SCREEN

PRIOR APPLICATION INFORMATION

The instant application claims the benefit of US Provisional Patent Application 63/348,701 , filed June 3, 2022 and entitled “NOVEL ANTIMICROBIAL COMPOUNDS ISOLATED USING A MACHINE LEARNING MODEL TRAINED ON A HIGH- THROUGHPUT ANTIBACTERIAL SCREEN”, the entire contents of which are incorporated herein by reference for all purposes.

BACKGROUND OF THE INVENTION

Multidrug-resistant (MDR) bacterial infections present a serious threat to public health. In 2019, the Centers for Disease Control and Prevention (CDC) reported that approximately 3 million people suffer from MDR infections, which results in about 35,000 deaths in the USA annually (1). One of many interdisciplinary actions to address the crisis of antibiotic resistance is the acceleration of early antibiotic discovery (2,3). To achieve this goal, many programs have searched for antibacterial compounds in small molecule libraries using high throughput screens (HTS). Target- based HTS search for inhibitory compounds of a known target for which an in vitro activity assay is available (4). These target-based approaches often fail to find active molecules against Gram-negative bacteria because most of the identified compounds do not penetrate the Gram-negative cell envelope (5). Whole cell-based screens, where small molecule libraries are examined for inhibition of bacterial growth can overcome this problem. However, the high costs and low success rate (1 -2%) of HTS approaches have discouraged these efforts.

With the recent application of artificial intelligence (Al) in biology (6-8), merging drug discovery with Al may rapidly predict active molecules in silico, with a substantial decrease in associated expenses (9). One key component of Al approaches in drug discovery is to obtain a computable representation of chemical molecules. In the field of machine learning (ML), convolutional neural networks (CNNs) can recognize these representations and detect patterns automatically through conducting convolutional operations (10). On the other hand, graph convolutional networks (GCNs) apply the principles of convolution on chemical structures, which are represented as non- Euclidean structured graphs in graph neural networks, in which nodes and edges of graphs represent atomic information (atomic number, formal charge, chirality, etc.) and bonding (bond type, conjugation, ring membership, etc.), respectively. Among the variants of GCNs, the directed-message passing neural network (D-MPNN) (11) successfully generates molecule-level representations by iterative message passing process on directed graphs. In brief, D-MPNN works by propagating atom and bond information in a directed manner during the message passing phase, resulting in a high-level feature (hidden state) for each atom in a molecule. In the readout phase, all hidden states of atoms are aggregated together and form a molecule-level feature vector, which can be fed into a feed-forward neural network (FFN) for the task-specific predictions.

SUMMARY OF THE INVENTION

According to an aspect of the invention, there is provided a method of screening a library of test compounds for candidate compounds having a specific biological activity comprising: providing a training library comprising a plurality of training compounds wherein:

(a) the chemical structure of each respective one training compound is known and is represented as Simplified Molecular Input Line Entry System (SMILES) strings in a structural data entry in the training library;

(b) the relative biological activity of each respective one training compound of the training library is known and is stored in a corresponding biological activity data entry in the training library; subjecting the structural data entry and the corresponding biological activity data entry of respective one training compounds of the training library to direct- message passing neural network (D-MPNN) analysis to detect at least one feature of training compounds of the training library corresponding to biological activity above a threshold level, thereby training a machine learning model for identifying candidate compounds; providing a testing library comprising a plurality of testing compounds wherein:

(i) the chemical structure of each respective one testing compound is known and is represented as SMILES strings in a structural data entry in the testing library; and

(ii) the relative biological activity of each respective one testing compound of the testing library is not known; applying the trained machine learning model to the structural data entry of respective one testing compounds of the testing library, thereby identifying candidate testing compounds comprising said at least one feature; and testing said candidate testing compounds for said biological activity.

As will be apparent to those of skill in the art, a typical HT screening, not using Al, will have a hit rate of 0.1 - 1 % depending on the threshold set. But, as demonstrated herein, if you run the Al algorithm with virtual libraries and then use the compounds that the Al predicted as active for a real screen in the lab, then you will get an increase in the hit rate. In our case the increase was ten-fold as between 10 to 20 % of the compounds we tested in the lab were active, as discussed herein.

As will be apparent to those of skill in the art, this method can be used with any type of high-throughput screen, for example, screening for antibacterial agents, antineoplastics, antifungals, enzymatic inhibitors and the like, essentially anything for which the screen is done in whole cells and has a phenotypic response: for example but by no means limited to growth live-dead and/or fluorescence due to a reporter system.

According to an aspect of the invention, there is provided an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5-396799; STL552768; STK802026; STOCK1 N-11202; STL513145; STL525490; AG-205/33771006; 8019- 5018; Z1205495496; Z1982493964; Z1029465088; and AS-69879.

In some embodiments, the bacteria is selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus and P. aeruginosa.

According to another aspect of the invention, there is provided a method for inhibiting growth of bacteria comprising: administering an effective amount of an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414- 0936; STK760075; NAT5-396799; STL552768; STK802026; STOCK1 N-11202; STL513145; STL525490; AG-205/33771006; 8019-5018; Z1205495496;

Z1982493964; Z1029465088; and AS-69879, to a region comprising the bacteria, said bacteria growth being inhibited compared to bacteria growing under identical growth conditions except for the presence of the antibacterial compound.

In some embodiments, the bacteria is selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus; and P. aeruginosa.

In some embodiments, growth inhibition is determined by comparing numbers of colony forming units from a specific volume, as discussed herein.

A method of treating a bacterial infection comprising: administering an effective amount of a compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5-396799; STL552768; STK802026; STOCK1 N-11202; STL513145; STL525490; AG- 205/33771006; 8019-5018; Z1205495496; Z1982493964; Z1029465088; and AS- 69879 to an individual who has or who is suspected of having a bacterial infection.

In some embodiments, the bacteria are selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus; and P. aeruginosa.

According to another aspect of the invention, there is provided an anti- methicillin-resistant Staphylococcus aureus compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984.

According to another aspect of the invention, there is provided a method for inhibiting growth of methicillin-resistant S. aureus bacteria comprising: administering an effective amount of an antibacterial compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984, to a region comprising the methicillin-resistant S. aureus bacteria, said methicillin-resistant S. aureus bacteria growth being inhibited compared to methicillin-resistant S. aureus bacteria growing under identical growth conditions except for the presence of the antibacterial compound.

In some embodiments, growth inhibition is determined by comparing the optical density or the colony forming units from a specific volume, as discussed herein.

According to another aspect of the invention, there is provided a method of treating a methicillin-resistant S. aureus bacterial infection comprising: administering an effective amount of a compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984, to an individual who has or who is suspected of having a methicillin-resistant S. aureus bacterial infection.

According to another aspect of the invention, there is provided a method of developing an improved antibacterial compound comprising: providing an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5- 396799; STL552768; STK802026; STOCK1 N-11202; STL513145; STL525490; AG- 205/33771006; 8019-5018; Z1205495496; Z1982493964; Z1029465088; and AS- 69879; and modifying said antibacterial compound such that one or more properties of antibacterial compound is improved.

As will be apparent to one of skill in the art, the improved property may be for example but by no means limited to improved anti-bacterial activity. Examples of improved anti-bacterial activity include but are by no means limited to: greater antibacterial activity against one or more bacterial strains, increased residency time, increased stability, increased solubility, increased organism range and the like. Other suitable improvements will be well known to those of skill in the art, as will methods for making such modifications.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 : Initial training and performance evaluation of the machine learning model. (A) High-throughput screening data generated by screening a compound library of 29,537 compounds against B. cenocepacia K56-2 wild-type. Using B-score ≤ -17.5 as threshold, the screening yielded 256 active compounds. Darkblue are inactive and red are active compounds. (B) A graph-based machine learning model, namely D-MPNN, was trained on atom and bond features of molecules. To further increase the model’s performance, additional global features were also incorporated into the model and the results were compared. Dataset was split into 80:10:10 ratio to train, validate and test the model. (C) ROC-AUC plot evaluating model performance after training. The model (binary classification model trained with RDKit descriptors) attained a ROC-AUC of 0.823.

