Search

In the construction of QSAR models for the prediction of molecular activity, feature selection is a common task aimed at improving the results and understanding of the problem. The selection of features allows elimination of irrelevant and redundant features, reduces the effect of dimensionality problems, and improves the generalization and interpretability of the models. In many feature selection applications, such as those based on ensembles of feature selectors, it is necessary to combine different selection processes. In this work, we evaluate the application of a new feature selection approach to the prediction of molecular activity, based on the construction of an undirected graph to combine base feature selectors. The experimental results demonstrate the efficiency of the graph-based method in terms of the classification performance, reduction, and redundancy compared to the standard voting method. The graph-based method can be extended to different feature selection algorithms and applied to other cheminformatics problems.

In silico metabolism prediction is a cheminformatic task of autonomously predicting the set of metabolic byproducts produced from a specified molecule and a set of enzymes or reactions. Here, we describe a novel machine learned in silico cytochrome P450 (CYP450) metabolism prediction suite, called CyProduct, that accurately predicts metabolic byproducts for a specified molecule and a human CYP450 isoform. It includes three modules: (1) CypReact, a tool that predicts if the query compound reacts with a given CYP450 enzyme, (2) CypBoM, a tool that accurately predicts the "bond site" of the reaction (i.e., which specific bonds within the query molecule react with the CYP isoform), and (3) MetaboGen, a tool that generates the metabolic byproducts based on CypBoM's bond-site prediction. CyProduct predicts metabolic biotransformation products for each of the nine most important human CYP450 enzymes. CypBoM uses an important new concept called "bond of metabolism" (BoM), which extends the traditional "site of metabolism" (SoM) by specifying the information about the set of chemical bonds that is modified or formed in a metabolic reaction (rather than the specific atom). We created a BoM database for 1845 CYP450-mediated Phase I reactions, then used this to train the CypBoM Predictor to predict the reactive bond locations on substrate molecules. CypBoM Predictor's cross-validated Jaccard score for reactive bond prediction ranged from 0.380 to 0.452 over the nine CYP450 enzymes. Over variants of a test set of 68 known CYP450 substrates and 30 nonreactants, CyProduct outperformed the other packages, including ADMET Predictor, BioTransformer, and GLORY, by an average of 200% (with respect to Jaccard score) in terms of predicting metabolites. The CyProduct suite and the data sets are freely available at https://bitbucket.org/wishartlab/cyproduct/src/master/.

During the drug development process, it is common to carry out toxicity tests and adverse effect studies, which are essential to guarantee patient safety and the success of the research. The use of in silico quantitative structure-activity relationship (QSAR) approaches for this task involves processing a huge amount of data that, in many cases, have an imbalanced distribution of active and inactive samples. This is usually termed the class-imbalance problem and may have a significant negative effect on the performance of the learned models. The performance of feature selection (FS) for QSAR models is usually damaged by the class-imbalance nature of the involved datasets. This paper proposes the use of an FS method focused on dealing with the class-imbalance problems. The method is based on the use of FS ensembles constructed by boosting and using two well-known FS methods, fast clustering-based FS and the fast correlation-based filter. The experimental results demonstrate the efficiency of the proposal in terms of the classification performance compared to standard methods. The proposal can be extended to other FS methods and applied to other problems in cheminformatics.

Machine-learned ranking models have been developed for the prediction of substrate-specific cross-coupling reaction conditions. Data sets of published reactions were curated for Suzuki, Negishi, and C-N couplings, as well as Pauson-Khand reactions. String, descriptor, and graph encodings were tested as input representations, and models were trained to predict the set of conditions used in a reaction as a binary vector. Unique reagent dictionaries categorized by expert-crafted reaction roles were constructed for each data set, leading to context-aware predictions. We find that relational graph convolutional networks and gradient-boosting machines are very effective for this learning task, and we disclose a novel reaction-level graph attention operation in the top-performing model.

