Bayesian Models for Screening and TB Mobile for Target Inference with Mycobacterium tuberculosis

Abstract

The search for compounds active against Mycobacterium tuberculosis is reliant upon high throughput screening (HTS) in whole cells. We have used Bayesian machine learning models which can predict anti-tubercular activity to filter an internal library of over 150,000 compounds prior to in vitro testing. We used this to select and test 48 compounds in vitro; 11 were active with MIC values ranging from 0.4 µM to 10.2 µM, giving a high hit rate of 22.9%. Among the hits, we identified several compounds belonging to the same series including five quinolones (including ciprofloxacin), three molecules with long aliphatic linkers and three singletons. This approach represents a rapid method to prioritize compounds for testing that can be used alongside medicinal chemistry insight and other filters to identify active molecules. Such models can significantly increase the hit rate of HTS, above the usual 1% or lower rates seen. In addition, the potential targets for the 11 molecules were predicted using TB Mobile and clustering alongside a set of over 740 molecules with known M. tuberculosis target annotations. These predictions may serve as a mechanism for prioritizing compounds for further optimization.

Introduction

The search for drugs to prevent or treat infectious diseases is an urgent research focus both in academia and across the pharmaceutical industry. In recent years there has been an increase in the efforts around high throughput screening (HTS) for Mycobacterium tuberculosis, in order to find compounds as therapeutics against tuberculosis (TB) ^1–6, A recent review of the state of TB research has summarized the limited pipeline of molecules in various drug discovery/development stages ⁷. Collaborative efforts that coordinate fragmented TB research efforts by individual groups will be critical to improve the chances of success in both identifying new targets and finding new molecules that could target them. Such efforts include the initiatives funded by NIAID, the Bill and Melinda Gates Foundation (BMGF) and the FP7 funded More Medicines For Tuberculosis (MM4TB) project.

The pipeline for TB therapeutics had not produced a new approved drug in over 40 years until the recently FDA-approved Bedaquiline ^8–10, although there are several candidates in the clinic ^{9, 11}. Only a tiny fraction of TB targets have been addressed with approved drugs or early leads ¹² and recent testing has targeted additional proteins (e.g. MmpL3 ¹³). The relative lack of success with target-based screening compared with whole cell phenotypic screening is a pattern observed for other antibacterial targets, reflecting the difficulty of target-based high-throughput screening for novel antibiotics ¹⁴. In pharmaceutical companies, computational approaches are widely used to aid in drug discovery, but these have not been as exhaustively applied or validated for TB research. For example, virtual screening of compound libraries is used as a complement to high-throughput screening in vitro for many diseases ^15–21. Computational approaches applied to TB have been generally used by specialists focused on a single target or series of compounds and rarely in combination with other computational tools ^{22, 23}. We recently exhaustively reviewed this topic ^{22, 24}, as computational methods are used in workflows by many pharmaceutical company project teams ¹⁸. We found several gaps when we look at how computational methods could be used in TB drug discovery including limited use of filtering for drug-likeness or lead-likeness ²⁵, target deconvolution ²⁶, lack of sequential virtual and biochemical screening and lack of in silico ADME/Tox model use²². A clear disconnect was noted between the generation, utilization, dissemination, sharing and reuse of computational models and the entire drug discovery process ²².

We have proposed using recently retrospectively validated Bayesian machine learning models for M. tuberculosis ^{25, 27, 28} for prospective compound evaluation. Three recent studies have also explored the optimization of these models by combining bioactivity and cytotoxicity data ^29–31 and delivered hit rates in excess of 20%. In the current study we have validated the use of three Bayesian machine learning models by prospectively selecting a small percentage of an in house library for testing. We have identified 11 compounds with in vitro activity and predicted their potential targets using ligand-based computational approaches.

Experimental Methods

Chemicals

Compounds were purchased from ChemBridge (San Diego, CA), ChemDiv (San Diego, CA), Maybridge/Thermo Fisher Scientific Inc. (Waltham, MA) and Sigma - Aldrich (St. Louis, Mo).

