Yang et al. [27] utilized virtual docking to suggest possible interactions between a established of 845 proteins and a set of 162 drugs that induced at minimum 1 of four ADRs. BML-284Lounkine et al. [28] predicted the activity of 656 promoted drugs on seventy three targets from the Novartis in vitro basic safety panel utilizing the similarity ensemble method (SEA). This was not a correct docking research for each se, in that SEA calculates the chemical similarity of each and every drug with each and every of the native ligands of the 73 targets. Two preceding endeavours, in specific, are equivalent to our recent research. 1st, Wallach and co-employees [29] used multiple levels of logistic regression to docking scores involving 730 drugs, 830 human protein targets and then utilized numerous levels of logistic regression to these information and knowledge on 506 ADRs, creating 32 ADR-pathway associations supported by the scientific literature (i.e. PubMed). Next, Xie et al. [thirty] produced a methodology that discovered 3D protein structures in the PDB that experienced equivalent ligand binding web sites to people of the main targets of Cholesteryl Ester Transferase Protein (CETP) inhibitors. Subsequently, they applied molecular docking to support rank get the atomic-stage interactions of the medicines with the putative off-targets. This investigation led to 204 structures with binding internet sites equivalent to CETP. This set of off-targets was then integrated into a network that incorporated numerous metabolic signal transduction and gene regulation pathway constituents, medicines, and scientific outcomes. From this network, they have been in a position to elucidate numerous ADRs recognized to be linked with the CETP inhibitors: the unfavorable influence of Torcetrapib on blood force noticed in Stage III scientific trials and the improved loss of life charges from an infection and most cancers. These scientific studies utilized the “first principles” method to circumvent the bias troubles in experimental knowledge outlined above, but none of these preceding efforts explain computational frameworks scalable to the data measurements essential for a large-precision, higher-throughput ADR screening panel for nascent compounds. A lot more not too long ago, Reardon [31] reported on a computational hard work that uses publicly available profiles of 600,000 chemical compounds and assesses their ability to bind to ,7000 chemical pockets on 570 human proteins. The recognized expression profiles of the proteins and receptors on human organs is then employed to forecast where in the body a given drug will most probably have results. Even though these efforts surely work at the essential scale, they do not report a technique to statistically associate the docking scores with ADR phenotypes, which is specifically the purpose of our operate below. Our doing work speculation is that it is beneficial to forecast ADRs as early in the lead identification period as feasible. Framework-dependent, substantial throughput, digital screening is already broadly applied in the early levels of drug discovery due to the fact of its minimal cost and substantial performance in pinpointing putative drug-candidate/drug-concentrate on interactions. Molecular docking-dependent screening scientific studies include fitting a massive library of N little molecules into the energetic internet sites of M goal protein buildings, to compute estimates of binding affinities. M and N can be really large. At the moment, the PDB has M.ninety K protein buildings, increasing at a fee of more than 7500 for every yr [26]. The combinatorics of the feasible chemical structural area occupied by tiny molecules is enormous, i.e. N < 1060 possible drug compounds [32]. These numbers, combined with the complexities of conformational sampling to find the best fit of the small molecule (i.e. ``pose'') in the target, and the computational cost of the scoring function itself, make high-throughput ADR screening ideal for high-performance computing. Zhang et al. [33] implemented a mixed parallel scheme using Message Passing Interface (MPI) and multithreading in a parallel molecular docking program, called VinaLC, by modifying the existing AutoDock Vina molecular docking program. One million flexible docking calculations took about 1.4 hours to finish on ,15 K CPUs. The docking accuracy of VinaLC has been validated against the DUD (Directory of Useful Decoys) database [34] by the re-docking of X-ray ligands and an enrichment study. The statistical results presented in their study [33] show VinaLC is one of the better performing docking codes on the DUD set of decoys/ligands, having a mean receiver operator characteristic area-under the curve (ROC AUC) of 0.64 (95th CI: 0.60-0.68). VinaLC identified 64.4% of the top scoring poses with an RMSD under the 2.0 A cutoff, while that for the best poses is 70.0%. For the best poses, all the targets have RMSD values within 10 A and . Overall, about half of the targets have RMSD values less than 1 A the VinaLC docking program performed well for re-docking the X-ray ligands back into the active site of the X-ray structures with the default setting for the grid sizes and exhaustiveness = 8. To improve the enrichment of the docking results, Zhang et al. [35] have also developed a massively parallel virtual screening pipeline using Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) rescoring and have shown improvements in the docking benchmark AUC to 0.71, on average. Overall, the results demonstrate that MM/GBSA rescoring has higher AUC values and consistently better early recovery of actives than VinaLC docking alone. A significant fraction of these molecules (e.g. drugs approved by regulatory agencies like the U.S. Food and Drug Administration) are annotated with known associated ADRs in public databases, such as SIDER. As in the prior work we cited, machine learning methods can identify statistical associations between these ADR outcomes and patterns in drug-protein binding as revealed by our VinaLC docking scores. The results can be used to build predictive models so the probabilities of certain ADRs can be predicted for a nascent or theoretical small molecule drug candidate that may not have undergone in vitro or clinical trial testing. This study potentially provides a technological and methodological path forward to large-scale, high-throughput, in silico, comprehensive ADR screening. Our results indicate that molecular docking performed with sufficiently detailed docking models on high performance computers (HPC) may provide reliable, costeffective, comprehensive high-throughput screening of a drug candidate for binding across many known on- and off-targets to predict clinically important ADRs.We extracted 4,020 Swiss-Prot protein knowledgebase UniProt ID numbers for proteins that were identified as drug targets