Search

In order to better foramize it, the notorious inverse-QSAR problem (finding structures of given QSAR-predicted properties) is considered in this paper as a two-step process including (i) finding "seed" descriptor vectors corresponding to user-constrained QSAR model output values and (ii) identifying the chemical structures best matching the "seed" vectors. The main development effort here was focused on the latter stage, proposing a new attention-based conditional variational autoencoder neural-network architecture based on recent developments in attention-based methods. The obtained results show that this workflow was capable of generating compounds predicted to display desired activity while being completely novel compared to the training database (ChEMBL). Moreover, the generated compounds show acceptable druglikeness and synthetic accessibility. Both pharmacophore and docking studies were carried out as "orthogonal"

Nowadays, drug discovery is inevitably intertwined with the usage of large compound collections. Understanding of their chemotype composition and physicochemical property profiles is of the highest importance for successful hit identification. Efficient polyfunctional tools allowing multifaceted analysis of constantly growing chemical libraries must be Big Data-compatible. Here, we present the freely accessible ChemSpace Atlas (https://chematlas.chimie.unistra.fr), which includes almost 40K hierarchically organized Generative Topographic Maps (GTM) accommodating up to 500 M compounds covering fragment-like, lead-like, drug-like, PPI-like, and NP-like chemical subspaces. They allow users to navigate and analyze ZINC, ChEMBL, and COCONUT from multiple perspectives on different scales: from a bird's eye view of the entire library to structural pattern detection in small clusters. Around 20 physicochemical properties and almost 750 biological activities can be visualized (associated with map zones), supporting activity profiling and analogue search. Moreover, ChemScape Atlas will be extended toward new chemical subspaces (e.g., DNA-encoded libraries and synthons) and functionalities (ADMETox profiling and property-guided de novo compound generation).

This work introduces CGRdb2.0─an open-source database management system for molecules, reactions, and chemical data. CGRdb2.0 is a Python package connecting to a PostgreSQL database that enables native searches for molecules and reactions without complicated SQL syntax. The library provides out-of-the-box implementations for similarity and substructure searches for molecules, as well as similarity and substructure searches for reactions in two ways─based on reaction components and based on the Condensed Graph of Reaction approach, the latter significantly accelerating the performance. In benchmarking studies with the RDKit database cartridge, we demonstrate that CGRdb2.0 performs searches faster for smaller data sets, while allowing for interactive access to the retrieved data.

Graph-based architectures are becoming increasingly popular as a tool for structure generation. Here, we introduce novel open-source architecture HyFactor in which, similar to the InChI linear notation, the number of hydrogens attached to the heavy atoms was considered instead of the bond types. HyFactor was benchmarked on the ZINC 250K, MOSES, and ChEMBL data sets against conventional graph-based architecture ReFactor, representing our implementation of the reported DEFactor architecture in the literature. On average, HyFactor models contain some 20% less fitting parameters than those of ReFactor. The two architectures display similar validity, uniqueness, and reconstruction rates. Compared to the training set compounds, HyFactor generates more similar structures than ReFactor. This could be explained by the fact that the latter generates many open-chain analogues of cyclic structures in the training set. It has been demonstrated that the reconstruction error of heavy molecules can be significantly reduced using the data augmentation technique. The codes of HyFactor and ReFactor as well as all models obtained in this study are publicly available from our GitHub repository: https://github.com/Laboratoire-de-Chemoinformatique/HyFactor.

International audience; The days when medicinal chemistry was limited to a few series of compounds of therapeutic interest are long gone. Nowadays, no human may succeed to acquire a complete overview of more than a billion existing or feasible compounds within which the potential “blockbuster drugs” are well hidden and yet only a few mouse clicks away. To reach these “hidden treasures”, we adapted the generative topographic mapping method to enable efficient navigation through the chemical space, from a global overview to a structural pattern detection, covering, for the first time, the complete ZINC library of purchasable compounds, relative to 1.6 million biologically relevant ChEMBL molecules. About 40 000 hierarchical maps of the chemical space were constructed. Structural motifs inherent to only one library were identified. Roughly 20 000 off-market ChEMBL compound families represent incentives to enrich commercial catalogs. Alternatively, 125 000 ZINC-specific compound classes, absent in structure–activity bases, are novel paths to explore in medicinal chemistry. The complete list of these chemotypes can be downloaded using the link https://forms.gle/B6bUJj82t9EfmttV6.

