Virtual Screening Performance Research Articles

ClassyPose: A Machine‐Learning Classification Model for Ligand Pose Selection Applied to Virtual Screening in Drug Discovery

Determining the target‐bound conformation of a drug‐like molecule is a crucial step in drug design, as it affects the outcome of virtual screening (VS), and paves the way for hit‐to‐lead and lead optimization. While most docking programs usually manage to produce at least a near‐native pose for a bioactive molecule inside its binding pocket, their integrated classical scoring functions (SFs) generally fail to prioritize this pose. Many studies have been carried out to tackle this SF problem, offering multiple pose refinement and/or classification methods, albeit with limitations. This study presents a new support vector machine model for pose classification, called “ClassyPose”, which predicts the probability that a receptor‐bound ligand conformation could be near‐native, without any additional pose optimization step. Trained on protein‐ligand extended connectivity features extracted from over 21 600 crystal and docking poses of diverse ligands, this model outperformed other machine‐learning algorithms and three existing SFs in terms of docking power, identifying the native ligand pose as top‐ranked solution for more than 90% of entries in two test sets. It also achieved high specificity (above 0.96), and improved VS performance when used for pose selection. This efficient, user‐friendly tool and all related data are available at https://github.com/vktrannguyen/Classy_Pose.

Evaluation of AlphaFold2 Structures for Hit Identification across Multiple Scenarios.

The introduction of AlphaFold2 (AF2) has sparked significant enthusiasm and generated extensive discussion within the scientific community, particularly among drug discovery researchers. Although previous studies have addressed the performance of AF2 structures in virtual screening (VS), a more comprehensive investigation is still necessary considering the paramount importance of structural accuracy in drug design. In this study, we evaluate the performance of AF2 structures in VS across three common drug discovery scenarios: targets with holo, apo, and AF2 structures; targets with only apo and AF2 structures; and targets exclusively with AF2 structures. We utilized both the traditional physics-based Glide and the deep-learning-based scoring function RTMscore to rank the compounds in the DUD-E, DEKOIS 2.0, and DECOY data sets. The results demonstrate that, overall, the performance of VS on AF2 structures is comparable to that on apo structures but notably inferior to that on holo structures across diverse scenarios. Moreover, when a target has solely AF2 structure, selecting the holo structure of the target from different subtypes within the same protein family produces comparable results with the AF2 structure for VS on the data set of the AF2 structures, and significantly better results than the AF2 structures on its own data set. This indicates that utilizing AF2 structures for docking-based VS may not yield most satisfactory outcomes, even when solely AF2 structures are available. Moreover, we rule out the possibility that the variations in VS performance between the binding pockets of AF2 and holo structures arise from the differences in their biological assembly composition.

On the relevance of query definition in the performance of 3D ligand-based virtual screening

Ligand-based virtual screening (LBVS) methods are widely used to explore the vast chemical space in the search of novel compounds resorting to a variety of properties encoded in 1D, 2D or 3D descriptors. The success of 3D-LBVS is affected by the overlay of molecular pairs, thus making selection of the template compound, search of accessible conformational space and choice of the query conformation to be potential factors that modulate the successful retrieval of actives. This study examines the impact of adopting different choices for the query conformation of the template, paying also attention to the influence exerted by the structural similarity between templates and actives. The analysis is performed using PharmScreen, a 3D LBVS tool that relies on similarity measurements of the hydrophobic/philic pattern of molecules, and Phase Shape, which is based on the alignment of atom triplets followed by refinement of the volume overlap. The study is performed for the original DUD-E+ database and a Morgan Fingerprint filtered version (denoted DUD-E+-Diverse; available in https://github.com/Pharmacelera/Query-models-to-3DLBVS), which was prepared to minimize the 2D resemblance between template and actives. Although in most cases the query conformation exhibits a mild influence on the overall performance, a critical analysis is made to disclose factors, such as the content of structural features between template and actives and the induction of conformational strain in the template, that underlie the drastic impact of the query definition in the recovery of actives for certain targets. The findings of this research also provide valuable guidance for assisting the selection of the query definition in 3D LBVS campaigns.Graphical

