Virtual Screening Performance Research Articles

Incorporating structural similarity into a scoring function to enhance the prediction of binding affinities

In this study, we developed a novel algorithm to improve the screening performance of an arbitrary docking scoring function by recalibrating the docking score of a query compound based on its structure similarity with a set of training compounds, while the extra computational cost is neglectable. Two popular docking methods, Glide and AutoDock Vina were adopted as the original scoring functions to be processed with our new algorithm and similar improvement performance was achieved. Predicted binding affinities were compared against experimental data from ChEMBL and DUD-E databases. 11 representative drug receptors from diverse drug target categories were applied to evaluate the hybrid scoring function. The effects of four different fingerprints (FP2, FP3, FP4, and MACCS) and the four different compound similarity effect (CSE) functions were explored. Encouragingly, the screening performance was significantly improved for all 11 drug targets especially when CSE = S4 (S is the Tanimoto structural similarity) and FP2 fingerprint were applied. The average predictive index (PI) values increased from 0.34 to 0.66 and 0.39 to 0.71 for the Glide and AutoDock vina scoring functions, respectively. To evaluate the performance of the calibration algorithm in drug lead identification, we also imposed an upper limit on the structural similarity to mimic the real scenario of screening diverse libraries for which query ligands are general-purpose screening compounds and they are not necessarily structurally similar to reference ligands. Encouragingly, we found our hybrid scoring function still outperformed the original docking scoring function. The hybrid scoring function was further evaluated using external datasets for two systems and we found the PI values increased from 0.24 to 0.46 and 0.14 to 0.42 for A2AR and CFX systems, respectively. In a conclusion, our calibration algorithm can significantly improve the virtual screening performance in both drug lead optimization and identification phases with neglectable computational cost.

A Multi-Objective Approach for Drug Repurposing in Preeclampsia.

Preeclampsia is a hypertensive disorder that occurs during pregnancy. It is a complex disease with unknown pathogenesis and the leading cause of fetal and maternal mortality during pregnancy. Using all drugs currently under clinical trial for preeclampsia, we extracted all their possible targets from the DrugBank and ChEMBL databases and labeled them as “targets”. The proteins labeled as “off-targets” were extracted in the same way but while taking all antihypertensive drugs which are inhibitors of ACE and/or angiotensin receptor antagonist as query molecules. Classification models were obtained for each of the 55 total proteins (45 targets and 10 off-targets) using the TPOT pipeline optimization tool. The average accuracy of the models in predicting the external dataset for targets and off-targets was 0.830 and 0.850, respectively. The combinations of models maximizing their virtual screening performance were explored by combining the desirability function and genetic algorithms. The virtual screening performance metrics for the best model were: the Boltzmann-Enhanced Discrimination of ROC (BEDROC)α=160.9 = 0.258, the Enrichment Factor (EF)1% = 31.55 and the Area Under the Accumulation Curve (AUAC) = 0.831. The most relevant targets for preeclampsia were: AR, VDR, SLC6A2, NOS3 and CHRM4, while ABCG2, ERBB2, CES1 and REN led to the most relevant off-targets. A virtual screening of the DrugBank database identified estradiol, estriol, vitamins E and D, lynestrenol, mifrepristone, simvastatin, ambroxol, and some antibiotics and antiparasitics as drugs with potential application in the treatment of preeclampsia.

Generating property-matched decoy molecules using deep learning.

MotivationAn essential step in the development of virtual screening methods is the use of established sets of actives and decoys for benchmarking and training. However, the decoy molecules in commonly used sets are biased meaning that methods often exploit these biases to separate actives and decoys, and do not necessarily learn to perform molecular recognition. This fundamental issue prevents generalization and hinders virtual screening method development.ResultsWe have developed a deep learning method (DeepCoy) that generates decoys to a user’s preferred specification in order to remove such biases or construct sets with a defined bias. We validated DeepCoy using two established benchmarks, DUD-E and DEKOIS 2.0. For all 102 DUD-E targets and 80 of the 81 DEKOIS 2.0 targets, our generated decoy molecules more closely matched the active molecules’ physicochemical properties while introducing no discernible additional risk of false negatives. The DeepCoy decoys improved the Deviation from Optimal Embedding (DOE) score by an average of 81% and 66%, respectively, decreasing from 0.166 to 0.032 for DUD-E and from 0.109 to 0.038 for DEKOIS 2.0. Further, the generated decoys are harder to distinguish than the original decoy molecules via docking with Autodock Vina, with virtual screening performance falling from an AUC ROC of 0.70 to 0.63.Availability and implementationThe code is available at https://github.com/oxpig/DeepCoy. Generated molecules can be downloaded from http://opig.stats.ox.ac.uk/resources.Supplementary informationSupplementary data are available at Bioinformatics online.

