Predicting activities without computing descriptors: graph machines for QSAR

Temperature-Dependent Density and Viscosity Prediction for Hydrocarbons: Machine Learning and Molecular Dynamics Simulations

Machine learning-based predictive models allow rapid and reliable prediction of material properties and facilitate innovative materials design. Base oils used in the formulation of lubricant products are complex hydrocarbons of varying sizes and structure. This study developed Gaussian process regression-based models to accurately predict the temperature-dependent density and dynamic viscosity of 305 complex hydrocarbons. In our approach, strongly correlated/collinear predictors were trimmed, important predictors were selected by least absolute shrinkage and selection operator (LASSO) regularization and prior domain knowledge, hyperparameters were systematically optimized by Bayesian optimization, and the models were interpreted. The approach provided versatile and quantitative structure-property relationship (QSPR) models with relatively simple predictors for determining the dynamic viscosity and density of complex hydrocarbons at any temperature. In addition, we developed molecular dynamics simulation-based descriptors and evaluated the feasibility and versatility of dynamic descriptors from simulations for predicting the material properties. It was found that the models developed using a comparably smaller pool of dynamic descriptors performed similarly in predicting density and viscosity to models based on many more static descriptors. The best models were shown to predict density and dynamic viscosity with coefficient of determination (R2) values of 99.6% and 97.7%, respectively, for all data sets, including a test data set of 45 molecules. Finally, partial dependency plots (PDPs), individual conditional expectation (ICE) plots, local interpretable model-agnostic explanation (LIME) values, and trimmed model R2 values were used to identify the most important static and dynamic predictors of the density and viscosity.

Fast and Accurate Prediction of Refractive Index of Organic Liquids with Graph Machines

The refractive index (RI) of liquids is a key physical property of molecular compounds and materials. In addition to its ubiquitous role in physics, it is also exploited to impart specific optical properties (transparency, opacity, and gloss) to materials and various end-use products. Since few methods exist to accurately estimate this property, we have designed a graph machine model (GMM) capable of predicting the RI of liquid organic compounds containing up to 16 different types of atoms and effective in discriminating between stereoisomers. Using 8267 carefully checked RI values from the literature and the corresponding 2D organic structures, the GMM provides a training root mean square relative error of less than 0.5%, i.e., an RMSE of 0.004 for the estimation of the refractive index of the 8267 compounds. The GMM predictive ability is also compared to that obtained by several fragment-based approaches. Finally, a Docker-based tool is proposed to predict the RI of organic compounds solely from their SMILES code. The GMM developed is easy to apply, as shown by the video tutorials provided on YouTube.

HiGNN: A Hierarchical Informative Graph Neural Network for Molecular Property Prediction Equipped with Feature-Wise Attention

Elucidating and accurately predicting the druggability and bioactivities of molecules plays a pivotal role in drug design and discovery and remains an open challenge. Recently, graph neural networks (GNNs) have made remarkable advancements in graph-based molecular property prediction. However, current graph-based deep learning methods neglect the hierarchical information of molecules and the relationships between feature channels. In this study, we propose a well-designed hierarchical informative graph neural network (termed HiGNN) framework for predicting molecular property by utilizing a corepresentation learning of molecular graphs and chemically synthesizable breaking of retrosynthetically interesting chemical substructure (BRICS) fragments. Furthermore, a plug-and-play feature-wise attention block is first designed in HiGNN architecture to adaptively recalibrate atomic features after the message passing phase. Extensive experiments demonstrate that HiGNN achieves state-of-the-art predictive performance on many challenging drug discovery-associated benchmark data sets. In addition, we devise a molecule-fragment similarity mechanism to comprehensively investigate the interpretability of the HiGNN model at the subgraph level, indicating that HiGNN as a powerful deep learning tool can help chemists and pharmacists identify the key components of molecules for designing better molecules with desired properties or functions. The source code is publicly available at https://github.com/idruglab/hignn.

