Several binary molecular fingerprints were compressed using an autoencoder neural network. We analyzed the impact of compression on fingerprint performance in downstream classification and regression tasks. Classifiers trained on compressed fingerprints were negligibly affected. Regression models benefitted from compression, especially of long fingerprints (Morgan, RDK). However, their performance dropped rapidly for compression levels exceeding 90 %. Property co-learning positively influenced the predictive power of the compressed fingerprints, with a mean score improvement up to 20 %, suggesting that autoencoder compression with property co-learning biases the molecular representation toward the predicted target, facilitating downstream training.

Dimensionality reduction (DR) techniques are used for various purposes such as exploratory data analysis. A commonly employed linear DR technique is principal component analysis (PCA), which is one of the most popular methods for DR. Owing to its linear nature, PCA enables the determination of axes in a low-dimensional space and the calculation of corresponding loading vectors. However, PCA cannot necessarily extract important features of non-linearly distributed data. This study presents a technique aimed at aiding the interpretation of data reduced through non-linear DR methods. In the proposed method, non-linear dimensionally reduced data was clustered via a density-based clustering method. Thereafter, the obtained cluster labels were classified by random forest (RF) classifiers. Further, feature importance (FI) of RF classifiers and Spearman's rank correlation coefficients between predictive probabilities to obtained clusters and original feature values were utilized for characterizing the visualized dimensionally reduced data. The results revealed that the proposed method can provide the interpretable FI-based images of the handwritten digits dataset. Moreover, the proposed method was also applied to the polymer dataset. The study found that incorporating signed FI was advantageous in achieving a meaningful interpretation. Furthermore, Gaussian process regression was utilized to produce intuitive FI-based heatmaps on a 2-dimensional space for greater ease of understanding. Additionally, to enhance the interpretability of the obtained clusters, a feature selection technique called Boruta was applied. The Boruta feature selection method worked effectively to interpret the obtained clusters with limited and commonly important features. Additionally, the study suggested that computing FI solely from substructure-based descriptors could further enhance the interpretability of the results. Finally, the automation of the proposed method was investigated, and through maximizing the target score based on the quality of both the DR and clustering, indicative results were automatically obtained for both the handwritten digits and polymer datasets.

Dissemination of novel research methods, especially in the form of chemoinformatics software, depends heavily on their ease of applicability for non-expert users with only a little or no programming skills and knowledge in computer science. Visual programming has become widely popular over the last few years, also enabling researchers without in-depth programming skills to develop tailored data processing pipelines using elements from a repository of predefined standard procedures. In this work, we present the development of a set of nodes for the KNIME platform implementing the QPhAR algorithm. We show how the developed KNIME nodes can be included in a typical workflow for biological activity prediction. Furthermore, we present best-practice guidelines that should be followed to obtain high-quality QPhAR models. Finally, we show a typical workflow to train and optimise a QPhAR model in KNIME for a set of given input compounds, applying the discussed best practices.

Understanding structure-activity landscapes is essential in drug discovery. Similarly, it has been shown that the presence of activity cliffs in compound data sets can have a substantial impact not only on the design progress but also can influence the predictive ability of machine learning models. With the continued expansion of the chemical space and the currently available large and ultra-large libraries, it is imperative to implement efficient tools to analyze the activity landscape of compound data sets rapidly. The goal of this study is to show the applicability of the n-ary indices to quantify the structure-activity landscapes of large compound data sets using different types of structural representation rapidly and efficiently. We also discuss how a recently introduced medoid algorithm provides the foundation to finding optimum correlations between similarity measures and structure-activity rankings. The applicability of the n-ary indices and the medoid algorithm is shown by analyzing the activity landscape of 10 compound data sets with pharmaceutical relevance using three fingerprints of different designs, 16 extended similarity indices, and 11 coincidence thresholds.

The present work was devoted to explore the quantitative structure-property relationships for gas-to-ionic liquid partition coefficients (log KILA ). A series of linear models were first established for the representative dataset (IL01). The optimal model was a four-parameter equation (1Ed) consisting of two electrostatic potential-based descriptors ( ΣVs,ind- and Vs,max ), one 2D matrix-based descriptor (J_D/Dt) and dipole moment (μ). All of the four descriptors introduced in the model can find the corresponding parameters, directly or indirectly, from Abraham's linear solvation energy relationship (LSER) or its theoretical alternatives, which endows the model good interpretability. Gaussian process was utilized to build the nonlinear model. Systematical validations, including 5-fold cross-validation for the training set, the validation for test set, as well as a more rigorous Monte Carlo cross-validation were performed to verify the reliability of the constructed models. Applicability domain of the model was evaluated, and the Williams plot revealed that the model can be used to predict the log KILA values of structurally diverse solutes. The other 13 datasets were also processed in the same way, and all of the linear models with expressions similar to equation 1Ed were obtained. These models, whether linear of nonlinear, represent satisfactory statistical results, which confirms the universality of the method adopted in this study in QSPR modeling of gas-to-IL partition.