Novel cocrystals of the potent 1,2,4-thiadiazole-based neuroprotector with carboxylic acids: virtual screening, crystal structures and solubility performance

Five new multicomponent solid forms of the biologically active 1,2,4-thiadiazole derivative (TDZH) with dicarboxylic and hydroxybenzoic acids have been discovered by combined virtual/experimental cocrystal screening.

Ligand-Binding-Site Refinement to Generate Reliable Holo Protein Structure Conformations from Apo Structures.

The first important step in a structure-based virtual screening is the judicious selection of a receptor protein. In cases where the holo protein receptor structure is unavailable, significant reduction in virtual screening performance has been reported. In this work, we present a robust method to generate reliable holo protein structure conformations from apo structures using molecular dynamics (MD) simulation with restraints derived from holo structure binding-site templates. We perform benchmark tests on two different datasets: 40 structures from a directory of useful decoy-enhanced (DUD-E) and 84 structures from the Gunasekaran dataset. Our results show successful refinement of apo binding-site structures toward holo conformations in 82% of the test cases. In addition, virtual screening performance of 40 DUD-E structures is significantly improved using our MD-refined structures as receptors with an average enrichment factor (EF), an EF1% value of 6.2 compared to apo structures with 3.5. Docking of native ligands to the refined structures shows an average ligand root mean square deviation (RMSD) of 1.97 Å (DUD-E dataset and Gunasekaran dataset) relative to ligands in the holo crystal structures, which is comparable to the self-docking (i.e., docking of the native ligand back to its crystal structure receptor) average, 1.34 Å (DUD-E dataset) and 1.36 Å (Gunasekaran dataset). On the other hand, docking to the apo structures yields an average ligand RMSD of 3.65 Å (DUD-E) and 2.90 Å (Gunasekaran). These results indicate that our method is robust and can be useful to improve virtual screening performance of apo structures.

Improvement in the screening performance of potential aryl hydrocarbon receptor ligands by using supervised machine learning

The aryl hydrocarbon receptor (AhR), which is a ligand-dependent transcription factor, plays a crucial role in the regulation of xenobiotic metabolism. There are a large number of artificial or natural molecules in the environment that can activate AhR. In this study, we developed a virtual screening procedure to identify potential ligands of AhR. One structure-based method and two ligand-based methods were used for the virtual screening procedure. The results showed that the precision rate (0.96) and recall rate (0.64) of our procedure were significantly higher than those of a procedure used in a previous study, which suggests that supervised machine learning techniques can greatly improve the performance of virtual screening. Moreover, a pesticide dataset including 777 frequently used pesticides was screened. Seventy-seven pesticides were identified as potential AhR ligands by all three screening methods, among which 12 have never been previously reported as AhR agonists. Two non-agonist AhR ligands and 14 of the 77 pesticides were randomly selected for testing by in vitro and in vivo assays. All 14 pesticides showed different degrees of AhR agonistic activity, and none of the two non-agonist AhR ligand pesticides showed AhR agonistic activity, which suggests that our procedure had good robustness. Four of the pesticides were reported as AhR agonists for the first time, suggesting that these pesticides may need further toxicity assessment. In general, our procedure is a rapid, powerful and computationally inexpensive tool for predicting chemicals with AhR agonistic activity, which could be useful for environmental risk prediction and management.

Virtual Coformer Screening by Crystal Structure Predictions: Crucial Role of Crystallinity in Pharmaceutical Cocrystallization.

One of the most popular strategies of the optimization of drug properties in the pharmaceutical industry appears to be a solid form changing into a cocrystalline form. A number of virtual screening approaches have been previously developed to allow a selection of the most promising cocrystal formers (coformers) for an experimental follow-up. A significant drawback of those methods is related to the lack of accounting for the crystallinity contribution to cocrystal formation. To address this issue, we propose in this study two virtual coformer screening approaches based on a modern cloud-computing crystal structure prediction (CSP) technology at a dispersion-corrected density functional theory (DFT-D) level. The CSP-based methods were for the first time validated on challenging cases of indomethacin and paracetamol cocrystallization, for which the previously developed approaches provided poor predictions. The calculations demonstrated a dramatic improvement of the virtual coformer screening performance relative to the other methods. It is demonstrated that the crystallinity contribution to the formation of paracetamol and indomethacin cocrystals is a dominant one and, therefore, should not be ignored in the virtual screening calculations. Our results encourage a broad utilization of the proposed CSP-based technology in the pharmaceutical industry as the only virtual coformer screening method that directly accounts for the crystallinity contribution.