Fast Prediction of the Equivalent Alkane Carbon Number Using Graph Machines and Neural Networks

The hydrophobicity of oils is a key parameter to design surfactant/oil/water (SOW) macro-, micro-, or nano-dispersed systems with the desired features. This essential physicochemical characteristic is quantitatively expressed by the equivalent alkane carbon number (EACN) whose experimental determination is tedious since it requires knowledge of the phase behavior of the SOW systems at different temperatures and for different surfactant concentrations. In this work, two mathematical models are proposed for the rapid prediction of the EACN of oils. They have been designed using artificial intelligence (machine-learning) methods, namely, neural networks (NN) and graph machines (GM). While the GM model is implemented from the SMILES codes of a 111-molecule training set of known EACN values, the NN model is fed with some σ-moment descriptors computed with the COSMOtherm software for the 111-molecule set. In a preliminary step, the leave-one-out algorithm is used to select, given the available data, the appropriate complexity of the two models. A comparison of the EACNs of liquids of a fresh set of 10 complex cosmetic and perfumery molecules shows that the two approaches provide comparable results in terms of accuracy and reliability. Finally, the NN and GM models are applied to nine series of homologous compounds, for which the GM model results are in better agreement with the experimental EACN trends than the NN model predictions. The results obtained by the GMs and by the NN based on σ-moments can be duplicated with the demonstration tool available for download as detailed in the Supporting Information.

Bio-based alternatives to volatile silicones: Relationships between chemical structure, physicochemical properties and functional performances

Emollient oils are ubiquitous ingredients of personal care products, especially skin care and hair care formulations. They offer excellent spreading properties and give end-use products a soft, pleasant and non-sticky after-feel. Emollients belong to various petro- or bio-based chemical families among which silicone oils, hydrocarbons and esters are the most prominent. Silicones have exceptional physicochemical and sensory properties but their high chemical stability results in very low biodegradability and a high bioaccumulation potential. Nowadays, consumers are increasingly responsive to environmental issues and demand more environmentally friendly products. This awareness strongly encourages cosmetics industries to develop bio-based alternatives to silicone oils. Finding effective silicon-free emollients requires understanding the molecular origin of emollience. This review details the relationships between the molecular structures of emollients and their physicochemical properties as well as the resulting functional performances in order to facilitate the design of alternative oils with suitable physicochemical and sensory properties. The molecular profile of an ideal emollient in terms of chemical function (alkane, ether, ester, carbonate, alcohol), optimal number of carbons and branching is established to obtain an odourless oil with good spreading on the skin. Since none of the carbon-based emollients alone can imitate the non-sticky and dry feel of silicone oils, it is judicious to blend alkanes and esters to significantly improve both the sensory properties and the solubilizing properties of the synergistic mixture towards polar ingredients (sun filters, antioxidants, fragrances). Finally, it is shown how modelling tools (QSPR, COSMO-RS and neural networks) can predict in silico the key properties of hundreds of virtual candidate molecules in order to synthesize only the most promising whose predicted properties are close to the specifications.

A New Machine-Learning Tool for Fast Estimation of Liquid Viscosity. Application to Cosmetic Oils

The viscosities of pure liquids are estimated at 25 °C, from their molecular structures, using three modeling approaches: group contributions, COSMO-RS σ-moment-based neural networks, and graph machines. The last two are machine-learning methods, whereby models are designed and trained from a database of viscosities of 300 molecules at 25 °C. Group contributions and graph machines make use of the 2D-structures only (the SMILES codes of the molecules), while neural networks estimations are based on a set of five descriptors: COSMO-RS σ-moments. For the first time, leave-one-out is used for graph machine selection, and it is shown that it can be replaced with the much faster virtual leave-one-out algorithm. The database covers a wide diversity of chemical structures, namely, alkanes, ethers, esters, ketones, carbonates, acids, alcohols, silanes, and siloxanes, as well as different chemical backbone, i.e., straight, branched, or cyclic chains. A comparison of the viscosities of liquids of an independent set of 22 cosmetic oils shows that the graph machine approach provides the most accurate results given the available data. The results obtained by the neural network based on sigma-moments and by the graph machines can be duplicated easily by using a demonstration tool based on the Docker technology, available for download as explained in the Supporting Information. This demonstration also allows the reader to predict, at 25 °C, the viscosity of any liquid of moderate molecular size (M < 600 Da) that contains C, H, O, or Si atoms, starting either from its SMILES code or from its σ-moments computed with the COSMOtherm software.

