Search

In recent years, there has been an uptick in interest in generative models for molecules in drug development. In the field of de novo molecular design, these models are used to make molecules with desired properties from scratch. This is occasionally used instead of virtual screening, which is limited by the size of the libraries that can be searched in practice. Rather than screening existing libraries, generative models can be used to build custom libraries from scratch. Using generative models, which may optimize molecules straight towards the desired profile, this time-consuming approach can be sped up. The purpose of this work is to show how current shortcomings in evaluating generative models for molecules can be avoided. We cover both distribution-learning and goal-directed generation with a focus on the latter. Three well-known targets were downloaded from ChEMBL: Janus kinase 2 (JAK2), epidermal growth factor receptor (EGFR), and dopamine receptor D2 (DRD2) (Bento et al. 2014). We preprocessed the data to get binary classification jobs. Before calculating a scoring function, the data is split into two halves, which we shall refer to as split 1/2. The ratio of active to inactive users. Our goal is to train three bioactivity models with equal prediction performance, one to be used as a scoring function for chemical optimization and the other two to be used as performance evaluation models. Our findings suggest that distribution-learning can attain near-perfect scores on many existing criteria even with the most basic and completely useless models. According to benchmark studies, likelihood-based models account for many of the best technologies, and we propose that test set likelihoods be included in future comparisons.

Background Computational methods support nowadays each stage of drug design campaigns. They assist not only in the process of identification of new active compounds towards particular biological target, but also help in the evaluation and optimization of their physicochemical and pharmacokinetic properties. Such features are not less important in terms of the possible turn of a compound into a future drug than its desired affinity profile towards considered proteins. In the study, we focus on metabolic stability, which determines the time that the compound can act in the organism and play its role as a drug. Due to great complexity of xenobiotic transformation pathways in the living organisms, evaluation and optimization of metabolic stability remains a big challenge. Results Here, we present a novel methodology for the evaluation and analysis of structural features influencing metabolic stability. To this end, we use a well-established explainability method called SHAP. We built several predictive models and analyse their predictions with the SHAP values to reveal how particular compound substructures influence the model’s prediction. The method can be widely applied by users thanks to the web service, which accompanies the article. It allows a detailed analysis of SHAP values obtained for compounds from the ChEMBL database, as well as their determination and analysis for any compound submitted by a user. Moreover, the service enables manual analysis of the possible structural modifications via the provision of analogous analysis for the most similar compound from the ChEMBL dataset. Conclusions To our knowledge, this is the first attempt to employ SHAP to reveal which substructural features are utilized by machine learning models when evaluating compound metabolic stability. The accompanying web service for metabolic stability evaluation can be of great help for medicinal chemists. Its significant usefulness is related not only to the possibility of assessing compound stability, but also to the provision of information about substructures influencing this parameter. It can assist in the design of new ligands with improved metabolic stability, helping in the detection of privileged and unfavourable chemical moieties during stability optimization. The tool is available at https://metstab-shap.matinf.uj.edu.pl/.

Deep generative models are attracting much attention in the field of de novo molecule design. Compared to traditional methods, deep generative models can be trained in a fully data-driven way with little requirement for expert knowledge. Although many models have been developed to generate 1D and 2D molecular structures, 3D molecule generation is less explored, and the direct design of drug-like molecules inside target binding sites remains challenging. In this work, we introduce DeepLigBuilder, a novel deep learning-based method for de novo drug design that generates 3D molecular structures in the binding sites of target proteins. We first developed Ligand Neural Network (L-Net), a novel graph generative model for the end-to-end design of chemically and conformationally valid 3D molecules with high drug-likeness. Then, we combined L-Net with Monte Carlo tree search to perform structure-based de novo drug design tasks. In the case study of inhibitor design for the main protease of SARS-CoV-2, DeepLigBuilder suggested a list of drug-like compounds with novel chemical structures, high predicted affinity, and similar binding features to those of known inhibitors. The current version of L-Net was trained on drug-like compounds from ChEMBL, which could be easily extended to other molecular datasets with desired properties based on users' demands and applied in functional molecule generation. Merging deep generative models with atomic-level interaction evaluation, DeepLigBuilder provides a state-of-the-art model for structure-based de novo drug design and lead optimization., DeepLigBuilder, a novel deep generative model for structure-based de novo drug design, directly generates 3D structures of drug-like compounds in the target binding site.

Accurate prediction of lipophilicity—logP—based on molecular structures is a well-established field. Predictions of logP are often used to drive forward drug discovery projects. Driven by the SAMPL7 challenge, in this manuscript we describe the steps that were taken to construct a novel machine learning model that can predict and generalize well. This model is based on the recently described Directed-Message Passing Neural Networks (D-MPNNs). Further enhancements included: both the inclusion of additional datasets from ChEMBL (RMSE improvement of 0.03), and the addition of helper tasks (RMSE improvement of 0.04). To the best of our knowledge, the concept of adding predictions from other models (Simulations Plus logP and logD@pH7.4, respectively) as helper tasks is novel and could be applied in a broader context. The final model that we constructed and used to participate in the challenge ranked 2/17 ranked submissions with an RMSE of 0.66, and an MAE of 0.48 (submission: Chemprop). On other datasets the model also works well, especially retrospectively applied to the SAMPL6 challenge where it would have ranked number one out of all submissions (RMSE of 0.35). Despite the fact that our model works well, we conclude with suggestions that are expected to improve the model even further.

pKa is an important property in the lead optimization process since the charge state of a molecule in physiologic pH plays a critical role in its biological activity, solubility, membrane permeability, metabolism, and toxicity. Accurate and fast estimation of small molecule pKa is vital during the drug discovery process. We present MolGpKa, a web server for pKa prediction using a graph-convolutional neural network model. The model works by learning pKa related chemical patterns automatically and building reliable predictors with learned features. ACD/pKa data for 1.6 million compounds from the ChEMBL database was used for model training. We found that the performance of the model is better than machine learning models built with human-engineered fingerprints. Detailed analysis shows that the substitution effect on pKa is well learned by the model. MolGpKa is a handy tool for the rapid estimation of pKa during the ligand design process. The MolGpKa server is freely available to researchers and can be accessed at https://xundrug.cn/molgpka.

Tuberculosis (TB) is an infectious disease and the leading cause of death globally. The rapidly emerging cases of drug resistance among pathogenic mycobacteria have been a global threat urging the need of new drug discovery and development. However, considering the fact that the new drug discovery and development is commonly lengthy and costly processes, strategic use of the cutting-edge machine learning (ML) algorithms may be very supportive in reducing both the cost and time involved. Considering the urgency of new drugs for TB, herein, we have attempted to develop predictive ML algorithms-based models useful in the selection of novel potential small molecules for subsequent in vitro validation. For this purpose, we used the GlaxoSmithKline (GSK) TCAMS TB dataset comprising a total of 776 hits that were made publicly available to the wider scientific community through the ChEMBL Neglected Tropical Diseases (ChEMBL-NTD) database. After exploring the different ML classifiers, viz. decision trees (DT), support vector machine (SVM), random forest (RF), Bernoulli Naive Bayes (BNB), K-nearest neighbors (k-NN), and linear logistic regression (LLR), and ensemble learning models (bagging and Adaboost) for training the model using the GSK dataset, we concluded with three best models, viz. Adaboost decision tree (ABDT), RF classifier, and k-NN models that gave the top prediction results for both the training and test sets. However, during the prediction of the external set of known anti-tubercular compounds/drugs, it was realized that each of these models had some limitations. The ABDT model correctly predicted 22 molecules as actives, while both the RF and k-NN models predicted 18 molecules correctly as actives; a number of molecules were predicted as actives by two of these models, while the third model predicted these compounds as inactives. Therefore, we concluded that while deciding the anti-tubercular potential of a new molecule, one should rely on the use of consensus predictions using these three models; it may lessen the attrition rate during the in vitro validation. We believe that this study may assist the wider anti-tuberculosis research community by providing a platform for predicting small molecules with subsequent validation for drug discovery and development.

