Bayesian Models Leveraging Bioactivity and Cytotoxicity Information for Drug Discovery

SUMMARY

Identification of unique leads represents a significant challenge in drug discovery. This hurdle is magnified in neglected diseases such as tuberculosis. We have leveraged public high-throughput screening (HTS) data, to experimentally validate virtual screening approach employing Bayesian models built with bioactivity information (single-event model) as well as bioactivity and cytotoxicity information (dual-event model). We virtually screen a commercial library and experimentally confirm actives with hit rates exceeding typical HTS results by 1-2 orders of magnitude. The first dual-event Bayesian model identified compounds with antitubercular whole-cell activity and low mammalian cell cytotoxicity from a published set of antimalarials. The most potent hit exhibits the in vitro activity and in vitro/in vivo safety profile of a drug lead. These Bayesian models offer significant economies in time and cost to drug discovery.

INTRODUCTION

Modern drug discovery must be more time- and cost-efficient in discovering novel therapeutics. These challenges are felt even more significantly in the search for neglected disease treatments, where public-private partnerships coordinate drug discovery with very limited resources. A prime example is tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), which infects approximately one-third of the world's population and results in 1.7-1.8 million deaths annually (Lienhardt et al., 2012). New drugs active against Mtb are urgently needed to combat a pandemic heavily affected by resistance to available therapies and co-infection with HIV/AIDS (Nuermberger et al., 2010). TB drug discovery is challenging, reflected in the lack of a new TB-focused therapeutic approved in over 40 years (Grosset et al., 2012, Sacchettini et al., 2008). One response has been to screen very large compound libraries (Ananthan et al., 2009, Maddry et al., 2009, Reynolds et al., 2012), hoping to deliver on the promise of chemical diversity (O'Connor et al., 2012). Phenotypic whole-cell high-throughput screens (HTS) of commercial libraries have searched for inhibitors of mycobacterial growth, at a cost of millions of dollars, with resultant low single-digit (or less) hit rates (Macarron et al., 2011, Magnet et al., 2010, Mak et al., 2012, Stanley et al., 2012). The campaigns have resulted in numerous hits, but resource constraints have limited follow-up to the few most promising compounds and/or compound series. Fortunately, one screen of the non-pathogenic Mycobacterium smegmatis unearthed a diarylquinoline hit that led to the clinical candidate bedaquiline (Andries et al., 2005), while another resulted in the early-phase candidate SQ109 (Lee et al., 2003). Although SQ109 arose directly from a library of congeners of the front-line drug ethambutol, HTS typically does not deliver a clinical candidate. Exhaustive optimization of a screening hit must occur, initially following whole-cell activity and then considering pharmacokinetics, pharmacodynamics, and safety to afford clinical candidates such as PA-824 (Stover et al., 2000). The remainder of current TB clinical trials arose from repurposing other antibacterials or rediscovering antituberculars from decades ago (Lienhardt et al., 2012). Despite these successful efforts, the expected failure of ~85% clinical candidates (Ledford 2011) and growth of TB drug resistance necessitate new clinical submissions, which ultimately require the discovery of novel hits and leads. We assert that the TB field should further leverage existing HTS data, focusing on not just the few most promising hits due to resource limitations, but the entire data set of actives and inactives.

We hypothesize that prior knowledge of Mtb actives and inactives, combined with machine learning models, can significantly focus compound selection and improve screening efficiency (Ekins and Freundlich 2011, Ekins et al., 2011, Ekins et al., 2010, Ekins et al., 2010), as practiced in the pharmaceutical industry (Prathipati et al., 2008), to improve the performance of virtual screening (Schneider 2010). These and other cheminformatics methods have been utilized in the TB field, although in our opinion not to the extent as in the pharmaceutical industry (Ekins et al., 2011). Thus, cheminformatics technologies such as virtual screening and structure based design have contributed to clinical submissions in the pharmaceutical industry (Volarath et al., 2007), but have yet to impact TB drug candidates (Barry et al., 2000, Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Koul et al., 2011).

An alternative cheminformatics approach to computational screening discriminates between the user-defined actives and inactives present in a screening dataset. This approach called Bayesian modeling, can then be utilized in an unsupervised or automated manner to predict the likelihood of a new molecule (absent from the training set) being a hit (using Bayes Theorem described in equation 1). We (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Sarker et al., 2012) and others (Periwal et al., 2011, Prathipati et al., 2008) have undertaken a systematic Bayesian machine learning modeling effort focused solely on Mtb bioactivity. Bayesian models were developed that learn from public efficacy data for both actives and inactives and correlate 2D compound structural features with antitubercular activity or lack thereof. We have consistently seen enrichment values from 4–10 fold over experimental HTS (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Sarker et al., 2012) for resultant models. Yet, prospective validation of these models (using test molecules predicted prior to testing in vitro) has been lacking until now. Here we validate these models with the prospective identification of novel antitubercular hits from a commercial library with a 14% hit rate (at least 1–2 orders of magnitude greater than empirical HTS (Ananthan et al., 2009, Gold et al., 2012, Maddry et al., 2009, Magnet et al., 2010, Mak et al., 2012, Reynolds et al., 2012, Stanley et al., 2012)).

