Search

Machine Learning (ML) techniques face significant challenges when predicting advanced chemical properties, such as yield, feasibility of chemical synthesis, and optimal reaction conditions. These challenges stem from the high-dimensional nature of the prediction task and the myriad essential variables involved, ranging from reactants and reagents to catalysts, temperature, and purification processes. Successfully developing a reliable predictive model not only holds the potential for optimizing high-throughput experiments but can also elevate existing retrosynthetic predictive approaches and bolster a plethora of applications within the field. In this review, we systematically evaluate the efficacy of current ML methodologies in chemoinformatics, shedding light on their milestones and inherent limitations. Additionally, a detailed examination of a representative case study provides insights into the prevailing issues related to data availability and transferability in the discipline.

Acute toxicity is one of the most challenging properties to predict purely with computational methods due to its direct relationship to biological interactions. Moreover, toxicity can be represented by different end points: it can be measured for different species using different types of administration, etc., and it is questionable if the knowledge transfer between end points is possible. We performed a comparative study of prediction multitask toxicity for a broad chemical space using different descriptors and modeling algorithms and applied multitask learning for a large toxicity data set extracted from the Registry of Toxic Effects of Chemical Substances (RTECS). We demonstrated that multitask modeling provides significant improvement over single-output models and other machine learning methods. Our research reveals that multitask learning can be very useful to improve the quality of acute toxicity modeling and raises a discussion about the usage of multitask approaches for regulation purposes. Our MultiTox models are freely available in OCHEM platform ( ochem.eu/multitox ) under CC-BY-NC license.

Firefly luciferase is an enzyme that has found ubiquitous use in biological assays in high-throughput screening (HTS) campaigns. The inhibition of luciferase in such assays could lead to a false positive result. This issue has been known for a long time, and there have been significant efforts to identify luciferase inhibitors in order to enhance recognition of false positives in screening assays. However, although a large amount of publicly accessible luciferase counterscreen data is available, to date little effort has been devoted to building a chemoinformatic model that can identify such molecules in a given data set. In this study we developed models to identify these molecules using various methods, such as molecular docking, SMARTS screening, pharmacophores, and machine learning methods. Among the structure-based methods, the pharmacophore-based method showed promising results, with a balanced accuracy of 74.2%. However, machine-learning approaches using associative neural networks outperformed all of the other methods explored, producing a final model with a balanced accuracy of 89.7%. The high predictive accuracy of this model is expected to be useful for advising which compounds are potential luciferase inhibitors present in luciferase HTS assays. The models developed in this work are freely available at the OCHEM platform at http://ochem.eu .

This article contributes a highly accurate model for predicting the melting points (MPs) of medicinal chemistry compounds. The model was developed using the largest published data set, comprising more than 47k compounds. The distributions of MPs in drug-like and drug lead sets showed that >90% of molecules melt within [50,250]°C. The final model calculated an RMSE of less than 33 °C for molecules from this temperature interval, which is the most important for medicinal chemistry users. This performance was achieved using a consensus model that performed calculations to a significantly higher accuracy than the individual models. We found that compounds with reactive and unstable groups were overrepresented among outlying compounds. These compounds could decompose during storage or measurement, thus introducing experimental errors. While filtering the data by removing outliers generally increased the accuracy of individual models, it did not significantly affect the results of the consensus models. Three analyzed distance to models did not allow us to flag molecules, which had MP values fell outside the applicability domain of the model. We believe that this negative result and the public availability of data from this article will encourage future studies to develop better approaches to define the applicability domain of models. The final model, MP data, and identified reactive groups are available online at http://ochem.eu/article/55638.

The dimethyl sulfoxide (DMSO) solubility data from Enamine and two UCB pharma compound collections were analyzed using 8 different machine learning methods and 12 descriptor sets. The analyzed data sets were highly imbalanced with 1.7-5.8% nonsoluble compounds. The libraries' enrichment by soluble molecules from the set of 10% of the most reliable predictions was used to compare prediction performances of the methods. The highest accuracies were calculated using a C4.5 decision classification tree, random forest, and associative neural networks. The performances of the methods developed were estimated on individual data sets and their combinations. The developed models provided on average a 2-fold decrease of the number of nonsoluble compounds amid all compounds predicted as soluble in DMSO. However, a 4-9-fold enrichment was observed if only 10% of the most reliable predictions were considered. The structural features influencing compounds to be soluble or nonsoluble in DMSO were also determined. The best models developed with the publicly available Enamine data set are freely available online at http://ochem.eu/article/33409 .