Figure 2: In vitro testing of top ranked predicted compounds from an FDA approved compound library. (A) Schematic of the screening protocol. 81 commercially available compounds (from the top 100) were screened. (B) The screening identified 21 bioactive compounds with positive predictive value (PPV) of 25.9%. Darkblue are inactive and red are active compounds. (C) Top 100 ranked compounds selected for empirical testing belong to different drug families. As expected, most of the compounds exhibiting bioactivity were antibiotics or antimicrobial compounds. (D) The ratio of OD _600nm and prediction scores were plotted against the predicted rank of the corresponding compounds. The results show a linear correlation between the prediction score and bioactivity. Predicted score is the probability of a compound being active as predicted by the ML model. Predicted rank is the order of the compounds based on the predicted score where compounds with the higher predicted scores ranked higher. Darkblue and red indicate compounds’ probability of being inactive and active, respectively. Results are average of at least three independent biological replicates. Figure 3: Enhanced sensitivity of the CRISPRi knockdown mutants indicated RpoB as the in vivo target of STL558147. (A) Chemical structures of STL558147 and Rifampicin. (B-D) Comparison of hypersensitive CRISRPi knockdown mutants to novobiocin (B), rifampicin (C) and STL558147 (D). As expected, CRISPRi knockdown mutants exhibited hypersusceptibility to their cognate antibiotics and suggested RpoB as the in vivo target of STL558147. Blue indicates more growth (less inhibition) and red indicates less growth (more inhibition). Results are average of at least three independent biological replicates.

Figure 4: Synergy maps of STL558147 and rifampicin in combination of other antibiotics against B. cenocepacia K56-2. Synergy plots of STL558147 (A) and rifampicin (B) with ceftazidime, colistin, and polymyxin B. The observed synergistic interactions of STL558147 were 2-3 times stronger than rifampicin and similar to the widely used synergistic antimicrobial combination of avibactam and ceftazidime (C). The synergy scores were calculated based on the widely used Bliss independence (52) and Loewe additivity (53) models. The most synergistic area in each combination is highlighted with a rectangular box inside the plot. Green (negative δ-scores) indicate antagonistic interactions and red (positive 5-scores) indicate synergistic interactions. Synergy scores >15 was considered synergistic, between -5 to 15 was considered additive and < -15 was considered antagonistic. Results are average of at least three independent biological replicates. Synergy scores are shown as mean ± SEM. Synergy scores were calculated using SynergyFinder 2.0 (33).

Figure 5: Screening of PHAR261659 analogs. PHAR261659 analogs with different side chains were selected based on lower predicted logP values. STL529920, a stereoisomer of PHAR261659 exhibited growth inhibitory activity against all six pathogens tested. The activity of compounds identified as growth inhibitory and non-growth inhibitory are shown in red and blue, respectively. Results are the average of three independent biological replicates. Error bars indicate mean ± SD. Figure 6: Determination of PHAR261659 mechanism of action with CRISPRi-Seq. A) Overview of the CRISPRi-Seq workflow. (B) Exposure of a Burkholderia cenocepacia K56-2 pooled essential gene knockdown mutant library to PHAR261659 indicated pth as the most significantly depleted mutants (from the pool). Blue and red dots represent enriched and depleted mutants, respectively. Dashed line indicates the significant threshold (a = 0.05). (C) Clonal exposure of the depleted mutants (from panel B) to PHAR261659 revealed enhanced sensitivity of a peptidyl- tRNA hydrolase (Pth) CRISPRi knockdown mutant compared to the control pgRNA_non-target and wild-type (K56-2 WT). Blue indicates growth and red indicates growth inhibition.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, the preferred methods and materials are now described. All publications mentioned hereunder are incorporated herein by reference.

Screening for novel antibacterial compounds in small molecule libraries has a low success rate. Here, we applied a machine learning (ML)-based virtual screening for antibacterial activity and evaluated its predictive power by experimental validation. We first binarized thirty-thousand compounds according to their growth inhibitory activity (hit rate 0.67%) against the antibiotic-resistant bacterium Burkholderia cenocepacia and described their molecular features with a directed-message passing neural network (D-MPNN). Then, we used the data to train an ML model that achieved a receiver operating characteristic (ROC) score of 0.823 on the test set. Finally, we predicted antibacterial activity in virtual libraries corresponding to 1 ,614 compounds from the Food and Drug Administration (FDA)-approved list and 224,205 natural products. A hit rate of 26% and 12%, respectively, was obtained when we tested the top ranked predicted compounds for growth inhibitory activity against B. cenocepacia, which represents at least a 12-fold increase from the previous hit rate. In addition, more than 51 % of the predicted antibacterial natural compounds inhibited ESKAPE pathogens showing that predictions expand beyond the organism-specific dataset to a broad range of bacteria. As discussed herein, the developed ML approach is used for compound prioritization prior to screening, thereby providing a method by which a library of compounds can be screened for any detectable biological activity to identify a subset of candidate compounds for testing, with a greatly increased hit rate compared to typical prior art high throughput screens.

As will be apparent to those of skill in the art, the application of Al methods to antibiotic discovery is gaining attention (12). Recently, deep learning was successfully used to discover antibiotic activity in compounds structurally unrelated to known antibiotics (13). A library of approximately 2,300 structurally diverse molecules was compiled, evaluated for growth inhibition of Escherichia coli, and the binarized dataset was used to train a D-MPNN model to predict antibacterial activity in several virtual molecule libraries. The identified compounds displayed antibiotic activity against Escherichia coli and other bacterial pathogens. Here, we aimed to leverage this approach by applying it to a HTS against the Gram-negative bacterium, Burkholderia cenocepacia, previously performed in our laboratory (14). B. cenocepacia is part of the Burkholderia cepacia complex (Bcc), naturally antibiotic-resistant bacteria that cause infections in immunocompromised individuals (15). Our goals were 1 ) to find new antibacterial compounds against B. cenocepacia and, 2) to test the predictive power of the deep learning model for broad range activity when trained on HTS datasets performed in antibiotic-resistant bacteria. We first evaluated the predictive ability of our primary model on an FDA-approved compound library and found that the ML approach increased the hit rate to approximately 25.925%. Subsequently, we applied the trained model to a natural product library and identified a panel of growth inhibitory compounds active against both Gram-positive and Gram-negative pathogens, demonstrating that predictions of antibacterial activity against B. cenocepacia, could be partially extrapolated to other bacterial pathogens. Two active compounds STL558147 and PHAR261659 had no previous record of antibacterial activity. STL558147 is structurally similar to rifampicin and has broad-spectrum activity, including the ESKAPE pathogens (16). The structure of PHAR2611659 does not resemble any compound with previous record of antibiotic activity. Our work extends the applicability of previously developed whole-cell HTS to the training of ML models, increasing the hit rate of subsequent screening campaigns.

As discussed above, the application of artificial intelligence in biomedical sciences has shown promising outcomes, from diagnosis (34) to disease prediction from medical records (35), to treatment response (36). Using high-throughput compound screening datasets to train ML models for drug discovery represents another application of machine learning in the biomedical science field. Specifically, the application of machine learning to antibiotic discovery has demonstrated promise as a next-generation strategy to address the crisis of antibiotic resistance (13, 37, 38).

Here, we demonstrate that ML approaches that use algorithmic solutions to identify novel structural classes of antibiotics can decrease the associated cost and time of HT-screening by allowing in silico exploration of vast, diverse chemical spaces that are otherwise unprocurable (37, 39). In the present study, we applied a machine learning model using a directed-message passing neural network (D-MPNN) that learns properties of compounds by sending messages along the atoms in a directed fashion (11). We trained the algorithm with a high-throughput screening dataset performed against B. cenocepacia K56-2 along with >200 computationally extracted compound features (14, 40). After evaluating the trained model using an FDA- approved library (Figure 2), we subsequently applied it to an unprecedented natural product library containing over 200,000 compounds and identified multiple growth inhibitory compounds (Tables 1 and 2). A similar Al algorithm was applied previously by a different research group (13). They first built a small molecule library curated for high structural diversity, tested the compounds for bacterial growth inhibition, and used the dataset to train the Al algorithm for predicting growth inhibitory activity in bacteria. In our case, we used a small molecule library that was larger in size but not previously curated for Al training. Our training dataset had random structural diversity and was imbalanced between the number of active and inactive compounds. The use of a dataset not previously designed for Al training could have reduced the Al algorithm’s ability to predict growth inhibitory activity in molecules with very different scaffolds. However, we were able to predict and validate compounds with growth inhibitory activity, although with a lower hit rate compared to the one previously reported (13), and find active molecules with novel scaffolds compared with current antibiotic compounds. As the predictive ability of machine learning algorithms decreases when using datasets with limited diversity (41) it was surprising to find that our previously developed HTS dataset could be leveraged for training an Al algorithm in the prediction of biological activity.

As will be appreciated by those of skill in the art, it is surprising that any HT screen dataset can be applied to train the Al algorithm to increase the hit rate of further screens. A possible reason is that large size small molecule libraries may contain structural diversity even though they are not specifically curated for that property.

