New inhibitors of Mycobacterium tuberculosis identified using systems chemical biology

With the rise in antibiotic resistance, new drugs are desperately needed against Mycobacterium tuberculosis (Mtb). Combining chemistry and genetics, we developed a new strategy for rapidly identifying many new small molecule candidates against Mtb and shedding light on their mechanisms of action (MOA), by performing large-scale chemical screening on pooled genetic libraries containing >100 barcoded strains hypomorphic for individual essential genes. We created barcoded hypomorphic strains for 474 of the ∼∼ 625 essential genes in Mtb and developed a multiplexed, whole-cell assay to measure strain abundance. Applying the approach with an activity-enriched, 3226 compound library and an unbiased 47,353 compound library, we characterized >8.5 million chemical-genetic interactions. Using machine learning, we identified >40 novel compounds against known MOAs, including new inhibitors of DNA gyrase, mycolic acid biosynthesis, and folate biosynthesis. By identifying highly specific chemical-genetic interactions, we identified new inhibitors of RNA polymerase and of a novel target, EfpA. Finally, we showed an inhibitor discovered by screening the hypomorphic strains could be optimized by medicinal chemistry to be active against wild-type Mtb. The results demonstrate that a systems chemical biology approach can empower discovery, prioritization, and development of compounds towards novel TB therapeutics.

T uberculosis (TB) imposes an enormous global burden, causing more than 1.6 million deaths annually (1). Although curative chemotherapies were developed in the mid-20th century (2), TB became the deadliest infectious disease worldwide in 2016, fueled partly by the HIV epidemic. At the same time, TB has become harder to treat due to increasing prevalence of multi-drug resistant (MDR) TB (defined as resistance to two front-line drugs, rifampin (RIF) and isoniazid (INH)) and the emergence of extensively (XDR) and totally drug resistant (TDR) TB (3). Controlling TB requires concerted public health and social efforts, coupled with new diagnostics and therapeutics. Despite the recent approval of two new drugs (bedaquiline and delamanid), TB drug development remains slow, hampered by a paucity of validated targets in the causative bacterium, Mycobacterium tuberculosis (Mtb), and a limited number of active compounds with favorable chemical characteristics that could serve as lead candidates. New targets and drug leads are desperately needed.
A major challenge has been the difficulty of identifying the molecular targets of compounds found to inhibit TB in wholecell screens to enable chemical optimization. In the few cases where mechanism-of-action (MOA) has been illuminated, the targets often belong to the same small set of membrane proteins (such as MmpL3 and DprE1). For the vast majority of the ∼625 essential genes in Mtb, no chemical inhibitors are known (4)(5)(6).
To accelerate drug discovery, we took a systems chemical biology approach involving multiplex chemical screening of genetic libraries of hundreds of mutant strains. We genetically engineered 474 hypomorphic Mtb strains, each depleted for a different essential gene product and each carrying a distinct genetic barcode. We then screened compounds against a pool of 100 strains (99 hypomorphs and wildtype), determined the sensitivity of each strain to each compound from the abundance of the barcodes, and made inferences about the MOA of compounds from the sensitivity pattern across depletion strains. The approach is based on the concept that strains depleted for a protein tend to be hypersensitive to drugs targeting the protein (7)-an idea first championed in yeast and applied in the discovery of platensimycin in Staphylococcus aureus (8) and recent screening of individual depleted strains of Mtb (9,10).
We validated the approach on a collection of 3226 small molecules enriched for anti-Mtb activity. We successfully predicted MOA of novel compounds against known targets based on similar sensitivity patterns to compounds with known MOA, resulting in the discovery of new inhibitors of DNA gyrase, cell wall biosynthesis, and folate biosynthesis. We also identified compounds against targets based on their ability to inhibit specific knockdown strains, leading to the identification of novel inhibitors of RNA polymerase (RNAP) and the essential efflux pump EfpA, a new and uncharacterized target. In the latter case, we were able to optimize the initial compound so that it inhibited not only the EfpA-depletion strain, but also wild-type Mtb.