Drug leads must not only be efficacious, but sufficiently non-cytotoxic to mammalian cells (Langdon et al., 2010). We have therefore created dual-event Bayesian models combining antitubercular activity and mammalian cell cytotoxicity. We demonstrate enhanced predictive power over models that exclude cytotoxicity. In addition, we apply a dual-event model to the discovery of Mtb inhibitors from a published library of antimalarial hits (Gamo et al., 2010), demonstrating a significant application to drug discovery, and report a potent small molecule TB drug lead exhibiting nanomolar growth inhibition of cultured mycobacteria and acceptable in vitro and in vivo mouse safety.

RESULTS

Validation of a Bayesian model for TB whole-cell activity

Mtb Bayesian models (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Periwal et al., 2011, Prathipati et al., 2008, Sarker et al., 2012) have lacked critical published validation as to their ability to prospectively predict novel actives. To prospectively test the previously generated Molecular Libraries Small Molecule Repository (MLSMR) dose response Bayesian model (Ekins et al., 2010) we virtually screened a commercial library of >25,000 compounds (Asinex, available in the CDD database (Hohman et al., 2009)). Principal Component Analysis (PCA) demonstrated that the commercial library members occupy similar chemical space as do the model's actives (Fig S1). The compounds were ranked by Bayesian score (range, -28.4 to 15.3), which relates to the likelihood of a compound being active through determination of its molecular features compared to the features in the model's actives and inactives. The more positive the value, the higher probability of being active. The top scoring 100 compounds (Bayesian score 9.4 to 15.3) were then selected. Ninety-nine of these were commercially available and tested for growth inhibition of Mtb (Table S1). Fourteen of these compounds exhibited an IC₅₀ < 25 μg/mL, affording a hit rate of 14%. The most potent molecule (SYN 22269076), featuring an IC₅₀ of 1.1 μg/mL or 3.2 μM (Fig. 1), represents a novel member of the pyrazolo[1,5-a]pyrimidine class present in the HTS data utilized to train the model (Ananthan et al., 2009).

Fig. 1.

The validation of a single-event Bayesian model for Mtb efficacy. Chemical structure of the most active compound (IC₅₀ = 1.1 μg/ml) – SYN 22269076 – from a virtual screen of a 25,000-member commercial library.

Dual-event Bayesian models

Previous Mtb Bayesian models (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Periwal et al., 2011, Prathipati et al., 2008, Sarker et al., 2012) did not account for molecule cytotoxicity and while machine learning methods have been created for various human toxicities (Ekins et al., 2010, Langdon et al., 2010, Zientek et al., 2010) these have not been combined with bioactivity endpoints. We created a dual-event Bayesian model merging in vitro cytotoxicity (CC₅₀) data for the compounds with Mtb dose response data (Ananthan et al., 2009, Maddry et al., 2009) used in our previous studies. We selected for active (IC₉₀ < 10 μg/mL) and non-cytotoxic molecules (selectivity index, SI = CC₅₀/IC₉₀ > 10) to construct the dual-event Bayesian model (SRI - MLSMR dose response and cytotoxicity model). Thus, this model has learned what molecular features amongst the training set are consistent with Mtb growth inhibition and lack of comparatively significant toxicity to Vero cells. This model had a leave-one-out Receiver Operator Characteristic (ROC – a general performance measure of models, where an ideal model has an ROC = 1 (Zientek et al., 2010)) value of 0.86 (Table S2). All statistics for this model were equivalent or superior to the previously published MLSMR single point and dose response models (Ekins et al., 2010), which have been extensively retrospectively validated (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010). This model was also shown to predict 7 of 9 first- and second-line TB drugs (Sacchettini et al., 2008) that were absent in the training set (Table S3). Notably, isoniazid and pyrazinamide are prodrugs (Konno et al., 1967, Rozwarski et al., 1998, Scorpio and Zhang 1996) and were not predicted to be active by the model, which only learns from explicit chemical structures and is likely ignorant of chemical reactivity. Using the functional class fingerprint 6 (FCFP-6 : each atom is described by its six nearest neighbors (Zientek et al., 2010)) descriptors, we identify those substructure descriptors contributing to Mtb activity and an acceptable SI (Fig. S2), including the oxazole 2-thioether, aryl/heteroaryloxyacetic acid, and quinolone 3-carboxylic acid cores. We also note chemical features inconsistent with both satisfactory activity and SI, such as thiazole 2-amides, 2-substituted pyrazoles, 2-substituted benzimidazoles, N-functionalized pyrrolidines, N-arylamides, and 2-substituted pyridines (Fig. S3).