Among the top predicted compounds, we experimentally validated multiple compounds with narrow to broad-range growth inhibitory activity (Tables 1 and 2). To identify compounds with growth inhibitory activity against different bacteria we used five of the six ESKAPE pathogens (56), which include Gram-positive and Gram- negative nosocomial pathogens with multiple antibiotic resistance and high virulence. The strains used were Acinetobacter baumannii 1225, Enterobacter cloacae ENT001_EB001 , Staphylococcus aureus ATCC33592, Klebsiella pneumoniae ESBL_120310 and Pseudomonas aeruginosa PAO1 . We also included B. cenocepacia K56-2 as this was the strain used in the original HTS screen. Compound STL558147 was active against all six species tested. Compound PHAR261659 was active against 4 species: A. baumannii 1225, E. cloacae ENT001_EB001 , P, aeruginosa PAO1 , K. pneumoniae ESBL_120310 but not against S. aureus ATCC33592 and B. cenocepacia K56-2. Other compounds had a more narrow range of activity and inhibited only the Gram-positive S. aureus AT CC33592 (STL546315, STL547239, NP-0192110, STK760075, STL552768). Although 0167-0032 and STK760075 are marked as irritants in the PubChem database, some clinically used antibiotics are also known to cause discomforts at higher concentrations. For example, clinically used common antibiotics meropenem (PubChem CID 441130), penicillin G (PubChem CID 22502), cephalexin (PubChem CID 27447), and ampicillin (PubChem CID 6249) are also labelled as irritants or allergenic by GHS (Globally Harmonized System of Classification and Labelling of Chemicals). Regardless, given their bioactivity against a wide range of clinical pathogens, these compounds can be intriguing candidates for further development. While our study identified multiple bioactive compounds (Tables 1 and 2), STL558147 was the most potent and had broad-spectrum activity (Tables 1 and 2). STL558147

(ZINC001286671837) obtained high probability score of being growth inhibitory (0.833) probably due to its structural similarity to rifampicin (Figure 3A) which was present in the training dataset (14). We are unsure about the origin of STL558147, but it is likely that the compound is synthesized by Amycolatopsis sp. (with unknown isolation source) similar to rifampicin.

With the rise of antimicrobial resistance, combinatorial antibiotic treatments can be an effective therapeutic strategy to prevent antimicrobial resistance. Combinatorial antibiotic strategies achieve the therapeutic effect at relatively lower concentration, decreasing the adverse and toxic effects of high antibiotic concentration and severely restrict the acquisition of drug resistance. Rifampicin is known to enhance the activity of other antibiotics when used in combination (29-32) which is known as synergy. Using B. cenocepacia K56-2 we could observe the synergistic growth inhibitory effect of rifampicin with ceftazidime and colistin but not with polymyxin B (Figure 4B). This is evidenced by the Bliss and Loewe additive scores (d-scores) d-Scores >15 are considered indicative of synergy and are shaded darker in Figure 5. An example of strong synergistic activity is provided with ceftazidime in combination with the B- lactamase inhibitor Avibactam (Figure 5A bottom). Similar to the ceftazidime- avibactam combination, STL558147 demonstrated strong synergistic interactions with ceftazidime, colistin, and Polymyxin B (Figure 4A). Colistin and polymyxin B disrupts the bacterial outer membrane (42), whereas ceftazidime interacts with penicillin binding protein 3 (PBP3) to inhibit cell wall synthesis (43). STL558147 has a relatively high minimum inhibitory concentration (MIC) (256μg/mL) compared to rifampicin (64μg/mL) against B. cenocepacia K56-2. However, STL558147 exerted 2-3 times stronger synergistic growth inhibitory activity with ceftazidime, colistin and polymyxin B at the same drug concentrations (Figure 4), warranting the further development of STL558147 to generate a more potent derivative.

While PHAR261659 exhibited a broad range of bioactivity, the compound was not active against B. cenocepacia K56-2 and methicillin-resistant S. aureus ATCC33592. After screening for analogs of PHAR261659, we found STL529920, a stereoisomer of PHAR261659, to be active against all six pathogens tested. This finding is not surprising as stereoisomers often differ in their biological activity primarily due to stereoselectivity, or specificity of biological systems during the drug- target interactions (44). The drug-target binding involves interactions between the target and its complementary site in the drug molecule. Such interactions may have steric constraints due to the three-dimensional spatial arrangement of the functional groups within the drug molecule and may alter the functionality and/or efficacy. For example, DR-3355, the S isomer of ofloxacin was found to be two times more potent than the S isomer primarily due to stronger binding interactions of the S isomer with its target, DNA gyrase (45).

Overall, our study accentuates the application of a deep learning model to increase the bioactive molecule discovery rate and expand our antibiotic repertoire. This highlights the generalizability of our model, trained with a HTS dataset, to predict molecules with antibacterial activity against a diverse panel of pathogens. Our model demonstrates the promise to train a deep learning model with heterologous screening dataset and extending its ability to predict bioactivity of molecules beyond the organism the screening dataset was generated from.

Our method utilizes a HTS dataset of an antibiotic-resistant bacterium to train a machine learning model in the discovery of antibacterial molecules of broad-range spectrum. We believe that our methodology can be applied to any previously obtained HTS datasets. Bioactivity predictions can be used to prioritize compounds to be tested, which will reduce the library size and increase the hit rate of subsequent screens.

According to an aspect of the invention, there is provided a method of screening a library of test compounds for candidate compounds having a specific biological activity comprising: providing a training library comprising a plurality of training compounds wherein:

(a) the chemical structure of each respective one training compound is known and is represented as Simplified Molecular Input Line Entry System (SMILES) strings in a structural data entry in the training library;

(b) the relative biological activity of each respective one training compound of the training library is known and is stored in a corresponding biological activity data entry in the training library; subjecting the structural data entry and the corresponding biological activity data entry of respective one training compounds of the training library to direct- message passing neural network (D-MPNN) analysis to detect at least one feature of training compounds of the training library corresponding to biological activity above a threshold level, thereby training a machine learning model for identifying candidate compounds; providing a testing library comprising a plurality of testing compounds wherein:

(i) the chemical structure of each respective one testing compound is known and is represented as SMILES strings in a structural data entry in the testing library; and

(ii) the relative biological activity of each respective one testing compound of the testing library is not known; applying the trained machine learning model to the structural data entry of respective one testing compounds of the testing library, thereby identifying candidate testing compounds comprising said at least one feature; and testing said candidate testing compounds for said biological activity. According to another aspect of the invention, there is provided an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5-396799; STL552768; STK802026; STOCK1 N-1 1202; STL513145; STL525490; AG-205/33771006; 8019- 5018; Z1205495496; Z1982493964; Z1029465088; and AS-69879.

In some embodiments, the bacteria is selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus; and P. aeruginosa.

According to another aspect of the invention, there is provided a method for inhibiting growth of bacteria comprising: administering an effective amount of an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414- 0936; STK760075; NAT5-396799; STL552768; STK802026; STOCK1 N-1 1202; STL513145; STL525490; AG-205/33771006; 8019-5018; Z1205495496;

Z1982493964; Z1029465088; and AS-69879, to a region comprising the bacteria, said bacteria growth being inhibited compared to bacteria growing under identical growth conditions except for the presence of the antibacterial compound.

In some embodiments, the bacteria is selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus; and P. aeruginosa.

In some embodiments, growth inhibition is determined by comparing the optical density or the colony forming units from a specific volume, as discussed herein.

As will be appreciated by one of skill in the art, wavelengths for determining optical density of bacterial cultures are well-known in the art.

As will be apparent to one of skill in the art, as used herein in the context of “bacteria growth” or “bacterial growth”, “region” refers to a location capable of and/or known to support bacterial growth.

According to another aspect of the invention, there is provided a method of treating a bacterial infection comprising: administering an effective amount of a compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5-396799; STL552768;

STK802026; STOCK1 N-11202; STL513145; STL525490; AG-205/33771006; 8019- 5018; Z1205495496; Z1982493964; Z1029465088; and AS-69879 to an individual who has or who is suspected of having a bacterial infection.

In some embodiments, the bacteria is selected from the group consisting of: B. cenocepacia; A. baumannii; E. cloacae; K. pneumoniae; S. aureus; methicillin- resistant S. aureus; and P. aeruginosa.

According to another aspect of the invention, there is provided an anti- methicillin-resistant Staphylococcus aureus compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984.

According to another aspect of the invention, there is provided a method for inhibiting growth of methicillin-resistant S. aureus bacteria comprising: administering an effective amount of an antibacterial compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984, to a region comprising the methicillin resistant S. aureus bacteria, said methicillin-resistant S. aureus bacteria growth being inhibited compared to a methicillin-resistant S. aureus bacteria growing under identical growth conditions except for the presence of the antibacterial compound.

In some embodiments, growth inhibition is determined by comparing the optical density or the colony forming units from a specific volume, as discussed herein.

A method of treating a methicillin-resistant S. aureus bacterial infection comprising: administering an effective amount of a compound selected from the group consisting of: STL546315; STL547239; NP-019210; NAT2-252601 ; STL007987; and STL512984, to an individual who has or who is suspected of having a methicillin- resistant S. aureus bacterial infection.