in DrugBank as of October 12, 2012. Mappings to 587 experimental structures in the Protein Data Bank were obtained using the pdbtosp.txt file (Nov 2, 2013) from which links PDB ID numbers to UniProt IDs. A set of quality control rules were then applied (Figure S1 in File S1) which further reduced the list of proteins down to a final set of 409 experimental PDB structures. If multiple structures were given for the same protein, structures were selected by the following criteria in priority order: (1) human species (2) X-ray crystal structure (3) higher structural resolution (smaller A). This set of PDBs included 33 structures belonging to 16 UniProt IDs that are a subset of a larger consensus in vitro toxicity panel. This panel consists of 44 targets that were presented as a minimum in vitro toxicology panel from a collaboration of four major pharmaceutical companies [6]. The structures of 906 FDA-approved small molecule compounds in SDF format were obtained from the ``Orange Book'' of approved products [36]. Drugs that have more than 20 rotatable bonds were not included because most of them are natural products. The 3D structures of target proteins and the small molecule compounds were then prepared for molecular docking calculations as described below. A set of 85 side effects were selected from the SIDER database (http://sideeffects.embl.de/ extracted on November 26, 2012) because they were as sociated with high morbidity, high case fatality ratio, and/or the need for extended hospitalization. Individual side effects were grouped into higher-level health outcome groupings to reduce noise and provide signals at the organ or system level. Individual side effects were identified as lowest level terms in the medical dictionary for regulatory activities (MedDRA) [37]. Following the work of Huang and co-workers [25], the side effects of interest were grouped into ten MedDRAdefined system organ classes: (1) Neoplasms, benign, malignant, and unspecified (``neoplasms''), (2) Blood and lymphatic system disorders (``bloodAndLymph''), (3) Immune system disorders (``immuneSystem''), (4) Endocrine disorders (``endocrineDisorders''), (5) Psychiatric disorders (``psychDisorders''), (6) Cardiac disorders (``cardiacDisorders''), (7) Vascular disorders (``vascularDisorders''), (8) Gastrointestinal disorders (``gastroDisorders''), (9) Hepatobiliary disorders (``hepatoDisorders''), and (10) Renal and urinary disorders (``renalDisorders''). A subset of 560 of the 906 compounds in our docking score set were found to have associations to at least one of the 85 side effects we consider. The complete list of side effects by organ class is presented in Table S1 in File S1. We produce a 560610 drug-ADR matrix where a `1'(`0') indicates the presence (absence) of one or more side effects in the group. At the end of the dataset creation stage, we have a total of 906 compounds (560 with ADR associations), 409 proteins, and 10 outcome groups, comprising 85 severe side effects. In order to compare the ADR prediction capability of ``offtarget'' effects, obtained by the molecular docking calculations, with that of experimentally derived ``on-target'' drug-protein associations, a 560 drug6555 target protein association matrix was extracted from DrugBank. More precisely, in order for a specific protein to be in the list of 555 proteins, it must be identified as a `target' in the DrugBank database of one or more of the 560 drugs in our dataset. The matrix is boolean-valued where a `1'(`0') indicates the presence (absence) of the association in DrugBank.21920898The 409 target protein structures retrieved from the PDB were processed for molecular docking calculations. The raw PDB files were processed by our in-house Protein Function Prediction (PFP) pipeline [38]. The structures of the protein targets were cleaned and protonated. “Cleaning” was defined by the following: alternate location “a” records for atoms were kept, and any ligands (i.e. atoms designated as `HETATM’ after the TER record in the PDB file that are not part of common ions) were deleted. Molecular modeling software (Schrodinger Inc.) was used to protonate the protein structure. In those cases where a known catalytic site was identified, the centroid coordinates for the active sites/binding sites of the protein targets were determined by CatSId (Catalytic Site Identification) [39], otherwise, these sites were determined by Sitemap [40]. A similarity to a known catalytic site was identified in 83 cases. Cofactors, metals, and crystallographic waters were removed from the protein structure when performing the docking calculation [34], [41], [42], [43].Missing residues in the active site were reconstructed. For structures with residues having multiple positions, the first one was used. These pre-treated protein target structures were further processed by the in-house program, preReceptor [35]. The program preReceptor provides interfaces to integrate several external programs for target protein preparation. The preReceptor program determines the dimensions of docking grids by utilizing the DMS [44] and SPHGEN programs [45]. The DMS program calculates the molecular surface of the target protein, and the SPHGEN program fills the active site of the target protein with non-overlapping spheres of uniform dimension. The dimensions of the docking grid for each protein were determined by first finding the distribution of spheres along the X, Y, and Z axes. The grid boundaries were set to the location where the density of spheres falls off drastically. In order to reduce the computer time, the docking grid determination was limited to portions of the target protein within 30 A of the centroid of the active site (60 A maximum diameter) because binding pockets typically are less than 40 A in diameter. The dimensions of the docking grids and centroid of active site were stored for the docking calculation in the next step. The AMBER force field f99SB [46] was employed in the calculations. Non-standard amino acids distant from the binding site were converted to alanine. Otherwise, non-standard amino acids were stored in the library, if present in the active site. [47]. The energy minimization of the protein target was carried out using MM/GBSA [35] implemented in the Sander program of the AMBER package [46]. The structures were minimized with whole-protein heavy atom (i.e. all atoms that are not hydrogens) constraints so the geometry of the active site remains unchanged. The PDB files of energy-minimized protein structures were converted to PDBQT files, which are used in the docking procedure. During the conversion, the non-polar hydrogen atoms are removed from the protein target structures. Parameters for non-standard amino acids were calculated by the Antechamber program from the AMBERTOOLS suite. The set of 906 approved drugs were processed by the in-house program, preLigand [35].