A self-attention based message passing neural network for predicting molecular lipophilicity and aqueous solubility

Efficient and accurate prediction of molecular properties, such as lipophilicity and solubility, is highly desirable for rational compound design in chemical and pharmaceutical industries. To this end, we build and apply a graph-neural-network framework called self-attention-based message-passing neural network (SAMPN) to study the relationship between chemical properties and structures in an interpretable way. The main advantages of SAMPN are that it directly uses chemical graphs and breaks the black-box mold of many machine/deep learning methods. Specifically, its attention mechanism indicates the degree to which each atom of the molecule contributes to the property of interest, and these results are easily visualized. Further, SAMPN outperforms random forests and the deep learning framework MPN from Deepchem. In addition, another formulation of SAMPN (Multi-SAMPN) can simultaneously predict multiple chemical properties with higher accuracy and efficiency than other models that predict one specific chemical property. Moreover, SAMPN can generate chemically visible and interpretable results, which can help researchers discover new pharmaceuticals and materials. The source code of the SAMPN prediction pipeline is freely available at Github (https://github.com/tbwxmu/SAMPN).

Chemoinformatics at IFP Energies Nouvelles: Applications in the Fields of Energy, Transport, and Environment

The objective of the present paper is to summarize chemoinformatics based research, and more precisely, the development of quantitative structure property relationships performed at IFP Energies nouvelles (IFPEN) during the last decade. A special focus is proposed on research activities performed in the "Thermodynamics and Molecular Simulation" department, i. e. the use of multiscale molecular simulation methods in responses to projects. Molecular simulation techniques can be envisaged to supplement dataset when experimental information lacks, thus the review includes a section dedicated to molecular simulation codes, development of intermolecular potentials, and some of their possible applications. Know-how and feedback from our experiences in terms of machine learning application for thermophysical property predictions are included in a section dealing with methodological aspects. The generic character of chemoinformatics is emphasized through applications in the fields of energy, transport, and environment, with illustrations for three IFPEN business units: "Transports", "Energy Resources", and "Processes". More precisely, the review focus on different challenges such as the prediction of properties for alternative fuels, the prediction of fuel compatibility with polymeric materials, the prediction of properties for surfactants usable in chemical enhanced oil recovery, and the prediction of guest-host interactions between gases and nanoporous materials in the frame of carbon dioxide capture or gas separation activities.

A renaissance of neural networks in drug discovery

INTRODUCTION

Neural networks are becoming a very popular method for solving machine learning and artificial intelligence problems. The variety of neural network types and their application to drug discovery requires expert knowledge to choose the most appropriate approach.

AREAS COVERED

In this review, the authors discuss traditional and newly emerging neural network approaches to drug discovery. Their focus is on backpropagation neural networks and their variants, self-organizing maps and associated methods, and a relatively new technique, deep learning. The most important technical issues are discussed including overfitting and its prevention through regularization, ensemble and multitask modeling, model interpretation, and estimation of applicability domain. Different aspects of using neural networks in drug discovery are considered: building structure-activity models with respect to various targets; predicting drug selectivity, toxicity profiles, ADMET and physicochemical properties; characteristics of drug-delivery systems and virtual screening.

EXPERT OPINION

Neural networks continue to grow in importance for drug discovery. Recent developments in deep learning suggests further improvements may be gained in the analysis of large chemical data sets. It's anticipated that neural networks will be more widely used in drug discovery in the future, and applied in non-traditional areas such as drug delivery systems, biologically compatible materials, and regenerative medicine.