Prototype or instance selection techniques is an important field of research in knowledge discovery, data mining, and machine learning. In QSAR, the use of prototype selection techniques in the preprocessing stage of the construction of the QSAR models favors the data set curation, improving the interpretability and accuracy of the models as well as the performance of the algorithms. In this paper, we propose an efficient method for prototype selection to be used in the preprocessing stage of the construction of QSAR classification models. The proposed method is able to generate very high reduction rates in the cardinality of the training set while maintaining or even increasing the accuracy of the classification models. The validation of the method has been carried out by means of the prediction of external molecules, demonstrating that the prediction of new molecules is also maintained or even improved. The method has been tested using 40 benchmark data sets of different sizes and balancing ratios; the results of the tests have demonstrated the wide applicability domain of the proposed method.

Dimensionality reduction of the data set representation for the construction of the quantitative structure-activity relationship classification models is an important research subject for the interpretability of the models and the computational cost efficiency of the classification algorithms. Feature selection techniques are appropriate as only a short number of relevant features should be used in the classification process because irrelevant and redundant features should be discarded, the same as the noninterpretable ones. In this paper, we propose an embedded feature selection technique for the construction of classification models using the rivality index neighborhood (RINH) algorithm. This technique uses a filter selection in the preprocessing stage considering the selectivity of the features as a selection criterion and a wrapper technique in the processing stage based on the improvement of the accuracy and reliability of the models generated using the RINH algorithm with LTN and GTN functions. The results obtained using the RINH algorithm with and without the selection of features and compared with those results obtained using 14 machine learning algorithms have demonstrated that the feature selection technique proposed in this paper is capable of clearly building more accurate and reliable models, reducing the data dimensionality around 90%, and generating high robust and interpretable models.

The prediction of adverse drug reactions in the discovery of new medicines is highly challenging. In the task of predicting the adverse reactions of chemical compounds, information about different targets is often available. Although we can focus on every adverse drug reaction prediction separately, multilabel approaches have been proven useful in many research areas for taking advantage of the relationship among the targets. However, when approaching the prediction problem from a multilabel point of view, we have to deal with the lack of information for some labels. This missing labels problem is a relevant issue in the field of cheminformatics approaches. This paper aims to predict the adverse drug reaction of commercial drugs using a multilabel approach where the possible presence of missing labels is also taken into consideration. We propose the use of multilabel methods to deal with the prediction of a large set of 27 different adverse reaction targets. We also propose the use of multilabel methods specifically designed to deal with the missing labels problem to test their ability to solve this difficult problem. The results show the validity of the proposed approach, demonstrating a superior performance of the multilabel method compared with the single-label approach in addressing the problem of adverse drug reaction prediction.

An unambiguous algorithm, added to the study of the applicability domain and appropriate measures of the goodness of fit and robustness, represent the key characteristics that should be ideally fulfilled for a QSAR model to be considered for regulatory purposes. In this paper, we propose a new algorithm (RINH) based on the rivality index for the construction of QSAR classification models. This index is capable of predicting the activity of the data set molecules by means of a measurement of the rivality between their nearest neighbors belonging to different classes, contributing with a robust measurement of the reliability of the predictions. In order to demonstrate the goodness of the proposed algorithm we have selected four independent and orthogonally different benchmark data sets (balanced/unbalanced and high/low modelable) and we have compared the results with those obtained using 12 different machine learning algorithms. These results have been validated using 20 data sets of different balancing and sizes, corroborating that the proposed algorithm is able to generate highly accurate classification models and contribute with valuable measurements of the reliability of the predictions and the applicability domain of the built models.

Prediction of the capability of a data set to be modeled by a statistical algorithm in the development of quantitative structure-activity relationship (QSAR) regression models is an important issue that allows researchers to avoid unnecessary tasks, wasted time, and/or the need to depurate the molecule composition of the data set in order to achieve an improvement of the model's accuracy. In this paper, we propose and formulate a new index that correlates with the performance of QSAR models. This index, the regression modelability index, requires very low computational cost and is based on the rivality between the nearest neighbors of the molecules in the data set. This rivality allows measurement of the capability of each molecule of the data set to be correctly predicted by a regression algorithm. In this study, using 40 data sets with very different characteristics regarding the number of molecules and activity values, we prove the high correlation between the proposed regression modelability index and the correlation coefficient in cross-validation ( Q 2 ), reaching r 2 values of 0.8. In addition, we describe the ability of this index to discover the outliers detected by the regression algorithms, allowing easy data set depuration in the first stages of the construction of QSAR regression models.