The discovery of new medications in a cost-effective manner has become the top priority for many pharmaceutical companies. Despite decades of innovation, many of their processes arguably remain relatively inefficient. One such process is the prediction of biological activity. This paper describes a new deep learning model, capable of conducting a preliminary screening of chemical compounds in-silico. The model has been constructed using a variation autoencoder to generate chemical compound fingerprints, which have been used to create a regression model to predict their LogD property and a classification model to predict binding in selected assays from the ChEMBL dataset. The conducted experiments demonstrate accurate prediction of the properties of chemical compounds only using structural definitions and also provide several opportunities to improve upon this model in the future.

In ligand-based drug design, quantitative structure-activity relationship (QSAR) models play an important role in activity prediction. One of the major end points of QSAR models is half-maximal inhibitory concentration (IC50). Experimental IC50 data from various research groups have been accumulated in publicly accessible databases, providing an opportunity for us to use such data in predictive QSAR models. In this study, we focused on using a ranking-oriented QSAR model as a predictive model because relative potency strength within the same assay is solid information that is not based on any mechanical assumptions. We conducted rigorous validation using the ChEMBL database and previously reported data sets. Ranking support vector machine (ranking-SVM) models trained on compounds from similar assays were as good as support vector regression (SVR) with the Tanimoto kernel trained on compounds from all the assays. As effective ways of data integration, for ranking-SVM, integrated compounds should be selected from only similar assays in terms of compounds. For SVR with the Tanimoto kernel, entire compounds from different assays can be incorporated.

Artificial Intelligence/Machine Learning (AI/ML) algorithms may speed up the design of DADNP systems formed by Antibacterial Drugs (AD) and Nanoparticles (NP). In this work, we used IFPTML = Information Fusion (IF) + Perturbation-Theory (PT) + Machine Learning (ML) algorithm for the first time to study of a large dataset of putative DADNP systems composed by >165 000 ChEMBL AD assays and 300 NP assays vs. multiple bacteria species. We trained alternative models with Linear Discriminant Analysis (LDA), Artificial Neural Networks (ANN), Bayesian Networks (BNN), K-Nearest Neighbour (KNN) and other algorithms. IFPTML-LDA model was simpler with values of Sp ≈ 90% and Sn ≈ 74% in both training (>124 K cases) and validation (>41 K cases) series. IFPTML-ANN and KNN models are notably more complicated even when they are more balanced Sn ≈ Sp ≈ 88.5%-99.0% and AUROC ≈ 0.94-0.99 in both series. We also carried out a simulation (>1900 calculations) of the expected behavior for putative DADNPs in 72 different biological assays. The putative DADNPs studied are formed by 27 different drugs with multiple classes of NP and types of coats. In addition, we tested the validity of our additive model with 80 DADNP complexes experimentally synthetized and biologically tested (reported in >45 papers). All these DADNPs show values of MIC < 50 μg mL-1 (cutoff used) better that MIC of AD and NP alone (synergistic or additive effect). The assays involve DADNP complexes with 10 types of NP, 6 coating materials, NP size range 5-100 nm vs. 15 different antibiotics, and 12 bacteria species. The IFPTML-LDA model classified correctly 100% (80 out of 80) DADNP complexes as biologically active. IFPMTL additive strategy may become a useful tool to assist the design of DADNP systems for antibacterial therapy taking into consideration only information about AD and NP components by separate., G. D. H acknowledges financial support from grants Minister of Science and Innovation (PID2019-104148GB-I00) and grant (IT1045-16) – 2016–2021 of Basque Government.

Machine learning methods are attracting considerable attention from the pharmaceutical industry for use in drug discovery and applications beyond. In recent studies, we and others have applied multiple machine learning algorithms and modeling metrics and, in some cases, compared molecular descriptors to build models for individual targets or properties on a relatively small scale. Several research groups have used large numbers of datasets from public databases such as ChEMBL in order to evaluate machine learning methods of interest to them. The largest of these types of studies used on the order of 1400 datasets. We have now extracted well over 5000 datasets from CHEMBL for use with the ECFP6 fingerprint and in comparison of our proprietary software Assay Central with random forest, k-nearest neighbors, support vector classification, naïve Bayesian, AdaBoosted decision trees, and deep neural networks (three layers). Model performance was assessed using an array of fivefold cross-validation metrics including area-under-the-curve, F1 score, Cohen's kappa, and Matthews correlation coefficient. Based on ranked normalized scores for the metrics or datasets, all methods appeared comparable, while the distance from the top indicated that Assay Central and support vector classification were comparable. Unlike prior studies which have placed considerable emphasis on deep neural networks (deep learning), no advantage was seen in this case. If anything, Assay Central may have been at a slight advantage as the activity cutoff for each of the over 5000 datasets representing over 570,000 unique compounds was based on Assay Central performance, although support vector classification seems to be a strong competitor. We also applied Assay Central to perform prospective predictions for the toxicity targets PXR and hERG to further validate these models. This work appears to be the largest scale comparison of these machine learning algorithms to date. Future studies will likely evaluate additional databases, descriptors, and machine learning algorithms and further refine the methods for evaluating and comparing such models.

Mining the steadily increasing amount of chemical and biological data is a key challenge in drug discovery. Graph databases offer viable alternatives for capturing interrelationships between molecules and for generating novel insights for design. In a graph database, molecules and their properties are mapped to nodes, while relationships are described by edges. Here, we introduce a graph database for navigation in chemical space, analogue searching, and structure-activity relationship (SAR) analysis. We illustrate this concept using hERG channel inhibitors from ChEMBL to extract SAR knowledge. This graph database is built using different relationships, namely 2D-fingerprint similarity, matched molecular pairs, topomer distances, and structure-activity landscape indices (SALI). Typical applications include retrieving analogues linked by single or multiple edge paths to the query compound as well as detection of nonadditive SAR features. Finally, we identify triplets of linked molecules for clustering. The speed of searching and analysis allows the user to interactively navigate the database and to address complex questions in real-time.

Alzheimer’s disease (AD) is the most common cause of dementia, affecting approximately 35 million people worldwide. The current treatment options for people with AD consist of drugs designed to slow the rate of decline in memory and cognition, but these treatments are not curative, and patients eventually suffer complete cognitive injury. With the substantial amounts of published data on targets for this disease, we proposed that machine learning software could be used to find novel small-molecule treatments that can supplement the AD drugs currently on the market. In order to do this, we used publicly available data in ChEMBL to build and validate Bayesian machine learning models for AD target proteins. The first AD target that we have addressed with this method is the serine–threonine kinase glycogen synthase kinase 3 beta (GSK3β), which is a proline-directed serine–threonine kinase that phosphorylates the microtubule-stabilizing protein tau. This phosphorylation prompts tau to dissociate from the microtubule and form insoluble oligomers called paired helical filaments, which are one of the components of the neurofibrillary tangles found in AD brains. Using our Bayesian machine learning model for GSK3β consisting of 2368 molecules, this model produced a five-fold cross validation ROC of 0.905. This model was also used for virtual screening of large libraries of FDA-approved drugs and clinical candidates. Subsequent testing of selected compounds revealed a selective small-molecule inhibitor, ruboxistaurin, with activity against GSK3β (avg IC50 = 97.3 nM) and GSK3α (IC50 = 695.9 nM). Several other structurally diverse inhibitors were also identified. We are now applying this machine learning approach to additional AD targets to identify approved drugs or clinical trial candidates that can be repurposed as AD therapeutics. This represents a viable approach to accelerate drug discovery and do so at a fraction of the cost of traditional high throughput screening.