We subsequently generated prospective predictions for a screen of a targeted human kinase inhibitor library against Mtb (Reynolds et al., 2012), using the dual-event model in addition to previously described Bayesian models omitting cytotoxicity (Ekins et al., 2010). Inclusion of the cytotoxicity parameter resulted in a virtual screen hit rate of 12.9% for the top 1000 compounds in this dataset, compared with 3.5–9.8% in the absence of cytotoxicity data (Fig. S4). This is over three times the enrichment compared to the random hit rate, nearly three times higher than the recently reported hit rate for the HTS screening (5.1%) for these compounds (Reynolds et al., 2012), and 8-fold larger than the 1.7% hit rate achieved with larger and more structurally diverse libraries (Ananthan et al., 2009, Maddry et al., 2009).

These promising preliminary results prompted us to generate a second dual-event model with the Tuberculosis Antimicrobial Acquisition and Coordinating Facility(TAACF)-CB2 library dose response and cytotoxicity data (Ananthan et al., 2009). The robustness of this model was examined by the calculation of the ROC under two conditions. First each molecule is left out and the training set used to predict the missing molecule. Second, half of the members of the training set was left out and the model was rebuilt. This was repeated 100 times at random (Zientek et al., 2010). The leave-one-out ROC for this model was 0.64 and leave-out-50% X 100 statistics were lower than previous models (0.59) (Ekins et al., 2010) but still acceptable, even though the training set overlaps well with the MLSMR and kinase scaffold libraries as determined by PCA (Figure S5). Using the FCFP-6 descriptors, we can identify those substructure descriptors consistent with both activity and lack of cytotoxicity (Fig. S6) including N-alkylated imidazole, 1-amino-3-chlorobenzene, diaminopyrimidine, 5-substituted-1,3,4-thiadiazol-2-amine, and tetrafluorophenylamide. Features inconsistent with both activity and lack of cytotoxicity included nitroolefin, 3-hydrazonoindolin-2-one, and 1-amino-2-chlorophenyl (Fig. S7).

Bayesian models increase efficiency of TB drug discovery screening

We have published on the potential to “pathogen hop” between inhibitors of P. falciparum and Mtb (Vilcheze et al., 2011). The dual-event TAACF-CB2 dose response and cytotoxicity Bayesian model was used to rank a previously published set of >13,000 potential antimalarial hits (Gamo et al., 2010), possessing low cytotoxicity and chemical diversity, that represent chemical tools/probes being explored by the infectious disease community. From the commercially available subset, the top forty-six molecules were visually inspected and seven were chosen as representative (i.e., chemotypes such as furanylamide, quinazoline, quinoline, triazine, aminothiazole, and arylsulfonamide). Five were identified as active against Mtb with minimum inhibitory concentration (MIC) values ≤ 2 μg/mL (71% hit rate; Table 2). Out of these five hits, three molecules represent novel antitubercular chemical structures and the remaining two hits have been reported to exhibit in vitro efficacy versus Mtb, but without published follow-up (Bruhin et al., 1969, Reynolds et al., 2012), making them promising starting points for drug discovery. Interestingly, the most active compound was TCMDC-125802, (E)-6-(2-((5-nitrofuran-2-yl)methylene)hydrazinyl)-N²,N⁴-diphenyl-1,3,5-triazine-2,4-diamine, with an MIC of 0.0625 μg/mL. This compound has drug-like properties with a molecular weight of 414.42 g/mol, 10 hydrogen bond acceptors, 3 hydrogen bond donors, 9 rotatable bonds, total polar surface area of 94.27 Ų, and a calculated logP of 2.86 (Oprea et al., 2001, Walters and Murcko 2002).

A search of the publicly available MLSMR 215,110-compound screening data (Maddry et al., 2009) was conducted for diaminotriazine hydrazones with positive growth inhibition (at 10 μM compound concentration) of cultured Mtb. One hundred and seven molecules were found featuring three nitrofuryl hydrazones, ten furyl hydrazones (no nitro- substitution), and nineteen nitrophenyl hydrazones. Notable among the 32 inactives were hydrazones with aryl, nitroaryl, and furyl substituents (Table S4). These molecules were not discussed previously (Maddry et al., 2009). A search of the literature uncovered a single report of the antitubercular activity of TCMDC-125802 and related triazines against Mtb from over 40 years ago (Bruhin et al., 1969), which came to light after choosing TCMDC-125802 from the scored antimalarials set. It was shown to have activity against cultured Mtb strains (Bruhin et al., 1969). However, it was reported to be inactive in a mouse model of acute infection, as judged solely by extension of survival compared to an untreated control. A close analog, where the two aniline moieties were replaced with i-propylamino groups, demonstrated potent in vitro and in vivo activity, but was qualitatively less active in vivo than the front-line drug isoniazid (INH) (Bruhin et al., 1969).