CDD Database and SRI datasets

The development of the CDD TB database (Collaborative Drug Discovery Inc. Burlingame, CA) has been previously described ²⁵. The Tuberculosis Antimicrobial Acquisition and Coordinating Facility (TAACF) and Molecular Libraries Small Molecule Repository (MLSMR) screening datasets ^2–4 were collected and uploaded in CDD TB from sdf files and mapped to custom protocols ³². All of the public M. tuberculosis datasets are available for free public read-only access and mining upon registration, making them a valuable molecule resource for researchers along with available contextual data on these samples from other non M. tuberculosis assays. These datasets are also publically available in PubChem ³³. The IDRI database and screening data used in modeling is proprietary.

Machine learning models for M. tuberculosis

We have previously described the generation and validation of Laplacian-corrected Bayesian classifier models ^{25, 27, 28} developed with single point screening and dose response data. In this study we have generated Laplacian-corrected Bayesian classifier models using Discovery Studio 2.5.5 ^34–38 Molecular function class fingerprints of maximum diameter 6 (FCFP_6) ³⁹, AlogP, molecular weight, number of rotatable bonds, number of rings, number of aromatic rings, number of hydrogen bond acceptors, number of hydrogen bond donors, and molecular fractional polar surface area were calculated from input sdf files using the “calculate molecular properties” protocol to distinguish between compounds that are active against M. tuberculosis and those that are inactive in this study. A Bayesian classifier model with the molecular descriptors described above was built using the “create Bayesian model” protocol and IDRI % inhibition at 20 µM for 1106 samples (308 active with >90% inhibition) ⁴⁰. Each model was validated using leave-one-out cross-validation. Each sample was left out one at a time, and a model built using the results of the samples, and that model used to predict the left-out sample. Once all the samples had predictions, a receiver operator curve (ROC) plot was generated, and the cross validated (XV) ROC area under the curve (AUC) calculated (Table 1). 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, in order to generate the XV ROC and AUC (Table 1). These models were also used for screening the “Infectious Disease Research Institute (IDRI) library” of 156,719 compounds with M. tuberculosis activity.

Table 1.

Mean (SD) leave one out and leave out 50% × 100 cross validation of M. tuberculosis Bayesian models (ROC = receiver operator characteristic)

Datasets (number of molecules)	Leave one out ROC	Leave out 50% × 100 External ROC Score	Leave out 50% × 100Internal ROC Score	Leave out 50% × 100 Concordance	Leave out 50% × 100 Specificity	Leave out 50% × 100 Sensitivity
IDRI % inhibition at 20uM (1106)	0.82	0.77 ± 0.02	0.78 ± 0.02	73.4 ± 2.98	77.3 ± 6.28	63.4 ± 7.62

M. tuberculosis assays for biological activity

Molecules were screened at a single concentration of 20 µM in Middlebrook 7H9 medium plus 10% v/v OADC (oleic acid, albumen, dextrose, catalase) and 0.05 % w/v Tween 80; actives were classified as having ≥90% inhibition of growth of M. tuberculosis H37Rv after 5 d ⁴⁰. MICs were determined in liquid medium ⁴¹; briefly a 10 point serial dilution of compounds was run and % growth of M. tuberculosis determined after 5 days incubation ⁴¹. Curves were generated using the Gompertz fit and MICs determined as minimal concentration required to inhibit growth completely.

Target prediction for IDRI compounds

Over 700 compounds with known M. tuberculosis targets were collated from the literature ⁴² and made available in the mobile application TB Mobile (Collaborative Drug Discovery Inc. Burlingame, CA) which is freely available for iOS and Android platforms ^{12, 43}. This dataset was recently updated to 745 compounds and covers over 70 targets. Molecules representing hits from screening in this study were input as queries in TB Mobile and the similarity of all molecules calculated in the application. The top most structurally similar compounds were used to infer M. tuberculosis targets. In most cases multiple targets are shown were the top 2–3 molecules had different targets. The 745 compounds with known M. tuberculosis targets and the hit compounds from this study were used to generate a Principal Component Analysis (PCA) using the interpretable descriptors used for machine learning model building previously in Discovery Studio (AlogP, molecular weight, number of rotatable bonds, number of rings, number of aromatic rings, number of hydrogen bond acceptors, number of hydrogen bond donors, and molecular fractional polar surface area). 1200 M. tuberculosis screening hits (actives and non-toxic only from the SRI screens ^29–31) were used to show how they covered the target-chemistry PCA space alongside the 745 compounds.