The article presents a Web-based platform for collecting and storing toxicological structural alerts from literature and for virtual screening of chemical libraries to flag potentially toxic chemicals and compounds that can cause adverse side effects. An alert is uniquely identified by a SMARTS template, a toxicological endpoint, and a publication where the alert was described. Additionally, the system allows storing complementary information such as name, comments, and mechanism of action, as well as other data. Most importantly, the platform can be easily used for fast virtual screening of large chemical datasets, focused libraries, or newly designed compounds against the toxicological alerts, providing a detailed profile of the chemicals grouped by structural alerts and endpoints. Such a facility can be used for decision making regarding whether a compound should be tested experimentally, validated with available QSAR models, or eliminated from consideration altogether. The alert-based screening can also be helpful for an easier interpretation of more complex QSAR models. The system is publicly accessible and tightly integrated with the Online Chemical Modeling Environment (OCHEM, http://ochem.eu). The system is open and expandable: any registered OCHEM user can introduce new alerts, browse, edit alerts introduced by other users, and virtually screen his/her data sets against all or selected alerts. The user sets being passed through the structural alerts can be used at OCHEM for other typical tasks: exporting in a wide variety of formats, development of QSAR models, additional filtering by other criteria, etc. The database already contains almost 600 structural alerts for such endpoints as mutagenicity, carcinogenicity, skin sensitization, compounds that undergo metabolic activation, and compounds that form reactive metabolites and, thus, can cause adverse reactions. The ToxAlerts platform is accessible on the Web at http://ochem.eu/alerts, and it is constantly growing.

Several applications, such as risk assessment within REACH or drug discovery, require reliable methods for the design of experiments and efficient testing strategies. Keeping the number of experiments as low as possible is important from both a financial and an ethical point of view, as exhaustive testing of compounds requires significant financial resources and animal lives. With a large initial set of compounds, experimental design techniques can be used to select a representative subset for testing. Once measured, these compounds can be used to develop quantitative structure-activity relationship models to predict properties of the remaining compounds. This reduces the required resources and time. D-Optimal design is frequently used to select an optimal set of compounds by analyzing data variance. We developed a new sequential approach to apply a D-Optimal design to latent variables derived from a partial least squares (PLS) model instead of principal components. The stepwise procedure selects a new set of molecules to be measured after each previous measurement cycle. We show that application of the D-Optimal selection generates models with a significantly improved performance on four different data sets with end points relevant for REACH. Compared to those derived from principal components, PLS models derived from the selection on latent variables had a lower root-mean-square error and a higher Q2 and R2. This improvement is statistically significant, especially for the small number of compounds selected.

Prediction of CYP450 inhibition activity of small molecules poses an important task due to high risk of drug-drug interactions. CYP1A2 is an important member of CYP450 superfamily and accounts for 15% of total CYP450 presence in human liver. This article compares 80 in-silico QSAR models that were created by following the same procedure with different combinations of descriptors and machine learning methods. The training and test sets consist of 3745 and 3741 inhibitors and noninhibitors from PubChem BioAssay database. A heterogeneous external test set of 160 inhibitors was collected from literature. The studied descriptor sets involve E-state, Dragon and ISIDA SMF descriptors. Machine learning methods involve Associative Neural Networks (ASNN), K Nearest Neighbors (kNN), Random Tree (RT), C4.5 Tree (J48), and Support Vector Machines (SVM). The influence of descriptor selection on model accuracy was studied. The benefits of "bagging" modeling approach were shown. Applicability domain approach was successfully applied in this study and ways of increasing model accuracy through use of applicability domain measures were demonstrated as well as fragment-based model interpretation was performed. The most accurate models in this study achieved values of 83% and 68% correctly classified instances on the internal and external test sets, respectively. The applicability domain approach allowed increasing the prediction accuracy to 90% for 78% of the internal and 17% of the external test sets, respectively. The most accurate models are available online at http://ochem.eu/models/Q5747 .

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 .