Benchmarking Refined and Unrefined AlphaFold2 Structures for Hit Discovery

The recently developed AlphaFold2 (AF2) algorithm predicts proteins' 3D structures from amino acid sequences. The open AlphaFold protein structure database covers the complete human proteome. Using an industry-leading molecular docking method (Glide), we investigated the virtual screening performance of 37 common drug targets, each with an AF2 structure and known holo and apo structures from the DUD-E data set. In a subset of 27 targets where the AF2 structures are suitable for refinement, the AF2 structures show comparable early enrichment of known active compounds (avg. EF 1%: 13.0) to apo structures (avg. EF 1%: 11.4) while falling behind early enrichment of the holo structures (avg. EF 1%: 24.2). With an induced-fit protocol (IFD-MD), we can refine the AF2 structures using an aligned known binding ligand as the template to improve the performance in structure-based virtual screening (avg. EF 1%: 18.9). Glide-generated docking poses of known binding ligands can also be used as templates for IFD-MD, achieving similar improvements (avg. EF 1% 18.0). Thus, with proper preparation and refinement, AF2 structures show considerable promise for in silico hit identification.

Beware of Simple Methods for Structure-Based Virtual Screening: The Critical Importance of Broader Comparisons.

We discuss how data unbiasing and simple methods such as protein-ligand Interaction FingerPrint (IFP) can overestimate virtual screening performance. We also show that IFP is strongly outperformed by target-specific machine-learning scoring functions, which were not considered in a recent report concluding that simple methods were better than machine-learning scoring functions at virtual screening.

A Hybrid Docking and Machine Learning Approach to Enhance the Performance of Virtual Screening Carried out on Protein-Protein Interfaces.

The modulation of protein-protein interactions (PPIs) by small chemical compounds is challenging. PPIs play a critical role in most cellular processes and are involved in numerous disease pathways. As such, novel strategies that assist the design of PPI inhibitors are of major importance. We previously reported that the knowledge-based DLIGAND2 scoring tool was the best-rescoring function for improving receptor-based virtual screening (VS) performed with the Surflex docking engine applied to several PPI targets with experimentally known active and inactive compounds. Here, we extend our investigation by assessing the vs. potential of other types of scoring functions with an emphasis on docking-pose derived solvent accessible surface area (SASA) descriptors, with or without the use of machine learning (ML) classifiers. First, we explored rescoring strategies of Surflex-generated docking poses with five GOLD scoring functions (GoldScore, ChemScore, ASP, ChemPLP, ChemScore with Receptor Depth Scaling) and with consensus scoring. The top-ranked poses were post-processed to derive a set of protein and ligand SASA descriptors in the bound and unbound states, which were combined to derive descriptors of the docked protein-ligand complexes. Further, eight ML models (tree, bagged forest, random forest, Bayesian, support vector machine, logistic regression, neural network, and neural network with bagging) were trained using the derivatized SASA descriptors and validated on test sets. The results show that many SASA descriptors are better than Surflex and GOLD scoring functions in terms of overall performance and early recovery success on the used dataset. The ML models were superior to all scoring functions and rescoring approaches for most targets yielding up to a seven-fold increase in enrichment factors at 1% of the screened collections. In particular, the neural networks and random forest-based ML emerged as the best techniques for this PPI dataset, making them robust and attractive vs. tools for hit-finding efforts. The presented results suggest that exploring further docking-pose derived SASA descriptors could be valuable for structure-based virtual screening projects, and in the present case, to assist the rational design of small-molecule PPI inhibitors.

DENVIS: Scalable and High-Throughput Virtual Screening Using Graph Neural Networks with Atomic and Surface Protein Pocket Features.