Discovery of Novel Hsp90 C-Terminal Inhibitors Using 3D-Pharmacophores Derived from Molecular Dynamics Simulations.

Hsp90 C-terminal domain (CTD) inhibitors are promising novel agents for cancer treatment, as they do not induce the heat shock response associated with Hsp90 N-terminal inhibitors. One challenge associated with CTD inhibitors is the lack of a co-crystallized complex, requiring the use of predicted allosteric apo pocket, limiting structure-based (SB) design approaches. To address this, a unique approach that enables the derivation and analysis of interactions between ligands and proteins from molecular dynamics (MD) trajectories was used to derive pharmacophore models for virtual screening (VS) and identify suitable binding sites for SB design. Furthermore, ligand-based (LB) pharmacophores were developed using a set of CTD inhibitors to compare VS performance with the MD derived models. Virtual hits identified by VS with both SB and LB models were tested for antiproliferative activity. Compounds 9 and 11 displayed antiproliferative activities in MCF-7 and Hep G2 cancer cell lines. Compound 11 inhibited Hsp90-dependent refolding of denatured luciferase and induced the degradation of Hsp90 clients without the concomitant induction of Hsp70 levels. Furthermore, compound 11 offers a unique scaffold that is promising for the further synthetic optimization and development of molecules needed for the evaluation of the Hsp90 CTD as a target for the development of anticancer drugs.

Dynamic-shared Pharmacophore Approach as Tool to Design New Allosteric PRC2 Inhibitors, Targeting EED Binding Pocket.

The Polycomb Repressive complex 2 (PRC2) maintains a repressive chromatin state and silences many genes, acting as methylase on histone tails. This enzyme was found overexpressed in many types of cancer. In this work, we have set up a Computer-Aided Drug Design approach based on the allosteric modulation of PRC2. In order to minimize the possible bias derived from using a single set of coordinates within the protein-ligand complex, a dynamic workflow was developed. In details, molecular dynamic was used as tool to identify the most significant ligand-protein interactions from several crystallized protein structures. The identified features were used for the creation of dynamic pharmacophore models and docking grid constraints for the design of new PRC2 allosteric modulators. Our protocol was retrospectively validated using a dataset of active and inactive compounds, and the results were compared to the classic approaches, through ROC curves and enrichment factor. Our approach suggested some important interaction features to be adopted for virtual screening performance improvement.

VsFilt: A Tool to Improve Virtual Screening by Structural Filtration of Docking Poses.

The ability of ligands to form crucial interactions with a protein target, characteristic for the substrate and/or inhibitors, could be considered a structural criterion for identifying potent binders among docked compounds. Structural filtration of predicted poses improves the performance of virtual screening and helps in recovering specifically bound ligands. Here, we present vsFilt-a highly automated and easy-to-use Web server for postdocking structural filtration. The new tool can detect various types of interactions that are known to be involved in the molecular recognition, including hydrogen and halogen bonds, ionic interactions, hydrophobic contacts, π-stacking, and cation-π interactions. A case study for poly(ADP-ribose) polymerase 1 ligands illustrates the utility of the software. The Web server is freely available at https://biokinet.belozersky.msu.ru/vsfilt.

Fast Rescoring Protocols to Improve the Performance of Structure-Based Virtual Screening Performed on Protein-Protein Interfaces.