Leishmaniasis refers to a complex of diseases, caused by the intracellular parasitic protozoans belonging to the genus Leishmania. Among the three types of disease manifestations, the most severe type is visceral leishmaniasis, which is caused by Leishmania donovani, and is diagnosed in more than 20,000 cases annually, worldwide. Because the current therapeutic options for disease treatment are associated with several limitations, the identification of new potential leads/drugs remains necessary. In this study, a combined approach was used, based on two different virtual screening (VS) methods, which were designed to select promising antileishmanial agents from among the entire sesquiterpene lactone (SL) dataset registered in SistematX, a web interface for managing a secondary metabolite database that is accessible by multiple platforms on the Internet. Thus, a ChEMBL dataset, including 3159 and 1569 structures that were previously tested against L. donovani amastigotes and promastigotes in vitro, respectively, was used to develop two random forest models, which performed with greater than 74% accuracy in both the cross-validation and test sets. Subsequently, a ligand-based VS assay was performed against the 1306 SistematX-registered SLs. In parallel, the crystal structures of three L. donovani target proteins, N-myristoyltransferase, ornithine decarboxylase, and mitogen-activated protein kinase 3, and a homology model of pteridine reductase 1 were used to perform a structure-based VS, using molecular docking, of the entire SistematX SL dataset. The consensus analysis of these two VS approaches resulted in the normalization of probability scores and identified 13 promising, enzyme-targeting, antileishmanial SLs from SistematX that may act against L. donovani. A combined approach based on two different virtual screening methods (structure-based and ligand-based) was performed using an in-house dataset composed of 1306 sesquiterpene lactones to identify potential antileishmanial (Leishmania donovani) structures.

Bruton's tyrosine kinase (BTK) has a crucial role in multiple cell signaling pathways including B-cell antigen receptor (BCR) and Fc receptor (FcR) signaling cascades, which has attracted much attention to find BTK inhibitors to treat autoimmune diseases. In this work, we constructed a Bayesian classification model for virtually seeking novel BTK inhibitors, which showed good performance in terms of screening efficiency and accuracy. Through searching for several chemical libraries including Chembl_17 (1,317,484 compounds), Chembridge (103,473 compounds), and Chemdiv (700,000 compounds) using this model followed by molecular docking and activity prediction, 52 compounds with novel scaffolds were acknowledged as potential BTK inhibitors, which could be promising starting points for further exploration. This study also provided a guide to construct an efficient and effective protocol for virtual screening by integrating machine learning methods.

Recommender systems (RSs), which underwent rapid development and had an enormous impact on e-commerce, have the potential to become useful tools for drug discovery. In this paper, we applied RS methods for the prediction of the antiviral activity class (active/inactive) for compounds extracted from ChEMBL. Two main RS approaches were applied: collaborative filtering (Surprise implementation) and content-based filtering (sparse-group inductive matrix completion (SGIMC) method). The effectiveness of RS approaches was investigated for prediction of antiviral activity classes ("interactions") for compounds and viruses, for which some of their interactions with other viruses or compounds are known, and for prediction of interaction profiles for new compounds. Both approaches achieved relatively good prediction quality for binary classification of individual interactions and compound profiles, as quantified by cross-validation and external validation receiver operating characteristic (ROC) score >0.9. Thus, even simple recommender systems may serve as an effective tool in antiviral drug discovery.

Although the NaV1.7 sodium channel is a promising drug target for pain, traditional screening strategies for discovery of NaV1.7 inhibitors are very painstaking and time-consuming. Herein, we aimed to build machine learning models for screening and design of potent and effective NaV1.7 sodium channel inhibitors. We customized the imbalanced data set from ChEMBL and BindingDB to train and filter the best classification model. Then, the whole-cell voltage-clamp was employed to validate the inhibitors. We assembled a molecular group optimization method by combining the Grammar Variational Autoencoder, classification model, and simulated annealing. We found that the RF-CDK model (random forest + CDK fingerprint) performs best in the imbalanced data set. Of the three compounds that may have inhibitory effects, nortriptyline has been experimentally verified. In the molecule optimization process, 40 molecules located in the applicability domain of RF-CDK were used as a starting point, among which 34 molecules evolved to molecules with greater molecular scores (MS). The molecule with the highest MS was derived from CHEMBL2325245. The model and method we developed for NaV1.7 inhibitors are also applicable to other targets.

Bioscientists reading papers or patents strive to discern the key relationships reported within a document “D“ where a bioactivity “A” with a quantitative result “R” (e.g., an IC50) is reported for chemical structure “C” that modulates (e.g., inhibits) a protein target “P”. A useful shorthand for this connectivity thus becomes DARCP. The problem at the core of this article is that the community has spent millions effectively burying these relationships in PDFs over many decades but must now spend millions more trying to get them back out. The key imperative for this is to increase the flow into structured open databases. The positive impacts will include expanded data mining opportunities for drug discovery and chemical biology. Over the last decade commercial sources have manually extracted DARCP from ≈300,000 documents encompassing ≈7 million compounds interacting with ≈10,000 targets. Over a similar time, the Guide to Pharmacology, BindingDB and ChEMBL have carried out analogues DARCP extractions. Although their expert-curated numbers are lower (i.e., ≈2 million compounds against ≈3700 human proteins), these open sources have the great advantage of being merged within PubChem. Parallel efforts have focused on the extraction of document-to-compound (D-C-only) connectivity. In the absence of molecular mechanism of action (mmoa) annotation, this is of less value but can be automatically extracted. This has been significantly accomplished for patents, (e.g., by IBM, SureChEMBL and WIPO) for over 30 million compounds in PubChem. These have recently been joined by 1.4 million D-C submissions from three major chemistry publishers. In addition, both the European and US PubMed Central portals now add chemistry look-ups from abstracts and full-text papers. However, the fully automated extraction of DARCLP has not yet been achieved. This stands in contrast to the ability of biocurators to discern these relationships in minutes. Unfortunately, no journals have yet instigated a flow of author-specified DARCP directly into open databases. Progress may come from trends such as open science, open access (OA), findable, accessible, interoperable and reusable (FAIR), resource description framework (RDF) and WikiData. However, we will need to await the technical applicability in respect to DARCP capture to see if this opens up connectivity.

The chemical sciences are producing an unprecedented amount of large, high-dimensional data sets containing chemical structures and associated properties. However, there are currently no algorithms to visualize such data while preserving both global and local features with a sufficient level of detail to allow for human inspection and interpretation. Here, we propose a solution to this problem with a new data visualization method, TMAP, capable of representing data sets of up to millions of data points and arbitrary high dimensionality as a two-dimensional tree (http://tmap.gdb.tools). Visualizations based on TMAP are better suited than t-SNE or UMAP for the exploration and interpretation of large data sets due to their tree-like nature, increased local and global neighborhood and structure preservation, and the transparency of the methods the algorithm is based on. We apply TMAP to the most used chemistry data sets including databases of molecules such as ChEMBL, FDB17, the Natural Products Atlas, DSSTox, as well as to the MoleculeNet benchmark collection of data sets. We also show its broad applicability with further examples from biology, particle physics, and literature., Comment: 33 pages, 14 figures, 1 table, supplementary information included