Computational methods for virtual screening can dramatically accelerate early-stage drug discovery by identifying potential hits for a specified target. Docking algorithms traditionally use physics-based simulations to address this challenge by estimating the binding orientation of a query protein-ligand pair and a corresponding binding affinity score. Over the recent years, classical and modern machine learning architectures have shown potential for outperforming traditional docking algorithms. Nevertheless, most learning-based algorithms still rely on the availability of the protein-ligand complex binding pose, typically estimated via docking simulations, which leads to a severe slowdown of the overall virtual screening process. A family of algorithms processing target information at the amino acid sequence level avoid this requirement, however, at the cost of processing protein data at a higher representation level. We introduce deep neural virtual screening (DENVIS), an end-to-end pipeline for virtual screening using graph neural networks (GNNs). By performing experiments on two benchmark databases, we show that our method performs competitively to several docking-based, machine learning-based, and hybrid docking/machine learning-based algorithms. By avoiding the intermediate docking step, DENVIS exhibits several orders of magnitude faster screening times (i.e., higher throughput) than both docking-based and hybrid models. When compared to an amino acid sequence-based machine learning model with comparable screening times, DENVIS achieves dramatically better performance. Some key elements of our approach include protein pocket modeling using a combination of atomic and surface features, the use of model ensembles, and data augmentation via artificial negative sampling during model training. In summary, DENVIS achieves competitive to state-of-the-art virtual screening performance, while offering the potential to scale to billions of molecules using minimal computational resources.

Target-Specific Machine Learning Scoring Function Improved Structure-Based Virtual Screening Performance for SARS-CoV-2 Drugs Development.

Leveraging machine learning has been shown to improve the accuracy of structure-based virtual screening. Furthermore, a tremendous amount of empirical data is publicly available, which further enhances the performance of the machine learning approach. In this proof-of-concept study, the 3CLpro enzyme of SARS-CoV-2 was used. Structure-based virtual screening relies heavily on scoring functions. It is widely accepted that target-specific scoring functions may perform more effectively than universal scoring functions in real-world drug research and development processes. It would be beneficial to drug discovery to develop a method that can effectively build target-specific scoring functions. In the current study, the bindingDB database was used to retrieve experimental data. Smina was utilized to generate protein-ligand complexes for the extraction of InteractionFingerPrint (IFP) and SimpleInteractionFingerPrint SIFP fingerprints via the open drug discovery tool (oddt). The present study found that randomforestClassifier and randomforestRegressor performed well when used with the above fingerprints along the Molecular ACCess System (MACCS), Extended Connectivity Fingerprint (ECFP4), and ECFP6. It was found that the area under the precision-recall curve was 0.80, which is considered a satisfactory level of accuracy. In addition, our enrichment factor analysis indicated that our trained scoring function ranked molecules correctly compared to smina’s generic scoring function. Further molecular dynamics simulations indicated that the top-ranked molecules identified by our developed scoring function were highly stable in the active site, supporting the validity of our developed process. This research may provide a template for developing target-specific scoring functions against specific enzyme targets.

SCORCH: Improving structure-based virtual screening with machine learning classifiers, data augmentation, and uncertainty estimation

IntroductionThe discovery of a new drug is a costly and lengthy endeavour. The computational prediction of which small molecules can bind to a protein target can accelerate this process if the predictions are fast and accurate enough. Recent machine-learning scoring functions re-evaluate the output of molecular docking to achieve more accurate predictions. However, previous scoring functions were trained on crystalised protein-ligand complexes and datasets of decoys. The limited availability of crystal structures and biases in the decoy datasets can lower the performance of scoring functions. ObjectivesTo address key limitations of previous scoring functions and thus improve the predictive performance of structure-based virtual screening. MethodsA novel machine-learning scoring function was created, named SCORCH (Scoring COnsensus for RMSD-based Classification of Hits). To develop SCORCH, training data is augmented by considering multiple ligand poses and labelling poses based on their RMSD from the native pose. Decoy bias is addressed by generating property-matched decoys for each ligand and using the same methodology for preparing and docking decoys and ligands. A consensus of 3 different machine learning approaches is also used to improve performance. ResultsWe find that multi-pose augmentation in SCORCH improves its docking power and screening power on independent benchmark datasets. SCORCH outperforms an equivalent scoring function trained on single poses, with a 1 % enrichment factor (EF) of 13.78 vs. 10.86 on 18 DEKOIS 2.0 targets and a mean native pose rank of 5.9 vs 30.4 on CSAR 2014. Additionally, SCORCH outperforms widely used scoring functions in virtual screening and pose prediction on independent benchmark datasets. ConclusionBy rationally addressing key limitations of previous scoring functions, SCORCH improves the performance of virtual screening. SCORCH also provides an estimate of its uncertainty, which can help reduce the cost and time required for drug discovery.

