Data-mining of potential antitubercular activities from molecular ingredients of traditional Chinese medicines

Data sets

Molecular data sets of ingredients of traditional Chinese medicines were retrieved from traditional Chinese medicines Integrated Database (TCMID) (Ruichao et al., 2013). TCMID constitutes one of the most comprehensive online resources for ingredients used in TCM. The database hosts information on over 25,210 pure molecules retrieved from literature and other data resources.

Computational models for antitubercular activity

The computational predictive models used in our analysis were based on the following two confirmatory screens conducted to identify novel inhibitors of Mycobacterium tuberculosis H37Rv, previously published by our group (Periwal et al., 2011; Periwal, Kishtapuram & Scaria, 2012). The computational models used are available online at http://vinodscaria.rnabiology.org/2C4C/models.

Briefly these models were based on two bioassays deposited in PubChem and carrying IDs AID 1332 and AID 449762. Both the assays were based on microdilution Alamar Blue assays. The former used 7H12 broth while the latter used 7H9 media. A total of 1,120 and 327,669 compounds were screened in the respective assays. The models were generated using a machine learning approach as described in Periwal et al. (2011) and Periwal, Kishtapuram & Scaria (2012). The AID 1332 assay model was generated based on the Random forest classification algorithm and was evaluated using a variety of statistical measures which include accuracy, Balanced Classification Rate (BCR) and Area under Curve (AUC). Balanced Classification Rate is an average of sensitivity and specificity which introduces a balance in the classification rate. The model had an accuracy of 82.57%, BCR value of 82.2% and AUC value of 0.87. The AID 449762 assay model was generated based on SMO (Sequential Minimization Optimization) algorithm and was found to be 80.52% accurate, with BCR value of 66.30% and AUC as 0.75.

In addition, we created an additional model to predict the molecules active against non-replicating drug tolerant Mycobacterium tuberculosis. The assay was deposited in PubChem with identifier AID 488890. A total of 3,24,437 compounds were screened for the activity. The model was generated using Random forest classification algorithm as described in the previous papers (Periwal et al., 2011; Periwal, Kishtapuram & Scaria, 2012; Jamal, Periwal & Scaria, 2013; Jamal, Periwal & Scaria, 2012) and had an accuracy of 76%, BCR value 85.2% and AUC 0.66.

Molecular descriptors

Molecular descriptors for each of the molecules were computed using PowerMV (Liu, Feng & Young, 2005), popular cheminformatics software widely used to compute molecular descriptors. A total of 179 molecular descriptors were computed for each molecule. Out of the total 179 molecular descriptors, a few descriptors were pruned using bespoke scripts written in Perl depending on whether they were used in creating the respective models. We pruned a total of 29 and 25 descriptors corresponding to AID 1332 and AID 449762 respectively, while 25 were pruned for the AID 488890 model.

Formats and format conversion

The molecules were downloaded in mol2 format and converted to SDF (Structural Data Format) format using Openbabel (O’Boyle et al., 2011). The molecular descriptors were converted to ARFF format compatible with Machine learning toolkit Weka (Bouckaert et al., 2010). We used custom scripts written in Perl for the format conversions. A complete list of scripts is also available at Crowd Computing for Cheminformatics (2C4C) repository at URL: http://vinodscaria.rnabiology.org/2C4C/models.

SMARTS filters

The SMARTS filter is employed to remove the molecules with fragments leading to toxicity or unwanted reactivity. We used a set of SMARTS filters for the consensus candidate anti-tubercular molecules. The online server SMARTSfilter (http://pasilla.health.unm.edu/tomcat/biocomp/smartsfilter) web application was used for all comparisons. The web application was used to filter out molecules, which match to any of the five undesirable SMARTS catalogs.

Mycobacterium tuberculosis permeability prediction

The small molecules could not be effective unless they are able to penetrate the cell wall. MycPermCheck (Merget et al., 2013) a computational tool to predict permeability of small molecules across Mycobacterium tuberculosis, was employed to filter the subset of potential active molecules.

Data mining

We used Weka, a popular and freely available Data Mining Software toolkit. Predictions were performed for the dataset across the two models corresponding to assays AID 1332 and AID 449762 independently. Further, molecules predicted active in both the datasets were collated and analyzed for additional properties including activity against non-replicating drug tolerant Mycobacterium tuberculosis and potential to permeate the Mycobacterium tuberculosis cell wall. Additional filters which discount molecules with toxic fingerprints were removed using SMARTS filters. The summary of the entire workflow of prioritization is depicted as a Schema (Fig. 1).

Summary of datasets and molecules

A total of 25,210 ingredients were downloaded from Traditional Chinese Medicines Integrated Database (TCMID). We could retrieve molecular information for only 12,018 of the ingredients in the form of SMILE notations and the rest were not considered for further analysis. The molecules considered along with their SMILES are detailed in Table S1. A total of 179 descriptors were calculated using PowerMV as described above. The descriptors were further pruned for each of the models as described in the Materials and Methods section using custom scripts in Perl. This corresponds to 150 and 154 descriptors respectively for models AID 1332 and AID 449762 and 154 for AID 488890. The models, descriptors and scripts for formatting the files are available at the Crowd Computing for Cheminformatics Model Repository (http://vinodscaria.rnabiology.org/2C4C/models).