Motivation Systematic identification of molecular targets among known drugs plays an essential role in drug repurposing and understanding of their unexpected side effects. Computational approaches for prediction of drug–target interactions (DTIs) are highly desired in comparison to traditional experimental assays. Furthermore, recent advances of multiomics technologies and systems biology approaches have generated large-scale heterogeneous, biological networks, which offer unexpected opportunities for network-based identification of new molecular targets among known drugs. Results In this study, we present a network-based computational framework, termed AOPEDF, an arbitrary-order proximity embedded deep forest approach, for prediction of DTIs. AOPEDF learns a low-dimensional vector representation of features that preserve arbitrary-order proximity from a highly integrated, heterogeneous biological network connecting drugs, targets (proteins) and diseases. In total, we construct a heterogeneous network by uniquely integrating 15 networks covering chemical, genomic, phenotypic and network profiles among drugs, proteins/targets and diseases. Then, we build a cascade deep forest classifier to infer new DTIs. Via systematic performance evaluation, AOPEDF achieves high accuracy in identifying molecular targets among known drugs on two external validation sets collected from DrugCentral [area under the receiver operating characteristic curve (AUROC) = 0.868] and ChEMBL (AUROC = 0.768) databases, outperforming several state-of-the-art methods. In a case study, we showcase that multiple molecular targets predicted by AOPEDF are associated with mechanism-of-action of substance abuse disorder for several marketed drugs (such as aripiprazole, risperidone and haloperidol). Availability and implementation Source code and data can be downloaded from https://github.com/ChengF-Lab/AOPEDF. Supplementary information Supplementary data are available at Bioinformatics online.

Nanoparticles (NPs) decorated with coating agents (polymers, gels, proteins, etc.) form Nanoparticle Drug Delivery Systems (DDNS), which are of high interest in nanotechnology and biomaterials science. There have been increasing reports of experimental data sets of biological activity, toxicity, and delivery properties of DDNS. However, these data sets are still dispersed and not as large as the datasets of DDNS components (NP and drugs). This has prompted researchers to train Machine Learning (ML) algorithms that are able to design new DDNS based on the properties of their components. However, most ML models reported up to date predictions of the specific activities of NP or drugs over a determined target or cell line. In this paper, we combine Perturbation Theory and Machine Learning (PTML algorithm) to train a model that is able to predict the best components (NP, coating agent, and drug) for DDNS design. In so doing, we downloaded a dataset of >30 000 preclinical assays of drugs from ChEMBL. We also downloaded an NP data set formed by preclinical assays of coated Metal Oxide Nanoparticles (MONPs) from public sources. Both the drugs and NP datasets of preclinical assays cover multiple conditions of assays that can be listed as two arrays, namely, cjdrug and cjNP. The cjdrug array includes >504 biological activity parameters (c0drug), >340 target proteins (c1drug), >650 types of cells (c2drug), >120 assay organisms (c3drug), and >60 assay strains (c4drug). On the other hand, the cjNP array includes 3 biological activity parameters (c0NP), 40 types of proteins (c1NP), 10 shapes of nanoparticles (c2NP), 6 assay media (c3NP), and 12 coating agents (c4NP). After downloading, we pre-processed both the data sets by separate calculation PT operators that are able to account for changes (perturbations) in the drug, coating agents, and NP chemical structure and/or physicochemical properties as well as for the assay conditions. Next, we carry out an information fusion process to form a final dataset of above 500 000 DDNS (drug + MONP pairs). We also trained other linear and non-linear PTML models using R studio scripts for comparative purposes. To the best of our knowledge, this is the first multi-label PTML model that is useful for the selection of drugs, coating agents, and metal or metal–oxide nanoparticles to be assembled in order to design new DDNS with optimal activity/toxicity profiles.

Linking fragments to generate a focused compound library for a specific drug target is one of the challenges in fragment-based drug design (FBDD). Hereby, we propose a new program named SyntaLinker, which is based on a syntactic pattern recognition approach using deep conditional transformer neural networks. This state-of-the-art transformer can link molecular fragments automatically by learning from the knowledge of structures in medicinal chemistry databases (e.g. ChEMBL database). Conventionally, linking molecular fragments was viewed as connecting substructures that were predefined by empirical rules. In SyntaLinker, however, the rules of linking fragments can be learned implicitly from known chemical structures by recognizing syntactic patterns embedded in SMILES notations. With deep conditional transformer neural networks, SyntaLinker can generate molecular structures based on a given pair of fragments and additional restrictions. Case studies have demonstrated the advantages and usefulness of SyntaLinker in FBDD., Linking fragments to generate a focused compound library for a specific drug target is one of the challenges in fragment-based drug design (FBDD).

Molecular optimization aims to improve the drug profile of a starting molecule. It is a fundamental problem in drug discovery but challenging due to (i) the requirement of simultaneous optimization of multiple properties and (ii) the large chemical space to explore. Recently, deep learning methods have been proposed to solve this task by mimicking the chemist's intuition in terms of matched molecular pairs (MMPs). Although MMPs is a typical and widely used strategy by medicinal chemists, it offers limited capability in terms of exploring the space of solutions. There are more options to modify a starting molecule to achieve desirable properties, e.g. one can simultaneously modify the molecule at different places including changing the scaffold. This study trains the same Transformer architecture on different datasets. These datasets consist of a set of molecular pairs which reflect different types of transformations. Beyond MMP transformation, datasets reflecting general transformations are constructed from ChEMBL based on two approaches: Tanimoto similarity (allows for multiple modifications) and scaffold matching (allows for multiple modifications but keep the scaffold constant) respectively. We investigate how the model behavior can be altered by tailoring the dataset while keeping the same model architecture. Our results show that the models trained on differently prepared datasets transform a given starting molecule in a way that it reflects the nature of the dataset used for training the model. These models could complement each other and unlock the capability for the chemists to pursue different options for improving a starting molecule.

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.

BackgroundDrug-target interactions (DTIs) are critical for drug repurposing and elucidation of drug mechanisms, and they are collected in large databases, such as ChEMBL, BindingDB, DrugBank and DrugTargetCommons. However, the number of studies providing this data (~0.1 million) likely constitutes only a fraction of all studies on PubMed that contain experimental DTI data. Finding such studies and extracting the experimental information is a challenging task, and there is a pressing need for machine learning for the extraction and curation of DTIs. To this end, we developed new text mining document classifiers based on the Bidirectional Encoder Representations from Transformers (BERT) algorithm. Because DTI data intimately depends on the type of assays used to generate it, we also aimed to incorporate functions to predict the assay format.ResultsOur novel method identified and extracted DTIs from 2.1 million studies not previously included in public DTI databases. Using 10-fold cross-validation, we obtained ~99% accuracy for identifying studies containing drug-target pairs. The accuracy for the prediction of assay format is ~90%, which leaves room for improvement in future studies.ConclusionThe BERT model in this study is robust and the proposed pipeline can be used to identify new and previously overlooked studies containing DTIs and automatically extract the DTI data points. The tabular output facilitates validation of the extracted data and assay format information. Overall, our method provides a significant advancement in machine-assisted DTI extraction and curation. We expect it to be a useful addition to drug mechanism discovery and repurposing.

Cardiovascular diseases (CDs) are a major concern in the human race and one of the leading causes of death worldwide. β-Adrenergic receptors (β1-AR and β2-AR) play a crucial role in the overall regulation of cardiac function. In the present study, structure-based virtual screening, machine learning (ML), and a ligand-based similarity search were conducted for the PubChem database against both β1- and β2-AR. Initially, all docked molecules were screened using the threshold binding energy value. Molecules with a better binding affinity were further used for segregation as active and inactive through ML. The pharmacokinetic assessment was carried out on molecules retained in the above step. Further, similarity searching of the ChEMBL and DrugBank databases was performed. From detailed analysis of the above data, four compounds for each of β1- and β2-AR were found to be promising in nature. A number of critical ligand-binding amino acids formed potential hydrogen bonds and hydrophobic interactions. Finally, a molecular dynamics (MD) simulation study of each molecule bound with the respective target was performed. A number of parameters obtained from the MD simulation trajectories were calculated and substantiated the stability between the protein-ligand complex. Hence, it can be postulated that the final molecules might be crucial for CDs subjected to experimental validation.