The in vitro potency of TCMDC-125802 prompted us to re-examine its activity and safety profiles in vitro and in vivo. TCMDC-125802 was synthesized on a gram scale (SI Text) and found to be bactericidal, exhibiting a minimum bactericidal concentration (MBC; the concentration of compound reducing the initial bacterial load by 2 log₁₀ units) (Saunders 1992) of 0.25-0.5 μg/mL (Fig. 2a). Over the course of three weeks, 64X the MIC (4.0 μg/mL) afforded a 6 log₁₀ drop in CFUs (Fig. 2b). A CC₅₀ of 4.0 μg/mL with African Green monkey kidney cells (Vero cells; ATCC) was determined, representing a selectivity index (CC₅₀/MIC) of 64. Additionally, we determined a CC₅₀ of 1.0 μg/mL with B6D2F1 mouse bone marrow-derived macrophages. The cytotoxicity data may be compared with that previously reported (5% growth inhibition of HepG2 cells in the presence of 10 μM TCMDC-125802) (Gamo et al., 2010). Significantly, TCMDC-125802, formulated in 0.5% methyl cellulose, demonstrated no overt toxicity in C57BL/6 mice for 7 days post 3 days dosing (30, 100, and 300 mg/kg by gavage). Subsequently, TCMDC-125802 was examined in a standard mouse model of acute Mtb infection. This model uses the highly susceptible gamma interferon-gene disrupted (GKO) C57BL/6 mouse (Lenaerts et al., 2003). Eight-to ten-week-old female specific pathogen-free C57BL/6-Ifngtm1ts (GKO) mice (Jackson Laboratories, Bar Harbor, Maine) were infected with Mtb Erdman strain via low dose aerosol exposure.

Fig. 2.

The in vitro efficacy profile of TCMDC-125802 identified by a dual-event Bayesian model for Mtb efficacy and cytotoxicity. a. Minimum bactericidal concentration (MBC) determination through quantification of the Mtb CFUs as modulated by various compound concentrations. Error bars denote standard deviations. b. Killing kinetics examined through the time dependence of Mtb CFUs in the presence of TCMDC-125802. INH – isoniazid. Error bars denote standard deviations.

Thirteen days post infection the mice were administered TCMDC-125802 300 mg/kg by gavage daily for nine consecutive days. One day post cessation of TCMDC-125802 dosing, the animals were euthanized. TCMDC-125802 was not found to reduce the bacillary load in mouse lungs and spleens as compared to the untreated control (Table S5).

DISCUSSION

Machine learning using Bayesian models has previously focused on Mtb activity and excluded cytotoxicity data (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010, Periwal et al., 2011, Prathipati et al., 2008, Sarker et al., 2012). We provide the first experimental validation of such models through the prospective prediction of novel actives and demonstration of their antitubercular efficacy with a hit rate of 14%. Finding new hits is, therefore, greatly enhanced over current screening methods as typical experimental HTS success rates are less than 1% (Payne et al., 2007) and for Mtb growth inhibition screens they are similar (Ananthan et al., 2009, Gold et al., 2012, Maddry et al., 2009, Magnet et al., 2010, Mak et al., 2012, Reynolds et al., 2012, Stanley et al., 2012). Interestingly, many of the prospective hits from a vendor library exhibited a pyrazolo[1,5-a]pyrimidine core and five actives were found with similar (but not identical) substructures from a kinase library screen (Reynolds et al., 2012). All exhibited acceptable SI values of >10. The most active compound, SYN 22269076 (Fig. 1), is notably absent from any of the public TB datasets, suggesting it may be promising to perform hit-to-lead optimization on this compound series.

