Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction

Abstract

Determining the aqueous solubility of molecules is a vital step in many pharmaceutical, environmental, and energy storage applications. Despite efforts made over decades, there are still challenges associated with developing a solubility prediction model with satisfactory accuracy for many of these applications. The goals of this study are to assess current deep learning methods for solubility prediction, develop a general model capable of predicting the solubility of a broad range of organic molecules, and to understand the impact of data properties, molecular representation, and modeling architecture on predictive performance. Using the largest currently available solubility data set, we implement deep learning-based models to predict solubility from the molecular structure and explore several different molecular representations including molecular descriptors, simplified molecular-input line-entry system strings, molecular graphs, and three-dimensional atomic coordinates using four different neural network architectures—fully connected neural networks, recurrent neural networks, graph neural networks (GNNs), and SchNet. We find that models using molecular descriptors achieve the best performance, with GNN models also achieving good performance. We perform extensive error analysis to understand the molecular properties that influence model performance, perform feature analysis to understand which information about the molecular structure is most valuable for prediction, and perform a transfer learning and data size study to understand the impact of data availability on model performance.

Building Machine Learning Small Molecule Melting Points and Solubility Models Using CCDC Melting Points Dataset

“…These potential outliers could be handled well by tree-based algorithms but may affect the performance of DNN greatly. Many studies using other representations of a molecule, such as weave module, 40 SMILES string, 41 and molecular graph 41,42 achieved generally a performance similar to that of Morgan fingerprint or 2D/3D descriptors on the prediction of solubility or other physicochemical properties. These new representations of molecules were not applied in this study.…”

Blinded Predictions and Post Hoc Analysis of the Second Solubility Challenge Data: Exploring Training Data and Feature Set Selection for Machine and Deep Learning Models

“…Indeed, a wide-variety of deep learning architectures have been applied to the problem of predicting solubility, including graph-based neural networks, − recurrent neural networks, transformers, , message-passing neural networks, deep belief networks, and others. Unlike other fields, it is not yet clear that deep learning algorithms offer a significant improvement over traditional machine learning approaches for solubility prediction . This may partly be due to a lack of accurate experimental solubility measurements for training, although strategies such as transfer learning , or multitask learning may help in some cases.…”

“…Unlike other fields, it is not yet clear that deep learning algorithms offer a significant improvement over traditional machine learning approaches for solubility prediction . This may partly be due to a lack of accurate experimental solubility measurements for training, although strategies such as transfer learning , or multitask learning may help in some cases.…”

“…Most recent research has been informed by the rapid advances in artificial intelligence being made in other fields. Indeed, a wide-variety of deep learning architectures have been applied to the problem of predicting solubility, including graph-based neural networks, − recurrent neural networks, transformers, , message-passing neural networks, deep belief networks, and others. Unlike other fields, it is not yet clear that deep learning algorithms offer a significant improvement over traditional machine learning approaches for solubility prediction .…”

Machine Learning with Enormous “Synthetic” Data Sets: Predicting Glass Transition Temperature of Polyimides Using Graph Convolutional Neural Networks

“…The “transfer learning” approach has been extensively tested in studies devoted to the prediction of the properties of small molecules. − Various DNN have appeared to be more accurate if they have been pretrained in advance on large amounts of “synthetic” (i.e., theoretically calculated) data, and only then fine-tuned with other more precise computational or experimental data. − For small molecules, these results were obtained using specifically developed quantum-chemical databases comprising up to hundreds of thousands of compounds and their properties (including, for example, the well-known quantum-chemical QM9 database, the Materials Project database, and the Open Quantum Materials Database).…”