According to another aspect of the invention, there is provided a method of developing an improved antibacterial compound comprising: providing an antibacterial compound selected from the group consisting of: PHAR261659; STL529920; STL558147; 0167-0032; 6414-0936; STK760075; NAT5- 396799; STL552768; STK802026; ST0CK1 N-11202; STL513145; STL525490; AG- 205/33771006; 8019-5018; Z1205495496; Z1982493964; Z1029465088; and AS- 69879; and modifying said antibacterial compound such that one or more properties of antibacterial compound is improved.

As will be apparent to one of skill in the art, the improved property may be for example but by no means limited to improved anti-bacterial activity. Examples of improved anti-bacterial activity include but are by no means limited to: greater antibacterial activity against one or more bacterial strains, increased residency time, increased stability, increased solubility, increased organism range and the like. Other suitable improvements will be well known to those of skill in the art, as will methods for making such modifications.

One example of such a modification is the modification of PHAR261659 to generate bioisosteres. Bioisosteres are compounds where one or more functional groups of a molecule are changed to other functional groups that are of similar size or electronic properties. To date, using the BROOD program from Openeye scientific, 34 bioisosteres of PHAR261659 have been generated. None of the generated compounds are commercially available, however Yan-Ni Sun and coworkers (57) have found a method to reliably synthesize tetrahydro-p-carbolines. This is important as tetrahydro-p-carbolines, as pictured below are the class of compounds that PHAR261659 falls into.

Thescaffold known as tetrahydro-p-carbolines.

The actual synthetic method is shown below:

The structure below is of interest because the carboxylic acid in the structure can easily be functionalized which would allow for the synthesis of analogs of PHAR261659.

It is further of note that this compound is stereoselective, meaning that stereochemistry would not have to be introduced later on in the synthesis.

The invention will now be further explained and/or elucidated by way of examples; however, the invention is not necessarily limited to or by the examples.

EXAMPLE 1 - Training a Deep Learning Model with a B. cenocepacia HTS Dataset

To train our deep learning model, we repurposed a HTS previously performed in our laboratory that searched for growth inhibitory compounds against Burkholderia cenocepacia K56-2 (14). The dataset used in the ML approach consisted of 29,537 compounds with residual growth (RG) values and average B-scores (17). The RG measures the ratio of bacterial growth in the presence and absence of the compounds. The B-score is a measure of relative potency that adjusts the RG for any screening artifacts resulting from well position (row and column) in the assay plate during the HTS. The B-score is inversely proportional to compound potency where negative B-scores indicate greater growth inhibitory activity of the compounds. To binarize the compounds, the previously established average B-score threshold of - 17.5 was chosen (14). Overall, 256 small molecules were classified as growth inhibitory (Figure 1A). Specifically, as known by those of skill in the art, a typical HTS screening threshold to consider a molecule active or inactive depends on the distribution of the activities and the standard deviation (SD) of the distribution. Typically compounds that are 3X SD from the mean are considered to be positive, or, in this case, inhibitory.

Next, the dataset was applied to a directed-message passing neural network (D-MPNN) approach (11) to model the training data and make predictions. The D- MPNN was previously used to train a binary classification deep learning model with 2,335 compounds for predicting antibiotic activity (13). We reasoned that using the D- MPNN training with a larger library (approximately 30,000 compounds) and previously would enhance the model’s ability to learn, generalize and make predictions across compound libraries outside of the training dataset. To maximize the utilization of inactive compounds, we did not enforce class balance while training. The trained model generated scores between 0 and 1 for each molecule, with 1 indicating the highest probability for growth inhibitory activity. In addition to the binary classification, we also conducted regression tasks on both the average B-score and RG, where the two potency measurements of activity against B. cenocepacia were used simultaneously to train multi-task models.

To enhance the accuracy of the predictions, the model was additionally supplemented with molecular fingerprints or descriptors in the first layer of FFN (Figure 1 B). We also applied two standard machine learning optimization strategies to increase the robustness of the models: Bayesian hyperparameter optimization (18) and ensembling (19). The iteration of Bayesian optimization was set to 20 iterations and the ensemble size was equal to 5 for each model. Four models with different combinations of molecule-level features (Table 3) were trained and evaluated for classification and regression tasks. In addition, two different splitting strategies were used to build the ML models, a scaffold split and a random split. As our model, D- MPNN is a graph-based model, each compound is first featurized in graph representations. However, to further improve the model, we can assist the model with additional global features. These global features include and are not limited to: chemical descriptors, Morgan fingerprints, MACCS keys, and the like.

As can be seen in Table 3, we made four feature combinations from three types of global features, and tested on them. The one that performs best on the test set then served as our main model

The scaffold split separates samples into subsets based on molecular scaffolds. From the 8 different combinations, model 6 (In the binary classification with scaffold-based data split settings, Model 6 (binary classification, scaffold split trained with RDKit descriptors) achieved the highest area under curve (AUG) of the precision recall curve PRC (PRC-AUC = 0.241 ; F1 Score = 0.104) on the test set and was therefore selected as the primary model for our subsequent experiments). Model 6 also attained the second highest ROC-AUC score (0.823) among other baselines in the same category (Figure 1C).

EXAMPLE 2 - Predicting and Confirming Growth Inhibitory Activity in Virtual Compound Libraries

To validate the model’s ability to make predictions on compounds outside of the training dataset, we employed Model 6 on an FDA-approved compound library containing 1 ,614 compounds that were not present in the training set and predicted their growth inhibitory activity. The model generated a single value between 0 and 1 for each molecule, indicating the probability of the molecule being active. The 100 top- ranked compounds contained a large fraction (-48.75%) of antibiotics. After removal of duplicated compounds, we tested 81 commercially available compounds for growth inhibitory activity against B. cenocepacia K56-2 (Figure 2A). These 81 compounds were also enriched in known antibiotics, antimicrobials, and antineoplastic agents. For experimental validation, we defined as inhibitory those compounds that inhibited at least 20% of normal growth. Therefore, our threshold of residual growth (RG) was 0.8. As will be apparent to those of skill in the art, this threshold is in effect an arbitrary number, which can be varied, depending on the desired sensitity, that is, the desired number of candidate compounds to be identified. We reasoned that establishing a permissive RG threshold would allow us to capture compounds with potential antimicrobial activity that could be further optimized for potency. 21 compounds were experimentally validated as growth-inhibitory, establishing a positive predictive value (PPV) of 26% (Figure 2B). Overall, 14 of the 21 compounds exhibiting growth inhibitory activity were known antibiotics or antimicrobial compounds (Figure 2C). A correlation (R ² = 0.54) was found between bioactivity and the predicted rank. Compounds that ranked higher based on the prediction score displayed stronger growth inhibitory activity (Figure 2D).

Using the same approach, we predicted growth inhibitory activity in a natural product virtual library containing 224,205 compounds. We ranked the compounds based on their predicted score and filtered out the compounds with previously reported antimicrobial activity or toxicity. From the 100 top-ranked compounds, 43 compounds were tested against B. cenocepacia K56-2. The screening yielded 4 previously uncharacterized small molecules with growth inhibitory activity against B. cenocepacia K56-2, achieving a PPV of 12% (Tables 1 and 2). Compared to the hit rate from conventional whole-cell based high-throughput screens (<1%) (20), our strategy improved the hit rate at least 12-fold, suggesting that our in silico approach can substantially minimize costs and time associated to compound screening.

To examine the predictive power of the deep learning model for a broad-range antibacterial activity we tested the 43 curated compounds from the natural product library against five ESKAPE pathogens: Acinetobacter baumannii, Enterobacter cloacae, Methicillin-resistant Staphylococcus aureus (MRSA), Klebsiella pneumoniae ESBL_120310 and Pseudomonas aeruginosa. The screening identified 22 compounds with growth inhibitory activity against at least one of the species tested (PPV = 51.16%; Tables 1 and 2). While 13 compounds were active against Gram- positive MRSA ATCC33592, three, nine, five and nine compounds were active against E. cloacae ENT001_EB001 , P. aeruginosa PA01 , K. pneumoniae ESBL_120310 and A. baumannii 225, respectively (Tables 1 and 2). Among the inhibitory compounds identified from the screening, STK760075 (ZINC000004524372), PHAR261659 (ZINC000008876405), 0167-0032 (ZINC000000254599), 6414-0936 (ZINC000000465709) and STL558147 (ZINC001286671837) displayed the widest range of activity against the pathogens tested (Tables 1 and 2). To the best of our knowledge, none of these compounds have previously been reported to have growth inhibitory activity. STK760075 exhibited activity against B. cenocepacia K56-2, MRSA ATCC33592, and P. aeruginosa PAO1 whereas PHAR261659 was active against E. cloacae ENT001_EB001 , P. aeruginosa PAO1 , K. pneumoniae ESBL_120310 and A. baumannii 1225 (Tables 1 and 2). Compound 6414-0936 exerted antibiotic activity against four of the six pathogens tested (except B. cenocepacia K56-2 and E. cloacae ENT001_EB001). 0167-0032 had inhibitory activity against all of the pathogens tested, except, K. pneumoniae ESBL_120310. Compound STL558147 was active against all six species tested.