The paper presents the VEGA Online web service, which includes a set of freely available tools deriving from the development of the VEGA suite of programs. In detail, the paper is focused on two tools: the VEGA Web Edition (WE) and the Score tool. The former is a versatile file format converter including relevant features for 2D/3D conversion, for surface mapping and for editing/preparing input files. The Score application allows rescoring docking poses and in particular includes the MLP Interactions Scores (MLPInS) for describing hydrophobic interactions. To the best of our knowledge, this web service is the only available resource by which one can calculate both the virtual log P of a given input molecule according to the MLP approach plus the corresponding MLP surface.

Computational de novo drug design is a challenging issue in medicine, and it is desirable to consider all of the relevant information of the biological systems in a disease state. Here, we propose a novel computational method to generate drug candidate molecular structures from patient gene expression profiles via deep learning, which we call DRAGONET. Our model can generate new molecules that are likely to counteract disease-specific gene expression patterns in patients, which is made possible by exploring the latent space constructed by a transformer-based variational autoencoder and integrating the substructures of disease-correlated molecules. We applied DRAGONET to generate drug candidate molecules for gastric cancer, atopic dermatitis, and Alzheimer's disease, and demonstrated that the newly generated molecules were chemically similar to registered drugs for each disease. This approach is applicable to diseases with unknown therapeutic target proteins and will make a significant contribution to the field of precision medicine.

The new high selective mAChRs M3 inhibitors with IC50 in nanomolecular ranges, which can be the prototypes for effective COPD and asthma treatment drugs, were discovered with computational approaches among trifluoromethyl containing hexahydropyrimidinones/thiones. Compounds [6-(4-ethoxy-3-methoxy-phenyl)-4-hydroxy-2-thioxo-4-(trifluoromethyl)hexahydropyrimidin-5-yl]-phenyl-methanone (THPT-1) and 5-benzoyl-6-(3,4-dimethoxyphenyl)-4-hydroxy-4-(trifluoromethyl)hexahydropyrimidin-2-one (THPO-4) have been proved to be a highly effective (with IC50 values of 1.62 ⋅ 10-7  M and 3.09 ⋅ 10-9  M, respectively) at the same concentrations significantly competitive inhibit the signal conduction through mAChR3 in comparison with ipratropium bromide, without significant effect on mAChR2, nicotinic cholinergic and adrenergic receptors.

Androgen receptor (AR) inhibition remains the primary strategy to combat the progression of prostate cancer (PC). However, all clinically used AR inhibitors target the ligand-binding domain (LBD), which is highly susceptible to truncations through splicing or mutations that confer drug resistance. Thus, there exists an urgent need for AR inhibitors with novel modes of action. We thus launched a virtual screening of an ultra-large chemical library to find novel inhibitors of the AR DNA-binding domain (DBD) at two sites: protein-DNA interface (P-box) and dimerization site (D-box). The compounds selected through vigorous computational filtering were then experimentally validated. We identified several novel chemotypes that effectively suppress transcriptional activity of AR and its splice variant V7. The identified compounds represent previously unexplored chemical scaffolds with a mechanism of action that evades the conventional drug resistance manifested through LBD mutations. Additionally, we describe the binding features required to inhibit AR DBD at both P-box and D-box target sites.

PD-1/PD-L1 is a critical druggable target for immunotherapy against sepsis. Chemoinformatics techniques involved the structure-based 3D pharmacophore model development followed by virtual screening of small molecule databases to identify the small molecules against PD-L1 pathway inhibition. Raltitrexed and Safinamide act as potent repurposed drugs, and three other Specs database compounds using in silico methods. These compounds were screened based on the pharmacophore fit score and binding affinity towards the active site of the PD-L1 protein. In silico pharmacokinetic profiling of these screened compounds was done to test their biological activity. Next, experimental validation of the best four virtually screened hits was done in vitro for its hemocompatibility and cytotoxicity. Among these, Raltitrexed, Safinamide and Specs compound (AK-968/40642641) effectively increased the proliferation of immune cells and IFN-γ production. These compounds can act as potent PDL-1 inhibitors for adjuvant therapy against sepsis.

