Drug-Target data.

DrugBank (version 4.0) [3] was used as a reference database to collect comprehensive information about drug-drug target information. The ‘drugbank.xml’ file was downloaded from download section of DrugBank (http://www.drugbank.ca/downloads); it was parsed by custom perl scripts to extract drug, along with its associated information like indication area, targets, SMILES string [4]. The indication area(s) associated with a drug is represented as free text in DrugBank, which poses algorithmic challenge for the process of automated association of drug with its indication area(s). In the current study, we have mapped diseases or indication area associated with drug to its corresponding ICD10 disease code [5], [6] (http://apps.who.int/classifications/icd10/browse/2010/en can be referred for detailed mapping between ICD10 disease code to related diseases).

The file ‘drug-disease_TTD2013.txt’, available from the download section of Therapeutic Target Database (TTD) [7], had been used for drug-disease mapping. This file can be used for unambiguous association of drug with its indication area(s). The files, ‘drug_links.csv’ and ‘TTD_crossmatching.txt’ (TTD), were used to retrieve mapping between DrugBank ID to TTD Drug ID. The data for the approved drugs along with its associated information, like drug target, ICD10 disease classification and SMILES string, was extracted from ‘drugbank.xml’ file. The data of the drugs was segregated into two group, anticancer drugs and other drugs, which is available as online supplementary material–‘DB_cancer.txt’ (see S1 Text) and ‘DB_others.txt’ (see S2 Text), respectively. DrugBank represents target information as UniProt ID, which was mapped into its corresponding Entrez Gene ID and Gene Symbol (based on mapping provided in ‘HUMAN_9606_idmapping_selected.tab’ and ‘gene_info’ files which can downloaded from ftp sites ftp://ftp.uniprot.org/pub/databases/uniprot/current_release/knowledgebase/idmapping/by_organism/HUMAN_9606_idmapping_selected.tab.gz, and ftp://ftp.ncbi.nlm.nih.gov/gene/DATA/gene_info.gz, respectively).

In the current study, ICD10 disease codes ‘C00 to C06’ were considered to represent oral cancer. While reviewing the information of anticancer drugs, we noticed that there are many drugs which are mapped against ICD10 disease code ‘C00-C96’, which is a non-specific disease code for malignant neoplasms. We could not find any drug in DrugBank database, which was indicated to treat oral cancer; therefore, we extended our search to the literature database (NCBI PubMed), and found evidence to control the growth of oral cancer cells by couple of drugs like erlotinib [8], [9], vandetanib [10] and gefitinib [8], [11]. The ICD10 disease code mapped for these drugs were updated manually, to include ‘C00-C06’ as drug indication in ‘DB_cancer.txt’ (see S1 Text). We realized that such a low representation of drugs for oral cancer treatment in public databases, would act as a bottleneck in downstream predictive data mining processes; this prompted us to extend our search beyond compound databases like DrugBank.

Nature is a gold-mine for treatment of various diseases, including cancer, which is evident from the fact that the majority of existing anticancer drugs are either natural products or their chemical derivatives [12]-[14]. We compiled the list of plant based anti-cancer natural compounds by manually mining literature databases like PubMed, and also used Google Scholar to search articles, not indexed with PubMed. A total of 269 articles were referred to collect the data about plant based natural compounds, active against over 25 different cancer types. We collected data for 377 compounds from these articles. The list of plant based compounds with anti-cancer activity was further annotated with associated attributes like PubChem Compound ID (cid), SMILES string, molecular targets. Target information was not present for all the compounds in the base set of articles (269 articles), therefore we further referred 315 more articles to collect target information of un-annotated compounds. The list of plant based natural anti-cancer compounds complied in the current study consists of 30 compounds with growth inhibitory activities against oral cancer cells. The list of plant based natural compounds active against various cancers obtained in the current study, can be found as online supplementary material–‘Natural_Anticancer_list.txt’ (see S3 Text), which contains links to research articles that were used to infer anti-cancer activities of compounds against particular cancer-type, and it also contains reference to articles which were used to infer compound to target association. This is a manually curated list, which can be of great use to researchers working in the field of plant based natural anti-cancer compounds. The data of ‘Natural_Anticancer_list.txt’ (see S3 Text) was further rearranged in a format similar to files obtained after mining DrugBank (see S1 and S2 Texts) to make it amenable for downstream data mining processes; this file can be found as online supplementary material–‘Nat_Anticancer.txt’ (see S4 Text).