Protein-protein interactions (PPIs) are attractive targets for drug design because of their essential role in numerous cellular processes and disease pathways. However, in general, PPIs display exposed binding pockets at the interface, and as such, have been largely unexploited for therapeutic interventions with low-molecular weight compounds. Here, we used docking and various rescoring strategies in an attempt to recover PPI inhibitors from a set of active and inactive molecules for 11 targets collected in ChEMBL and PubChem. Our focus is on the screening power of the various developed protocols and on using fast approaches so as to be able to apply such a strategy to the screening of ultralarge libraries in the future. First, we docked compounds into each target using the fast "pscreen" mode of the structure-based virtual screening (VS) package Surflex. Subsequently, the docking poses were postprocessed to derive a set of 3D topological descriptors: (i) shape similarity and (ii) interaction fingerprint similarity with a co-crystallized inhibitor, (iii) solvent-accessible surface area, and (iv) extent of deviation from the geometric center of a reference inhibitor. The derivatized descriptors, together with descriptor-scaled scoring functions, were utilized to investigate possible impacts on VS performance metrics. Moreover, four standalone scoring functions, RF-Score-VS (machine-learning), DLIGAND2 (knowledge-based), Vinardo (empirical), and X-SCORE (empirical), were employed to rescore the PPI compounds. Collectively, the results indicate that the topological scoring algorithms could be valuable both at a global level, with up to 79% increase in areas under the receiver operating characteristic curve for some targets, and in early stages, with up to a 4-fold increase in enrichment factors at 1% of the screened collections. Outstandingly, DLIGAND2 emerged as the best scoring function on this data set, outperforming all rescoring techniques in terms of VS metrics. The described methodology could help in the rational design of small-molecule PPI inhibitors and has direct applications in many therapeutic areas, including cancer, CNS, and infectious diseases such as COVID-19.

Discovery of Novel Tankyrase Inhibitors through Molecular Docking-Based Virtual Screening and Molecular Dynamics Simulation Studies.

Tankyrase enzymes (TNKS), a core part of the canonical Wnt pathway, are a promising target in the search for potential anti-cancer agents. Although several hundreds of the TNKS inhibitors are currently known, identification of their novel chemotypes attracts considerable interest. In this study, the molecular docking and machine learning-based virtual screening techniques combined with the physico-chemical and ADMET (absorption, distribution, metabolism, excretion, toxicity) profile prediction and molecular dynamics simulations were applied to a subset of the ZINC database containing about 1.7 M commercially available compounds. Out of seven candidate compounds biologically evaluated in vitro for their inhibition of the TNKS2 enzyme using immunochemical assay, two compounds have shown a decent level of inhibitory activity with the IC50 values of less than 10 nM and 10 μM. Relatively simple scores based on molecular docking or MM-PBSA (molecular mechanics, Poisson-Boltzmann, surface area) methods proved unsuitable for predicting the effect of structural modification or for accurate ranking of the compounds based on their binding energies. On the other hand, the molecular dynamics simulations and Free Energy Perturbation (FEP) calculations allowed us to further decipher the structure-activity relationships and retrospectively analyze the docking-based virtual screening performance. This approach can be applied at the subsequent lead optimization stages.

Consensus Docking in Drug Discovery

Background: Molecular docking is probably the most popular and profitable approach in computer-aided drug design, being the staple technique for predicting the binding mode of bioactive compounds and for performing receptor-based virtual screening studies. The growing attention received by docking, as well as the need for improving its reliability in pose prediction and virtual screening performance, has led to the development of a wide plethora of new docking algorithms and scoring functions. Nevertheless, it is unlikely to identify a single procedure outperforming the other ones in terms of reliability and accuracy or demonstrating to be generally suitable for all kinds of protein targets. Methods: In this context, consensus docking approaches are taking hold in computer-aided drug design. These computational protocols consist in docking ligands using multiple docking methods and then comparing the binding poses predicted for the same ligand by the different methods. This analysis is usually carried out calculating the root-mean-square deviation among the different docking results obtained for each ligand, in order to identify the number of docking methods producing the same binding pose. Results: The consensus docking approaches demonstrated to improve the quality of docking and virtual screening results compared to the single docking methods. From a qualitative point of view, the improvement in pose prediction accuracy was obtained by prioritizing ligand binding poses produced by a high number of docking methods, whereas with regards to virtual screening studies, high hit rates were obtained by prioritizing the compounds showing a high level of pose consensus. Conclusion: In this review, we provide an overview of the results obtained from the performance assessment of various consensus docking protocols and we illustrate successful case studies where consensus docking has been applied in virtual screening studies.

Improving structure-based virtual screening performance via learning from scoring function components.

Scoring functions (SFs) based on complex machine learning (ML) algorithms have gradually emerged as a promising alternative to overcome the weaknesses of classical SFs. However, extensive efforts have been devoted to the development of SFs based on new protein-ligand interaction representations and advanced alternative ML algorithms instead of the energy components obtained by the decomposition of existing SFs. Here, we propose a new method named energy auxiliary terms learning (EATL), in which the scoring components are extracted and used as the input for the development of three levels of ML SFs including EATL SFs, docking-EATL SFs and comprehensive SFs with ascending VS performance. The EATL approach not only outperforms classical SFs for the absolute performance (ROC) and initial enrichment (BEDROC) but also yields comparable performance compared with other advanced ML-based methods on the diverse subset of Directory of Useful Decoys: Enhanced (DUD-E). The test on the relatively unbiased actives as decoys (AD) datasetalso proved the effectiveness of EATL. Furthermore, the idea of learning from SF components to yield improved screening power can also be extended to other docking programs and SFs available.