Monte Carlo QSAR models for predicting organophosphate inhibition of acetycholinesterase

A series of 278 organophosphate compounds acting as acetylcholinesterase inhibitors has been studied. The Monte Carlo method was used as a tool for building up one-variable quantitative structure-activity relationship (QSAR) models for acetylcholinesterase inhibition activity based on the principle that the target endpoint is treated as a random event. As an activity, bimolecular rate constants were used. The QSAR models were based on optimal descriptors obtained from Simplified Molecular Input-Line Entry System (SMILES) used for the representation of molecular structure. Two modelling approaches were examined: (1) 'classic' training-test system where the QSAR model was built with one random split into a training, test and validation set; and (2) the correlation balance based QSAR models were built with two random splits into a sub-training, calibration, test and validation set. The DModX method was used for defining the applicability domain. The obtained results suggest that studied activity can be determined with the application of QSAR models calculated with the Monte Carlo method since the statistical quality of all build models was very good. Finally, structural indicators for the increase and the decrease of the bimolecular rate constant are defined. The possibility of using these results for the computer-aided design of new organophosphate compounds is presented.

On the rational formulation of alternative fuels: melting point and net heat of combustion predictions for fuel compounds using machine learning methods

We report the development of predictive models for two fuel specifications: melting points (T(m)) and net heat of combustion (Δ(c)H). Compounds inside the scope of these models are those likely to be found in alternative fuels, i.e. hydrocarbons, alcohols and esters. Experimental T(m) and Δ(c)H values for these types of molecules have been gathered to generate a unique database. Various quantitative structure-property relationship (QSPR) approaches have been used to build models, ranging from methods leading to multi-linear models such as genetic function approximation (GFA), or partial least squares (PLS) to those leading to non-linear models such as feed-forward artificial neural networks (FFANN), general regression neural networks (GRNN), support vector machines (SVM), or graph machines. Except for the case of the graph machines method for which the only inputs are SMILES formulae, previously listed approaches working on molecular descriptors and functional group count descriptors were used to develop specific models for T(m) and Δ(c)H. For each property, the predictive models return slightly different responses for each molecular structure. Therefore, models labelled as 'consensus models' were built by averaging values computed with selected individual models. Predicted results were then compared with experimental data and with predictions of models in the literature.

Machine learning methods for property prediction in chemoinformatics: Quo Vadis?

This paper is focused on modern approaches to machine learning, most of which are as yet used infrequently or not at all in chemoinformatics. Machine learning methods are characterized in terms of the "modes of statistical inference" and "modeling levels" nomenclature and by considering different facets of the modeling with respect to input/ouput matching, data types, models duality, and models inference. Particular attention is paid to new approaches and concepts that may provide efficient solutions of common problems in chemoinformatics: improvement of predictive performance of structure-property (activity) models, generation of structures possessing desirable properties, model applicability domain, modeling of properties with functional endpoints (e.g., phase diagrams and dose-response curves), and accounting for multiple molecular species (e.g., conformers or tautomers).

High throughput screening of CO2 solubility in aqueous monoamine solutions

Post-combustion Carbon Capture and Storage technology (CCS) is viewed as an efficient solution to reduce CO(2) emissions of coal-fired power stations. In CCS, an aqueous amine solution is commonly used as a solvent to selectively capture CO(2) from the flue gas. However, this process generates additional costs, mostly from the reboiler heat duty required to release the carbon dioxide from the loaded solvent solution. In this work, we present thermodynamic results of CO(2) solubility in aqueous amine solutions from a 6-reactor High Throughput Screening (HTS) experimental device. This device is fully automated and designed to perform sequential injections of CO(2) within stirred-cell reactors containing the solvent solutions. The gas pressure within each reactor is monitored as a function of time, and the resulting transient pressure curves are transformed into CO(2) absorption isotherms. Solubility measurements are first performed on monoethanolamine, diethanolamine, and methyldiethanolamine aqueous solutions at T = 313.15 K. Experimental results are compared with existing data in the literature to validate the HTS device. In addition, a comprehensive thermodynamic model is used to represent CO(2) solubility variations in different classes of amine structures upon a wide range of thermodynamic conditions. This model is used to fit the experimental data and to calculate the cyclic capacity, which is a key parameter for CO(2) process design. Solubility measurements are then performed on a set of 50 monoamines and cyclic capacities are extracted using the thermodynamic model, to asses the potential of these molecules for CO(2) capture.