Maximizing the Performance of Similarity-Based Virtual Screening Methods by Generating Synergy from the Integration of 2D and 3D Approaches.

Methods for the pairwise comparison of 2D and 3D molecular structures are established approaches in virtual screening. In this work, we explored three strategies for maximizing the virtual screening performance of these methods: (i) the merging of hit lists obtained from multi-compound screening using a single screening method, (ii) the merging of the hit lists obtained from 2D and 3D screening by parallel selection, and (iii) the combination of both of these strategies in an integrated approach. We found that any of these strategies led to a boost in virtual screening performance, with the clearest advantages observed for the integrated approach. On test sets for virtual screening, covering 50 pharmaceutically relevant proteins, the integrated approach, using sets of five query molecules, yielded, on average, an area under the receiver operating characteristic curve (AUC) of 0.84, an early enrichment among the top 1% of ranked compounds (EF1%) of 53.82 and a scaffold recovery rate among the top 1% of ranked compounds (SRR1%) of 0.50. In comparison, the 2D and 3D methods on their own (when using a single query molecule) yielded AUC values of 0.68 and 0.54, EF1% values of 19.96 and 17.52, and SRR1% values of 0.20 and 0.17, respectively. In conclusion, based on these results, the integration of 2D and 3D methods, via a (balanced) parallel selection strategy, is recommended, and, in particular, when combined with multi-query screening.

GBO-kNN a new framework for enhancing the performance of ligand-based virtual screening for drug discovery

Virtual Screening (VS) in drug discovery campaigns deals with enormously large databases containing massive redundant and/or irrelevant features. Thus, the pre-selection process of virtual screening requires developing a fast and precise classification methodology for filtering such a huge database. This paper proposed a framework depending on a wrapper selection approach for features selection. It consists of 1) an optimizer: Gradient-Based Optimizer (GBO), that hybridized with 2) the classifier: k-nearest neighbor (k-NN). The performance of the introduced framework, GBO-kNN, is estimated using real-world benchmark datasets; the QSAR Biodegradation which is composed of 41 features, and the Monoamine Oxidase (MAO) that consisted of 1665 features. the GBO-kNN framework is compared against seven recent swarm intelligence algorithms: Hybrid Harris Hawks Optimization Algorithm (HHO), Grey Wolf Optimization Algorithm (GWO), Butterfly Optimization Algorithm (BOA), Dragonfly Algorithm (DA), Moth-Flame Optimization Algorithm (MFO), Sine Cosine Algorithm (SCA), and Salp Swarm Algorithm (SSA) for reaching maximum classification accuracy. The experimental results exhibited that the proposed process, in comparison with other algorithms, achieved high effective accuracy of 98.8% over the high dimensional dataset (MAO) and a moderate effective performance over the low dimensional dataset (QSAR Biodegradation). The proposed framework source code is available at https://codeocean.com/capsule/9906421/tree/v1.

Insect RDL Receptor Models for Virtual Screening: Impact of the Template Conformational State in Pentameric Ligand-Gated Ion Channels.