International audience; The Generalized Born (GB) solvent model is offering the best accuracy/computing effort ratio yet requires drastic simplifications to estimate of the Effective Born Radii (EBR) in bypassing a too expensive volume integration step. EBRs are a measure of the degree of burial of an atom and not very sensitive to small changes of geometry: in molecular dynamics, the costly EBR update procedure is not mandatory at every step. This work however aims at implementing a GB model into the Sampler for Multiple Protein−Ligand Entities (S4MPLE) evolutionary algorithm with mandatory EBR updates at each step triggering arbitrarily large geometric changes. Therefore, a quantitative structure−property relationship has been developed in order to express the EBRs as a linear function of both the topological neighborhood and geometric occupancy of the space around atoms. A training set of 810 molecular systems, starting from fragment-like to drug-like compounds, proteins, host−guest systems, and ligand−protein complexes, has been compiled. For each species, S4MPLE generated several hundreds of random conformers. For each atom in each geometry of each species, its "standard" EBR was calculated by numeric integration and associated to topological and geometric descriptors of the atom neighborhood. This training set (EBR, atom descriptors) involving >5 M entries was subjected to a boot-strapping multilinear regression process with descriptor selection. In parallel, the strategy was repurposed to also learn atomic solvent-accessible areas (SA) based on the same descriptors. Resulting linear equations were challenged to predict EBR and SA values for a similarly compiled external set of >2000 new molecular systems. Solvation energies calculated with estimated EBR and SA match "standard" energies within the typical error of a force-field-based approach (a few kilocalories per mole). Given the extreme diversity of molecular systems covered by the model, this simple EBR/SA estimator covers a vast applicability domain.

The ability to efficiently synthesize desired compounds can be a limiting factor for chemical space exploration in drug discovery. This ability is conditioned not only by the existence of well-studied synthetic protocols but also by the availability of corresponding reagents, so-called building blocks (BB). In this work, we present a detailed analysis of the chemical space of 400K purchasable BB. The chemical space was defined by corresponding synthons – fragments contributed to the final molecules upon reaction. They allow an analysis of BB physicochemical properties and diversity, unbiased by the leaving and protective groups in actual reagents. The main classes of BB were analyzed in terms of their availability, rule-of-two-defined quality, and diversity. Available BBs were eventually compared to a reference set of biologically relevant synthons derived from ChEMBL fragmentation, in order to illustrate how well they cover the actual medicinal chemistry needs. This was performed on a newly constructed universal generative topographic map of synthon chemical space, allowing to visualize both libraries and analyze their overlapping and library-specific regions.

Modern QSAR approaches have wide practical applications in drug discovery for designing potentially bioactive molecules. If such models are based on the use of 2D descriptors, important information contained in the spatial structures of molecules is lost. The major problem in constructing models using 3D descriptors is the choice of a putative bioactive conformation, which affects the predictive performance. The multi-instance (MI) learning approach considering multiple conformations in model training could be a reasonable solution to the above problem. In this study, we implemented several multi-instance algorithms, both conventional and based on deep learning, and investigated their performance. We compared the performance of MI-QSAR models with those based on the classical single-instance QSAR (SI-QSAR) approach in which each molecule is encoded by either 2D descriptors computed for the corresponding molecular graph or 3D descriptors issued for a single lowest energy conformation. The calculations were carried out on 175 data sets extracted from the ChEMBL23 database. It is demonstrated that (i) MI-QSAR outperforms SI-QSAR in numerous cases and (ii) MI algorithms can automatically identify plausible bioactive conformations.

CGRtools is an open-source Python library aimed to handle molecular and reaction information. It is the sole library developed so far which can process condensed graph of reaction (CGR) handling. CGR provides the possibility for advanced operations with reaction information and could be used for reaction descriptor calculation, structure-reactivity modeling, atom-to-atom mapping comparison and correction, reaction center extraction, reaction balancing, and some other related tasks. Unlike other popular libraries, CGRtools is fully written in Python with minor dependencies on other libraries and cross-platform. Reaction, molecule, and CGR objects in CGRtools support native Python methods and are comparable with the help of operations "equal to", "less than", and "bigger than". CGRtools supports common structural formats. CGRtools is distributed via an L-GPL license and available on https://github.com/cimm-kzn/CGRtools .