While 0167-0032 (PubChem CID 1560156) and STK760075 (PubChem CID 1549520) are annotated as irritants in the PubChem database, no known toxicity data is available for other compounds with broad-range activity. For this study, we selected STL558147 and PHAR 261659 for further characterization since they exhibited the strongest growth inhibitory activity against most pathogens tested (Tables 1 and 2).

EXAMPLE 3 - Antibacterial Properties of STL558147

To confirm the identity of STL558147, we performed nuclear magnetic resonance (NMR) spectroscopy and compared the spectrum with that of STL558147 provided by the supplier. We found that STL558147 and the antibiotic rifampicin share a similar scaffold (Maximum common substructure Tanimoto score 0.6389) but are different compounds (Figure 3A).

Rifampicin targets the [3 subunit of the bacterial DNA-dependent RNA polymerase (21 , 22), encoded by rpoB. To test if STL558147 targets RpoB, we explored the link between RpoB expression and susceptibility to STL558147, as target depletion often sensitizes cells to cognate antimicrobials (23). We created four knockdown mutants of the rpoBC operon in B. cenocepacia K56-2 (K562_RS01210- 5) using CRISPR interference (CRISPRi) (24). The CRISPRi system developed for Burkholderia comprises of a chromosomally integrated dCas9 from Streptococcus pyogenes placed under the control of a rhamnose-inducible promoter and plasmid- borne target-specific single guide RNA (sgRNA) driven by a constitutively active synthetic promoter PJ23119 (25). Addition of rhamnose induces dCas9 expression which binds to sgRNA and sterically blocks the transcription of the target specified by the sgRNA. Four B. cenocepacia K56-2 gyrB CRISPRi knockdown mutants (K562_RS02180) were included as a negative control. When grown at the rhamnose concentration that inhibits 50% of growth (Rha IC ₅₀ ) compared to wild-type, the gyrB CRISPRi mutant was more susceptible to novobiocin than K56-2 and non-targeting controls (Figure 3B). This was expected as GyrB is the target of novobiocin (26). Similarly, the rpoBC knockdown mutant exhibited enhanced sensitivity against both rifampicin (Figure 3C) and STL558147 (Figure 3D). The hypersensitivity of the rpoBC CRISPRi mutant against both rifampicin and STL55817 suggests that STL558147 has similar mechanism of action to rifampicin and RpoB is the likely in vivo target of STL558147.

Antibiotic combinatorial therapy has been a common strategy to enhance the efficacy of the antibiotics (27, 28). Rifampicin is known to have synergistic interactions with colistin and meropenem in vitro against Pseudomonas spp., Acinetobacter spp., and carbapenemase-producing Enterobacteriaceae (29-32). To elucidate if STL558147 exerts synergistic interactions with clinically relevant antibiotics against B. cenocepacia K56-2, we performed a microdilution checkerboard assay. Using the Bliss interaction score and Loewe additivity score, we considered scores >15 as synergistic and < -15 as antagonistic (33). We observed a strong synergistic interaction when STL558147 was combined with either ceftazidime, colistin or polymyxin B (Figure 4A) similar to the synergistic interaction observed to ceftazidime and avibactam combination (Figure 4B). Compared to rifampicin, the observed synergistic activity of STL558147 was 2-3 times stronger than the synergistic activity observed for rifampicin with ceftazidime, colistin and polymyxin B (Figure 4C). The observed synergistic activity of rifampicin was weaker than the commonly used drug pair ceftazidime and avibactam (Figure 4). When combined STL558147 with these antibiotics against B. cenocepacia K56-2, we observed a similar growth inhibitory activity at 8-fold lower STL558147 concentration (compared to STL558147 MIC) (Figure 4A). We observed no interaction between STL558147 and rifampicin or rifabutin. This is expected, as both rifampicin and rifabutin target RpoB which is also the likely in vivo target of STL558147. On the contrary, the combination of STL559147 with ciprofloxacin is antagonistic.

EXAMPLE 4 - PHAR261659 Exhibited Broad-spectrum Antibacterial Activity

The screening of the 43 compounds from the top 100 ranked natural products revealed that PHAR261659, a compound with no previously reported antibiotic activity, has antibacterial activity against a broad range of pathogens (Tables 1 and 2). Specifically, PHAR261659 exhibited growth inhibitory activity against A. baumannii 1225, E. cloacae ENT001_EB001 and K. pneumoniae ESBL_120310 (Tables 1 and 2). However, we did not observe growth inhibitory activity against B. cenocepacia K56-2 and S. aureus AT CC33592 (MRSA). Moreover, while PHAR261659 displayed bioactivity at the screening concentration (50μM), the compound was not soluble beyond 128μM. However, we found 15 unique analogs with modified side chains and lower logP values in the natural product library and tested then against B. cenocepacia K56-2 and the ESKAPE pathogens. While PHAR261659 did not show bioactivity against B. cenocepacia K56-2 and methicillin-resistant S. aureus ATCC33592 (Tables 1 and 2), five and ten analogs exhibited growth inhibitory activity against these pathogens (Figure 5). Particularly, STL529920 (ZINC000008876407), a stereoisomer of PHAR261659, was active against all six pathogens tested (Figure 5). Both STL529920 and PHAR261659 were predicted to be active by our deep learning algorithm with scores of 0.48 and 0.49, respectively.

EXAMPLE 5 - Determination of MOA of STL529920: Essential gene mutant strains that are depleted in essential proteins become hypersensitive to their cognate antibiotics (2). To elucidate the mechanism of action of

STL529920 we exposed a pooled CRISPR nterference (CRISPRi) mediated

Burkhoderia cenocepacia K56-2 essential gene mutant library to 50μM of STL529920.

Upon sequencing of the sample (CRISPRi-Seq; Figure 6A), we observed significant depletion of a few mutants. Peptidyl-tRNA hydrolase (pth) was the most significantly depleted among them (Figure 6B), suggesting that PHAR261659 may inhibit bacterial growth by targeting bacterial essential protein, Peptidyl-tRNA hydrolase (pth). To validate the CRISPRi-Seq results, we performed enhanced susceptibility assays by clonally exposing the depleted mutants (from CRISPRi-Seq) against STL529920. A pth CRISPRi mutants was the most sensitive to STL529920 compared to the non- genome targeting control and K56-2 wild-type (Figure 6C). Other mutants that were partially sensitive to STL529920 were aminopeptidase P, mnmA, rpsA, and waaA. Products of the pth, aminopeptidase P, mnmA, rpsA genes involved in protein translation suggesting that STL529920 inhibits bacterial growth by inhibiting translation, specifically by inhibiting peptidyl-tRNA hydrolase.

EXAMPLE 6 - Methods

Deep learning model details

The code of D-MPNN used in this study is implemented in the package Chemprop (46), which was built based on the architecture proposed by Gilmer etal., named message passing neural network (MPNN) (47). In general, MPNNs take atom and bond features as inputs and aggregate the features through a message passing phase and a readout phase (47). Different from the message passing mechanism of MPNN, in which molecular messages are centered on atoms, D-MPNN propagates molecular information through neighboring bonds with directions. As D-MPNN aggregates messages associated with directed bonds, it can avoid unnecessary loops and totters. This property allows D-MPNN become more efficient during training and is able to construct a more informative featurization with less noise. The readout phase of D-MPNN follows the same paradigm of typical MPNNs, in which hidden states for atoms in a molecule are aggregated together and form a molecule-level representation.

Supplementary molecular features

Due to the limited ability of extracting molecule-level features in the message passing phase, especially in case of large molecules, we choose to supplement the learned graphical representations from D-MPNN with calculated molecule-level features. In this study, we investigated three types of molecular features as additional auxiliary to D-MPNN architecture: binary Morgan fingerprints, count-based Morgan fingerprints, and 200 chemical descriptors extracted with the chemoinformatics software RDKit (40). However, as discussed above, we tested 4 combinations of features and selected the features based on their individual performance. The reason for using Morgan fingerprints is that Morgan fingerprints are embedded with the topological characteristics of molecules and is one of the most common molecular representations for property prediction. The reason for using 200 RDKit descriptors is that they are comprehensive cheminformatics descriptors which include a broad spectrum of chemical properties at the molecular level, thus providing a rich source of chemical information on multiple aspects. To avoid the effect of large range additional features dominating other features, normalization was applied to the additional features to scale the values to a fixed range before entering into networks. In particular, min-max scaling was applied to normalize count-based Morgan fingerprints and 200 RDKit descriptors were normalized by fitting to the cumulative density functions.