Compound-Target Data sources.

ChEMBL—Compound Database. ChEMBL is a freely available database of drug-like bioactive compounds [15]. The compound information present in this database is linked with bioactivity measurements, which are manually extracted from primary published literature. In the current study, we have utilized compound repository of ChEMBL (version 19.0) to be used for prediction of anti-cancer activity. We downloaded MySQL dump of ChEMBL and created a local database (ftp://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/chembl_19_mysql.tar.gz).

In the current study, we used perl libraries DBI and DBH for interfacing with ChEMBL database, created in locally installed MySQL. Perl scripts were written to extract data from the ChEMBL database. We extracted SMILES string along with ChEMBL id from the database with the help of following SQL query—“select c.canonical_smiles, m.chembl_id from compound_structures c, molecule_dictionary m where c.molregno = m.molregno”. A total of 1404752 compounds (i.e. ~1.4 million compounds) along with their SMILES strings were extracted from the database.

STITCH—Chemical-Protein Interaction Database. STITCH is a chemical-protein interaction database which integrates information about interactions from metabolic pathways, crystal structures, binding experiments and drug—target relationships [16]. In the current study, we have downloaded latest dataset from the STITCH database (version 4.0). Following files were downloaded from the download section of STITCH:

1. http://stitch.embl.de/download/protein_chemical.links.v4.0/9606.protein_chemical.links.v4.0.tsv.gz → Chemical-Protein Interaction data which contain over 4.5 million records. Chemicals are derived from the PubChem compound database, and proteins are represented by Ensembl protein identifiers.
2. http://stitch.embl.de/download/chemicals.v4.0.tsv.gz → Contains STITCH compound’s chemical structure information in the form of SMILES string. It contains 82841024 (i.e. ~ 82.84 million) compound records.
3. ftp://ftp.ncbi.nlm.nih.gov/gene/DATA/gene2ensembl.gz → Contains mapping between Ensembl protein identifier to NCBI-Entrez Gene ID.

Benchmark Dataset.

Benchmark dataset prepared for prediction of anti-cancer activity by Li et al. [17] was used in the current study. This dataset is from the NCI-60 Developmental Therapeutics Program (DTP) project [18]. The details of protocol used to create the benchmark dataset, can be found in primary published article [17]. The dataset consist of more than 18,000 compounds, divided into active and inactive anticancer compounds. The benchmark dataset can be downloaded from http://bsb.kiz.ac.cn/site_media/download/CDRUG/Benchmark.rar.

CDRUG.

CDRUG is an analytical method for prediction of anticancer activity of chemical compound [17]. In the current study, we have downloaded and used the latest standalone version of CDRUG for anticancer activity prediction. This tool takes a list of SMILES string of query compounds as an input and generates ranked list consisting of various scores and p value. In the current study, we have considered the cutoff p value of ≤ 0.05, as criteria to select compounds with anticancer activity. The algorithmic details of CDRUG can be found in primary publication [17].

Support Vector Machine (SVM) Classifier.

In the current study, we have built SVM based model for the prediction of anticancer activity of chemical compound. Support Vector Machines are a useful tool for data classification, which has found its application in wide range of domains including computational biology. We have used software LIBSVM (version 3.18) [19] in our current study for SVM based classification. The SVM based classification task starts with the process of “model building”, in which data is divided into training and testing sets. Each instance in the training set contains one “target value” or “class label” (in our case it is either 1 or 0; where ‘1’ represents compound has anti-cancer activity and ‘0’, otherwise), and several “attributes” or “features”. The goal of SVM [20], [21] is to rigorously build a model (based on instances from training data) which predicts the target values / class labels of the instances from test data, given only attributes in the test data. In the current study, we selected ‘C-SVM’ (Multi-class classification) as SVM type, and radial basis function (RBF) as a kernel type for building anti-cancer activity prediction model. RBF kernel was chosen on the basis of its popularity, robustness, and the fact that other kernels available with LIBSVM are special cases of RBF under certain parameter [22], [23].