The theoretical prediction of drug-decorated nanoparticles (DDNPs) has become a very important task in medical applications. For the current paper, Perturbation Theory Machine Learning (PTML) models were built to predict the probability of different pairs of drugs and nanoparticles creating DDNP complexes with anti-glioblastoma activity. PTML models use the perturbations of molecular descriptors of drugs and nanoparticles as inputs in experimental conditions. The raw dataset was obtained by mixing the nanoparticle experimental data with drug assays from the ChEMBL database. Ten types of machine learning methods have been tested. Only 41 features have been selected for 855,129 drug-nanoparticle complexes. The best model was obtained with the Bagging classifier, an ensemble meta-estimator based on 20 decision trees, with an area under the receiver operating characteristic curve (AUROC) of 0.96, and an accuracy of 87% (test subset). This model could be useful for the virtual screening of nanoparticle-drug complexes in glioblastoma. All the calculations can be reproduced with the datasets and python scripts, which are freely available as a GitHub repository from authors., The APC was funded by IKERDATA, S.L. under grant 3/12/DP/2021/00102, This work is supported by the “Collaborative Project in Genomic Data Integration (CICLOGEN)” PI17/01826 funded by the Carlos III Health Institute, from the Spanish National Plan for Scientific and Technical Research and Innovation 2013–2016 and the European Regional Development Funds (FEDER)—“A way to build Europe”. This project was also supported by the General Directorate of Culture, Education and University Management of Xunta de Galicia (Ref. ED431G/01, ED431D 2017/16), the “Galician Network for Colorectal Cancer Research” (Ref. ED431D 2017/23), Competitive Reference Groups (Ref. ED431C 2018/49), and the Spanish Ministry of Economy and Competitiveness via funding of the unique installation BIOCAI (UNLC08-1E-002, UNLC13-13-3503) and the European Regional Development Funds (FEDER). Lastly, the authors also acknowledge research grants from the Ministry of Economy and Competitiveness, MINECO, Spain (FEDER CTQ2016-74881-P), the Basque government (IT1045-16), and the kind support of Ikerbasque, Basque Foundation for Science and Zitek.

We present a new quantitative ligand-based bioactivity prediction approach employing a multifingerprint similarity search algorithm, enabling the polypharmacological profiling of small molecules. Quantitative bioactivity predictions are made on the basis of the statistical distributions of multiple Tanimoto similarity θ values, calculated through 13 different molecular fingerprints, and of the variation of the measured biological activity, reported as ΔpIC50, for all of the ligands sharing a given protein drug target. The application data set comprises as much as 4241 protein drug targets as well as 418 485 ligands selected from ChEMBL (release 25) by employing a set of well-defined filtering rules. Several large internal and external validation studies were carried out to demonstrate the robustness and the predictive potential of the herein proposed method. Additional comparative studies, carried out on two freely available and well-known ligand-target prediction platforms, demonstrated the reliability of our proposed approach for accurate ligand-target matching. Moreover, two applicative cases were also discussed to practically describe how to use our predictive algorithm, which is freely available as a user-friendly web platform. The user can screen single or multiple queries at a time and retrieve the output as a terse html table or as a json file including all of the information concerning the explored similarities to obtain a deeper understanding of the results. High-throughput virtual reverse screening campaigns, allowing for a given query compound the quick detection of the potential drug target from a large collection of them, can be carried out in batch on demand.

Artificial intelligence (AI) algorithms are dramatically redefining the current drug discovery landscape by boosting the efficiency of its various steps. Still, their implementation often requires a certain level of expertise in AI paradigms and coding. This often prevents the use of these powerful methodologies by non-expert users involved in the design of new biologically active compounds. Here, the random matrix discriminant (RMD) algorithm, a high-performance AI method specifically tailored for the identification of new ligands, was implemented in a new fully automated tool, PyRMD. This ligand-based virtual screening tool can be trained using target bioactivity data directly downloaded from the ChEMBL repository without manual intervention. The software automatically splits the available training compounds into active and inactive sets and learns the distinctive chemical features responsible for the compounds' activity/inactivity. PyRMD was designed to easily screen millions of compounds in hours through an automated workflow and intuitive input files, allowing fine tuning of each parameter of the calculation. Additionally, PyRMD features a wealth of benchmark metrics, to accurately probe the model performance, which were used here to gauge its predictive potential and limitations. PyRMD is freely available on GitHub (https://github.com/cosconatilab/PyRMD) as an open-source tool.

Generative neural networks have shown promising results in de novo drug design. Recent studies suggest that one of the efficient ways to produce novel molecules matching target properties is to model SMILES sequences using deep learning in a way similar to language modeling in natural language processing. In this paper, we present a survey of various machine learning methods for SMILES-based language modeling and propose our benchmarking results on a standardized subset of ChEMBL database.

In drug discovery, Deep Learning algorithms are emerging as a potential method to generate novel chemical structures since they can speed up the traditional process and decrease expenditure. Recurrent architectures are amongst the most promising methods for computational de novo drug design. One current challenge consists in finding the optimal architecture and parameters for the recurrent network that assures the generation of valid molecules that span the chemical space. In this work we perform an evaluation on Recurrent Neural Networks which can learn the syntax of molecular representation in terms of SMILES notation. We optimize the computational framework based on the recurrent architecture and its hyper-parameters. Moreover, we evaluate the performance of two types of encoding and spatial arrangement of molecules: Embedding and One-hot Encoding, and datasets with and without stereo-chemical information, respectively. The proposed model showed improved performance when compared to the current literature, both in terms of percentage of valid generated SMILES and diversity with 98.7% and 0.88, for the ChEMBL dataset, respectively. Even when considering the ZINC biogenic library, with stereochemical information, the values were 94.5% and 0.90. The obtained results reveal the potential of the recurrent architectures in learning the SMILES syntax and adding novelty to generate promising compounds.

Retroviral infections, such as HIV, are, until now, diseases with no cure. Medicine and pharmaceutical chemistry need and consider it a huge goal to define target proteins of new antiretroviral compounds. ChEMBL manages Big Data features with a complex data set, which is hard to organize. This makes information difficult to analyze due to a big number of characteristics described in order to predict new drug candidates for retroviral infections. For this reason, we propose to develop a new predictive model combining perturbation theory (PT) bases and machine learning (ML) modeling to create a new tool that can take advantage of all the available information. The PTML model proposed in this work for the ChEMBL data set preclinical experimental assays for antiretroviral compounds consists of a linear equation with four variables. The PT operators used are founded on multicondition moving averages, combining different features and simplifying the difficulty to manage all data. More than 140 000 preclinical assays for 56 105 compounds with different characteristics or experimental conditions have been carried out and can be found in ChEMBL database, covering combinations with 359 biological activity parameters (c0), 55 protein accessions (c1), 83 cell lines (c2), 64 organisms of assay (c3), and 773 subtypes or strains. We have included 150 148 preclinical experimental assays for HIV virus, 1188 for HTLV virus, 84 for simian immunodeficiency virus, 370 for murine leukemia virus, 119 for Rous sarcoma virus, 1581 for MMTV, etc. We also included 5277 assays for hepatitis B virus. The developed PTML model reached considerable values in sensibility (73.05% for training and 73.10% for validation), specificity (86.61% for training and 87.17% for validation), and accuracy (75.84% for training and 75.98% for validation). We also compared alternative PTML models with different PT operators such as covariance, moments, and exponential terms. Finally, we made a comparison between literature ML models with our PTML model and also artificial neural network (ANN) nonlinear models. We conclude that this PTML model is the first one to consider multiple characteristics of preclinical experimental antiretroviral assays combined, generating a simple, useful, and adaptable instrument, which could reduce time and costs in antiretroviral drugs research.