Drug-induced liver injury (DILI) is one of the major causes of drug withdrawals, acute liver injury and blackbox warnings. Clinical diagnosis of DILI is a huge challenge due to the complex pathogenesis and lack of specific biomarkers. In recent years, machine learning methods have been used for DILI risk assessment, but the model generalization does not perform satisfactorily. In this study, we constructed a large DILI data set and proposed an integration strategy based on hybrid representations for DILI prediction (HR-DILI). Benefited from feature integration, the hybrid graph neural network models outperformed single representation-based models, among which hybrid-GraphSAGE showed balanced performance in cross-validation with AUC (area under the curve) as 0.804±0.019. In the external validation set, HR-DILI improved the AUC by 6.4 %-35.9 % compared to the base model with a single representation. Compared with published DILI prediction models, HR-DILI had better and balanced performance. The performance of local models for natural products and synthetic compounds were also explored. Furthermore, eight key descriptors and six structural alerts associated with DILI were analyzed to increase the interpretability of the models. The improved performance of HR-DILI indicated that it would provide reliable guidance for DILI risk assessment.

Asthma and COPD are characterized by complex pathophysiology associated with chronic inflammation, bronchoconstriction, and bronchial hyperresponsiveness resulting in airway remodeling. A possible comprehensive solution that could fully counteract the pathological processes of both diseases are rationally designed multi-target-directed ligands (MTDLs), combining PDE4B and PDE8A inhibition with TRPA1 blockade. The aim of the study was to develop AutoML models to search for novel MTDL chemotypes blocking PDE4B, PDE8A, and TRPA1. Regression models were developed for each of the biological targets using "mljar-supervised". On their basis, virtual screenings of commercially available compounds derived from the ZINC15 database were performed. A common group of compounds placed within the top results was selected as potential novel chemotypes of multifunctional ligands. This study represents the first attempt to discover the potential MTDLs inhibiting three biological targets. The obtained results prove the usefulness of AutoML methodology in the identification of hits from the big compound databases.

Dual specificity protein kinase threonine/Tyrosine kinase (TTK) is one of the mitotic kinases. High levels of TTK are detected in several types of cancer. Hence, TTK inhibition is considered a promising therapeutic anti-cancer strategy. In this work, we used multiple docked poses of TTK inhibitors to augment training data for machine learning QSAR modeling. Ligand-Receptor Contacts Fingerprints and docking scoring values were used as descriptor variables. Escalating docking-scoring consensus levels were scanned against orthogonal machine learners, and the best learners (Random Forests and XGBoost) were coupled with genetic algorithm and Shapley additive explanations (SHAP) to determine critical descriptors for predicting anti-TTK bioactivity and for pharmacophore generation. Three successful pharmacophores were deduced and subsequently used for in silico screening against the NCI database. A total of 14 hits were evaluated in vitro for their anti-TTK bioactivities. One hit of novel chemotype showed reasonable dose-response curve with experimental IC50 of 1.0 μM. The presented work indicates the validity of data augmentation using multiple docked poses for building successful machine learning models and pharmacophore hypotheses.

In this study we present deimos, a computational methodology for optimal grouping, applied on the read-across prediction of engineered nanomaterials' (ENMs) toxicity-related properties. The method is based on the formulation and the solution of a mixed-integer optimization program (MILP) problem that automatically and simultaneously performs feature selection, defines the grouping boundaries according to the response variable and develops linear regression models in each group. For each group/region, the characteristic centroid is defined in order to allocate untested ENMs to the groups. The deimos MILP problem is integrated in a broader optimization workflow that selects the best performing methodology between the standard multiple linear regression (MLR), the least absolute shrinkage and selection operator (LASSO) models and the proposed deimos multiple-region model. The performance of the suggested methodology is demonstrated through the application to benchmark ENMs datasets and comparison with other predictive modelling approaches. However, the proposed method can be applied to property prediction of other than ENM chemical entities and it is not limited to ENMs toxicity prediction.

Predicting the likely biological activity (or property) of compounds is a fundamental and challenging task in the drug discovery process. Current computational methodologies aim to improve their predictive accuracies by using deep learning (DL) approaches. However, non-DL based approaches for small- and medium-sized chemical datasets have demonstrated to be most suitable for. In this approach, an initial universe of molecular descriptors (MDs) is first calculated, then different feature selection algorithms are applied, and finally, one or several predictive models are built. Herein we demonstrate that this traditional approach may miss relevant information by assuming that the initial universe of MDs codifies all relevant aspects for the respective learning task. We argue that this limitation is mainly because of the constrained intervals of the parameters used in the algorithms that compute MDs, parameters that define the Descriptor Configuration Space (DCS). We propose to relax these constraints in an open CDS approach, so that a larger universe of MDs can be initially considered. We model the generation of MDs as a multicriteria optimization problem and tackle it with a variant of the standard genetic algorithm. As a novel component, the fitness function is computed by aggregating four criteria via the Choquet integral. Experimental results show that the proposed approach generates a meaningful DCS by improving state-of-the-art approaches in most of the benchmarking chemical datasets accounted for.