The RDL receptor is one of the most relevant protein targets for insecticide molecules. It belongs to the pentameric ligand-gated ion channel (pLGIC) family. Given that the experimental structures of pLGICs are difficult to obtain, homology modeling has been extensively used for these proteins, particularly for the RDL receptor. However, no detailed assessments of the usefulness of homology models for virtual screening (VS) have been carried out for pLGICs. The aim of this study was to evaluate which are the determinant factors for a good VS performance using RDL homology models, specially analyzing the impact of the template conformational state. Fifteen RDL homology models were obtained based on different pLGIC templates representing the closed, open, and desensitized states. A retrospective VS process was performed on each model, and their performance in the prioritization of active ligands was assessed. In addition, the three best-performing models among each of the conformations were subjected to molecular dynamics simulations (MDS) in complex with a representative active ligand. The models showed variations in their VS performance parameters that were related to the structural properties of the binding site. VS performance tended to improve in more constricted binding cavities. The best performance was obtained with a model based on a template in the closed conformation. MDS confirmed that the closed model was the one that best represented the interactions with an active ligand. These results imply that different templates should be evaluated and the structural variations between their channel conformational states should be specially examined, providing guidelines for the application of homology modeling for VS in other proteins of the pLGIC family.

Impact of different protonation states on virtual screening performance against cruzain.

The cysteine protease cruzain is a Chagas disease target, exploited in computational studies. However, there is no consensus on the protonation states of the active site residues Cys25, His162, and Glu208 at the enzyme's active pH range. We evaluated the impact of different protonation states of these residues on docking calculations. Through a retrospective study with cruzain inhibitors and decoys, we compared the performance of virtual screening using four grids, varying protonation states of Cys25, His162, and Glu208. Based on enrichment factors and ROC plots, docking with the four grids affected compound ranking and the overall charge of top-ranking compounds. Different grids can be complementary and synergistic, increasing the odds of finding different ligands with diverse chemical properties.

Virtual Screening with Gnina 1.0.

Virtual screening—predicting which compounds within a specified compound library bind to a target molecule, typically a protein—is a fundamental task in the field of drug discovery. Doing virtual screening well provides tangible practical benefits, including reduced drug development costs, faster time to therapeutic viability, and fewer unforeseen side effects. As with most applied computational tasks, the algorithms currently used to perform virtual screening feature inherent tradeoffs between speed and accuracy. Furthermore, even theoretically rigorous, computationally intensive methods may fail to account for important effects relevant to whether a given compound will ultimately be usable as a drug. Here we investigate the virtual screening performance of the recently released Gnina molecular docking software, which uses deep convolutional networks to score protein-ligand structures. We find, on average, that Gnina outperforms conventional empirical scoring. The default scoring in Gnina outperforms the empirical AutoDock Vina scoring function on 89 of the 117 targets of the DUD-E and LIT-PCBA virtual screening benchmarks with a median 1% early enrichment factor that is more than twice that of Vina. However, we also find that issues of bias linger in these sets, even when not used directly to train models, and this bias obfuscates to what extent machine learning models are achieving their performance through a sophisticated interpretation of molecular interactions versus fitting to non-informative simplistic property distributions.

Validation of a Field-Based Ligand Screener Using a Novel Benchmarking Data Set for Assessing 3D-Based Virtual Screening Methods.

Ligand-based methods play a crucial role in virtual screening when the 3D structure of the target is not available. This study discusses the results of a validation study of the CSD field-based ligand screener using a novel benchmarking data set containing 56 targets. The data set was created starting from the target UniProt IDs in a previously published data set (i.e., the AZ data set), by mining ChEMBL to find known active molecules for these targets and by using DUD-E to generate property-matched decoys of the identified actives. Several experiments were performed to assess the virtual screening performance of the new method. One of its strengths is that it can use an overlay of multiple flexible ligands as a query without the need to run several parallel calculations with one ligand at a time. Here, we discuss how changes to different parameter settings or adoption of different query models can influence the final performance compared to the performance when using the experimentally observed overlay of ligands. We have also generated the enrichment scores based on three external benchmark data sets to enable the comparison with existing methods previously validated using these data sets. Here, we present results for the standard DUD-E data set, the DUD-E+ data set, as well as the DUD_Lib_VS_1.0 data set which was designed for ligand-based virtual screening validation and hence is more suitable for this type of methods.