A unique aspect of the dual-event machine learning models reported herein lies in their combination of bioactivity and cytotoxicity features. To the best of our knowledge, this is the first report of such a strategy, resulting in the selection of active molecules, e.g. TCMDC-125802, with promising efficacy and cytotoxicity profiles at a much higher success rate than with traditional HTS. The ability to identify non-cytotoxic hits is significant because HTS campaigns often fail to find whole-cell actives devoid of cytotoxicity (Payne et al., 2007). Toxicity-related events lead to more than one-third of all clinical trial failures and 90% of the withdrawals of approved drugs (Schuster et al., 2005). While learning from large cytotoxicity datasets has been described previously by us (Ekins et al., 2010) and others (Langdon et al., 2010), this has not been applied in the TB field. A dual-event Bayesian model was utilized to score an antimalarial library (Gamo et al., 2010) and the top 23 molecules that were commercially available contained only five molecules (21.7%) with significant cytotoxicity (≥40% HepG2 growth inhibition at 10 μM compound). It is important to note this library was biased toward compounds with relatively low cytotoxicity (Gamo et al., 2010). A performance comparison amongst our models as to their ability to identify hits from a kinase-focused library demonstrates a potential advantage for dual-event versus single-event models (Fig. S4), that warrants further study. Adding the cytotoxicity criteria to define an active hit, narrows down the number of hits used in model training. It is also possible we are removing spurious hits that work by multiple mechanisms, thus failing to discriminate sufficiently between Mtb and model mammalian cell lines.

Significantly, the dual-event model identified a TB drug lead TCMDC-125802, which exhibited promising in vitro bactericidal activity against Mtb, acceptable mammalian cellular cytotoxicity, and in vivo mouse safety. While the compound did not show activity at 300 mg/kg dosing in a single mouse model of acute infection, the value of a small molecule which is safe in vivo and possesses an excellent in vitro activity profile is significant for lead optimization. Additionally, the chemical similarity of it to the corresponding di-(i-propylamino) variant with demonstrated in vivo efficacy (Bruhin et al., 1969) should engender confidence that novel analogs will be found to have efficacy in accepted TB animal models. These chemical optimization efforts will leverage existing data around the core structure published previously (Bruhin et al., 1969) and from the MLSMR screen (Maddry et al., 2009) (Table S4) to probe absorption, distribution, metabolism, and excretion (ADME) properties that may likely be responsible for the lack of activity of TCMDC-125802 in the GKO mouse model. Additionally, none of the previous Mtb machine learning studies has derived such an active antitubercular with demonstrated in vivo safety.

Future efforts will seek to discern the Mtb target(s) of TCMDC-125802 through ongoing mechanistic studies. It may share a common mechanism with nitrofurantoin, (Tanimoto similarity 0.68; an approved antibacterial for uncomplicated urinary tract infections (Garau 2008) with modest activity (MIC = 12 μg/mL) against M. bovis BCG `(Murugasu-Oei and Dick 2000)) that undergoes bacteria-induced reduction of the nitro group via one or more nitroreductases (Whiteway et al., 1998) to a toxic nitroso or hydroxylamine derivative (Sandegren et al., 2008). One must also consider nitroimidazoles such as PA-824 (Stover et al., 2000) (Tanimoto similarity 0.63), which exhibits anaerobic activity via the release of reactive nitrogen species and aerobic efficacy through an undetermined mechanism (Singh et al., 2008).

SIGNIFICANCE

We propose the successes reported herein using single and dual-event Bayesian models for whole-cell Mtb activity to accelerate the discovery of novel hits and leads will be readily translated to other therapeutic areas. In so doing, we expect to achieve further enhancements through strategies such as consensus modeling (Ganguly et al., 2006) and combining datasets. The detailed study of the effect of training sets and model parameters on actives enrichment may lead to multi-event Bayesian models that focus on compound attributes important for in vitro and/or in vivo efficacy (Ekins et al., 2010). Since drug discovery is time-intensive and very costly, machine learning approaches can increase the efficiency of screening and should be implemented prior to future HTS campaigns (Ballel et al., 2005, Nathan 2011, Sacchettini et al., 2008). These campaigns represent multi-million dollar investments that will be more fully utilized through Bayesian models generated from previously generated data. This will also spare resources for more expensive and critical downstream studies to select candidates for clinical testing.

EXPERIMENTAL PROCEDURES

Compounds

Compounds were purchased from Asinex, Enamine, Life Chemicals, and Ryan Scientific and assayed as supplied without further quality assessment. TCMDC-125802 was also synthesized (Supplementary Methods). The NIAID Southern Research Institute screening datasets (Ananthan et al., 2009, Maddry et al., 2009, Reynolds et al., 2012), Asinex library (N >25,000) and antimalarial compounds (N >13,000) (Gamo et al., 2010) were downloaded from the CDD TB database (Collaborative Drug Discovery Inc. Burlingame, CA)(Ekins et al., 2010) and used for computational analysis (Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010).