Finally, we demonstrate the expansion of this strategy to screen a library of 47,353 compounds against 151 depletion strains. This screen showed ∼10-fold greater sensitivity over screening wild-type Mtb alone, and led to the identification of ∼1500 new small molecules, each associated with activity against specific subsets of Mtb depletion strains. By optimizing the EfpA inhibitor so that it inhibited not only the EfpAdepletion strain, but also wild-type Mtb, we showed that compounds discovered through mutant hypersensitivity could be developed to have pharmacologically-relevant potency against Mtb.
The new screening approach and new chemical hits should help accelerate TB drug development.

Development of a multiplexed chemical genetic screening platform in Mtb
We constructed knockdown strains targeting 474 of the ∼625 essential genes defined in Mtb-including genes that are absolutely essential, are severely growth defective, or have essential domains (11). Hypomorphic strains of Mtb H37Rv were created by placing an essential gene under conditional proteolytic control using a tetracycline-repressible SspB:DAS degradation system (12,13), or conditional transcriptional control through promoter replacement for some membrane proteins (Table S1). For proteolytic control, proteins were fused to a carboxy-terminal DAS-tag, which targets the protein for degradation via SspB-dependent shuttling. Up to five different engineered strains for each gene were created, with varying degrees of knockdown based on the strength of the promoter driving sspB expression and thus the degree of degradation (Fig. S1A). In total, we created 2014 strains.
Each strain was also engineered to carry a 20-nucleotide genetic barcode flanked by common primer-binding sites (Fig. S1B) for targeted gene identification by multiplexed polymerase chain reaction (PCR) and next generation sequencing (NGS). Sequencing reads of the barcode were intended to serve as a proxy for strain census in a pooled context.
We developed an assay to measure the abundance of each strain in a pool, by PCR amplifying and sequencing the barcodes (Fig. 1A) (14)(15)(16)(17). We confirmed that barcode counts were an accurate proxy for strain abundance, by mixing strains at known abundances spanning three orders of magnitude (r = 0.93 for log-transformed barcode count replicates; r = 0.95 for known cell abundance and barcode counts) (Fig. 1B).
We performed high-throughput chemical screen of a pool containing 100 strains (99 hypomorphs and wild-type H37Rv). Using the barcode quantification, we determined the growth rate of each pooled strain in 384-microwell broth culture over two weeks. From the multiple strains constructed for each gene, we selected the strain with the strongest degradation with proliferation most similar to the wild-type strain. This growth-matched pool was used for all subsequent multiplexed experiments.
We optimized screening assay parameters in 384-well format, including compound exposure time (14 days), genomic DNA extraction (10% dimethylsulfoxide (DMSO) and 95 • C for 15 min), and PCR conditions (20 cycles, annealing at 65 • C) that maximized the robust Z -score of log-transformed counts for a RIF dilution series (Table S2) and minimized random noise as determined by strain-wise coefficient of variance (CV) (Table S3).
An appropriate sequencing depth (500 reads per strain per well) was determined by performing repeated simulations of a 10,000-well high-throughput screening dataset with active compounds present at 1%, balancing cost, accuracy, sensitivity, and inference of fitness (via a generalized linear model). We developed our analysis protocol into a pipeline, ConCen-susGLM, that calculates, for each unique strain, compound, and concentration combination, a log 2 (fold change) (LFC, interpreted as chemical-genetic interaction score) of counts relative to the DMSO control screening wells, and a p-value (Wald test).

Screening a library of known Mtb-active compounds
We assembled a library of 3226 small-molecules enriched for compounds with activity against wild-type Mtb. The library was drawn from a pool of 1128 compounds we had identified in internal screening efforts as having activity against wild-type Mtb H37Rv (18)(19)(20), 5611 compounds reported by Southern Research Institute (21)(22)(23), and 177 compounds reported by GlaxoSmithKline (24). After removing duplicates, we assembled a library of 2000 available compounds. We supplemented this collection with 1226 compounds from Selleck's Pharmacologically Active Compound Library, to include known antibiotics to serve as positive controls and as ground truth data for machine learning. To confirm the reported Mtb activity, we screened the library against GFP-expressing wild-type Mtb H37Rv (22) and found that 1312 (45%) had an MIC 90 < 64 µM.