ChEMBL biological activities prediction for 1-5-bromofur-2-il-2-bromo-2-nitroethene (G1) is a difficult task for cytokine immunotoxicity. The current study presents experimental results for G1 interaction with mouse Th1/Th2 and pro-inflammatory cytokines using a cytometry bead array (CBA). In the in vitro test of CBA, the results show no significant differences between the mean values of the Th1/Th2 cytokines for the samples treated with G1 with respect to the negative control, but there are moderate differences for cytokine values between different periods (24/48 h). The experiments show no significant differences between the mean values of the pro-inflammatory cytokines for the samples treated with G1, regarding the negative control, except for the values of tumor necrosis factor (TNF) and Interleukin (IL6) between the group treated with G1 and the negative control at 48 h. Differences occur for these cytokines in the periods (24/48 h). The study confirmed that the antimicrobial G1 did not alter the Th1/Th2 cytokines concentration in vitro in different periods, but it can alter TNF and IL6. G1 promotes free radicals production and activates damage processes in macrophages culture. In order to predict all ChEMBL activities for drugs in other experimental conditions, a ChEMBL data set was constructed using 25 biological activities, 1366 assays, 2 assay types, 4 assay organisms, 2 organisms, and 12 cytokine targets. Molecular descriptors calculated with Rcpi and 15 machine learning methods were used to find the best model able to predict if a drug could be active or not against a specific cytokine, in specific experimental conditions. The best model is based on 120 selected molecular descriptors and a deep neural network with area under the curve of the receiver operating characteristic of 0.904 and accuracy of 0.832. This model predicted 1384 G1 biological activities against cytokines in all ChEMBL data set experimental conditions.

Several recent reports have shown that long short-term memory generative neural networks (LSTM) of the type used for grammar learning efficiently learn to write Simplified Molecular Input Line Entry System (SMILES) of druglike compounds when trained with SMILES from a database of bioactive compounds such as ChEMBL and can later produce focused sets upon transfer learning with compounds of specific bioactivity profiles. Here we trained an LSTM using molecules taken either from ChEMBL, DrugBank, commercially available fragments, or from FDB-17 (a database of fragments up to 17 atoms) and performed transfer learning to a single known drug to obtain new analogs of this drug. We found that this approach readily generates hundreds of relevant and diverse new drug analogs and works best with training sets of around 40,000 compounds as simple as commercial fragments. These data suggest that fragment-based LSTM offer a promising method for new molecule generation.

International audience; The Virtual Screening (VS) study described herein aimed at detecting novel Bromodomain BRD4 binders and relied on knowledge from public databases (ChEMBL, REAXYS) to establish a battery of predictive models of BRD activity for in silico selection of putative ligands. Beyond the actual discovery of new BRD ligands, this represented an opportunity to practically estimate the actual usefulness of public domain "Big Data" for robust predictive model building. Obtained models were used to virtually screen a collection of 2 million compounds from the Enamine company collection. This industrial partner then experimentally screened a subset of 2992 molecules selected by the VS procedure for their high likelihood to be active. Twenty nine confirmed hits were detected after experimental testing, representing 1% of the selected candidates. As a general conclusion, this study emphasizes once more that public structure-activity databases are nowadays key assets in drug discovery. Their usefulness is however limited by the state-of-the-art knowledge harvested so far by published studies. Target-specific structure activity information is rarely rich enough, and its heterogeneity makes it extremely difficult to exploit in rational drug design. Furthermore, published affinity measures serving to build models selecting compounds to be experimentally screened may not be well correlated with the experimental hit selection criterion (in practice, often imposed by equipment constraints). Nevertheless, a robust 2.6-fold increase in hit rate with respect to an equivalent, random screening campaign showed that machine learning is able to extract some real knowledge in spite of all the noise in structure-activity data.

Nano-systems for cancer co-therapy including vitamins or vitamin derivatives have showed adequate results to continue with further research studies to better understand them. However, the number of different combinations of drugs, vitamins, nanoparticle types, coating agents, synthesis conditions, and system types (nanocapsules, micelles, etc.) to be tested is very large generating a high cost in experimentations. In this context, there are reports of large datasets of preclinical assays of compounds (like in the ChEMBL database) and increasing but yet limited reports of experimental measurements of nano-systems per se. On the other hand, Machine Learning is gaining momentum in Nanotechnology and Pharmaceutical Sciences as a tool for rational design of new drugs and drug-release nano-systems. In this work, we propose to combine Perturbation Theory principles and Machine Learning to develop a PTML model for rational selection of the components of cancer co-therapy drug–vitamin release nano-systems (DVRNs). In doing so, we apply information fusion techniques with 2 data sets: (1) a large ChEMBL dataset of >36 000 preclinical assays of vitamin derivatives and a new dataset of >1000 outcomes of DVRNs, collected herein from the literature for the first time. The ChEMBL dataset used covers a considerable number of assay conditions (cjvit) each one with multiple levels. These conditions included >504 biological activity parameters (c0vit), >340 types of proteins (c1vit), >650 types of cells (c2vit), >120 assay organisms (c3vit), >60 assay strains (c4vit). Regarding the DVRNs, there are 25 different types of nano-systems (njn), with up to 16 conditions (cjn) including also different levels such as 8 biological activity parameters (c0n), 9 raw nanomaterials (c4n), 15 assay cells (c11n), etc. In the first stage, we used Moving Average operators to quantify the perturbations (deviations) in all input variables with respect to the conditions. After that, we used multiplicative PT operators to carry out data fusion, and dimension reduction, and Linear Discriminant Analysis (LDA) to seek the PTML model. The best PTML model found showed values of specificity, sensitivity, and accuracy in the range of 83–88% in training and external validation series for >130 000 cases (DVRNs vs. ChEMBL data pairs) formed after data fusion. To the best of our knowledge, this is the first general purpose model for the rational design of DVRNs for cancer co-therapy.

Here we report PPB2 as a target prediction tool assigning targets to a query molecule based on ChEMBL data. PPB2 computes ligand similarities using molecular fingerprints encoding composition (MQN), molecular shape and pharmacophores (Xfp), and substructures (ECfp4) and features an unprecedented combination of nearest neighbor (NN) searches and Naı̈ve Bayes (NB) machine learning, together with simple NN searches, NB and Deep Neural Network (DNN) machine learning models as further options. Although NN(ECfp4) gives the best results in terms of recall in a 10-fold cross-validation study, combining NN searches with NB machine learning provides superior precision statistics, as well as better results in a case study predicting off-targets of a recently reported TRPV6 calcium channel inhibitor, illustrating the value of this combined approach. PPB2 is available to assess possible off-targets of small molecule drug-like compounds by public access at http://gdb.unibe.ch .

In classical medicinal chemistry, nitrile groups were commonly considered as bioisosteres of carbonyl, hydroxyl and carboxyl groups, as well as halogen atoms. However, there is a lack of in-depth understanding about the structural and energetic characteristics of nitrile groups in protein–ligand interactions. Here, we have surveyed the Protein Data Bank and ChEMBL databases with the goal of characterizing such protein–ligand interactions for nitrile-containing compounds. We discuss the versatile roles of nitrile groups in improving binding affinities, and give special attention to examples of displacing and mimicking binding-site waters by nitrile groups. We expect that this review article will further inspire medicinal chemists to exploit nitrile groups rationally in structure-based drug design.