The process of classification with SVM involves following steps:

1. Model building: In the current study, we have used benchmark dataset [17] (see the section Benchmark Dataset) for building SVM prediction model. The rationale behind the selection of dataset common to that, used by CDRUG [17], was to compare prediction outcomes of two methods (CDRUG and SVM classifier) build from the same underlying dataset. The process of building model involves following sub-steps:
   1. Feature extraction of training compounds and transformation of feature vector into SVM input format.
   2. Cross validation based parameter estimation and building model with best parameters.
2. Prediction of query compounds:
   1. Data processing of query compound(s).
   2. Prediction of anti-cancer activity of query compound(s).

Feature Extraction. In the current study, the features were derived from the entities in the compound, which are responsible for defining its reaction mechanism, and are the contributing factor towards its activity. These entities can be of organic (i.e. ‘functional groups’) or inorganic (i.e. ‘metal ions’) in nature. Functional groups present in organic molecules had been used in the past to predict drug-target interaction networks [24], wherein authors had used 28 functional groups to characterize drugs. In addition to the functional group, metals also play a very important role in determining the activity of drugs, especially in the field of cancer drug, such as cisplatin, which can be regarded as a pioneer in the field of metal based anti-cancer drug [25]. The functional groups and metals present in a compound can be visualized as building block or substructure of a compound. SMARTS is a very powerful language for describing such molecular substructures [26]. SMARTS strings are typically used for substructure searching, to identify molecules based on pattern matching, either a singular string or as a group of SMARTS strings. In the current study, we rigorously prepared SMARTS strings of over 300 functional groups (including common metallic forms found in various drugs). We have followed the guidelines given by Daylight [26], while preparing these SMARTS strings.

Features were extracted from the training compounds, from the Benchmark dataset [17]. The dataset consist of over 18,000 compounds (positive- and negative-set) in SMILES format (refer to: http://bsb.kiz.ac.cn/site_media/download/CDRUG/Benchmark.rar). In the current study, we have used open-source python library Pybel [27] for finding substructures encoded as a SMARTS string in a query compound. Python script was written to automate the task of matching the list of SMARTS stings against the benchmark dataset (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Pseudocode for substructure matching process.

https://doi.org/10.1371/journal.pone.0141719.g002

On reviewing the extracted features of all compounds (positive and negative dataset), we observed that many of the substructures from our initial list of SMARTS string were not present in either of the dataset (i.e. positive- or negative-set), and therefore, they were excluded from the further downstream analysis process. The final list of SMARTS strings along with corresponding representative substructure (functional groups or metal ion) consisted of 228 SMARTS strings, which can be found as online supplementary material–‘SMARTS_pattern.txt’ (see S5 Text). At the end of this exercise, we obtained feature matrix of dimension M Γ N matrix; where ‘M’ corresponds to the number of compounds in benchmark dataset and ‘N’ corresponds to number of features/substructures (i.e. 228) used to prepare feature vector of a compound. This feature vector was transformed into a SVM format as given below:

<label> <index1>:<value1> <index2>:<value2>…

.

.

.

Where, each line contains an instance and is ended by a '\n' character. The <label> is an integer indicating the class label (1→Compound with anti-cancer activity and 0→Compound without anti-cancer activity). The pair <index>:<value> gives a feature (attribute) value: <index> is an integer starting from 1 and <value> is a real number (In the current study, <value> can be [0,1], where 0→indicates feature is absent in the compound, and 1→indicates feature is present in the compound). Indices must be in ascending order [19].

Parameter Estimation and Model Building. The RBF kernel has two parameters C and γ; for a given prediction problem, the value of these parameters is not known beforehand, and therefore, some kind of parameter search has to be done to estimate values of these parameters. The main objective of parameter search is to find good (C, γ), so that the prediction model will accurately predict activity of unknown compounds. Generally poorly optimized models tend to suffer with an overfitting problem, which refers to the condition when prediction model / classifier shows high accuracy with training data, but its accuracy drops drastically when used to predict unknown test data. Cross-validation is a technique which is applied to overcome the overfitting problem. In n-fold cross-validation, training dataset is divided into n subsets of equal size. Sequentially one subset is tested using the model, trained on the remaining n-1 subsets. In this way, each instance of the whole training set is predicted once, so that, the cross-validation accuracy is the percentage of data which are correctly classified.