The knowledge of the capacity of a data set to be modeled in the first stages of the building of quantitative structure-activity relationship (QSAR) prediction models is an important issue because it might reduce the effort and time necessary to select or reject data sets and in refining the data set's composition. The modelability index (MODI) is based on the counting of the first nearest neighbor belonging to the molecules of the data set and is a standardized measurement assumed in the QSAR community. In this paper, we revisit the calculation of the modelability index, proposing a more formal formulation that extends the calculation to the first nearest neighbors that belong to each existing class in the data set. In addition, this new formulation allows the calculation of the rivality index, as a measurement of the presence of correctly classifiable molecules and activity cliffs. By weighting the rivality index considering the cardinality of the neighborhood of each molecule of the data set, the calculated weighted modelability index is highly correlated with the correct classification rate (QSAR_CCR) obtained in the building of QSAR models using different classification algorithms. The results obtained with the weighted modelability index show correlations of r 2 higher than 0.9, slopes close to 1, and bias close to zero for different algorithms.

In silico metabolism prediction requires first predicting whether a specific molecule will interact with one or more specific metabolizing enzymes, then predicting the result of each enzymatic reaction. Here, we provide a computational tool, CypReact, for performing this first task of reactant prediction. Specifically, CypReact takes as input an arbitrary molecule (specified as a SMILES string or a standard SDF file) and any one of the nine of the most important human cytochrome P450 (CYP450) enzymes-CYP1A2, CYP2A6, CYP2B6, CYP2C8, CYP2C9, CYP2C19, CYP2D6, CYP2E1, or CYP3A4-and accurately predicts whether the query molecule will react with that given CYP450 enzyme. Tests of CypReact, conducted over a data set of 1632 molecules (each considered a "plausible" reactant) show that it is very effective, with a (cross-validation) AUROC (area under the receiver operating characteristic curve) of 0.83-0.92. We also show that CypReact performs significantly better than other reactant prediction tools such as ADMET Predictor and (a reactant-predicting extension of) SMARTCyp, whose average AUROCs are 0.75 and 0.53, respectively. We then applied the learned CypReact models to a previously unseen set of molecules and found that our CypReact did even better and still significantly surpassed the performance of SMARTCyp and ADMET Predictor. These results suggest that CypReact could be an important component of a suite of in silico metabolism prediction tools for accurately predicting the products of Phase I, Phase II, and microbial metabolism in humans. CypReact is available at https://bitbucket.org/Leon_Ti/cypreact .

The appropriate selection of a chemical space represented by the data set, the selection of its chemical data representation, the development of a correct modeling process using a robust and reproducible algorithm, and the performance of an exhaustive training and external validation determine the usability and reproducibility of a quantitative structure-activity relationship (QSAR) classification model. In this paper, we show that the use of relative versus absolute data in the representation of the data sets produces better classification models when the other processes are not modified. Relative data considers a reference frame to measure the chemical characteristics involved in the classification model, refining the data set representation and smoothing the lack of chemical information. Three data sets with different characteristics have been used in this study, and classifications models have been built applying the support vector machine algorithm. For randomly selected training and test sets, values of accuracy and area under the receiver operating characteristic curve close to 100% have been obtained for the generation of the models and external validations in all cases.