Bayesian optimization

To maximize the performance, we performed Bayesian hyperparameter optimization using Hyperopt package (48). We optimized four hyperparameters in D- MPNN: the number of the message passing steps, hidden size of the neural network, the number of FFN layers, and dropout probability. The Bayesian hyperparameter optimization uses the results of prior trials to make informed decisions for what parameter values to try in the next trial. Compared to the grid search method, this method requires fewer iterations to find the optimal hyperparameters.

Ensembling

Ensembling, a commonly used technique in machine learning for improving model's performance, was also applied in our training process. The ensemble method combines the predicted results from several identically structured models with different initial weights. That is, models are trained independently and separately, then the prediction values were averaged with equal weights, resulting the final prediction (19).

Deep Learning Model Training and Prediction

The deep learning model was trained with a HTS dataset that used the Canadian Compound Collection (CCC) against B. cenocepacia K56-2 (14). The training dataset consisted of 29,537 compounds. The textual molecular representation from the SMILES strings were transformed into numeric representations to generate molecular features during training. The SMILES strings are converted to numerical features by using algorithms integrated in cheminformatics software RDKit. Similar to Selin et al., (14), we used Residual Growth (RG) (< 0.7 RG) and average B-Score (< - 17.5) to call growth inhibitory activity of the compounds against B. cenocepacia K56- 2.

For the classification task, we used the average B-Score -17.5 as the bioactivity threshold. Compounds with average B-Score of less than -17.5 were considered as growth inhibitory, which resulted in 256 active compounds. Although the data is severely imbalanced, to maximize the utilization of data, we decided against enforcing class balance on data during training. With the determined threshold, the active compounds are labeled as 1 while the rest are labeled as 0, then used as targets for classification training. After training, the D-MPNN output was a single value between 0 and 1 for each molecule, indicating the probability of the compound to have growth inhibitory activity. For the regression task of D-MPNN both the average B-Score and RG were used in the training together but evaluated separately in the results.

Data Splits

In this work, data was divided into training, validation, and testing sets follows the ratio 80:10:10 with the data split strategy named scaffold split. Different from the random split that is commonly used in ML, the scaffold split is specifically designed for the QSPR/QSAR tasks, where the Murcko scaffold for each molecule is calculated and used during the splitting process (49). Assigning compounds to data bins based on Scaffold scores enforces validation and testing sets to have more molecular diversities and reduces similarities between data sets. Therefore, the scaffold split strategy has a more realistic and challenging evaluation comparing to the random split (11). Besides the scaffold split, we also trained models on data bins that were randomly partitioned as comparison.

Evaluation Methods

We evaluated our binary classification models by three metrics: ROC-AUC, PRC-AUC, and F1 score after binarizing the results. The cut-off of binarization for each model was determined by the Youden's J statistic on the ROC curve of validation set, which is defined as:

After obtaining optimum cut-offs, we then computed F1 scores to further evaluate the performance of classifiers. F1 Scores were calculated using the following formula:

Bacterial Strains and Growth Conditions

Unless otherwise indicated, all the strains were grown at 37°C in Luria-Bertani (LB) media with shaking (230 rpm). Compound Screening

Overnight cultures of the strains were back diluted to OD _600nm of 0.036 in fresh LB medium. 100μL of cells were then arrayed onto flat-bottomed 96-well plates (Greiner Bio-One or Sarstedt, Inc.) containing 100μL of 100μM compounds, resulting in total volume of 200μL at final compound concentration of 50μM and final OD _600nm of the strains 0.018. All compounds were dissolved in neat DMSO. The plates were incubated at 37°C for 5 hours and OD _600nm readings were taken using BioTeK Synergy 2 plate reader. Residual Growth (RG) was calculated based on the ratio of growth in the presence of compounds and growth in the DMSO control. The identity of the STL558147 was confirmed with nuclear magnetic resonance (NMR) spectroscopy. The obtained NMR spectrum was compared with those of STL558147 (provided by the supplier) and rifampicin.

CRISPRi Mutant Construction and Rhamnose IC ₅₀ Determination

CRISPRi knockdown mutants targeting the rpoBC were created as previously mentioned (25). The conditional growth phenotype was determined by performing growth kinetics. Overnight cultures of the mutants were back diluted to OD _600nm 0.01 and arrayed onto 96-well plate containing LB broth with trimethoprim 100μg/mL and with or without 1 % rhamnose at 37°C with shaking (230 rpm). OD _600nm readings were taken using BioTek Synergy 2 plate reader at 1 -hour interval for 22-24h.

Rhamnose concentration that inhibits 50% growth of the mutants (Rha IC ₅₀ ) compared to the wild-type was determined by growing the mutants in a rhamnose gradient at 37°C with shaking (230 rpm) in 96-well format. OD _600nm reading was taken after 20 hours of growth and dose-response curve was created with Graphpad PRISM version 6.0.0. Rha IC ₅₀ values were calculated from the rhamnose dose-response curve using the Hill coefficient of the equation.

Enhanced Sensitivity Assay

Overnight cultures of the CRISPRi mutants were diluted to 1 :100 and subcultured with Rha IC ₅₀ (in LB with trimethoprim 100μg/mL) for 4h at 37°C with shaking (230 rpm). After 4h, OD _600nm of the cultures were adjusted to 0.01 and grown in 96-well format containing LB broth supplemented with trimethoprim 100μg/mL, rhamnose (Rha IC ₅₀ ) and various concentrations of STL558147, novobiocin, or rifampicin. The plates were incubated for 20-22h at 37°C with shaking (230 rpm). OD _600nm readings were taken using BioTek Synergy 2 microplate reader.

Checkerboard Assay and Synergy Calculation

Checkerboard assay was performed as described before (50, 51 ). Briefly, overnight cultures of the studied strains were back diluted to the equivalent 0.5 McFarland standard. Diluted cultures were further diluted to 1 :100 and inoculated into 96-well plates containing two-dimensional gradient of the STL558147 and various antibiotics, starting from half of the MICs. Synergy scores were calculated using SynergyFinder 2.0 web-application (33). Effect of the drug combinations (synergistic or antagonistic) were calculated comparing the observed responses against expected response computed using a reference model. Bliss (52), Loewe (53), highest single agent (HSA) (54) and zero interaction potency (ZIP) (55) were used as reference models to determine the degree of interaction. The drug combinations with synergy scores above 15 calculated by all models were considered synergistic; between -5 and 15 were considered additive, whereas synergy scores below -15 were considered antagonistic. Synergy scores represent the mean response deviated from the reference model due to interactions between the combined drugs.

While the preferred embodiments of the invention have been described above, it will be recognized and understood that various modifications may be made therein, and the appended claims are intended to cover all such modifications which may fall within the spirit and scope of the invention. Table 1: Compounds from the natural product library that exhibited broad-spectrum growth inhibitory activity.

Note: The numbers show the residual growth (RG) of the organisms after 5h of exposure to the compounds in LB media. RG is the ratio of growth (measured as optical density) in the presence and absence of the compound. Results are average of at least three

Table 2: Compounds from the natural product library that exhibited growth inhibitory activity against Gram-positive Methicillin- resistant Staphylococcus aureus (MRSA) only. Note: The numbers show the residual growth (RG) of the organisms after 5h of exposure to the compounds in LB media. RG is the ratio of growth (measured as optical density) in the presence and absence (DMSO) of compound. Results are average of at

least three independent biological replicates. Compounds with RG < 0.8 were considered active and >0.8 were inactive (shown in bold).

Table 3: Deep learning models trained with different combinations of molecule- level features.

REFERENCES

(1 ) CDC. Antibiotic Resistance Threats in the United States; U.S. Department of Health and Human Services, CDC: Atlanta, GA, 2019.

(2) Billington, J. K. A New Product Development Partnership Model for Antibiotic

Resistance. Am. J. Law Med. 2016, 42 (2-3), 487-523. https://doi.Org/10.1177/0098858816658277.

(3) Brown, E. D.; Wright, G. D. Antibacterial Drug Discovery in the Resistance Era. Nature 2016, 529 (7586), 336-343. https://doi.org/10.1038/nature17042.

(4) Payne, D. J.; Gwynn, M. N.; Holmes, D. J.; Pompliano, D. L. Drugs for Bad Bugs: Confronting the Challenges of Antibacterial Discovery. Nat. Rev. Discov. 2007, 6 (1 ), 29-40. https://doi.org/10.1038/nrd2201.