We then screened the compound library against our pool of 100 Mtb strains in duplicate (log-transformed Pearson's r = 0.93). Because fitness differences between a given hypomorph and wild-type Mtb can depend on compound concentration, we screened at four library concentrations: 1.1, 3.3, 10, and 30 µM (chosen based on measured MIC 90 values for the entire library).
The screen readily identified chemical-genetic interactions between compounds and hypomorphic strains for their known targets. For example, the fluoroquinolones selectively inhibited the DNA gyrase α-subunit (GyrA) hypomorph, a known interaction that occurs at extreme depletion of GyrA (Fig.  1C). RIF inhibited the hypomorph for the β-subunit of its known target, RNAP (RpoB) (Fig. 1C). The tryptophan syn- and allows accurate estimation of strain relative abundance with Pearson's r = 0.95 (right panel). (C) Chemical genetic interaction profiles from screen data showed expected hypersensitivity for compounds of known MOA. Profiles show the LFC (relative to DMSO negative controls) of each strain at each concentration tested, with wild-type Mtb H37Rv highlighted in green and a mutant of interest highlighted in orange. Error bars of highlighted strains show 95% confidence interval of the mean. Examples shown are the GyrA hypomorph which is hypersensitive to the fluoroquinolone sarafloxacin (top left panel), the TrpG hypomorph which is hypersensitive to trimethoprim (bottom left), the RpoB hypomorph which is hypersensitive to RIF (top right), and the TrpA hypomorph which is hypersensitive to BRD-4592 (bottom right). (D) ROC curve of primary data against a confirmatory secondary growth assay. We retested more than 100 compounds using a resazurin-based colorimetric assay. Taking 50% inhibition in the secondary assay as ground truth, we demonstrated the primary assay as predictive of real activity that could be detected by more conventional growth methods. thase inhibitor BRD-4592 (19) selectively inhibited the tryptophan synthase α-subunit (TrpA) hypomorph (Fig. 1C). Interestingly, TMP, a folate biosynthesis inhibitor, demonstrated a clear interaction with the hypomorph for the folate pathway enzyme glutamine amidotransferase (TrpG) rather than the DHFR hypomorph (Fig. S2A).
We retested a sample of 112 specific hits (LFC < −2; p < 10 −10 ; specificity defined by activity against < 10 strains) for activity against wild type Mtb, each compound's strongest hypomorph interactor, and several other hypomorph strains that served as negative controls, using an orthogonal growth assay based on colorimetric resazurin detection. Because growth rates in different assays cannot be directly compared, we constructed a receiver operating characteristic (ROC) curve to determine how well inhibitory activity in the primary multiplex assay predicted activity in the secondary resazurin assay. Defining a confirmatory result in the secondary assay by a stringent cut-off of 50% growth inhibition compared to DMSO controls (Fig. 1D), the ROC area under the curve (AUC) was 0.73, indicating a high true positive rate in the primary assay with a well-controlled false positive rate. Given the complexity of the primary screen, we were reassured that 1375 (52%) of the 2664 extremely statistically significant interactions were confirmed in the secondary assay.
We recognize the potential value of the current primary data and chemical genetic interaction profiles identified in this study, exceeding those described here and the confines of our own laboratories and expertise, and thus are making data freely available as a resource to the community at https://broadinstitute.org/infectious-diseasemicrobiome/cgtb.

Identification of new inhibitors with specified MOAs by machine learning from known drugs
We next sought to use the resulting chemical-genetic interaction profiles to identify novel compounds with known MOAs. We applied supervised machine learning, leveraging the fact that 107 compounds in the compound library had known antimicrobial MOA (Table S4).