Dual-Event Bayesian Model Building

Bayesian classification is a simple probabilistic classification model based on Bayes’ theorem (equation 1):

$$p(h\mid d)=\frac{P(d\mid h)P\left(h\right)}{P\left(d\right)}$$ (eqn 1)

Where h is the hypothesis or model, d is the observed data, p(h) is the prior belief (probability of hypothesis h before observing any data), p(d) is the data evidence (marginal probability of the data), p(d|h) is the likelihood (probability of data d if hypothesis h is true) and p(h|d) is the posterior probability (probability of hypothesis h being true given the observed data d). Bayesian statistics take into consideration the complexity of the model as well as the likelihood of a model, such that it automatically picks the simplest model that can explain the observed data and prevents overfitting. In the Bayesian modeling software within Discovery Studio (Accelrys, San Diego, CA), the learned models are created with a learn-by-example paradigm: the user marks the sample data that is of interest (good or active), and then the system learns to distinguish them from background data (i.e those that are inactive). The learning process generates a large set of Boolean features (e.g. and, not, or etc) from the input descriptors, and then collects the frequency of occurrence of each feature in the good subset and in all data samples. To apply the model, the features of the sample are generated, and a weight is calculated for each feature using a Laplacian-adjusted probability estimate to account for the different sampling frequencies of different features. The weights are summed to provide a probability estimate, which is a relative predictor of the likelihood of that sample being from the good subset (e.g. a more positive value).

Dual-event Bayesian classifier models were created for: 1. the MLSMR dose response and cytotoxicity data (Maddry et al., 2009) for 2,273 compounds (165 active with IC₉₀ < 10 μg/mL and selectivity SI > 10 for Vero cells) and 2. the TAACF CB2 dose response and cytotoxicity data for 1,783 compounds (1,006 active with IC₉₀ < 10 μg/mL and selectivity SI > 10 for Vero cells) as described(Ekins and Freundlich 2011, Ekins et al., 2010, Ekins et al., 2010) using Discovery Studio 2.5.5 (Bender et al., 2007, Hassan et al., 2006, Klon et al., 2006, Prathipati et al., 2008, Rogers et al., 2005). Models were validated using leave-one-out cross-validation in which each sample was left out one at a time, a model was built using the remaining samples, and that model utilized to predict the left-out sample. Each model was internally validated and receiver operator characteristic (ROC) plots generated, and the cross-validated ROC area under the curve (XV ROC AUC) calculated. All models generated were additionally evaluated by leaving out 50% of the data and rebuilding the model 100 times using a custom protocol for validation, to generate the ROC AUC, concordance, specificity and selectivity.

Principal Component Analysis

Principal Component Analysis (PCA) in Discovery Studio 2.5.5 (Accelrys, San Diego, CA) was used to compare the molecular descriptor space for the dose response data for the three datasets from SRI (Ananthan et al., 2009, Maddry et al., 2009, Reynolds et al., 2012) as well to compare the actives from the MSLMR dataset and the Asinex database (using ALogP; molecular weight; number of hydrogen bond donors (HBD); number of hydrogen bond acceptors (HBA); number of rotatable bonds (RBN); number of rings; number of aromatic rings and molecular fractional polar surface area (PSA) calculated in the software).

Additional Cheminformatics

Molecular properties of TCMDC-125802 were determined in MolPrime (Molecular Materials Informatics Inc., Montreal, Canada))(Clark et al., 2012). Molecular similarity was calculated using MDL fingerprints and the Tanimoto Similarity algorithm in Discovery Studio Version 3.5.5 (Accelrys, San Diego, CA).

Mtb growth inhibition assays

SRI

The Mtb HTS assay was modified from that previously described (Collins and Franzblau 1997) by using black, clear-bottom, 384-well microtiter plates and 7H12 broth. It should be noted that the initial screening concentration and final units for the reported inhibition were driven by how the libraries were supplied (see (Ananthan et al., 2009, Maddry et al., 2009, Reynolds et al., 2012)) as some were μM and some were μg/mL.

Typically, compounds stocks of 10 μg/mL in 100% DMSO were diluted in assay media and 25 μL of these diluted compounds were transferred to 384-well plates. Amikacin was included in the positive control wells in every assay plate in two concentrations, 0.13 and 2.5 μg/mL. The low concentration is the approximate MIC and is an indicator of proper assay performance of each plate. The high concentration completely inhibits growth and is used in lieu of uninoculated medium (background) to calculate percent inhibition by the test compounds for each plate. Plates containing test compounds (320 compounds/plate) and positive control compounds were transferred into the BSL3 facility for bacteria addition and incubation. The Mtb stock H37Rv was diluted to 1-2 × 10⁵ CFU/mL in the assay medium, Middlebrook 7H12 broth (7H9 broth supplemented with 0.1% casitone, 5.6 μg/mL palmitate, 0.5% bovine serum albumin and 4 μg/mL catalase) and 25 μL was plated over the compounds. Positive and negative control wells were included in each plate. Amikacin was included in one of the compound wells as an internal control in dose-response runs. Plates were placed in stacks of two inside double low density polyethylene bags and incubated for 7 days at 37 °C with approximately 90% humidity. After 7 days of incubation, end point reagent (two parts Alamar blue (Trek Diagnostics) + 1.5 parts 18.2% Tween 80 (Difco) diluted in Milli-Q water) was added to all wells in a volume of 9 μL per well. The plates were returned to the incubator for an additional 18-20 hours. Plates were sealed and bottom read for fluorescence using a Perkin Elmer Envision plate reader at 535 nm excitation and 590 nm emission.