The 745 compounds with known M. tuberculosis targets and the hit compounds from this study were also clustered (100 clusters) using MDL fingerprints in Discovery Studio, and the position of the screening hits in specific clusters identified along with the targets of the other molecules in these clusters. In cases when a hit was a singleton, the identity of targets for clusters around a hit was noted. This clustering approach can also be used to infer targets alongside TB Mobile.

Results

IDRI – Bayesian model

A Bayesian model was generated with whole cell M. tuberculosis data for 1106 previously described TAACF and MLSMR actives and inactives [34]. The leave one out ROC was 0.82 and this decreased slightly (0.77) with internal validation with leave out 50% × 100 (Table 2). The concordance (73.4%), specificity (77.3%) and selectivity (63.4%) were in line with the other models described previously (Table 1) ^{25, 27, 29, 30}. Using the FCFP-6 descriptors we can identify those substructure descriptors that contribute to the M. tuberculosis activity in the training set including imidazole, benzothiazole and quinolone, (Figure 1) and those that are not present in active compounds including acetamide, thioether, pyrrole, phenylether and piperazine (Figure 2).

Table 2.

Hits from IDRI library selected with Bayesian models. Promiscuity is calculated from activity of compound in PubChem as this represents a very large database of compounds and bioassays. Tanimoto similarity is reported when compared to the CDD database which contains over 300,000 compounds screened in vitro against Mtb. Target prediction was performed with TB Mobile and clustering as described in the Materials and Methods.

Compound	Structure	Bayesian Model & Score	MIC90 (µM)	Promiscuity	NIH PubChem Notes	Highest Tanimoto similarity in CDD public TB datasets	Prediction with TB Mobile	Prediction with Clustering
IDR- 0157809		MLSMR 20.01	5.0	0.18	Active in 3 out of 17 bioassays including against Escherichia coli Pseudomonas aeroginosa and Staphylococcus aureus	95%	GyrA (Rv0006) GyrB (Rv0005)	GyrA (Rv0006) GyrB (Rv0005) MurD (Rv2155c) cluster 84
IDR- 018217*		MLSMR 21.29	4.4	0	No Data	99%	GyrA (Rv0006) GyrB (Rv0005)	GyrA (Rv0006) GyrB (Rv0005) MurD (Rv2155c) cluster 84
IDR- 0159662		IDRI 20.45	9.4	0	Tested in 7 assays inactive in all - no bacterial species tested	97%	InhA (Rv 1484) ThiL	InhA (Rv 1484) cluster 80
IDR- 0168354		IDRI 19.41	4.8	0	Tested in 7 assays -inactive in all - no bacterial species tested	93%	InhA (Rv 1484) ThiL (Rv2977c)	InhA (Rv 1484) cluster 80
IDR- 0171075		IDRI 20.47	10.2	0	Tested in 1 assay - inactive in all - no bacterial species tested	97%	KasA (Rv2245)	InhA (Rv 1484) cluster 80
IDR- 0173634		MLSMR 17.91	2.2	0	No Data	100%*	InhA (Rv 1484)	GyrA (Rv0006) cluster 53
IDR- 0229683*		MLSMR 27.39	0.4	0.76	Ciprofloxacin - active in 6698 out of 8773 bioaasays including against M tuberculosis	80%	GyrA (Rv0006)	GyrA (Rv0006) GyrB (Rv0005) murD (Rv2155c) cluster 84
IDR- 0198303		IDRI 20.79	2.4	0	Tested in 1 bioassay - inactive	89%	FolP1 (Rv3608c) FolP2 (Rv1207) Rv1885c	Singleton in cluster 89 Dxs1 (Rv2682c) cluster 88 MbyA (Rv2384) cluster 90
IDR- 0204586		MLSMR+ cytotox 38.11	6.3	0	No Data	100%*	PtpA (Rv2234)	Molecule a singleton in cluster 72 RplJ (Rv0651) TlyA (Rv1694) cluster 71 Alr (Rv3423c) Cluster 73
IDR- 0218592		MLSMR 21.71	9.9	0	No Data	64%	ThiL (Rv2977c)	RpoB (Rv0667) cluster 95
IDR- 0236229*		MLSMR 22.48	8.7	0.793991	Perfloxacin - active in 185 out of 233 bioassays including Mycobacteriuim leprae	100%	GyrA (Rv0006) GyrB (Rv0005)	GyrA (Rv0006) GyrB (Rv0005) MurD (Rv2155c) cluster 84