Two inductive knowledge transfer approaches - multitask learning (MTL) and Feature Net (FN) - have been used to build predictive neural networks (ASNN) and PLS models for 11 types of tissue-air partition coefficients (TAPC). Unlike conventional single-task learning (STL) modeling focused only on a single target property without any relations to other properties, in the framework of inductive transfer approach, the individual models are viewed as nodes in the network of interrelated models built in parallel (MTL) or sequentially (FN). It has been demonstrated that MTL and FN techniques are extremely useful in structure-property modeling on small and structurally diverse data sets, when conventional STL modeling is unable to produce any predictive model. The predictive STL individual models were obtained for 4 out of 11 TAPC, whereas application of inductive knowledge transfer techniques resulted in models for 9 TAPC. Differences in prediction performances of the models as a function of the machine-learning method, and of the number of properties simultaneously involved in the learning, has been discussed.

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.

Selecting most rigorous quantitative structure-activity relationship (QSAR) approaches is of great importance in the development of robust and predictive models of chemical toxicity. To address this issue in a systematic way, we have formed an international virtual collaboratory consisting of six independent groups with shared interests in computational chemical toxicology. We have compiled an aqueous toxicity data set containing 983 unique compounds tested in the same laboratory over a decade against Tetrahymena pyriformis. A modeling set including 644 compounds was selected randomly from the original set and distributed to all groups that used their own QSAR tools for model development. The remaining 339 compounds in the original set (external set I) as well as 110 additional compounds (external set II) published recently by the same laboratory (after this computational study was already in progress) were used as two independent validation sets to assess the external predictive power of individual models. In total, our virtual collaboratory has developed 15 different types of QSAR models of aquatic toxicity for the training set. The internal prediction accuracy for the modeling set ranged from 0.76 to 0.93 as measured by the leave-one-out cross-validation correlation coefficient ( Q abs2). The prediction accuracy for the external validation sets I and II ranged from 0.71 to 0.85 (linear regression coefficient R absI2) and from 0.38 to 0.83 (linear regression coefficient R absII2), respectively. The use of an applicability domain threshold implemented in most models generally improved the external prediction accuracy but at the same time led to a decrease in chemical space coverage. Finally, several consensus models were developed by averaging the predicted aquatic toxicity for every compound using all 15 models, with or without taking into account their respective applicability domains. We find that consensus models afford higher prediction accuracy for the external validation data sets with the highest space coverage as compared to individual constituent models. Our studies prove the power of a collaborative and consensual approach to QSAR model development. The best validated models of aquatic toxicity developed by our collaboratory (both individual and consensus) can be used as reliable computational predictors of aquatic toxicity and are available from any of the participating laboratories.

Several popular machine learning methods--Associative Neural Networks (ANN), Support Vector Machines (SVM), k Nearest Neighbors (kNN), modified version of the partial least-squares analysis (PLSM), backpropagation neural network (BPNN), and Multiple Linear Regression Analysis (MLR)--implemented in ISIDA, NASAWIN, and VCCLAB software have been used to perform QSPR modeling of melting point of structurally diverse data set of 717 bromides of nitrogen-containing organic cations (FULL) including 126 pyridinium bromides (PYR), 384 imidazolium and benzoimidazolium bromides (IMZ), and 207 quaternary ammonium bromides (QUAT). Several types of descriptors were tested: E-state indices, counts of atoms determined for E-state atom types, molecular descriptors generated by the DRAGON program, and different types of substructural molecular fragments. Predictive ability of the models was analyzed using a 5-fold external cross-validation procedure in which every compound in the parent set was included in one of five test sets. Among the 16 types of developed structure--melting point models, nonlinear SVM, ASNN, and BPNN techniques demonstrate slightly better performance over other methods. For the full set, the accuracy of predictions does not significantly change as a function of the type of descriptors. For other sets, the performance of descriptors varies as a function of method and data set used. The root-mean squared error (RMSE) of prediction calculated on independent test sets is in the range of 37.5-46.4 degrees C (FULL), 26.2-34.8 degrees C (PYR), 38.8-45.9 degrees C (IMZ), and 34.2-49.3 degrees C (QUAT). The moderate accuracy of predictions can be related to the quality of the experimental data used for obtaining the models as well as to difficulties to take into account the structural features of ionic liquids in the solid state (polymorphic effects, eutectics, glass formation).

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.