Performance of virtual screening against GPCR homology models: Impact of template selection and treatment of binding site plasticity.

Rational drug design for G protein-coupled receptors (GPCRs) is limited by the small number of available atomic resolution structures. We assessed the use of homology modeling to predict the structures of two therapeutically relevant GPCRs and strategies to improve the performance of virtual screening against modeled binding sites. Homology models of the D2 dopamine (D2R) and serotonin 5-HT2A receptors (5-HT2AR) were generated based on crystal structures of 16 different GPCRs. Comparison of the homology models to D2R and 5-HT2AR crystal structures showed that accurate predictions could be obtained, but not necessarily using the most closely related template. Assessment of virtual screening performance was based on molecular docking of ligands and decoys. The results demonstrated that several templates and multiple models based on each of these must be evaluated to identify the optimal binding site structure. Models based on aminergic GPCRs showed substantial ligand enrichment and there was a trend toward improved virtual screening performance with increasing binding site accuracy. The best models even yielded ligand enrichment comparable to or better than that of the D2R and 5-HT2AR crystal structures. Methods to consider binding site plasticity were explored to further improve predictions. Molecular docking to ensembles of structures did not outperform the best individual binding site models, but could increase the diversity of hits from virtual screens and be advantageous for GPCR targets with few known ligands. Molecular dynamics refinement resulted in moderate improvements of structural accuracy and the virtual screening performance of snapshots was either comparable to or worse than that of the raw homology models. These results provide guidelines for successful application of structure-based ligand discovery using GPCR homology models.

Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets.

Predicting protein-ligand interactions using artificial intelligence (AI) models has attracted great interest in recent years. However, data-driven AI models unequivocally suffer from a lack of sufficiently large and unbiased datasets. Here, we systematically investigated the data biases on the PDBbind and DUD-E datasets. We examined the model performance of atomic convolutional neural network (ACNN) on the PDBbind core set and achieved a Pearson R2 of 0.73 between experimental and predicted binding affinities. Strikingly, the ACNN models did not require learning the essential protein-ligand interactions in complex structures and achieved similar performance even on datasets containing only ligand structures or only protein structures, while data splitting based on similarity clustering (protein sequence or ligand scaffold) significantly reduced the model performance. We also identified the property and topology biases in the DUD-E dataset which led to the artificially increased enrichment performance of virtual screening. The property bias in DUD-E was reduced by enforcing the more stringent ligand property matching rules, while the topology bias still exists due to the use of molecular fingerprint similarity as a decoy selection criterion. Therefore, we believe that sufficiently large and unbiased datasets are desirable for training robust AI models to accurately predict protein-ligand interactions.

Applying Machine Learning to Ultrafast Shape Recognition in Ligand-Based Virtual Screening.

Ultrafast Shape Recognition (USR), along with its derivatives, are Ligand-Based Virtual Screening (LBVS) methods that condense 3-dimensional information about molecular shape, as well as other properties, into a small set of numeric descriptors. These can be used to efficiently compute a measure of similarity between pairs of molecules using a simple inverse Manhattan Distance metric. In this study we explore the use of suitable Machine Learning techniques that can be trained using USR descriptors, so as to improve the similarity detection of potential new leads. We use molecules from the Directory for Useful Decoys-Enhanced to construct machine learning models based on three different algorithms: Gaussian Mixture Models (GMMs), Isolation Forests and Artificial Neural Networks (ANNs). We train models based on full molecule conformer models, as well as the Lowest Energy Conformations (LECs) only. We also investigate the performance of our models when trained on smaller datasets so as to model virtual screening scenarios when only a small number of actives are known a priori. Our results indicate significant performance gains over a state of the art USR-derived method, ElectroShape 5D, with GMMs obtaining a mean performance up to 430% better than that of ElectroShape 5D in terms of Enrichment Factor with a maximum improvement of up to 940%. Additionally, we demonstrate that our models are capable of maintaining their performance, in terms of enrichment factor, within 10% of the mean as the size of the training dataset is successively reduced. Furthermore, we also demonstrate that running times for retrospective screening using the machine learning models we selected are faster than standard USR, on average by a factor of 10, including the time required for training. Our results show that machine learning techniques can significantly improve the virtual screening performance and efficiency of the USR family of methods.