^*not cytotoxic based on Vero cell toxicity data in CDD.

Figure 1.

Good features identified in the IDRI Bayesian Model

Figure 2.

Bad features identified in the IDRI Bayesian Model

IDRI - Prospective testing of the Bayesian Models

The previously published MLSMR dose response model ²⁵, MLSMR dose response and cytotoxicity model ²⁹ and the IDRI Bayesian model were used to rank the “IDRI library” of 156,719 compounds for M. tuberculosis activity. This library can be considered leadlike based on the mean Molecular weight (344.5), log P (3.3), hydrogen bond donors (1.0), hydrogen bond acceptors (3.6) and other properties (Supplemental Figure 1). After ranking the library with the Bayesian score derived from each Bayesian model, the top 1000 compounds were selected and analyzed. There was minimal overlap between all three models and compounds in the top scoring 1000 (Figure 3). The MLSMR models overlapped to the greatest degree (over 20% of the top 1000). Forty eight compounds were selected from these ranked lists and tested in vitro; 11 of these were classed as hits (22.9% hit rate) as they possessed anti-tubercular activity with MIC <10 µM (Table 2). To illustrate the diversity of hits (Table 2), this included five quinolones including ciprofloxacin, three azole containing molecules with long aliphatic linkers and three singletons. Six compounds were found with the MLSMR dose response model, four were found with the IDRI model and one with the MLSMR dose response and cytotoxicity model. The Tanimoto similarity of the 11 compounds were compared to all the publically accessible TB related datasets and these ranged from 64–100%.

Figure 3.

Venn diagram showing the overlap of IDRI library compounds selected with the MLSMR dose response model, the MLSMR dose response and cytotoxicity model and the IDRI model for M. tuberculosis whole cell activity.

Target prediction for IDRI compounds

The PCA model of compounds with annotated M. tuberculosis targets represents the target-chemistry space and 88.7% of the variance is explained by the 3 principal components. The 11 hit compounds from this study were also added to this set and show they are clustered in a relatively narrow region (Figure 4A). Similarly, the hits from previous SRI screens only partially cover the target space (Figure 4B). Clustering the 11 hits with the 745 compounds with annotated target information enabled complementary target predictions with those based on molecular similarity performed with TB mobile (Table 2, Supplemental Figure 2). The known gyrase inhibitor class, fluoroquinolones were well predicted by both target inference methods. The remaining compounds had divergent predictions apart from the azoles, which were predicted to be InhA inhibitors.

Figure 4.

Principal Component Analysis of 745 compounds with A. known M. tuberculosis targets (Blue) from TB Mobile and 11 screening hits (yellow) and B. 1200 active and non toxic compounds from SRI screens (yellow)

Discussion

Drug discovery is time consuming and very costly ^{44, 45} such that any tools that can point out liabilities earlier will have considerable value ^{21, 46, 47}. The need for new anti-tubercular therapies is unquestioned in the face of drug resistance that has progressed to the point of the identification of totally drug resistant strains in India ⁴⁸ and a call for re-opening the TB sanatoria that were closed more than 60 years ago ⁴⁹. To address the challenge of drug resistance in TB infection, many groups have turned to HTS campaigns with chemically diverse libraries of small molecules to identify novel starting points for drug discovery ^{1, 50}. The TB community must now ask how to mine efficiently and leverage this growing database to provide new drug candidates, in the face of well known complications such as latency and persistence ⁵¹ and the numerous issues associated with typical HTS data ⁵². To help answer this question, we have identified a significant opportunity for the tuberculosis drug discovery community to harness pharmaceutical industry-tested computational methodologies ²². Subsequently, we turned in part to the cheminformatics methods which occupy an important place in the industrial drug discovery workflow. Ligand- and protein-based methods, for example, have been used as a complement to high-throughput screening in vitro ¹⁵. In order to validate the predictions from such methods we are required to test molecules for their whole cell TB activity.

