10521 (Curran Associates Inc., Red Hook, NY, USA, 2018).
Claims
Claims:
1. A compound of Formula (I):R2-C(O)-NR1R7(Formula (I))or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof, wherein:R1is H or C1-C6alkyl;R2is benzyl substituted with up to four R3, wherein each R3is independently selected from hydroxyl or halogen, or R2is CH2CR4R5R6, wherein R4is H or hydroxyl, R5is hydroxy, halogen or CF3 and R6is benzyl;R7is C1-C6alkyl, optionally substituted with one or more substituents independently selected from hydroxyl, halogen, benzyl or 5- or 6-membered heterocycle comprising S, N or O, wherein the benzyl or the 5- or 6-membered heterocycle is optionally substituted with one or more substituents independently selected from C1-C6alkyl, hydroxyl, halogen or CF3, or R7is benzyl or thiol, optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
2. The compound of claim 1, wherein R2is benzyl substituted with three R3, wherein each R3is independently selected from hydroxyl or chloro.
3. The compound of claim 1, wherein R2is CH2CR4R5R6, wherein R4is H or hydroxyl, R5is CF3 and R6is benzyl.
4. The compound of any one of claims 1 to 3, wherein R7is C1-C4alkyl, optionally substituted with one or more substituents independently selected from fluoro, hydroxyl or benzyl, wherein the benzyl is optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
5. The compound of any one of claims 1 to 3, wherein R7is benzyl or thiol, optionally substituted with one or more substituents independently selected from hydroxyl, halogen or CF3.
6. The compound of any one of claims 1 to 5, wherein R1is H.
7. The compound of claim 1, wherein the compound of Formula (I) is:or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof.
8. The compound of claim 1, wherein the compound is:or a pharmaceutically acceptable salt, solvate, and / or prodrug thereof.
9. The compound of any one of claims 1 to 8, wherein the prodrug is an ester with a hydroxyl or amine group.
10. The compound of any one of claims 1 to 9, for inhibiting growth of a Staphylococcus aureus bacterium.
11. A composition comprising a compound of any one of claims 1 to 9, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, and a pharmaceutically acceptable carrier or vehicle.
12. A method for inhibiting growth of a Staphylococcus aureus bacterium, comprising contacting the bacterium with an effective amount of a compound of anyone of claims 1 to 9, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof.
13. A method of treating a Staphylococcus aureus bacterial infection comprising administering an effective amount of a compound of any one of claims 1 to 9, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, to a subject in need thereof.
14. Use of a compound of any one of claims 1 to 9, or a pharmaceutically acceptable salt pharmaceutically acceptable salt, solvate and / or prodrug thereof, to inhibit growth of a Staphylococcus aureus bacterium.
15. Use of a compound of any one of claims 1 to 9, or a pharmaceutically acceptable salt pharmaceutically acceptable salt, solvate and / or prodrug thereof, to treat or prevent a Staphylococcus aureus bacterial infection.
16. Use of any compound of any one of claims 1 to 9, or a pharmaceutically acceptable salt, solvate and / or prodrug thereof, in preparation of a medicament for treating or prevent a Staphylococcus aureus bacterial infection.