HIV-1 protease plays an important role in the processing of virus infection. Protease is an effective therapeutic target for the treatment of HIV-1. Our data set is based on a selection of 4855 HIV-1 protease inhibitors (PIs) from ChEMBL. A series of 15 classification models for predicting the active inhibitors were built by machine learning methods, including k-nearest neighors (K-NN), decision tree (DT), random forest (RF), support vector machine (SVM), and deep neural network (DNN). The molecular structures were characterized by (1) fingerprint descriptors including MACCS fingerprints and PubChem fingerprints and (2) physicochemical descriptors calculated by CORINA Symphony. The prediction accuracies of all of the models are more than 70% on the test set; the best accuracy of 83.07% was obtained by model 4A, which was built by the SVM method based on MACCS fingerprint descriptors. Nine consensus models were built with three kinds of different descriptors, which combined all of the machine learning methods using the “consensus prediction”. Model C3a developed with MACCS fingerprint descriptors showed the highest accuracy on both training set (91.96%) and test set (83.15%). An external validation set including 35 989 compounds from DUD database and 239 active inhibitors from the recent literature was used to verify the performance of our model. The best prediction accuracy of 98.37% was obtained by model 3C, which was built by RF based on CORINA Symphony descriptors. In addition, from the analysis of molecular descriptors, it shows that the aromatic system and atoms related to hydrogen bonding provide important contributions to the bioactivity of PIs.

The large amount of data that has been collected so far for G protein-coupled receptors requires machine learning (ML) approaches to fully exploit its potential. Our previous ML model based on gradient boosting used for prediction of drug affinity and selectivity for a receptor subtype was compared with explicit information on ligand-receptor interactions from induced-fit docking. Both methods have proved their usefulness in drug response predictions. Yet, their successful combination still requires allosteric/orthosteric assignment of ligands from datasets. Our ligand datasets included activities of two members of the secretin receptor family: GCGR and GLP-1R. Simultaneous activation of two or three receptors of this family by dual or triple agonists is not a typical kind of information included in compound databases. A precise allosteric/orthosteric ligand assignment requires a continuous update based on new structural and biological data. This data incompleteness remains the main obstacle for current ML methods applied to class B GPCR drug discovery. Even so, for these two class B receptors, our ligand-based ML model demonstrated high accuracy (5-fold cross-validation Q2 &gt, 0.63 and Q2 &gt, 0.67 for GLP-1R and GCGR, respectively). In addition, we performed a ligand annotation using recent cryogenic-electron microscopy (cryo-EM) and X-ray crystallographic data on small-molecule complexes of GCGR and GLP-1R. As a result, we assigned GLP-1R and GCGR actives deposited in ChEMBL to four small-molecule binding sites occupied by positive and negative allosteric modulators and a full agonist. Annotated compounds were added to our recently released repository of GPCR data.

A large proportion of lead compounds are derived from natural products. However, most natural products have not been fully tested for their targets. To help resolve this problem, a model using transfer learning was built to predict targets for natural products. The model was pre-trained on a processed ChEMBL dataset and then fine-tuned on a natural product dataset. Benefitting from transfer learning and the data balancing technique, the model achieved a highly promising area under the receiver operating characteristic curve (AUROC) score of 0.910, with limited task-related training samples. Since the embedding distribution difference is reduced, embedding space analysis demonstrates that the model’s outputs of natural products are reliable. Case studies have proved our model’s performance in drug datasets. The fine-tuned model can successfully output all the targets of 62 drugs. Compared with a previous study, our model achieved better results in terms of both AUROC validation and its success rate for obtaining active targets among the top ones. The target prediction model using transfer learning can be applied in the field of natural product-based drug discovery and has the potential to find more lead compounds or to assist researchers in drug repurposing.

New drugs introduced in the market should not only be able to cure the targeted indications, but also should have no or minimal side effects to our body. Traditional drug discovery including in-vitro and in-vivo methods is a challenging, costly and time-consuming process. Thus, in-silico method was introduced where the Drug-Target Interaction (DTI) is predicted using computation methods. A reliable DTI prediction model would save a lot of time and costs for drug discovery since not all drug candidates but only the actives candidates would be further tested in the wet labs. In this paper, various binary classifier models have been developed to predict the activeness or inactiveness of compounds (drugs) from ChEMBL database against 11 activity classes (targets). The prediction methods used in this study are Support Vector Machine (SVM), Naive Bayes (NB), Feed-forward Sequential Model (SEQ) and Convolutional Neural Networks (CNN). The molecular descriptors used to convert the SMILES chemical notation into binary fingerprint are Extended Connectivity Fingerprints 4 (ECFP4) and Maccs Key (MACCS). The results shows that the 11 predictive models have potential to be applied in real-world situation to predict the interaction of compounds against 11 target proteins as it is found that most machine learning and deep learning methods works better with data represented in ECFP4 compared to MACCS molecular descriptor.

In the chemical industry, the generation of novel molecular structures with beneficial pharmacological and physicochemical properties in de novo molecular design is a critical problem. The advent of deep learning and neural generative models has recently enabled significant achievements in constructing molecular design models in de novo design. Consequently, studies on new generative models continue to generate molecules that exhibit more useful chemical properties. In this study, we propose a method for de novo design that utilizes generative adversarial networks based on reinforcement learning for realistic molecule generation. This method learns to reproduce the training data distribution of simplified molecular-input line-entry system strings. The proposed method is demonstrated to effectively generate novel molecular structures from five benchmark results using a real-world public dataset, ChEMBL. The code is available at https://github.com/dudwojae/SMILES-MaskGAN.

Malaria is a disease affecting hundreds of millions of people across the world, mainly in developing countries and especially in sub-Saharan Africa. It is the cause of hundreds of thousands of deaths each year and there is an ever-present need to identify and develop effective new therapies to tackle the disease and overcome increasing drug resistance. Here, we extend a previous study in which a number of partners collaborated to develop a consensus in silico model that can be used to identify novel molecules that may have antimalarial properties. The performance of machine learning methods generally improves with the number of data points available for training. One practical challenge in building large training sets is that the data are often proprietary and cannot be straightforwardly integrated. Here, this was addressed by sharing QSAR models, each built on a private data set. We describe the development of an open-source software platform for creating such models, a comprehensive evaluation of methods to create a single consensus model and a web platform called MAIP available at https://www.ebi.ac.uk/chembl/maip/. MAIP is freely available for the wider community to make large-scale predictions of potential malaria inhibiting compounds. This project also highlights some of the practical challenges in reproducing published computational methods and the opportunities that open-source software can offer to the community.

Machine learning scoring functions for protein-ligand binding affinity have been found to consistently outperform classical scoring functions when trained and tested on crystal structures of bound protein-ligand complexes. However, it is less clear how these methods perform when applied to docked poses of complexes. We explore how the use of docked rather than crystallographic poses for both training and testing affects the performance of machine learning scoring functions. Using the PDBbind Core Sets as benchmarks, we show that the performance of a structure-based machine learning scoring function trained and tested on docked poses is lower than that of the same scoring function trained and tested on crystallographic poses. We construct a hybrid scoring function by combining both structure-based and ligand-based features, and show that its ability to predict binding affinity using docked poses is comparable to that of purely structure-based scoring functions trained and tested on crystal poses. We also present a new, freely available validation set-the Updated DUD-E Diverse Subset-for binding affinity prediction using data from DUD-E and ChEMBL. Despite strong performance on docked poses of the PDBbind Core Sets, we find that our hybrid scoring function sometimes generalizes poorly to a protein target not represented in the training set, demonstrating the need for improved scoring functions and additional validation benchmarks.