We have developed and utilized machine learning models for M. tuberculosis ^{25, 27, 28} using large publically accessible HTS data sets ^{3, 4}. During retrospective validation of these models we observed at least 4–10 fold enrichment in identifying TB actives in the top scoring molecules ²⁸. These results indicated that using whole cell screening data from one laboratory for computational models can be used to benefit other laboratories via predictions of their compounds of interest and narrowing down the number of compounds to be tested in vitro ²⁸. We have recently updated our approach to incorporate cytotoxicity data into the models ^29–31. These previously published Bayesian models had considerably higher hit rates than random HTS screening ^{29, 30}. One study virtually screened over 82,000 molecules, 550 were tested in vitro and 124 actives were identified in total (22.5% hit rate) ³⁰. A second study virtually screened over 38,000 molecules, tested 106 in vitro and identified 17 actives (22.5% hit rate) ²⁹. In the current study we utilized several previously published models as well as a newly constructed model generated with new data from >1000 previously published active compounds. Three Bayesian models were ultimately used to screen the in house library of 156,719 molecules and 48 (0.03%) of the compounds were tested in vitro resulting in 11 hits. Again this confirmed the hit rates previously observed with a value of 22.9%. In our experience and as a point of contrast, the whole cell HTS hit rate for the IDRI group has varied from 0.6 – 2% depending on the assay (unpublished).

Using several Bayesian models for M. tuberculosis activity to prioritize compounds from a screening library of this size is by far the largest such analysis we have performed to date to our knowledge ^29–31. The results obtained further validate the hypothesis that Bayesian models ^{25, 27–31, 42} identify subsets of compound libraries enriched with active compounds, therefore requiring the testing of far fewer compounds. Future research will involve investigating open source descriptors and algorithms that can enable deploying such models more widely ^{53, 54}. This Bayesian modeling and virtual screening approach is also applicable to other neglected diseases.

This study also further utilized a recently developed mobile application for inferring potential M. tuberculosis targets for the 11 hits (Table 2, Supplemental Figure 2). This application draws together known small molecules and their annotated targets as well as other information relevant to the pathway targeted, essentiality, human ortholog etc. ^{12, 42}. Generating predictions with this application was also complemented by using clustering of the known compounds with targets and assessing which clusters the 11 compounds were in. The fluoroquinolone compounds are not surprisingly predicted as gyrase inhibitors using both target inference methods, apart from IDR-0173634 which is also predicted as a potential inhibitor of InhA. Although azoles are well known cytochrome P450 inhibitors ^{55, 56} they are predominantly predicted as targeting InhA. These and the remaining 3 singleton compounds with different predicted targets with no concordance with the target inference methods would be worthy of testing in vitro. Our analysis of the 11 hits suggest, as one would expect, that they are covering a very narrow section of chemistry and target space. In particular we have multiple fluoroquinolones and azoles (Table 2) so these may essentially count as a single data point in each case. Our approach (using known compounds with Mtb targets) to infer potential targets for similar compounds is more conservative than methods which would use similarity to compounds known to be active against targets in other organisms. Such target prediction efforts would help us to prioritize targets to test.

In conclusion we have presented an approach using multiple Bayesian models to prioritize compounds for testing which identified active compounds. These in turn were used with TB Mobile ¹² and clustering as mechanisms for predicting potential targets for compounds in M. tuberculosis, thereby serving as an approach for further identifying the best compounds for optimization. Such computational workflows leveraging prior knowledge further our aim of optimally using and integrating the data and resources available to us in order to accelerate drug discovery for M. tuberculosis ²².