Prediction of potential anti-tubercular hits

The 12,018 molecules obtained from TCMID were analyzed for the antitubercular activity using the computational predictive models as described above. The AID 1332 and AID 449762 models predicted 2,363 compounds and 5,864 compounds respectively as potentially active anti-tubercular. Of these molecules, a total of 1,472 molecules were predicted potential actives by both the models based on molecular descriptors and were considered for further analysis (Table S2).

Briefly we used a popular approach for filtering molecules with undesirable properties. These included briefly using SMARTS filters. Molecules which passed the filtering step were further evaluated for their effect against drug-tolerant and slow growing Mycobacterium. Molecules were further evaluated for their potential permeability with respect to the Mycobacterial cell wall.

SMARTS filter for filtering undesirable structures

We used a set of five SMARTS filters to remove the molecules matching to any of these filters. Such substructure based filtering approach has been extensively used to prioritize molecules by filtering unwanted or potential false positives in cheminformatics screens (Singla et al., 2013). The SMARTS filters included 5 independent approaches namely Glaxo, PAINS, Oprea, Blake and ALARM-NMR used in tandem. Pan Assay Interference Compounds (PAINS) describes a set of substructures known to be promiscuous and have issues in high throughput assays (Baell & Holloway, 2010), while the Glaxo filter describes unsuitable hits or unsuitable natural products (Hann et al., 2009). ALARM NMR assay to detect reactive molecules by nuclear magnetic resonance (ALARM-NMR) set filters for molecules which are reactive false positives in high-throughput assays by oxidizing or alkylating a protein target (Huth et al., 2005). The Glaxo, Oprea and Blake filters were based on specific fitness properties. The Glaxo method involves classification of the molecules into different chemical categories based on the presence of acids, bases, electrophiles and nucleophiles in the molecule. Prior to the categorization the molecules are filtered for non-drug like properties and to remove inappropriate functional groups (unsuitable leads and unsuitable natural products) (Hann et al., 2009).

Out of a total of 1,472 molecules, 160 molecules passed all the filters. A total of 63.1% (929) molecules failed the ALARM NMR filter, while 49.9% (734) failed to pass Oprea filter. Similarly 49% (722) failed to pass the PAINS filter. The detailed schema showing the number of molecules failed by each filter is depicted in Fig. 2. A similar comparison of the complete set of 12,018 TCMID compounds revealed that only 1,539 compounds passed all the filters. We observed that most of the molecules did not pass through ALARM NMR (60.7%, 7,295) molecules followed by Oprea filter (52.4%, 6,303) molecules and 5,799, 48.3% molecules could not pass through the PAINS filter.

Molecules potentially active against non-replicating drug tolerant Mycobacterium tuberculosis

A total of 160 compounds filtered through SMARTSfilter were tested using a computational predictive model for potential activity against non-replicative Mycobacterium tuberculosis. The model predicted 19 compounds as active to act as potential inhibitors of non-replicating drug tolerant Mycobacterium tuberculosis. The detailed description about 19 compounds is given in Table 1. The table also shows the permeability probability of the molecules to pass through Mtb cell wall.

Table 1:

List of 19 compounds predicted as active against non replicating antibiotic tolerant Mycobacterium tuberculosis.