ALADDIN: Docking Approach Augmented by Machine Learning for Protein Structure Selection Yields Superior Virtual Screening Performance.

Protein flexibility and solvation pose major challenges to docking algorithms and scoring functions. One established strategy for addressing these challenges is to use multiple protein conformations for docking (all‐against‐all ensemble docking). Recent studies have shown that the performance of ensemble docking can be improved by selecting the most relevant protein structures for docking. In search for a robust approach to protein structure selection, we have come up with an integrated mAchine Learning AnD DockINg approach (ALADDIN). ALADDIN employs a battery of random forest classifiers to select, individually for each compound of interest, from an ensemble of protein structures, the single most suitable protein structure for docking. ALADDIN outperformed the best single‐structure docking runs, ensemble docking and a similarity‐based docking approach on three out of four investigated targets, with up to 0.15, 0.11 and 0.16 higher area under the receiver operating characteristic curve (AUC) values, respectively. Only in the case of cytochrome P450 3A4, ALADDIN, like any of the other tested approaches, failed to obtain decent performance. ALADDIN can be particularly useful for structure‐based virtual screening of malleable proteins, including kinases, some viral enzymes and anti‐targets.

Electrostatic-field and surface-shape similarity for virtual screening and pose prediction

We introduce a new method for rapid computation of 3D molecular similarity that combines electrostatic field comparison with comparison of molecular surface-shape and directional hydrogen-bonding preferences (called “eSim”). Rather than employing heuristic “colors” or user-defined molecular feature types to represent conformation-dependent molecular electrostatics, eSim calculates the similarity of the electrostatic fields of two molecules (in addition to shape and hydrogen-bonding). We present detailed virtual screening performance data on the standard 102 target DUD-E set. In its moderately fast screening mode, eSim running on a single computing core is capable of processing over 60 molecules per second. In this mode, eSim performed significantly better than all alternate methods for which full DUD-E data were available (mean ROC area of 0.74, p < 10^{-9}, by paired t-test, compared with the best performing alternate method). In addition, for 92 targets of the DUD-E set where multiple ligand-bound crystal structures were available, screening performance was assessed using alternate ligands or sets thereof (in their bound poses) as similarity targets. Using the joint alignment of five ligands for each protein target, mean ROC area exceeded 0.82 for the 92 targets. Design-focused application of ligand similarity methods depends on accurate predictions of geometric molecular relationships. We comprehensively assessed pose prediction accuracy by curating nearly 400,000 bound ligand pose pairs across the DUD-E targets. Overall, beginning from agnostic initial poses, we observed an 80% success rate for RMSD le 2.0 Å among the top 20 predicted eSim poses. These examples were split roughly 50/50 into cases with high direct atomic overlap (where a shared scaffold exists between a pair) and low direct atomic overlap (where where a ligand pair has dissimilar scaffolds but largely occupies the same space). Within the high direct atomic overlap subset, the pose prediction success rate was 93%. For the more challenging subset (where dissimilar scaffolds are to be aligned), the success rate was 70%. The eSim approach enables both large-scale screening and rational design of ligands and is rooted in physically meaningful, non-heuristic, molecular comparisons.

Improving (Q)SAR predictions by examining bias in the selection of compounds for experimental testing

ABSTRACTExisting data on structures and biological activities are limited and distributed unevenly across distinct molecular targets and chemical compounds. The question arises if these data represent an unbiased sample of the general population of chemical-biological interactions. To answer this question, we analyzed ChEMBL data for 87,583 molecules tested against 919 protein targets using supervised and unsupervised approaches. Hierarchical clustering of the Murcko frameworks generated using Chemistry Development Toolkit showed that the available data form a big diffuse cloud without apparent structure. In contrast hereto, PASS-based classifiers allowed prediction whether the compound had been tested against the particular molecular target, despite whether it was active or not. Thus, one may conclude that the selection of chemical compounds for testing against specific targets is biased, probably due to the influence of prior knowledge. We assessed the possibility to improve (Q)SAR predictions using this fact: PASS prediction of the interaction with the particular target for compounds predicted as tested against the target has significantly higher accuracy than for those predicted as untested (average ROC AUC are about 0.87 and 0.75, respectively). Thus, considering the existing bias in the data of the training set may increase the performance of virtual screening.