In the current study, we performed an exhaustive grid—search on C and γ using 5-fold cross-validation. After feature extraction and data transformation of the benchmark dataset (see section Feature Extraction), we first did a coarse grid search for finding best C and γ using 5-fold cross-validation. We first started with coarse grid search with an exponentially growing sequence of C and γ (C = 2⁻⁵, 2⁻⁴, 2⁻³…, 2¹⁴, 2¹⁵ and γ = 2⁻¹⁵, 2⁻¹⁴….2⁴, 2³), which gave us best parameters (C = 2² and γ = 2⁻²) with cross-validation accuracy of 80.99% (Fig 3). The parameters with cross-validation accuracy of over 80.5% are distinctly marked with green color in grid space of Fig 3, we next focused on fine grid search in this region.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Coarse Grid Search for C and γ for parameter estimation.

https://doi.org/10.1371/journal.pone.0141719.g003

The fine grid search was conducted with a growing sequence of C and γ (C = 2⁻¹, 2^−0.75, 2⁻⁵⁰…2^5.50, 2^5.75, 2⁶ and γ = 2⁰, 2^−0.75….2^−4.50, 2^−4.75, 2⁻⁵), which gave us best parameters (C = 2^1.5 and γ = 2^−1.5) with cross-validation accuracy of 81.18% (Fig 4). Whole training set (i.e. the transformed benchmark dataset with feature vectors) was used for building a final classifier with the best parameters (C = 2^1.5 and γ = 2^−1.5). The intermediate files generated during grid search, along with final classifier ‘cancer.model’ can be found as online supplementary material ‘Model_Build.zip’ (S6 Text). In the current study, the classifier ‘cancer.model’ was used in the subsequent SVM based prediction of anticancer activity. The exhaustive grid based parameter search was done with the help of the python script ‘grid.py’ available with LIBSVM package [19]. Computationally grid search is memory and CPU intensive task, in a parallel mode, it took almost 10 days to complete this task in 4 GB Intel^® Core^™ i5 desktop installed with Linux operating system.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Fine Grid Search for C and γ for parameter estimation.

https://doi.org/10.1371/journal.pone.0141719.g004

Prediction Process. The prediction of anticancer activity with SVM classifier ‘cancer.model’ for query compounds involves following steps:

1. Read list of ‘n’ number of query compounds.
2. Set initial index i = 1.
3. Preparation of feature vector for i^th query compound (as explained in section Feature Extraction). The feature vector D_i[x1, x2….x228] for a i^th query compound, would be a binary vector representing the presence or absence of functional group/substructure in a query compound.
4. Check if ‘i’ is less than ‘n’, If yes then i = i+1 and go to step 3, else go to step 5.
5. Transform feature matrix into SVM input format and save as file “svm_input.dat”.
6. Predict with the following command:
   1. ./svm-predict svm_input.dat cancer.model <output_name>

Prediction of Anticancer Activity.

The compounds collected from ChEMBL and STITCH database (see section Compound-Target Data Sources) were given as input to the SVM classifier for the prediction of anticancer activity. There were over 82.84 million compounds from STITCH, and over 1.4 million compounds from ChEMBL databases. Various analytical steps involved in preprocessing (like feature extraction) and SVM prediction, have certain physical memory and CPU requirement which is determined by the size of a dataset and complexity of underlying algorithm, because of these constraints, it was not possible to analyze the whole dataset of over 84 million compounds all at once. After a couple of initial trial runs of prediction workflow with varied sized subsets of the compound dataset, we were able to find upper threshold of 2.6 million compounds which can be analyzed in the desktop with 4GB memory (with 4 cores).