(5) Zgurskaya, H. I.; Lopez, C. A.; Gnanakaran, S. Permeability Barrier of Gram- Negative Cell Envelopes and Approaches To Bypass It. ACS Infect. Dis. 2015, 1 (11), 512-522. https://d0i.0rg/l 0.1021 /acsinfecdis.5b00097.

(6) Angermueller, C.; Parnamaa, T.; Parts, L.; Stegle, O. Deep Learning for

Computational Biology. Mol. Syst. Biol. 2016, 12 (7), 878. https://d0i.0rg/l 0.15252/msb.2O156651 .

(7) Tang, B.; Pan, Z.; Yin, K.; Khateeb, A. Recent Advances of Deep Learning in Bioinformatics and Computational Biology. Front. Genet. 2019, 10, 214. https://d0i.0rg/l 0.3389/fgene.2019.00214.

(8) Villoutreix, P. What Machine Learning Can Do for Developmental Biology. Development 2021 , 148 (1 ), dev188474. https://doi.org/10.1242/dev.188474.

(9) Chan, H. C. S.; Shan, H.; Dahoun, T.; Vogel, H.; Yuan, S. Advancing Drug Discovery via Artificial Intelligence. Trends Pharmacol. Sci. 2019, 40 (8), 592-604. https://d0i.0rg/l 0.1016/j. tips.2019.06.004.

(10) Zhou, J.; Cui, G.; Zhang, Z.; Yang, C.; Liu, Z.; Wang, L.; Li, C.; Sun, M. Graph Neural Networks: A Review of Methods and Applications. ArXivI 81208434 Cs Stat 2019. (11) Yang, K.; Swanson, K.; Jin, W.; Coley, C.; Eiden, P.; Gao, H.; Guzman-Perez,

A.; Hopper, T.; Kelley, B.; Mathea, M.; Palmer, A.; Settels, V.; Jaakkola, T.; Jensen, K.; Barzilay, R. Analyzing Learned Molecular Representations for Property Prediction. J. Chem. Inf. Model. 2019, 59 (8), 3370-3388. https://d0i.0rg/l 0.1021 /acs.jcim.9b00237.

(12) Melo, M. C. R.; Maasch, J. R. M. A.; de la Fuente-Nunez, C. Accelerating Antibiotic Discovery through Artificial Intelligence. Common. Biol. 2021 , 4 (1 ), 1050. https://d0i.0rg/l 0.1038/S42003-021 -02586-0.

(13) Stokes, J. M.; Yang, K.; Swanson, K.; Jin, W.; Cubillos-Ruiz, A.; Donghia, N.

M.; MacNair, C. R.; French, S.; Carfrae, L. A.; Bloom-Ackerman, Z.; Tran, V. M.; Chiappino-Pepe, A.; Badran, A. H.; Andrews, I. W.; Chory, E. J.; Church, G. M.; Brown, E. D.; Jaakkola, T. S.; Barzilay, R.; Collins, J. J. A Deep Learning Approach to Antibiotic Discovery. Cell 2020, 180 (4), 688-702. e13. https://d0i.0rg/l 0.1016/j. cell.2020.01 .021 .

(14) Selin, C.; Stietz, M. S.; Blanchard, J. E.; Gehrke, S. S.; Bernard, S.; Hall, D. G.; Brown, E. D.; Cardona, S. T. A Pipeline for Screening Small Molecules with Growth Inhibitory Activity against Burkholderia Cenocepacia. PLOS ONE 2015, 10 (6), e0128587. https://doi.Org/10.1371 /journal. pone.0128587.

(15) Loutet, S. A.; Valvano, M. A. A Decade of Burkholderia Cenocepacia Virulence

Determinant Research. Infect. Immun. 2010, 78 (10), 4088-4100. https://doi.Org/10.1128/IAI.00212-10.

(16) Rice, L. B. Federal Funding for the Study of Antimicrobial Resistance in Nosocomial Pathogens: No ESKAPE. J. Infect. Dis. 2008, 197 (8), 1079-1081. https://d0i.0rg/l 0.1086/533452.

(17) Brideau, C.; Gunter, B.; Pikounis, B.; Liaw, A. Improved Statistical Methods for Hit Selection in High-Throughput Screening. J. Biomol. Screen. 2003, 8 (6), 634-647. https://doi.Org/10.1177/1087057103258285.

(18) Shahriari, B.; Swersky, K.; Wang, Z.; Adams, R. P.; de Freitas, N. Taking the Human Out of the Loop: A Review of Bayesian Optimization. Proc. IEEE 2016, 104 (1 ), 148-175. https://doi.org/10.1109/JPROC.2015.2494218.

(19) Dietterich, T. G. Ensemble Methods in Machine Learning. In Multiple Classifier Systems; Goos, G., Hartmanis, J., van Leeuwen, J., Series Eds.; Lecture Notes in Computer Science; Springer Berlin Heidelberg: Berlin, Heidelberg, 2000; Vol. 1857, pp 1-15. https://doi.org/10.1007/3-540-45014-9_1.

(20) Zhu, T.; Cao, S.; Su, P.-C.; Patel, R.; Shah, D.; Chokshi, H. B.; Szukala, R.; Johnson, M. E.; Hevener, K. E. Hit Identification and Optimization in Virtual Screening: Practical Recommendations Based Upon a Critical Literature Analysis. J. Med. Chem. 2013, 56 (17), 6560-6572. https://doi.org/10.1021/jm301916b.

(21 ) Campbell, E. A.; Korzheva, N.; Mustaev, A.; Murakami, K.; Nair, S.; Goldfarb,

A.; Darst, S. A. Structural Mechanism for Rifampicin Inhibition of Bacterial RNA Polymerase. Cell 2001 , 104 (6), 901-912. https://doi.org/10.1016/S0092-

8674(01 )00286-0.

(22) Hartmann, G.; Honikel, K. O.; Knusel, F.; Nuesch, J. The Specific Inhibition of the DNA-Directed RNA Synthesis by Rifamycin. Biochim. Biophys. Acta 1967, 145 (3), 843-844. https://doi.Org/10.1016/0005-2787(67)90147-5.

(23) Cardona, S. T.; Selin, C.; Gislason, A. S. Genomic Tools to Profile Antibiotic

Mode of Action. Crit. Rev. Microbiol. 2015, 4, 465-472. https://d0i.0rg/l 0.3109/1040841 X.2013.866073.

(24) Qi, L. S.; Larson, M. H.; Gilbert, L. A.; Doudna, J. A.; Weissman, J. S.; Arkin, A. P.; Lim, W. A. Repurposing CRISPR as an RNA-Guided Platform for Sequence- Specific Control of Gene Expression. Cell 2013, 152 (5), 1173-1183. https://d0i.0rg/l 0.1016/j.cell .2013.02.022 [doi],

(25) Hogan, A. M.; Rahman, A. S. M. Z.; Lightly, T. J.; Cardona, S. T. A Broad-Host- Range CRISPRi Toolkit for Silencing Gene Expression in Burkholderia. ACS Synth. Biol. 2019, 8 (10), 2372-2384. https://doi.org/10.1021/acssynbio.9b00232.

(26) Lewis, R. J.; Tsai, F. T.; Wigley, D. B. Molecular Mechanisms of Drug Inhibition of DNA Gyrase. BioEssays News Rev. Mol. Cell. Dev. Biol. 1996, 18 (8), 661-671. https://d0i.0rg/l 0.1002/bies.950180810.

(27) Coates, A. R. M.; Hu, Y.; Holt, J.; Yeh, P. Antibiotic Combination Therapy against Resistant Bacterial Infections: Synergy, Rejuvenation and Resistance Reduction. Expert Rev. Anti Infect. Ther. 2020, 18 (1 ), 5-15. https://doi.Org/10.1080/14787210.2020.1705155.

(28) Tängden, T. Combination Antibiotic Therapy for Multidrug-Resistant Gram-

Negative Bacteria. Ups. J. Med. Sci. 2014, 119 (2), 149-153. https://d0i.0rg/l 0.3109/03009734.2014.899279.

(29) Biancofiore, G.; Tascini, C.; Bisa, M.; Gemignani, G.; Bindi, M. L.; Leonildi, A.; Giannotti, G.; Menichetti, F. Colistin, Meropenem and Rifampin in a Combination Therapy for Multi-Drug-Resistant Acinetobacter Baumannii Multifocal Infection. A Case Report. Minerva Anestesiol. 2007, 73 (3), 181-185.

(30) Morelli, P.; Ferrario, A.; Tordato, F.; Piazza, A.; Casari, E. Successful

Treatment of Post-Neurosurgical Multidrug-Resistant Pseudomonas Aeruginosa Meningo-Encephalitis with Combination Therapy of Colistin, Rifampicin and Doripenem. J. Antimicrob. Chemother. 2014, 69 (3), 857-859. https://d0i.0rg/l 0.1093/jac/dkt437.