In chemography, grid-based maps sample molecular descriptor space by injecting a set of nodes, and then linking them to some regular 2D grid representing the map. They include self-organizing maps (SOMs) and generative topographic maps (GTMs). Grid-based maps are predictive because any compound thereupon projected can "inherit" the properties of its residence node(s)-node properties themselves "inherited" from node-neighboring training set compounds. This Article proposes a formalism to define the trustworthiness of these nodes as "providers" of structure-activity information captured from training compounds. An empirical four-parameter node trustworthiness (NT) function of density (sparsely populated nodes are less trustworthy) and coherence (nodes with training set residents of divergent properties are less trustworthy) is proposed. Based upon it, a trustworthiness score T is used to delimit the applicability domain (AD) by means of a trustworthiness threshold TT. For each parameter setup, success of ensuing inside-AD predictions is monitored. It is seen that setup-specific success levels (averaged over large pools of prediction challenges) are highly covariant, irrespectively of the targets of prediction challenges, of the (classification or regression) type of problems, of the specific parametrization, and even of the nature (GTM or SOM) of underlying maps. Thus, success levels determined on the basis of regression problems (445 target-specific affinity QSAR sets) on GTMs and levels returned by completely unrelated classification problems (319 target-specific active-/inactive-labeled sets) on SOMs were seen to correlate to a degree of 70%. Therefore, a common, general-purpose setup of the herein proposed parametric AD definition was shown to generally apply to grid-based map-driven property prediction problems.

Here, we describe a concept of conjugated models for several properties (activities) linked by a strict mathematical relationship. This relationship can be directly integrated analytically into the ridge regression (RR) algorithm or accounted for in a special case of "twin" neural networks (NN). Developed approaches were applied to the modeling of the logarithm of the prototropic tautomeric constant (logK

In this paper, we demonstrate that the kernel target alignment (KTA) parameter can efficiently be used to estimate the relevance of molecular descriptors for QSAR modeling on a given data set, i.e., as a modelability measure. The efficiency of KTA to assess modelability was demonstrated in two series of QSAR modeling studies, either varying different descriptor spaces for one same data set, or comparing various data sets within one same descriptor space. Considered data sets included 25 series of various GPCR binders with ChEMBL-reported pKi values, and a toxicity data set. Employed descriptor spaces covered more than 100 different ISIDA fragment descriptor types, and ChemAxon BCUT terms. Model performances (RMSE) were seen to anticorrelate consistently with the KTA parameter. Two other modelability measures were employed for benchmarking purposes: the Jaccard distance average over the data set (Div), and a measure related to the normalized mean absolute error (MAE) obtained in 1-nearest neighbors calculations on the training set (Sim = 1 - MAE). It has been demonstrated that both Div and Sim perform similarly to KTA. However, a consensus index combining KTA, Div and Sim provides a more robust correlation with RMSE than any of the individual modelability measures.

Predicting the activity profile of a molecule or discovering structures possessing a specific activity profile are two important goals in chemoinformatics, which could be achieved by bridging activity and molecular descriptor spaces. In this paper, we introduce the "Stargate" version of the Generative Topographic Mapping approach (S-GTM) in which two different multidimensional spaces (e.g., structural descriptor space and activity space) are linked through a common 2D latent space. In the S-GTM algorithm, the manifolds are trained simultaneously in two initial spaces using the probabilities in the 2D latent space calculated as a weighted geometric mean of probability distributions in both spaces. S-GTM has the following interesting features: (1) activities are involved during the training procedure; therefore, the method is supervised, unlike conventional GTM; (2) using molecular descriptors of a given compound as input, the model predicts a whole activity profile, and (3) using an activity profile as input, areas populated by relevant chemical structures can be detected. To assess the performance of S-GTM prediction models, a descriptor space (ISIDA descriptors) of a set of 1325 GPCR ligands was related to a B-dimensional (B = 1 or 8) activity space corresponding to pKi values for eight different targets. S-GTM outperforms conventional GTM for individual activities and performs similarly to the Lasso multitask learning algorithm, although it is still slightly less accurate than the Random Forest method.

This paper is devoted to the analysis and visualization in 2-dimensional space of large data sets of millions of compounds using the incremental version of generative topographic mapping (iGTM). The iGTM algorithm implemented in the in-house ISIDA-GTM program was applied to a database of more than 2 million compounds combining data sets of 36 chemicals suppliers and the NCI collection, encoded either by MOE descriptors or by MACCS keys. Taking advantage of the probabilistic nature of GTM, several approaches to data analysis were proposed. The chemical space coverage was evaluated using the normalized Shannon entropy. Different views of the data (property landscapes) were obtained by mapping various physical and chemical properties (molecular weight, aqueous solubility, LogP, etc.) onto the iGTM map. The superposition of these views helped to identify the regions in the chemical space populated by compounds with desirable physicochemical profiles and the suppliers providing them. The data sets similarity in the latent space was assessed by applying several metrics (Euclidean distance, Tanimoto and Bhattacharyya coefficients) to data probability distributions based on cumulated responsibility vectors. As a complementary approach, data sets were compared by considering them as individual objects on a meta-GTM map, built on cumulated responsibility vectors or property landscapes produced with iGTM. We believe that the iGTM methodology described in this article represents a fast and reliable way to analyze and visualize large chemical databases.