The number of compounds in the dataset, obtained from STITCH (~82.84 million) were marginally above the required threshold of 2.6 million supported by system configuration used in the current study. The compound dataset from STITCH was, therefore, divided into manageable chunks, each consisting of ~2.6 million compounds, with the help of Linux ‘split’ command. The data chunks, thus obtained were given as input for anti-cancer activity prediction by SVM classifier (see section Prediction Process for the steps involved in prediction by SVM classifier). After all the data chunks were processed by SVM classifier, prediction results obtained from the individual chunks were consolidated as a single output file. Compound dataset collected from ChEMBL consisted of 1.4 million compounds, which were given as input for anti-cancer activity prediction by SVM classifier (see section Prediction Process for the steps involved in prediction by SVM classifier).

Association of Predicted Anticancer Compounds with Targets.

Target information for the compounds derived from the STITCH database [16] were compiled by parsing the file ‘9606.protein_chemical.links.v4.0.tsv.gz’, downloaded during the data collection stage of the current study (see section STITCH—Chemical-Protein Interaction Database). STITCH database stores target information as an Ensembl protein identifier, we have used gene2ensembl file downloaded from NCBI to covert Ensembl ID to the corresponding Entrez Gene ID. The chemical-protein interactions in STITCH database are arranged into four categories, viz. (i) experimental: interaction information collected from experimental databases like ChEMBL [15], (ii) database: interaction information is collected from manually curated drug-traget databases like DrugBank [3], (iii) textmining: interactions gathered from literature based on co-occurrence text-mining and Natural Language Processing (NLP), and (iv) predicted: interaction information predicted as explained by Kuhn et al. [16]. In the current study, we have considered only those chemical-protein interactions, which are of categories ‘experimental’ or ‘database’, with an objective of including interactions collected from the most reliable sources. The interaction information of 449,666 compounds from STITCH database was collected, which satisfy the source constraint (experimental/database), and were, thus used internally for extracting target information of STITCH compounds which were predicted to have anticancer activity by SVM classifier.

Target information for the ChEMBL compounds predicted to have anticancer activity by the SVM classifier was retrieved by querying ChEMBL database. Bioactivity data was mined by querying ‘activities’ table, and protein assays with IC50 value of ≤ 1 μM or 1000 nM, was used as a criteria to associate target(s) with the compound. Detailed steps involved in extraction of target information of ChEMBL compounds can be found in perl script ‘chemdb_targets_pl.txt’, available as online supplementary material (S13 Text). UniProt ids were mapped into corresponding Entrez Gene ID with the help of idmapping.tab file (available for ftp download at ftp://ftp.uniprot.org/pub/databases/uniprot/current_release/knowledgebase/idmapping/by_organism/HUMAN_9606_idmapping_selected.tab.gz).

The compounds present in ChEMBL and STITCH are not mutually exclusive, therefore, there was a possibility of existence of duplicate compounds, albeit with a different ids (ChEMBL ID & STITCH compound id) in the list of predicted anti-cancer compounds. In the current analysis, duplicate compounds were identified based on similarity in their SMILES strings. The methods like calcfp() along with bitwise OR operator ‘|’, implemented in python library PyBel [27] were used to calculate the Tanimato coefficient [28]. A Tanimato coefficient of 1 was used for the identification of duplicate compounds. The targets of duplicate compounds were merged by taking the union of the targets between duplicate compounds, for an instance, compound A (targets 1,2,3), compound B (targets 2,5,6,8) and compound C (targets 3,9) were identified as identical (Tanimato coefficient of 1), such compounds were merged as a single record with a target list as 1,2,3,5,6,8,9.

Computation of Feature Weights.