(31 ) Pournaras, S.; Vrioni, G.; Neou, E.; Dendrinos, J.; Dimitroulia, E.; Poulou, A.; Tsakris, A. Activity of Tigecycline Alone and in Combination with Colistin and Meropenem against Klebsiella Pneumoniae Carbapenemase (KPC)-Producing Enterobacteriaceae Strains by Time-Kill Assay. Int. J. Antimicrob. Agents 2011 , 37 (3), 244-247. https://doi.org/10.1016/j-ijantimicag.2010.10.031 .

(32) Urban, C.; Mariano, N.; Rahal, J. J. In Vitro Double and Triple Bactericidal Activities of Doripenem, Polymyxin B, and Rifampin against Multidrug-Resistant Acinetobacter Baumannii, Pseudomonas Aeruginosa, Klebsiella Pneumoniae, and Escherichia Coli. Antimicrob. Agents Chemother. 2010, 54 (6), 2732-2734. https://d0i.0rg/l 0.1128/AAC.01768-09.

(33) lanevski, A.; Giri, A. K.; Aittokallio, T. SynergyFinder 2.0: Visual Analytics of Multi-Drug Combination Synergies. Nucleic Acids Res. 2020, 48 (W1 ), W488-W493. https://d0i.0rg/l 0.1093/nar/gkaa216.

(34) McKinney, S. M.; Sieniek, M.; Godbole, V.; Godwin, J.; Antropova, N.;

Ashrafian, H.; Back, T.; Chesus, M.; Corrado, G. S.; Darzi, A.; Etemadi, M.; Garcia- Vicente, F.; Gilbert, F. J.; Halling-Brown, M.; Hassabis, D.; Jansen, S.; Karthikesalingam, A.; Kelly, C. J.; King, D.; Ledsam, J. R.; Melnick, D.; Mostofi, H.; Peng, L.; Reicher, J. J.; Romera-Paredes, B.; Sidebottom, R.; Suleyman, M.; Tse, D.; Young, K. C.; De Fauw, J.; Shetty, S. International Evaluation of an Al System for Breast Cancer Screening. Nature 2020, 577 (7788), 89-94. https://d0i.0rg/l 0.1038/S41586-019-1799-6.

(35) Artzi, N. S.; Shilo, S.; Hadar, E.; Rossman, H.; Barbash-Hazan, S.; Ben- Haroush, A.; Balicer, R. D.; Feldman, B.; Wiznitzer, A.; Segal, E. Prediction of Gestational Diabetes Based on Nationwide Electronic Health Records. Nat. Med. 2020, 26 (1 ), 71-76. https://doi.org/10.1038/s41591 -019-0724-8.

(36) Zitnik, M.; Nguyen, F.; Wang, B.; Leskovec, J.; Goldenberg, A.; Hoffman, M. M.

Machine Learning for Integrating Data in Biology and Medicine: Principles, Practice, and Opportunities. Inf. Fusion 2019, 50, 71-91. https://d0i.0rg/l 0.1016/j. i nff us.2018.09.012. (37) Macesic, N.; Polubriaginof, F.; Tatonetti, N. P. Machine Learning: Novel Bioinformatics Approaches for Combating Antimicrobial Resistance. Curr. Opin. Infect. Dis. 2017, 30 (6), 511-517. https://doi.org/10.1097/QCC.0000000000000406.

(38) Richter, M. F.; Drown, B. S.; Riley, A. P.; Garcia, A.; Shirai, T.; Svec, R. L.;

Hergenrother, P. J. Predictive Compound Accumulation Rules Yield a Broad- Spectrum Antibiotic. Nature 2017, 545 (7654), 299-304. https://doi.Org/10.1038/nature22308.

(39) Lavecchia, A. Machine-Learning Approaches in Drug Discovery: Methods and

Applications. Drug Discov. Today 2015, 20 (3), 318-331. https://d0i.0rg/l 0.1016/j.drudis.2O14.10.012.

(40) Landrum, G. RDKit: Open-Source Cheminformatics; 2006.

(41 ) Glavatskikh, M.; Leguy, J.; Hunault, G.; Cauchy, T.; Da Mota, B. Dataset’s Chemical Diversity Limits the Generalizability of Machine Learning Predictions. J. Cheminformatics 2019, 11 (1 ), 69. https://d0i.0rg/l 0.1186/s13321 -019-0391 -2.

(42) Yu, Z.; Qin, W.; Lin, J.; Fang, S.; Qiu, J. Antibacterial Mechanisms of Polymyxin and Bacterial Resistance. BioMed Res. Int. 2015, 2015, 1-11. https://doi.Org/10.1155/2015/679109.

(43) Hayes, M. V.; Orr, D. C. Mode of Action of Ceftazidime: Affinity for the Penicillin-Binding Proteins of Escherichia Coli K12, Pseudomonas Aeruginosa and Staphylococcus Aureus. J. Antimicrob. Chemother. 1983, 12 (2), 119-126. https://d0i.0rg/l 0.1093/jac/12.2.119.

(44) Hutt, A. J.; O’Grady, J. Drug Chirality: A Consideration of the Significance of the Stereochemistry of Antimicrobial Agents. J. Antimicrob. Chemother. 1996, 37 (1 ), 7-32. https://doi.Org/10.1093/jac/37.1 .7.

(45) Morrissey, I.; Hoshino, K.; Sato, K.; Yoshida, A.; Hayakawa, I.; Bures, M. G.; Shen, L. L. Mechanism of Differential Activities of Ofloxacin Enantiomers. Antimicrob. Agents Chemother. 1996, 40 (8), 1775-1784. https://doi.Org/10.1128/AAC.40.8.1775.

(46) Yang, K.; Swanson, K.; Jin, W.; Coley, C.; Eiden, P.; Gao, H.; Guzman-Perez, A.; Hopper, T.; Kelley, B.; Mathea, M.; Palmer, A.; Settels, V.; Jaakkola, T.; Jensen, K.; Barzilay, R. Chemprop: Message Passing Neural Networks for Molecule Property Prediction; 2019.

(47) Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; Dahl, G. E. Neural Message Passing for Quantum Chemistry. ArXiv170401212 Cs 2017.

(48) Bergstra, J.; Komer, B.; Eliasmith, C.; Yamins, D.; Cox, D. D. Hyperopt: A Python Library for Model Selection and Hyperparameter Optimization. Comput. Sci. Discov. 2015, 8 (1 ), 014008. https://doi.Org/10.1088/1749-4699/8/1/014008.

(49) Wu, Z.; Ramsundar, B.; Feinberg, E. N.; Gomes, J.; Geniesse, C.; Pappu, A. S.; Leswing, K.; Pande, V. MoleculeNet: A Benchmark for Molecular Machine Learning. Chem. Sci. 2018, 9 (2), 513-530.

(50) Garcia, Lynne S. Clinical Microbiology Procedures Handbook, Volumes 1 -3 (3rd Edition). American Society for Microbiology (ASM) 2016, 3 (24).

(51 ) Mittal, N.; Tesfu, H. H.; Hogan, A. M.; Cardona, S. T.; Sorensen, J. L. Synthesis and Antibiotic Activity of Novel Acylated Phloroglucinol Compounds against Methicillin-Resistant Staphylococcus Aureus. J. Antibiot. (Tokyo) 2019. https://d0i.0rg/l 0.1038/S41429-019-0153-4. (52) Bliss, C. I. THE TOXICITY OF POISONS APPLIED JOINTLY1 . Ann. Appl. Biol. 1939, 26 (3), 585-615. https://doi.Org/10.1111/j.1744-7348.1939.tb06990.x.

(53) Loewe, S. The Problem of Synergism and Antagonism of Combined Drugs. Arzneimittelforschung. 1953, 3 (6), 285-290. (54) Berenbaum, M. C. What Is Synergy? Pharmacol. Rev. 1989, 41 (2), 93-141.

(55) Yadav, B.; Wennerberg, K.; Aittokallio, T.; Tang, J. Searching for Drug Synergy in Complex Dose-Response Landscapes Using an Interaction Potency Model. Comput. Struct. B iotech nol. J. 2015, 13, 504-513. https://doi.Org/10.1016/j . csbj .2015.09.001 . (56) Rice L. B. (2008). Federal funding for the study of antimicrobial resistance in nosocomial pathogens: no ESKAPE. J. Infect. Dis. 197, 1079-1081. 10.1086/533452 (57) Yan-Ni Sun, Cui-Ling Wang, Ning Zhang, Zheng Wang, Zhu-Lan Liu, Jian-Li Liu, Synthesis of tetrahydro-[3-carbolines from phthalic anhydrides and tryptamine, Chinese Chemical Letters, Volume 25, Issue 11 , 2014, Pages 1503-1506, https://d0i.0rg/l 0.1016/j.cclet.2O14.06.012.