| Techniques |
This page help users to understand about the datasets and methodology used to developed the various QSAR models implemented in this web-servers. We have developed three kind of models e.g., (1) QSAR based, (2) Similarity based and (3) Hybrid (QSAR+Similarity). Description of these models are mentioned below.
Datasets
PubChem bioassay ID: 1750
Total Number of molecules: 96
Principle: Principle of this assay was to monitor the disruption of fluorescence polarization of XIAP-BIR3 domain with rhodamine tagged 7-mer N-terminal SMAC peptide. In this assay, total 138 chemical molecules were screened out of which 96 were found to be active and remaining 42 were inactive.
This assay provided the IC50 values and we have been converted the IC50 values to pIC50 {-log10[IC50(μM) x 10-6]}. For our QSAR studies, we took 96 active molecules and generated their 3D coordinates (originally they were 2D format) using “gen3D” module of Open Babel software40.
QSAR model development
We adopted three strategies to develop our QSAR models.
(1). In first strategy, we have calculated chemical desriptors (2D, 3D and 10 types of fingerprints) using PaDEL software. Then, we selected the best chemical chemical desriptors using "Best First" algorithm of Weka machine learning tool. Selected decriptors were used for the development of QSAR models by various machine learning algorithm like, SVM, ANN, KNN, SLR etc.
(2). In second strategy, we did not use best selested descriptor as such for machine learning, as we did in first strategy, instead we transform them to principle components (PCs) and then used them for machine learning and QSAR model development.
(3). Tn third strategy, we seven docking used energy parameters calculkated by AutoDock software and further used for QSAR model developments.
TOP
Similarity Approach
Molecules with same structure generally posses the same activity, based on this principle we have been calculated the similarity matrix of Tanimoto coefficients (TCs) based on atom pair similarity between two compounds. We used “cmp.similarity” function of R-package ChemmineR50. After, calculation of similarity of each compound with every other compound in the training dataset, we applied a varying cut-offs of TC to identify the similar compounds of a query compound. Then query compound assigned with pIC50, which is average of pIC50 of its similar compounds at that particular cut-off of TC. This cycle was repeated for whole training set compounds and their new pIC50 was assigned in this way.
Hybrid Approach
Three type of hybrid models were developed:
(a). Hybrid model of PaDEL descriptors and docking parameters: In this model, we have combined 15 selected PaDEL descriptors and 7 energy parameters from docking.
(b). Hybrid model of PaDEL descriptors and principle components (PCs): Here we combined 15 selected PaDEL descriptors and 6 principle components.
(c). we hybridized the similarity-based method with PaDEL descriptor based QSAR models to improve its performance in terms of coverage as well as correlation. In this model, first query molecule is searched for its similar molecule at particular TC cut-off and assigned with average pIC50 value of its respective similar molecules. If for any query molecule, similar molecules were not obtained at particular TC cut-off, then its pIC50 value was predicted by the QSAR model based on PaDEL descriptors.
TOP
Screening of Zinc library
Screening of DrugBank library
We have downloaded FDA approved drugs (1375) from DrugBank website. Then, we have calculated the chemical desriptors using PaDEL of these drugs. Finally, pIC50 values were predicted using QSAR model based on PaDEL descriptors(discussed above) as in case of ZINC library molecules.
TOP