We visualized the 400-dimensional interaction profiles ( Fig. 2A) for each compound through reduction to two dimensions using a t-distributed stochastic neighbor embedding (t-SNE) (26). Compounds known to have the same target clustered together, independent of their chemical structural similarities. This is illustrated with the protein synthesis inhibitors tylosin, erythromycin, and clindamycin (Fig. 2B).
The 107 compounds with known antimicrobial MOA included inhibitors of DNA gyrase (n = 14), mycolic acid synthesis (n = 6), and folate biosynthesis (n = 12). We trained separate binary classifiers using Lasso regression (27) to determine the chemical-genetic interaction profiles associated with each class, using 60% of the data as a training set and 40% as a test set.
DNA gyrase inhibitors. The 14 known DNA gyrase inhibitors in the compound collections were all fluoroquinolones. The regularized regression coefficients in the model suggested that the single most discriminatory feature was strongly decreased fitness of the GyrA hypomorph (Fig. S3A).
Among the compound library, the model predicted 55 non-quinolone DNA gyrase inhibitors (Fig. 3A), including novobiocin, a known inhibitor of the DNA gyrase β-subunit (GyrB). Both the GyrB and GyrA hypomorphs were hypersensitive to novobiocin, suggesting that the GyrA hypomorph could act as a sentinel strain for DNA gyrase inhibition in general.
To test these predictions, we obtained 52 of these compounds and assayed their ability to inhibit Mtb DNA gyrase supercoiling and decatenation activity in vitro ( Fig. 3B and Fig. S3B). Of these, 27 (52%) showed significant activity. We also tested 25 randomly-chosen compounds predicted not to be DNA gyrase inhibitors; none showed activity. The classifier thus significantly enriched for DNA gyrase inhibitors (p = 2 × 10 −7 , Fisher's exact test; ROC AUC = 0.89).
Of the 55 non-quinolone compounds predicted to have DNA gyrase inhibitory activity, only one scaffold, ethacridine (Fig. 3C), has been previously reported to have activity against Mtb DNA gyrase (28). Interestingly, one of the predicted compounds, tryptanthrin, which has been the subject of extensive antibacterial and antitrypanosomal work (29,30), previously had no understood MOA. We found that it inhibits in vitro decatenation at 160 µM and supercoiling at higher concentrations (Fig. S3B). This activity may account for its antimicrobial activity. Interestingly, the quinolone derivatives fleroxacin and nalidixic acid were predicted not to be DNA gyrase inhibitors in Mtb; this prediction was borne out in vitro.

Mycolic acid biosynthesis inhibitors.
We next trained a model to predict new mycolic acid (MA) biosynthesis inhibitors by using the clinical antitubercular prodrugs INH and ETH, whose active forms both competitively inhibit InhA (31), a key enzyme in MA biosynthesis. Even though the strains tested did not include an InhA hypomorph, our analysis still yielded an excellent predictive model. The most discriminatory feature was increased relative fitness of the MshC hypomorph (Fig. S3C). MshC catalyzes the incorporation of cysteine into mycothiol, an antioxidant unrelated to MA biosynthesis. Consistent with this result, mycothiol biosynthesis deficiency has been shown to confer resistance to INH and ETH, probably through suppression of cysteine-induced respiration leading to reactive oxygen species (ROS)-mediated death (32).
Our model predicted seven compounds to be MA inhibitors (Fig. 3D), of which six were hydrazone derivatives of INH and one, BRD-9942, was an indenedione. All seven compounds inhibited wild-type Mtb growth at 50 µM or less in a broth microdilution assay, allowing us to test the effect of these compounds on wild-type Mtb. In a radiolabeling assay of MA biosynthesis, three INH hydrazones and BRD-9942 caused a detectable reduction in exogenous 14C-acetate incorporation into MA (Fig. 3E), validating that they act in the same pathway as INH. Further, a broth microdilution assay showed that an overexpressor of inhA (33) was resistant to the six INH-hydrazones, but not BRD-9942 (Fig. 3F, Fig.  S3D).