Compound no.	Compound structure	Name	English name	Latin name	Permeability probability	Sources with antitubercular activities
1.		F lemichapparin b	Climbing Jewelvine	Derris scandens	0.993
2.		Murrayafoline A	Taiwan Common Jasminorange, Indian Common Jasminorange, Euchretaleaf Common Jasminorange, Narrowfruit Glycosmis Root	Murraya crenulata, Murraya koenigii, Murraya euchrestifolia, Glycosmis stenocarpa	0.98
3.		2-hexenyl benzoate	Common Tea, Szechwan Tangshen	Camellia sinensis, Codonopsis tangshen	0.855
4.		Anonaine	Hindu Lotus Large Rhizome, Bullocksheart Custardapple, Custard Apple, Chinaberry-tree Flower, Uncinate Tailgrape	Nelumbo nucifera, Annona reticulata, Annona squamosa, Melia azedarach, Artabotrys uncinatus	0.52
5.		Orchinol	Frog Orchid, European Gymnadenia, Liriop Equivalent plant: Liriope spicata var prolifera	Coeloglossum viride [Syn. Coeloglossum viride var. bracteatum], Gymnadenia albida, Ophiopogon japonicus	0.407
6.		1-phenyl-1- pentanone	Chuanxiong rhizome, Szechuan lovage root, Chuanxiong (Wallich Ligusticum) Equivalent plant: Cnidium officinale	Radix chuanxiong; G Rhizoma Chuanxiong, Ligusticum chuanxiong	0.338
7.		Brassilexin	India Mustard	Brassica juncea	0.295
8.		Bisacumol	Zedoary Turmeric Equivalent plant: Curcuma kwangsiensis, Common Turmeric Equivalent plant: Curcuma aromatica	Curcuma zedoaria, Curcuma longa	0.104
9.		Totarol	Longleaf Podocarpus Leaf Equivalent plant: Podocarpus macrophyllus var maki, Water Nightshade	Podocarpus macrophyllus, Solanum torvum	0.037	Solanum torvum
10.		Cyclostachine a	Hairspike Pepper	Piper trichostachyon	0.029	Piper trichostachyon
11.		Isolobinine	Indian Tobacco, Chinese Lobelia	Lobelia inflata, Lobelia chinensis	0.018
12.		Urinatetralin	Common Leafflower	Phyllanthus urinaria	0.012	Phyllanthus urinaria (Nair & Abraham, 2008)
13.		2-methoxy-1h- pyrrole			0.004
14.		Gmelofuran	Medicinal Breynia Leaf	Breynia officinalis	0.00
15.		Petasalbin methyl ether	Japanese Butterbur	Petasites japonicus	0.00	Petasites japonicus
16.		Verruculotoxin			0.00
17.		Hinokiol	Yellowish Rabdosia	Isodon flavidus	0.00
18.		Thymine	Przewalsk Fritillary, Anhui Fritillary, Ussuri Fritillary	Fritillaria przewalskii, Fritillaria anhuiensis, Fritillaria ussuriensis	0.00	Fritillaria przewalskii (Chang & Paul, 2001)
19.		n-methylcorydaldine	Fendler’s Meadowrue, Bracteate Poppy, Asiatic Moonseed Root, Lotusleaftung	Thalictrum fendleri, Papaver bracteatum, Menispermum dauricum, Hernandia sonora	0.00	Hernandia sonora

DOI: 10.7717/peerj.476/table-1

Pharmacophore search in 19 molecules found to be potentially active against non-replicating drug tolerant Mycobacterium tuberculosis

Since pharmacophore represents the features which play a key role in the recognition of ligand by the target molecule, we generated the pharmacophore features for the 19 molecules identified to be active against non-replicating drug tolerant Mycobacterium tuberculosis using PharmaGist software (Schneidman-Duhovny et al., 2008). PharmaGist computes the pharmacophore model by doing multiple flexible alignments of the input molecules. We have reported the three highest-scoring pharmacophores which can be used for the discovery of novel drug entities (Fig. 3).

Mycobacterium tuberculosis permeability prediction

We employed MycPermCheck to predict molecular permeability to Mycobacterial cell wall to estimate the potential permeability of the prioritized molecules. All the 160 molecules which passed the five SMARTSfilters were further evaluated for their ability to penetrate Mtb cell wall. Analysis revealed 9 molecules with highest probability (>0.98) to permeate the Mycobacterium cell wall barrier (Table S3).

Literature search suggests evidence of the sources and molecules used with antitubercular properties

We further searched for the role of the plant sources of the molecules with regard to their use or known information on antibacterial or antitubercular activities. We found several molecules in herbs to have antitubercular effects. These are Petasites japonicus (John, 1977), Piper trichostachyon (Wagner & Wolff, 1977), Solanum torvum (Silva et al., 2011), Fritillaria przewalskii (Chang & Paul, 2001), Hernandia sonora (Udino et al., 1999) and Phyllanthus urinari (Nair & Abraham, 2008). In addition, many of the herbs have been shown to have hepatoprotective activities, which include Annona reticulata (Thattakudian Sheik Uduman, Sundarapandian & Muthumanikkam, 2011; Mohamed Saleem et al., 2008), Annona squamosa (Thattakudian Sheik Uduman, Sundarapandian & Muthumanikkam, 2011; Mohamed Saleem et al., 2008), and Camellia sinensis (Issabeagloo & Taghizadieh, 2012). This offers a new opportunity for new drug development considering that most of the established first-line drugs used in the treatment of tuberculosis are hepatotoxic (Yew & Leung, 2006; Liu et al., 2008). We also found these molecules, Hinokiol (Chen et al., 2009), Totarol (Jaiswal et al., 2007), Murrayafoline A (Choi et al., 2006) and 2-hexenyl benzoate (Yantao, Zeng & Zhang) have been known to show antitubercular effects. The 9 molecules, which were identified by MycPermCheck, are able to penetrate the cell wall of Mycobacterium tuberculosis (Table S3). We tried to find if any of these molecules have been observed to show antitubercular or antimicrobial activity and we found 6 molecules that include Thalfinine (Zhou, Xie & Yan, 2011), Azulene (Kurti & Uldrich, 1958), Erypoegin E (Sato et al., 2006), Pongapinone A (Rani et al., 2013), Achilleol C (Iinuma et al., 1996) and Murrayafoline A (Choi et al., 2006).