The ability to interpret the predictions made by quantitative structure-activity relationships (QSARs) offers a number of advantages. While QSARs built using nonlinear modeling approaches, such as the popular Random Forest algorithm, might sometimes be more predictive than those built using linear modeling approaches, their predictions have been perceived as difficult to interpret. However, a growing number of approaches have been proposed for interpreting nonlinear QSAR models in general and Random Forest in particular. In the current work, we compare the performance of Random Forest to those of two widely used linear modeling approaches: linear Support Vector Machines (SVMs) (or Support Vector Regression (SVR)) and partial least-squares (PLS). We compare their performance in terms of their predictivity as well as the chemical interpretability of the predictions using novel scoring schemes for assessing heat map images of substructural contributions. We critically assess different approaches for interpreting Random Forest models as well as for obtaining predictions from the forest. We assess the models on a large number of widely employed public-domain benchmark data sets corresponding to regression and binary classification problems of relevance to hit identification and toxicology. We conclude that Random Forest typically yields comparable or possibly better predictive performance than the linear modeling approaches and that its predictions may also be interpreted in a chemically and biologically meaningful way. In contrast to earlier work looking at interpretation of nonlinear QSAR models, we directly compare two methodologically distinct approaches for interpreting Random Forest models. The approaches for interpreting Random Forest assessed in our article were implemented using open-source programs that we have made available to the community. These programs are the rfFC package ( https://r-forge.r-project.org/R/?group_id=1725 ) for the R statistical programming language and the Python program HeatMapWrapper [ https://doi.org/10.5281/zenodo.495163 ] for heat map generation.

Isomorphism measures based on the maximum common subgraph (MCS) calculation are widely used in computational chemistry for classifying, screening, and predicting properties and biological activity within chemical databases. The development of a weighted hierarchical structure based on the MCS is described in this paper. Furthermore, a 2D representation model is proposed as the proper tool for the study and preliminary analysis of molecule data sets. The development process of the weighted MCS tree is open to the use of different approaches. By taking into account different molecular descriptors, similarity and distance measures in the weighted MCS tree, the relationships between the molecular property or the activity, and the variables considered for the building and display of the weighted MCS tree can be observed. Besides that, the representation model based on snowflake diagrams allows to display of those relationships as well as shows any existing degeneration, in order to detect any possible outlier that could be obtained during the development of predictive models and to extract new variables that can be used in the building of quantitative structure-activity relationship models.

In chemoinformatics, searching for compounds which are structurally diverse and share a biological activity is called scaffold hopping. Scaffold hopping is important since it can be used to obtain alternative structures when the compound under development has unexpected side-effects. Pharmaceutical companies use scaffold hopping when they wish to circumvent prior patents for targets of interest. We propose a new method for scaffold hopping using inductive logic programming (ILP). ILP uses the observed spatial relationships between pharmacophore types in pretested active and inactive compounds and learns human-readable rules describing the diverse structures of active compounds. The ILP-based scaffold hopping method is compared to two previous algorithms (chemically advanced template search, CATS, and CATS3D) on 10 data sets with diverse scaffolds. The comparison shows that the ILP-based method is significantly better than random selection while the other two algorithms are not. In addition, the ILP-based method retrieves new active scaffolds which were not found by CATS and CATS3D. The results show that the ILP-based method is at least as good as the other methods in this study. ILP produces human-readable rules, which makes it possible to identify the three-dimensional features that lead to scaffold hopping. A minor variant of a rule learnt by ILP for scaffold hopping was subsequently found to cover an inhibitor identified by an independent study. This provides a successful result in a blind trial of the effectiveness of ILP to generate rules for scaffold hopping. We conclude that ILP provides a valuable new approach for scaffold hopping.

In this paper, we present the advantages of using data fusion of similarity and dissimilarity measurements for the development of quantitative structure-activity relationship models. Nonisomorphic fragments extracted in the matching process were considered to obtain dissimilarity values employed for correcting similarity measurements, thus, leading to finer chemical information. The purpose was to correlate similarity and dissimilarity matrices with pharmacological activities of drugs (the inhibitory capacity presented by 30 benzoxazinone derivatives for the NPY Y5 receptor). Wiener and hyper-Wiener descriptors computed over distance and weighted distance matrices were used for the calculation of dissimilarity values. A comparison with classical and fingerprints-based similarity was also carried out. The best approaches were achieved by means of dissimilarity and of fusion data spaces that take into account isomorphic and nonisomorphic information (Q2 = 0.88, SECV = 0.18, slope = 1.06, and intercept = 0.09). The study of anomalous behavior presented by some compounds was also undertaken.