The prodrug INH is activated by the catalase KatG, which cleaves its N-N bond (34). Because the INH hydrazones contain this bond, it seemed likely that they are activated in the same way. We confirmed this prediction by showing that the loss-of-function mutant Mtb katGS315T (katG − ) (35) was resistant to the INH hydrazones (Fig. S3D).

Folate biosynthesis inhibitors.
We also built a predictor of folate biosynthesis inhibitors based on the chemical-genetic interaction profiles of the sulfonamides, which are known dihydropteroate synthase (DHPS) inhibitors; the most discriminatory feature was inhibition of the TrpG hypomorph (Fig. S4A). TrpG is involved in both folate and tryptophan (Trp) biosynthesis (Fig. S4B), catalyzing formation of both 4-amino-4-deoxychorismate (a folate precursor) and 2-amino-2-deoxyisochorismate (a Trp precursor).
The model predicted 43 compounds to be folate biosynthesis inhibitors (Fig. 3G), of which we tested seven, spanning several chemotypes. All showed inhibitory activity not only against the TrpG mutant but also wild-type Mtb at 200 µM or less. Hypothesizing that these compounds might be inhibitors of either folate or Trp biosynthesis, we tested whether their effects on wild-type Mtb were abolished by supplementation with Trp, folate, or the folate pathway intermediate paraamino benzoic acid (PABA). For reference, we also tested three known folate biosynthesis inhibitors with distinct activities: methotrexate (MTX, a DHFR inhibitor), para-amino salicylic acid (PAS, a prodrug that inhibits DHFR), and sulfanilamide (a dihydropteroate synthase (DHPS) inhibitor) (Fig. S4C). The two nitrothiophene compounds, BRD-2550 (Fig. S4D) and BRD-9819, behaved similarly in the supplementation assay (Fig. 3H) to MTX, with their effects abolished by both PABA and folate supplementation; this implied an analogous MOA to MTX. Notably, unlike any of the reference compounds, four predicted folate inhibitors including BRD-8884 (Fig. S4D) had effects that could only be rescued by folate (Fig. 3H), suggesting that they inhibit a novel, late step in the pathway. The effect of the final compound BRD-7721 (Fig. 3H), a 3-indolepropionic acid (3-IPA) ester (Fig.  S4D), was only rescued by Trp supplementation, indicating that it is a Trp biosynthesis inhibitor. (3-IPA was recently identified as antimycobacterial (36); notably, 3-IPA's antitubercular effect is also abolished by Trp supplementation). By retraining on compounds whose effects were abolished by tryptophan versus folate or PABA supplementation, we found New inhibitors of Mycobacterium tuberculosis identified using systems chemical biology that the collateral strain depleted for dihydroneopterin aldolase (DHNA, encoded by folB), in conjunction with TrpG, could distinguish inhibitors of the two respective pathways (Fig. S4E-F). Altogether, this analysis identified novel inhibitors of folate biosynthesis, an important pathway that has been successfully exploited against many other pathogens, but to date, not Mtb.

Identification of new inhibitors against a specific target based on strain sensitivity
In the three examples above, we identified new compounds by training classifiers with data from known drugs with a specific MOA. We next sought to identify new compounds in situations where we lacked such information, by instead searching for compounds that selectively inhibited strains depleted for a specific target of interest.
We searched for inhibitors of the well-characterized and valuable target, RNAP, whose extremely potent clinical inhibitor RIF is credited with shortening TB chemotherapy. The RIF derivatives in our screen inhibited growth of all strains, providing no chemical-genetic signal for a machine learning model. However, we had tested RIF in the screen as a positive control across a wider concentration range and noted that the hypomorph of RpoB, a subunit of RNAP, was specifically inhibited at lower concentrations (Fig. 1C).