This study is an exhaustive analysis of the neighborhood behavior over a large coherent data set (ChEMBL target/ligand pairs of known Ki, for 165 targets with >50 associated ligands each). It focuses on similarity-based virtual screening (SVS) success defined by the ascertained optimality index. This is a weighted compromise between purity and retrieval rate of active hits in the neighborhood of an active query. One key issue addressed here is the impact of Tversky asymmetric weighing of query vs candidate features (represented as integer-value ISIDA colored fragment/pharmacophore triplet count descriptor vectors). The nearly a 3/4 million independent SVS runs showed that Tversky scores with a strong bias in favor of query-specific features are, by far, the most successful and the least failure-prone out of a set of nine other dissimilarity scores. These include classical Tanimoto, which failed to defend its privileged status in practical SVS applications. Tversky performance is not significantly conditioned by tuning of its bias parameter α. Both initial "guesses" of α = 0.9 and 0.7 were more successful than Tanimoto (at its turn, better than Euclid). Tversky was eventually tested in exhaustive similarity searching within the library of 1.6 M commercial + bioactive molecules at http://infochim.u-strasbg.fr/webserv/VSEngine.html , comparing favorably to Tanimoto in terms of "scaffold hopping" propensity. Therefore, it should be used at least as often as, perhaps in parallel to Tanimoto in SVS. Analysis with respect to query subclasses highlighted relationships of query complexity (simply expressed in terms of pharmacophore pattern counts) and/or target nature vs SVS success likelihood. SVS using more complex queries are more robust with respect to the choice of their operational premises (descriptors, metric). Yet, they are best handled by "pro-query" Tversky scores at α > 0.5. Among simpler queries, one may distinguish between "growable" (allowing for active analogs with additional features), and a few "conservative" queries not allowing any growth. These (typically bioactive amine transporter ligands) form the specific application domain of "pro-candidate" biased Tversky scores at α < 0.5.

Machine learning (SVM and JRip rule learner) methods have been used in conjunction with the Condensed Graph of Reaction (CGR) approach to identify errors in the atom-to-atom mapping of chemical reactions produced by an automated mapping tool by ChemAxon. The modeling has been performed on the three first enzymatic classes of metabolic reactions from the KEGG database. Each reaction has been converted into a CGR representing a pseudomolecule with conventional (single, double, aromatic, etc.) bonds and dynamic bonds characterizing chemical transformations. The ChemAxon tool was used to automatically detect the matching atom pairs in reagents and products. These automated mappings were analyzed by the human expert and classified as "correct" or "wrong". ISIDA fragment descriptors generated for CGRs for both correct and wrong mappings were used as attributes in machine learning. The learned models have been validated in n-fold cross-validation on the training set followed by a challenge to detect correct and wrong mappings within an external test set of reactions, never used for learning. Results show that both SVM and JRip models detect most of the wrongly mapped reactions. We believe that this approach could be used to identify erroneous atom-to-atom mapping performed by any automated algorithm.

A benchmark of several popular methods, Associative Neural Networks (ANN), Support Vector Machines (SVM), k Nearest Neighbors (kNN), Maximal Margin Linear Programming (MMLP), Radial Basis Function Neural Network (RBFNN), and Multiple Linear Regression (MLR), is reported for quantitative-structure property relationships (QSPR) of stability constants logK1 for the 1:1 (M:L) and logbeta2 for 1:2 complexes of metal cations Ag+ and Eu3+ with diverse sets of organic molecules in water at 298 K and ionic strength 0.1 M. The methods were tested on three types of descriptors: molecular descriptors including E-state values, counts of atoms determined for E-state atom types, and substructural molecular fragments (SMF). Comparison of the models was performed using a 5-fold external cross-validation procedure. Robust statistical tests (bootstrap and Kolmogorov-Smirnov statistics) were employed to evaluate the significance of calculated models. The Wilcoxon signed-rank test was used to compare the performance of methods. Individual structure-complexation property models obtained with nonlinear methods demonstrated a significantly better performance than the models built using multilinear regression analysis (MLRA). However, the averaging of several MLRA models based on SMF descriptors provided as good of a prediction as the most efficient nonlinear techniques. Support Vector Machines and Associative Neural Networks contributed in the largest number of significant models. Models based on fragments (SMF descriptors and E-state counts) had higher prediction ability than those based on E-state indices. The use of SMF descriptors and E-state counts provided similar results, whereas E-state indices lead to less significant models. The current study illustrates the difficulties of quantitative comparison of different methods: conclusions based only on one data set without appropriate statistical tests could be wrong.