Searching chemical databases for possible drug leads is often one of the main activities conducted during the early stages of a drug development project. This article shows that spherical harmonic molecular shape representations provide a powerful way to search and cluster small-molecule databases rapidly and accurately. Our clustering results show that chemically meaningful clusters may be obtained using only low order spherical harmonic expansions. Our database search results show that using low order spherical harmonic shape-based correlation techniques could provide a practical and efficient way to search very large 3D molecular databases, hence leading to a useful new approach for high throughput 3D virtual screening. The approach described is currently being extended to allow the rapid search and comparison of arbitrary combinations of molecular surface properties.

There is a pressing need for accurate in silico methods to predict the toxicity of molecules that are being introduced into the environment or are being developed into new pharmaceuticals. Predictive toxicology is in the realm of structure activity relationships (SAR), and many approaches have been used to derive such SAR. Previous work has shown that inductive logic programming (ILP) is a powerful approach that circumvents several major difficulties, such as molecular superposition, faced by some other SAR methods. The ILP approach reasons with chemical substructures within a relational framework and yields chemically understandable rules. Here, we report a general new approach, support vector inductive logic programming (SVILP), which extends the essentially qualitative ILP-based SAR to quantitative modeling. First, ILP is used to learn rules, the predictions of which are then used within a novel kernel to derive a support-vector generalization model. For a highly heterogeneous dataset of 576 molecules with known fathead minnow fish toxicity, the cross-validated correlation coefficients (R2CV) from a chemical descriptor method (CHEM) and SVILP are 0.52 and 0.66, respectively. The ILP, CHEM, and SVILP approaches correctly predict 55, 58, and 73%, respectively, of toxic molecules. In a set of 165 unseen molecules, the R2 values from the commercial software TOPKAT and SVILP are 0.26 and 0.57, respectively. In all calculations, SVILP showed significant improvements in comparison with the other methods. The SVILP approach has a major advantage in that it uses ILP automatically and consistently to derive rules, mostly novel, describing fragments that are toxicity alerts. The SVILP is a general machine-learning approach and has the potential of tackling many problems relevant to chemoinformatics including in silico drug design.

In this paper we propose a new method based on measurements of the structural similarity for the clustering of chemical databases. The proposed method allows the dynamic adjustment of the size and number of cells or clusters in which the database is classified. Classification is carried out using measurements of structural similarity obtained from the matching of molecular graphs. The classification process is open to the use of different similarity indexes and different measurements of matching. This process consists of the projection of the obtained measures of similarity among the elements of the database in a new space of similarity. The possibility of the dynamic readjustment of the dimension and characteristic of the projection space to adapt to the most favorable conditions of the problem under study and the simplicity and computational efficiency make the proposed method appropriate for its use with medium and large databases. The clustering method increases the performance of the screening processes in chemical databases, facilitating the recovery of chemical compounds that share all or subsets of common substructures to a given pattern. For the realization of the work a database of 498 natural compounds with wide molecular diversity extracted from SPECS and BIOSPECS B.V. free database has been used.

In this paper we present an algorithm for the generation of molecular graphs with a given value of the Wiener index. The high number of graphs for a given value of the Wiener index is reduced thanks to the application of a set of heuristics taking into account the structural characteristics of the molecules. The selection of parameters as the interval of values for the Wiener index, the diversity and occurrence of atoms and bonds, the size and number of cycles, and the presence of structural patterns guide the processing of the heuristics generating molecular graphs with a considerable saving in computational cost. The modularity in the design of the algorithm allows it to be used as a pattern for the development of other algorithms based on different topological invariants, which allow for its use in areas of interest, say as involving combinatorial databases and screening in chemical databases.