We identified 20 compounds that strongly inhibited (p < 10 −10 , LFC < −2) the RpoB hypomorph, with it being among the two most inhibited strains for at least one dose (Fig. 4A). Recognizing that many mechanisms might lead to inhibition of an RNAP mutant, we tested the compounds to determine if any had activity in an in vitro RNA synthesis assay (37,38). Importantly, three of the compounds (Fig. S5) showed direct inhibition of RNAP (Fig. 4B). Thus, compounds that specifically target the RpoB hypomorph are highly enriched for molecules that directly target RpoB itself. This example showed that new inhibitors can be identified based on activity against individual hypomorphic strains.

Screening a large chemical library
Our initial screen involved a library of 3226 compounds enriched for molecules with activity against wild-type Mtb. We next extended the approach to a large unbiased library of 47,353 compounds. This library was assembled from internal Broad Institute chemical libraries, including sub-libraries from the Broad diversity oriented synthesis collection (39), National Institutes of Health's Molecular Libraries Probe Production Centers Network library, and purchased from WuXi, filtered for physico-chemical properties including molecular weight < 550, calculated log(D) at pH 7.4 < 4, polar surface area < 110, charge at pH 7.4 > −1.
We screened these compounds (50 µM) against 152 strains (151 hypomorphs and wild-type), including 94 of the 100 strains used in the initial screen. Of the 7,245,009 potential chemical-genetic interactions tested, 95,685 (1.3%) scored as hits based on the criteria used in the initial screen. To validate the results, we repeated the assay using the 152-strain pool on 1331 compounds, which had specific activity against fewer than ten strains, and confirmed 78% of the chemicalgenetic interactions previously identified as inhibitory (LFC < 0), resulting in an ROC AUC of 0.74.
We identified 436 compounds that were active against wild-type Mtb (0.9% hit rate); in contrast, 10-fold more compounds (4403; 9%) were active against at least one of the 152 strains. This 10% hit rate is ∼3-fold lower than the hit rate observed with the smaller, Mtb active enriched compound library (29%). Among the 3967 compounds that did not hit wild-type Mtb, 73% were highly specific (1-10 hypomorphic strains hit), 11% were moderately specific (11-50 strains hit) and the remaining 16% were relatively non-specific (more than 50 strains hit). This distribution shows greater specificity than for compounds in the small screen (35%, 31%, and 34%), likely due to the enrichment for compounds with wild-type Mtb activity in the smaller library (Fig. S6A). Notably, the pattern of strain sensitivity showed considerable diversity: 59% of the strains tested in the smaller screen were the strongest hit for at least one compound at 30 µM, and the proportion was 75% in the larger screen-suggesting that compounds can be found against much of the target space. For the 94 strains that were common to both screens, there was strong correlation between the number of hit compounds against each strain in the two screens (log-transformed Pearson's r = 0.8) (Fig. S6B).
The larger screen demonstrates that our chemical-genomic approach should be applicable to any chemical library. Moreover, it should identify many more chemical hits than screening wild-type Mtb alone and associate them with activity against specific hypomorphs.

Developing wild-type activity of a compound discovered through hypomorph screening
While our approach identifies ∼10-fold more compounds than screening against wild-type Mtb and sheds light about the compounds' targets, its utility for drug development depends on the ability to optimize hits against the hypomorphs to molecules with pharmacologically relevant activity against wild-type Mtb. We therefore sought to demonstrate that it was possible to perform such optimization.
To identify the attractive hits for optimization, we excluded compounds with activity against wild-type Mtb, ranked compound-strain interactions based on specificity (how few other strains were significantly inhibited by the compound), susceptibility of strains at each concentration, and whether the compound contained chemically promising scaffolds. The highest ranked compound was BRD-8000, which across all four concentrations specifically inhibited the growth of the hypomorph of EfpA (Fig. 4C), a largely uncharacterized, but essential efflux pump in Mtb. In a separate broth microdilution assay, we confirmed that BRD-8000 showed strong activity against the EfpA hypomorph (MIC 90 = 6 µM). By contrast, the compound showed no wild-type activity in the screen and only weak activity in the broth microdilution assay (MIC 90 ≥ 50 µM) (Fig. 4D).