The data collected from DrugBank, and literature related by natural compounds (see section Drug-Target data) was used to train partial least squares regression model. Data from the file ‘DB_others.txt’ (see S2 Text), ‘DB_cancer.txt’ (see S1 Text) and ‘Nat_Anticancer.txt’ (see S4 Text) was combined and transformed into the training matrix to be used for PLS regression modeling. The compounds were divided into three classes based on their ICD10 codes and were represented as response variable/target value/class: ‘1’ → compounds with activity against oral cancer, ‘0’ → compounds with activity against cancer types other than oral cancer and ‘-1’ → compounds without anti-cancer activity. Compounds like resveratrol, which are active against various cancers, including oral cancer, were represented with two instances in the training matrix with two target values ‘0’ and ‘1’; this was done to ensure that no information is lost about the activity of a compound. The training data consisted of 33 compounds with activity against oral cancer (class → ‘1’). In a previously published study, we have identified a list of potential therapeutic targets for oral cancer, based on evidences gathered from the integrative study, and ability of these therapeutic targets to affect diverse cancer hallmarks involved in carcinogenesis [30]. These targets are not well represented in existing therapeutic compounds known for oral cancer, primarily because, very less number of drugs/therapeutic compounds are currently known to be active against oral cancer. In order to address this limitation, we have included an additional instance with targets reported in the previous study [30]. The data matrix used for training the partial least squares regression model can be found in the file–‘PLS_Matrix.txt’ (see S14 Text), which is available as online supplementary material. The first two columns of ‘PLS_Matrix.txt’ corresponds to compound id, and a class label (1/0/-1), respectively, and the rest of the columns indicates molecular targets represented as Entrez Gene ID; the presence or absence of a target among target profile of a compound is represented by 1 or 0, respectively, in ‘PLS_Matrix.txt’. The mapping between compound ids to their names can be found in the file ‘Compound_index.txt’ (see S15 Text), which is available as online supplementary material.

The PLS classifier is shown as below, (iv) where,

‘y’ is the training score for that compound;

where “1” represents a compound with anticancer activity against oral cancer cells, “0” represents a compound with generic anticancer activity or activity against cancers other than oral cancer, “-1” represents a compound lacking any anticancer activity.

‘w’ is the weight assigned to each target gene.

‘i’ is the index (or compound id) of the compound from P_Matrix.txt

‘a’, ‘b’, ‘c’, ‘d’, … are each gene’s compound specific feature value from P_Matrix.txt, where, “1” indicates that the compound targets gene, and “0”, is otherwise.

As implied by the equation (Eq iv), the model is trained by setting the weights for each target gene, such that, those weights when combined with individual feature values (1→Presence, 0→Absence of target gene for the particular compound), across all compounds from training matrix (‘P_Matrix.txt’) should achieve the closest correspondence of the score/class (1, 0, -1) for each training compound. The ‘leave one out’ cross-validation was used to internally train the regression model. The feature weights were extracted with the help of command “loading.weights()”. The weight obtained for a target indicates its contribution towards the activity of compound for oral cancer treatment. Feature weights were computed with the help of following R code:

1. > library(pls)
2. > data <- read.table('PLS_Matrix.txt', head = T, sep = "\t")
3. > y <- as.integer(data$class)
4. > x <- subset(data,select = -c(Record_ID,class))
5. > X <- as.matrix(x)
6. > weights <- as.numeric(loading.weights(plsr(y~X,ncomp = 1,validation = ‘LOO’)))

The weight of the target gene/protein can be found in ‘PLS_Weights.txt’ (see S16 Text) and ‘PLS_Weights_gs.txt’ (targets represented as Gene Symbol, see S17 Text), available as online supplementary material.

Computation of Oral Cancer specific score (OC_Score).

Compounds sharing indication area, invariably have a similar target profile, because at the mechanistic level, they are targeting biological processes involved in the genesis / progression of the same disease. In the current study, we have computed a target based statistic (or score) which could help in screening compounds with activity against oral cancer. The oral cancer specific feature weights (‘PLS_Weights.txt’ see S16 Text) were used to compute score, which we named as ‘OC_Score’. (V) where,

‘n’ is the number of targets for a query compound.

‘w[i]’ feature weight for i^th target.

The steps involved in computation of OC_Score, for a query compound from the training dataset is as follows:

1. Collect target profile of a query compound.
2. Extract weights for targets of a query compound from ‘PLS_Weights.txt’.
3. Compute OC_Score for the query compound by adding up the weights extracted in step 2.

*Scores thus computed for the training compounds can be found in the file ‘Score_distribution.txt’ (see S18 Text) available as online supplementary material.

Computation of OC_Score with an example is illustrated in Table 2. OC_Score is generally higher and positive when the potential of a compound to treat oral cancer is greater, and vice-versa.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Computation of OC_Score.

https://doi.org/10.1371/journal.pone.0141719.t002

Annotation of List of Potential Compounds.