A comprehensive comparative assessment of 3D molecular similarity tools in ligand-based virtual screening.

Three-dimensional (3D) molecular similarity, one major ligand-based virtual screening (VS) method, has been widely used in the drug discovery process. A variety of 3D molecular similarity tools have been developed in recent decades. In this study, we assessed a panel of 15 3D molecular similarity programs against the DUD-E and LIT-PCBA datasets, including commercial ROCS and Phase, in terms of screening power and scaffold-hopping power. The results revealed that (1) SHAFTS, LS-align, Phase Shape_Pharm and LIGSIFT showed the best VS capability in terms of screening power. Some 3D similarity tools available to academia can yield relatively better VS performance than commercial ROCS and Phase software. (2) Current 3D similarity VS tools exhibit a considerable ability to capture actives with new chemotypes in terms of scaffold hopping. (3) Multiple conformers relative to single conformations will generally improve VS performance for most 3D similarity tools, with marginal improvement observed in area under the receiving operator characteristic curve values, enrichment factor in the top 1% and hit rate in the top 1% values showed larger improvement. Moreover, redundancy and complementarity analyses of hit lists from different query seeds and different 3D similarity VS tools showed that the combination of different query seeds and/or different 3D similarity tools in VS campaigns retrieved more (and more diverse) active molecules. These findings provide useful information for guiding choices of the optimal 3D molecular similarity tools for VS practices and designing possible combination strategies to discover more diverse active compounds.

Can molecular dynamics simulations improve the structural accuracy and virtual screening performance of GPCR models?

The determination of G protein-coupled receptor (GPCR) structures at atomic resolution has improved understanding of cellular signaling and will accelerate the development of new drug candidates. However, experimental structures still remain unavailable for a majority of the GPCR family. GPCR structures and their interactions with ligands can also be modelled computationally, but such predictions have limited accuracy. In this work, we explored if molecular dynamics (MD) simulations could be used to refine the accuracy of in silico models of receptor-ligand complexes that were submitted to a community-wide assessment of GPCR structure prediction (GPCR Dock). Two simulation protocols were used to refine 30 models of the D3 dopamine receptor (D3R) in complex with an antagonist. Close to 60 μs of simulation time was generated and the resulting MD refined models were compared to a D3R crystal structure. In the MD simulations, the receptor models generally drifted further away from the crystal structure conformation. However, MD refinement was able to improve the accuracy of the ligand binding mode. The best refinement protocol improved agreement with the experimentally observed ligand binding mode for a majority of the models. Receptor structures with improved virtual screening performance, which was assessed by molecular docking of ligands and decoys, could also be identified among the MD refined models. Application of weak restraints to the transmembrane helixes in the MD simulations further improved predictions of the ligand binding mode and second extracellular loop. These results provide guidelines for application of MD refinement in prediction of GPCR-ligand complexes and directions for further method development.

Lean-Docking: Exploiting Ligands' Predicted Docking Scores to Accelerate Molecular Docking.

In structure-based virtual screening (SBVS), a binding site on a protein structure is used to search for ligands with favorable nonbonded interactions. Because it is computationally difficult, docking is time-consuming and any docking user will eventually encounter a chemical library that is too big to dock. This problem might arise because there is not enough computing power or because preparing and storing so many three-dimensional (3D) ligands requires too much space. In this study, however, we show that quality regressors can be trained to predict docking scores from molecular fingerprints. Although typical docking has a screening rate of less than one ligand per second on one CPU core, our regressors can predict about 5800 docking scores per second. This approach allows us to focus docking on the portion of a database that is predicted to have docking scores below a user-chosen threshold. Herein, usage examples are shown, where only 25% of a ligand database is docked, without any significant virtual screening performance loss. We call this method "lean-docking". To validate lean-docking, a massive docking campaign using several state-of-the-art docking software packages was undertaken on an unbiased data set, with only wet-lab tested active and inactive molecules. Although regressors allow the screening of a larger chemical space, even at a constant docking power, it is also clear that significant progress in the virtual screening power of docking scores is desirable.

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.