To optimize BRD-8000, we began by focusing on the fact that it is a mixture of stereoisomers. We found that the (S,S)-trans stereoisomer was the active form (BRD-8000.1; Table 1), with a wild-type MIC 90 = 12.5 µM, which is greater than four-fold lower than that of the corresponding BRD-8000 mixture. We further increased the potency by moving the pyridyl bromine from the 5-to the 6-position to obtain BRD-8000.2 (MIC 90 = 3 µM). The initial and optimized compounds showed similar differential activity (30-fold) with respect to inhibition of the EfpA hypomorph ( Table 1), suggesting that the optimized compound remained on-target.
With the increased potency of BRD-8000.2, we obtained genetic evidence for EfpA as the target by generating resistant mutants in wild-type Mtb H37Rv. We obtained more than 30 independent resistant clones (MIC 90 > 50 µM, Fig. S7A), each with a C955A mutation in efpA (coding for EfpA V319F ), occurring at a resistance frequency of ∼ 10 −8 .
We sought to functionally confirm that the BRD-8000 series targets EfpA by taking advantage of the fact that it does not inhibit growth of Mycobacterium smegmatis (Msm), likely because EfpA is not essential in Msm (40). We assayed the compound's effect on EfpA efflux activity in Msm by loading cells with the dye ethidium bromide (EtBr), a substrate of the pump, and then measuring efflux, since EtBr fluorescence is 20-fold higher when the dye is intracellular than extracel-lular (41). We determined that both BRD-8000 and BRD-8000.2 directly inhibited efflux in living cells, with an IC 50 of 37 µM and 15 µM, respectively (Fig. S7B). This inhibition was stereospecific as the inactive (R,R)-trans isomer of BRD-8000.2 has an efflux IC 50 of 500 µM (Fig. 4F). BRD-8000.2 is not a competitive substrate for EfpA, as its inhibition has an inverse relationship with initial EtBr concentration, suggesting an uncompetitive or non-competitive inhibitory mechanism (Fig. S7C, Fig. S7D). We also found that the IC 50 for the compound was substantially higher in an efpA knockout strain of Msm than in wild-type Msm (Fig. S7E), indicating that BRD8000.2 has an EfpA-specific higher-potency interaction; episomal complementation of the knockout with the Mtb efpA gene restored the lower IC 50 (Fig. S7F).
We found that BRD-8000.2 is bactericidal (Fig. S7G), has low human toxicity (hepatocyte IC 50 = 100 µM), and does not act on bacteria lacking an essential EfpA homolog, including conventional species such as Staphylococcus aureus, Pseudomonas aeruginosa, and Klebsiella pneumoniae. These properties, including the narrow spectrum of activity, are favorable for therapeutic potential for BRD-8000.2 against Mtb.
More broadly, the results for BRD-8000.2 and EfpA demonstrate the potential of chemical screening of hypomorph libraries to yield promising small molecule candidates with wild-type activity that would otherwise elude discovery and development.

Discussion
We developed a powerful and rapid systems chemical biology strategy both to discover many potential compounds for drug development against Mtb and to gain insights about their MOA. Using a pool of barcoded Mtb strains that are each depleted for a different essential gene product, we performed multiplexed screens with small-molecule libraries and used high-throughput barcode sequencing to estimate the census of each strain following chemical perturbations. Based on predictions from the chemical-genetic profiles of primary screening data, we readily identified MOA for 45 molecules. Importantly, these hits included new scaffolds against both known and new targets. With chemical optimization, we generated a new inhibitor of a new target, EfpA, active in wild-type Mtb. We also demonstrated that screening of genetic libraries identified ∼10-fold more hits than obtained by screening wild-type Mtb alone. Chemical-genetic screens were first reported in yeast, taking advantage of the ability to deplete essential targets through heterozygosity for gene knockouts (42). This approach is not possible in bacteria, because they are haploid. Instead, chemical-genetic screens have been performed in E. coli using strains carrying deletions of non-essential genes (7,(43)(44)(45). But, this approach is limited to non-essential genes, which are generally not drug targets. In yeast and E. coli, it is sometimes possible to make inferences about the MOA of drugs against the products of essential genes based on the sensitivity of strains deleted for non-essential genes, by using extensive prior knowledge of genetic interaction networks in these exhaustively studied model organisms. By contrast, Mtb lacks these advantages: 45% of gene functions cannot be assigned by homology, and 27% of its genes are totally uncharacterized (46,47).