Each assay run contained one plate of inoculated medium (sterility control), another plate containing inoculated medium (growth control), and a 96-well plate with ethambutol at the approximate MIC (0.5 μg/mL) and 20 times the MIC (10 μg/mL). In addition to fluorometric reads, these plates were read at an absorbance of 615 nm (the approximate peak wavelength for oxidized dye) and were used to monitor the quality of the Alamar blue as well as adequate growth of the organism. Expected absorbance readings were about 0.8 and 0.2 for a good dye reagent (medium only) and growth control wells, respectively. The ethambutol plate was used to help confirm that a contaminating organism was not present after incubation since mycobacteria are generally susceptible while other genera are resistant.

Data were analyzed using IDBS Activity Base (IDBS, Guildford, UK). Results of the single-dose screen were expressed as percent inhibition (% Inhibition), which was calculated as: 100*((Median Cell Ctrl – High Dose Ctrl Drug) – (Test well-High Dose Ctrl Drug))/(Median Cell Ctrl – High Dose Ctrl Drug). The dose-response data was analyzed using a four parameter logistic fit (Excel Fit equation 205) with the maximum and minimum locked at 100 and 0. From these curves, TB IC₉₀ and TB IC₅₀ values were calculated for Mtb.

UMDNJ–NJMS

Each compound was dissolved in DMSO at a final concentration of 12 mg/mL and serial dilutions were performed to generate test concentrations ranging from 32 μg/mL to 0.488 ng/mL. M. tuberculosis strain H37Rv at the mid-logarithmic stage of growth (OD₅₈₀ = 0.4) was diluted 1:100 and 0.1 mL was added to each well of a 96-well plate along with 0.1 mL of test compound solution. After 6 days of incubation at 37°C, Alamar Blue (Invitrogen, Grand Island, NY) reagent was added along with 12.5 μL of 20% Tween 80 (Sigma, St. Louis, MO) to evaluate bacterial cell viability. Plates were scanned 24 h later at 570 nm with a reference wavelength of 600 nm utilizing a Biotek Instruments ELX 808. Inoculum control wells of untreated H37Rv were used to create a survival inhibition curve with each assay. Rifampicin was used as a positive control (MIC = 0.0125 - 0.05 μg/mL).

Minimum bactericidal concentration (MBC) determination for TCMDC-125802

Following a literature protocol (Xie et al., 2005), M. tuberculosis H37Rv grown to the exponential phase (A₆₀₀=0.5) was adjusted to 5 × 10⁵ CFU/mL in 2 mL of Middlebrook 7H9 medium supplemented with 10% ADS, 0.25% glycerol and 0.5% Tween 80. Bacterial cultures were incubated 37 °C by shaking at low rpm after treatment with various concentrations of TCMDC-125802 (0.125, 0.25, 0.5, 1.0, 2.0, 4.0 μg/mL) in DMSO. Following nine days of incubation, bacterial cultures were serially diluted with sterile PBS-Tween 80, plated on Middlebrook 7H11 plates and CFUs were enumerated following 21-day incubation at 37 °C. Bacterial CF Us were represented as a mean ± standard deviation of triplicate samples per experimental condition.

Killing kinetics for TCMDC-125802 and INH

M. tuberculosis H37Rv was grown in 7H9 medium supplemented with ADS, 0.5% glycerol, and 0.05% Tween 80 to the exponential phase (A₆₀₀=0.5) at 37 °C. TCMDC-125802 or INH was dissolved i n DMSO at the appropriate concentration and added to 2 × 10⁷ CFU/mL of actively growing bacterial cultures. 100 μL of culture was collected at 2, 9, 14, and 21 days post-treatment, 10-fold serially diluted with sterile PBS-Tween 80 and placed on Middlebrook 7H11 agar plates (Sigma-Aldrich). Bacillary CFU was enumerated after 21 days incubation at 37 °C. The killing curves were plott ed using GraphPad Prism 4 (GraphPad Software, La Jolla, CA).