The estimation of accuracy and applicability of QSAR and QSPR models for biological and physicochemical properties represents a critical problem. The developed parameter of "distance to model" (DM) is defined as a metric of similarity between the training and test set compounds that have been subjected to QSAR/QSPR modeling. In our previous work, we demonstrated the utility and optimal performance of DM metrics that have been based on the standard deviation within an ensemble of QSAR models. The current study applies such analysis to 30 QSAR models for the Ames mutagenicity data set that were previously reported within the 2009 QSAR challenge. We demonstrate that the DMs based on an ensemble (consensus) model provide systematically better performance than other DMs. The presented approach identifies 30-60% of compounds having an accuracy of prediction similar to the interlaboratory accuracy of the Ames test, which is estimated to be 90%. Thus, the in silico predictions can be used to halve the cost of experimental measurements by providing a similar prediction accuracy. The developed model has been made publicly available at http://ochem.eu/models/1 .

The estimation of the accuracy of predictions is a critical problem in QSAR modeling. The "distance to model" can be defined as a metric that defines the similarity between the training set molecules and the test set compound for the given property in the context of a specific model. It could be expressed in many different ways, e.g., using Tanimoto coefficient, leverage, correlation in space of models, etc. In this paper we have used mixtures of Gaussian distributions as well as statistical tests to evaluate six types of distances to models with respect to their ability to discriminate compounds with small and large prediction errors. The analysis was performed for twelve QSAR models of aqueous toxicity against T. pyriformis obtained with different machine-learning methods and various types of descriptors. The distances to model based on standard deviation of predicted toxicity calculated from the ensemble of models afforded the best results. This distance also successfully discriminated molecules with low and large prediction errors for a mechanism-based model developed using log P and the Maximum Acceptor Superdelocalizability descriptors. Thus, the distance to model metric could also be used to augment mechanistic QSAR models by estimating their prediction errors. Moreover, the accuracy of prediction is mainly determined by the training set data distribution in the chemistry and activity spaces but not by QSAR approaches used to develop the models. We have shown that incorrect validation of a model may result in the wrong estimation of its performance and suggested how this problem could be circumvented. The toxicity of 3182 and 48774 molecules from the EPA High Production Volume (HPV) Challenge Program and EINECS (European chemical Substances Information System), respectively, was predicted, and the accuracy of prediction was estimated. The developed models are available online at http://www.qspr.org site. © 2008 American Chemical Society.

Earlier (Kireeva et al. Mol. Inf. 2012, 31, 301–312), we demonstrated that generative topographic mapping (GTM) can be efficiently used both for data visualization and building of classification models in the initial D-dimensional space of molecular descriptors. Here, we describe the modeling in two-dimensional latent space for the four classes of the BioPharmaceutics Drug Disposition Classification System (BDDCS) involving VolSurf descriptors. Three new definitions of the applicability domain (AD) of models have been suggested: one class-independent AD which considers the GTM likelihood and two class-dependent ADs considering respectively, either the predominant class in a given node of the map or informational entropy. The class entropy AD was found to be the most efficient for the BDDCS modeling. The predominant class AD can be directly visualized on GTM maps, which helps the interpretation of the model.

We herewith present a novel approach to predict protein–ligand binding modes from the single two-dimensional structure of the ligand. Known protein–ligand X-ray structures were converted into binary bit strings encoding protein–ligand interactions. An artificial neural network was then set up to first learn and then predict protein–ligand interaction fingerprints from simple ligand descriptors. Specific models were constructed for three targets (CDK2, p38-α, HSP90-α) and 146 ligands for which protein–ligand X-ray structures are available. These models were able to predict protein–ligand interaction fingerprints and to discriminate important features from minor interactions. Predicted interaction fingerprints were successfully used as descriptors to discriminate true ligands from decoys by virtual screening. In some but not all cases, the predicted interaction fingerprints furthermore enable to efficiently rerank cross-docking poses and prioritize the best possible docking solutions.