In principle, libraries of hypomorphic strains with vary-ing levels of knockdown provide a way to studying chemicalgenetic interactions involving essential genes. In 2009, Donald et al. characterized 58 known Staphylococcus aureus inhibitors by screening them against 101 strains with significant knockdowns of essential genes (48). However, this approach has not been widely used.
Here, we show that a genetic knock-down library of essential genes can be constructed in Mtb, can be used for chemical screening, and can shed light on MOA even without well-understood networks (8). We used the resulting chemical genetic interaction profiles to predict new inhibitors of known targets, finding and validating 41 new chemical scaffolds that target new DNA gyrase, MA biosynthesis, and folate biosynthesis, based on the similarity of their profiles with those of drugs against these known targets. We found that some MOAs were predicted by the behavior of predominantly one strain -defined as the pathway sentinel strains-such as the GyrA hypomorph for DNA gyrase. It may be possible, therefore, to define a minimal subset of pathway sentinels that allow prediction of any MOA. (Such an approach is used to infer the transcriptional response of cancer cells to drug treatment, based on the response of ∼1000 genes, out of the ∼20,000 in the human genome (49)). In some cases however, collateral information from multiple strains aided MOA assignment, as illustrated by the refinement of TrpG-based folate and tryptophan biosynthesis inhibitor prediction using collateral strain behavior (the DHNA hypomorph). Additionally, collateral mutants allowed MOA prediction in the absence of the hypomorph corresponding to the actual target or of a sentinel strain, as illustrated by the ability to identify MA biosynthesis inhibitors in the absence of an InhA hypomorph. Determining the optimal balance of strains to screen will require increased understanding of pathway sentinels, core vulnerabilities, and Mtb's genetic interaction networks, all of which will facilitate iterative improvement of MOA prediction.
In a complementary approach that underscores the power of our systems chemical biology strategy, we also predicted MOA of targets without relying on the genetic interaction profiles of known drugs. By prioritizing of compounds based on the identity of and selectivity for particular hypersensitive mutants, we found three new inhibitors of RNAP; RNAP is the high value target of the first-line drug RIF and new inhibitors have been elusive. Excitingly, we also found an inhibitor of an entirely novel target in our discovery of the EfpAinhibitor BRD-8000. While the initial hit had little wild-type Mtb activity, subsequent optimization yielded a molecule, BRD-8000.2, with clinically relevant potency against wildtype Mtb, on par with current antitubercular drugs. These results illustrate the value of the primary screening data for early elucidation of MOA, the ability to access novel smallmolecule candidates that would elude traditional methods of screening wild-type Mtb, and the ability to identify inhibitors of novel targets even in the absence of well-characterized target function.
The ability to leverage the much greater hit rate from screening pooled hypomorphs and to chemically optimize hits to obtain wild-type activity has the potential to revolutionize the process of Mtb drug discovery, by making it possible to identify and prioritize both targets and compounds based on chemical-genetic interaction profiles. We are hopeful that the approach will substantially expand the number and target diversity of drug leads for development and eventual human clinical testing. The next steps involve completing the genetic libraries to encompass all essential genes; large-scale screening of diverse chemical libraries; development of new analytical methods to infer MOA; and further efforts to optimize hits against hypomorphs to clinically relevant drug leads.