Isolation of bone marrow derived macrophages

Bone marrow-derived macrophages (BMM) were isolated from the femurs of 8-week-old female B6D2F1 mice (Jackson Laboratories, Bar Harbor, ME) as described (Ehrt et al., 2001). Macrophages were differentiated in Dulbecco's Modified Eagle Medium (Gibco, Grand Island, NY) supplemented with 10% Fetal Bovine Serum, 1% sodium pyruvate, 1% L-glutamine, 20% L929 cell-conditioned medium and 1% penicillin and streptomycin and cultured in a 5% CO₂ incubator for 7 days for further differentiation.

Cellular toxicity assays

Vero Cells

Vero cells (African Green monkey kidney epithelial cells; ATCC) were plated at 1×10⁵ cells per well in a 96 well plate and incubated for 2–3 hours to allow cells to settle. TCMDC-125802 was dissolved in DMSO at a final concentration of 12 mg/mL. Serial dilutions were performed to generate test concentrations ranging from 256 to 0.125 μg/mL. TCMDC-125802 was then added to plated cells resulting in final test concentrations of 128 to 0.0625 μg/mL. To evaluate bacterial cell viability, 20 μL of a 1:20 MTS/PMS (Promega, Madison, WI) reagent was added to each well after 72 hours of incubation at 37 °C and the plate was then incubated for an additional 3 hours. The plate was scanned at an absorbance of 490 nm utilizing a Molecular Devices SpectraMax M5 microplate reader. The CC₅₀ was extrapolated by plotting absorbance at 490 nm versus concentration of cells of untreated Vero cells in control plates.

Mouse Bone Marrow-Derived Macrophages

The toxicity of TCMDC-125802 for the macrophages was determined by the 3-(4,5-dimethyltiazol-2-yl)-2,5-diphenyltetrazoliuim bromide (MTT) cytotoxicity assay with the Vybrant MTT Cell Proliferation Assay Kit (Molecular Probes, Eugene, OR). Briefly, cells were plated at 10⁵ cells/well in a 96-well plate in triplicate. Following 3 h incubation with MTT at 37 °C in a 5% CO ₂ incubator, the tested molecule in various concentrations was added to the cells and the cytotoxicity determination was performed at 48 h post-treatment. Following 3 h incubation, solubilization/stop solution was added to each well. One hour later, the absorbances at 570 nm were read with a VersaMax ELISA microplate reader (Molecular Devices, Sunnyvale, CA). The values were normalized with those of medium control.

Maximum Tolerated Dose (MTD) assessment

TCMDC-125802 (formulated in 0.5% methyl cellulose) was administered as a single dose (30, 100, and 300 mg/kg) by gavage to eight-week-old C57BL/6 mice for three consecutive days. These mice were then observed over the period of seven days for adverse effects.

Acute model of infection

A low dose aerosol exposure of M. tuberculosis Erdman was used to infect twelve-week-old interferon-γ knockout C57BL/6 (GKO) mice (Lenaerts et al., 2003). Three groups of five mice were used: group 1 – positive control of INH dissolved in 1% methyl cellulose (25 mg/kg), group 2 – untreated negative control, and group 3 – methyl cellulose (0.5%) formulated TCMDC-125802 (synthetic material) via oral gavage. Thirteen days post infection the positive control and drug-treated mice were administered INH and TCMDC-125802, respectively, for 9 consecutive days (until day 21). One-day post cessation of TCMDC-125802 dosing (day 22), the animals were euthanized. The number of viable organisms was determined by serial dilution of organ homogenates on nutrient Middlebrook 7H11 agar plates (Becton, Dickinson and Company, Sparks MD.) The plates were incubated at 37 °C for 4 weeks prior to the counting of viable M. tuberculosis colonies.

Highlights.

• Including information on efficacy and cytotoxicity improves performance of in silico drug discovery methods

• Antitubercular hits, leads, and chemical probes identified using Bayesian models

• Most advanced compound provides significant in vitro activity and in vivo safety

• Demonstrated orders of magnitude higher hit rates than current HTS practices

Table 1.

Dual-event Bayesian Mtb model predictions and experimental verification in vitro with subset of the antimalarial hit library

Compound TCMDC#	Chemical Structure	Bayesian Score (Rank)	MIC (μg/mL)	% Inhibition HepG2 @ 10 μM compound
123868		5.73	>32	40
125802		5.63	0.0625	5
124192		5.27	2.0	4
124334		5.20	2.0	4
123856		5.09	1.0	83
123640		4.66	>32	10
124992		4.55	1.0	9

The Compound TCMDC # was assigned by GlaxoSmithKline, as detailed in their publication (Gamo et al., 2010). The Bayesian score was calculated utilizing the TAACF-CB2 dose response and cytotoxicity model. The MIC for each compound was determined versus Mtb H37Rv. The HepG2 cytotoxicities were obtained from Gamo and colleagues (Gamo et al., 2010).