Additional information like the associated GI50 value, physicochemical properties, etc. was gathered for the list of potential compounds. GI50 is the concentration of test compound required to cause a 50% reduction in the proliferation of cancer cells. PubChem bioactivity database assigns bioactivity outcome (as ‘active’) using a ≤ 50μM cutoff based on readouts such as IC50, GI50, EC50 etc. [31]. In the current study, we have mined GI50 values for the list of potential compounds for oral cancer treatment (from ChEMBL/PubChem bioassay db), and used the same cutoff of ≤ 50μM to associate active GI50 assays against them. Compounds from STITCH database are derived from PubChem, therefore, their compound ids can be converted into corresponding PubChem Compound ID (or cid) [16]. Therapeutic compounds collected from DrugBank and natural sources (see section Drug-Target Data) were associated with PubChem cid. Bioassay data from PubChem BioAssay, can be collected with the help PUG/REST structured URLs in the following format: http://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/cid/<ListOfCIDs>/assaysummary/CSV

BioAssay data for compounds from ChEMBL were collected by querying ChEMBL database. For an instance, GI50 values of active bioassays associated with resveratrol (chembl_id:CHEMBL165), were obtained with the help of following SQL query: “select a.assay_id, b.standard_type, b.standard_value, b.standard_units, a.description from assays a, activities b where a.assay_id = b.assay_id AND b.standard_type = 'GI50' AND ((b.standard_value< = 50 AND b.standard_units LIKE 'u%') OR (b.standard_value< = 50000 AND b.standard_units LIKE 'n%')) AND b.molregno = (select molregno from molecule_dictionary where chembl_id = ‘CHEMBL165');”

Various physicochemical parameters for the list of potential compounds were computed with the help of python library ‘PyBel’ [27]. The parameters, thus computed, were used for estimation of Lipinski’s ‘rule of 5’ which predicts that poor absorption or permeation is more likely when there are more than 5 hydrogen-bond donors, 10 hydrogen-bond acceptors, the molecular weight is greater than 500 and the Log P is greater than 5 [32]. Topological polar surface area (TPSA) was another physicochemical property which was calculated in the current study. TPSA is considered to be a good predictor of oral bioavailability. The molecule with TPSA of ≥ 140 Ų are likely to exhibit poor intestinal absorption [33]. ChEMBL tags compound with value ‘Y/N’ (for the field “MedChem Friendly”) based on the absence / presence of functional groups which are not desirable from the medicinal chemistry perspective (refer link https://www.ebi.ac.uk/chemblntd/glossary). We have collected SMARTS string of the undesirable functional groups used by ChEMBL, and used them to screen the list of potential compounds. This screening was conducted with the help of python script ‘Check_MedChemFriendly_py.txt’ (see S19 Text) available online as supplementary material. The smiles strings of the list of potential compounds obtained in the current study were given as a input to CDRUG [17], with an objective to flag out possible false positives.

The common names of the compound derived from PubChem/ChEMBL were retrieved from respective databases. PubChem allows retrieval of compound names for an input list of compound ids (PubChem CID), with the help of PUG/REST structured URLs in following format: https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/cid/<ListOfCIDs>/description/XML. The XML file consists of compound name enclosed between <Title>cname</Title> tags, the perl library ‘XML::Simple’ was used to parse xml file and collect common names of the list of compounds with PubChem cid. The common names of the compounds from ChEMBL database can be obtained by querying ChEMBL database table ‘molecule_dictionary’ and is available in the field ‘pref_name’. For an instance, common name for compound id ‘CHEMBL165’ (compound id for resveratrol) can be retrieved with the help of SQL query ‘select pref_name from molecule_dictionary where chembl_id = “CHEMBL165”’.

The NCBI-PubMed was queried to get overview of supporting publication evidences for potential compounds, identified in the current study. The PubMed search was conducted with the help of specific queries, structured to retrieve relevant articles which mention the therapeutic role of a compound with respect to oral cancer. For an instance, following PubMed search query can be used to retrieve research articles relevant to the therapeutic role of resveratrol with respect to oral cancer: “(Resveratrol [TIAB] AND mouth neoplasms[MH] AND (Therapeutic [TIAB] OR Therapy [TIAB] OR Treatment[TIAB])) NOT Review[PT]”.