Molecular property prediction is an essential task in drug discovery. Most computational approaches with deep learning techniques either focus on designing novel molecular representation or combining with some advanced models together. However, researchers pay fewer attention to the potential benefits in massive unlabeled molecular data (e.g., ZINC). This task becomes increasingly challenging owing to the limitation of the scale of labeled data. Motivated by the recent advancements of pretrained models in natural language processing, the drug molecule can be naturally viewed as language to some extent. In this paper, we investigate how to develop the pretrained model BERT to extract useful molecular substructure information for molecular property prediction. We present a novel end-to-end deep learning framework, named Mol-BERT, that combines an effective molecular representation with pretrained BERT model tailored for molecular property prediction. Specifically, a large-scale prediction BERT model is pretrained to generate the embedding of molecular substructures, by using four million unlabeled drug SMILES (i.e., ZINC 15 and ChEMBL 27). Then, the pretrained BERT model can be fine-tuned on various molecular property prediction tasks. To examine the performance of our proposed Mol-BERT, we conduct several experiments on 4 widely used molecular datasets. In comparison to the traditional and state-of-the-art baselines, the results illustrate that our proposed Mol-BERT can outperform the current sequence-based methods and achieve at least 2% improvement on ROC-AUC score on Tox21, SIDER, and ClinTox dataset.

Currently, a considerable amount of research is being carried out on the development of new drugs. Therefore, it is necessary to improve existing approaches to search for new chemical compounds with pharmacological activity. Such studies have been carried out for several years. During this period, a considerable amount of information is accumulated, which is quite difficult for one or even a group of experienced specialists to process and analyze. With such a volume of data, obtaining a sufficient degree of accuracy of the neural network data model is possible. The article discusses the methodology for conducting preclinical laboratory research. The Preclinical Validation Study method was conducted using the Ocular Disease Intelligent Recognition (ODIR) database of approximately 4,000 color fundus images and specialist diagnostic results. As a result, the accuracy of the constructed model was obtained together with the optimal hyperparameters. The article also considers the analysis of the data source ChEMBL (chemical database of bioactive molecules with drug-like properties), based on which it is planned to develop a method for predicting the presence of the pharmacological activity of a chemical substance. As a result of the study, a frequency prediction of the chemical compounds’ activity concerning five types of targets was developed.

The discovery and development of novel pharmaceuticals is an area of active research mainly due to the large investments required and long payback times. As of 2016, the development of a novel drug candidate required up to $ USD 2.6 billion in investment for only 10% rate of approval by the FDA. To help decreasing the costs associated with the process, a number of in silico approaches have been developed with relatively low success due to limited predicting performance. Here, we introduced a machine learning-based algorithm as an alternative for a more accurate search of new pharmacological candidates, which takes advantage of Recurrent Neural Networks (RNN) for active molecule prediction within large databases. Our approach, termed PharmaNet was implemented here to search for ligands against specific cell receptors within 102 targets of the DUD-E database, which contains 22886 active molecules. PharmaNet comprises three main phases. First, a SMILES representation of the molecule is converted into a raw molecular image. Second, a convolutional encoder processes the data to obtain a fingerprint molecular image that is finally analyzed by a Recurrent Neural Network (RNN). This approach enables precise predictions of the molecules’ target on the basis of the feature extraction, the sequence analysis and the relevant information filtered out throughout the process. Molecule Target prediction is a highly unbalanced detection problem and therefore, we propose that an adequate evaluation metric of performance is the area under the Normalized Average Precision (NAP) curve. PharmaNet largely surpasses the previous state-of-the-art method with 97.7% in the Receiver Operating Characteristic curve (ROC-AUC) and 65.5% in the NAP curve. We obtained a perfect performance for human farnesyl pyrophosphate synthase (FPPS), which is a potential target for antimicrobial and anticancer treatments. We decided to test PharmaNet for activity prediction against FPPS by searching in the CHEMBL data set. We obtained three (3) potential inhibitors that were further validated through both molecular docking and in silico toxicity prediction. Most importantly, one of this candidates, CHEMBL2007613, was predicted as a potential antiviral due to its involvement on the PCDH17 pathway, which has been reported to be related to viral infections.

A key enzyme in human immunodeficiency virus type 1 (HIV-1) life cycle, integrase (IN) aids the integration of viral DNA into the host DNA, which has become an ideal target for the development of anti-HIV drugs. A total of 1785 potential HIV-1 IN inhibitors were collected from the databases of ChEMBL, Binding Database, DrugBank, and PubMed, as well as from 40 references. The database was divided into the training set and test set by random sampling. By exploring the correlation between molecular descriptors and inhibitory activity, it is found that the classification and specific activity data of inhibitors can be more accurately predicted by the combination of molecular descriptors and molecular fingerprints. The calculation of molecular fingerprint descriptor provides the additional substructure information to improve the prediction ability. Based on the training set, two machine learning methods, the recursive partition (RP) and naive Bayes (NB) models, were used to build the classifiers of HIV-1 IN inhibitors. Through the test set verification, the RP technique accurately predicted 82.5% inhibitors and 86.3% noninhibitors. The NB model predicted 88.3% inhibitors and 87.2% noninhibitors with correlation coefficient of 85.2%. The results show that the prediction performance of NB model is slightly better than that of RP, and the key molecular segments are also obtained. Additionally, CoMFA and CoMSIA models with good activity prediction ability both were constructed by exploring the structure-activity relationship, which is helpful for the design and optimization of HIV-1 IN inhibitors.

This paper presents results from a rapid-response industry-academia collaboration for virtual screening of chemical, natural and virtual drug ligands towards identifying potential therapeutics for COVID-19. Compared to resource-intensive traditional approaches of either conducting high- throughput screening in a lab or in-silico molecular dynamics simulations on supercomputers, we have developed an open- source framework that leverages artificial intelligence (AI) to accurately and quickly predict the binding potential of a drug ligand with a target protein. We have trained a novel molecular-highway graph neural network architecture using the entirety of the BindingDB database to predict the probability of a drug ligand binding to a protein target. Our approach achieves a prodigious 98.3% accuracy with its predictions. Through this paper, we disseminate our source code and use the AI model to screen both public (ChEMBL, DrugBank) and proprietary databases. Compared to other AI-based methods, our approach outperforms the state-of-the-art on the following metrics - (i) number of molecules currently undergoing active clinical trials, (ii) number of antiviral drugs correctly identified, (iii) accuracy despite not needing active-site priors, and (iv) ability to screen more compounds in unit time.

Biomedical information mining is increasingly recognized as a promising technique to accelerate drug discovery and development. Especially, integrative approaches which mine data from several (open) data sources have become more attractive with the increasing possibilities to programmatically access data through Application Programming Interfaces (APIs). The use of open data in conjunction with free, platform-independent analytic tools provides the additional advantage of flexibility, re-usability, and transparency. Here, we present a strategy for performing ligand-based in silico drug repurposing with the analytics platform KNIME. We demonstrate the usefulness of the developed workflow on the basis of two different use cases: a rare disease (here: Glucose Transporter Type 1 (GLUT-1) deficiency), and a new disease (here: COVID 19). The workflow includes a targeted download of data through web services, data curation, detection of enriched structural patterns, as well as substructure searches in DrugBank and a recently deposited data set of antiviral drugs provided by Chemical Abstracts Service. Developed workflows, tutorials with detailed step-by-step instructions, and the information gained by the analysis of data for GLUT-1 deficiency syndrome and COVID-19 are made freely available to the scientific community. The provided framework can be reused by researchers for other in silico drug repurposing projects, and it should serve as a valuable teaching resource for conveying integrative data mining strategies.