Inter-laboratory comparison of water solubility methods applied to difficult-to-test substances

Outline and background for the EU-OS solubility prediction challenge

In June 2022, EU-OS came to the decision to make public a solubility data set of 100+K compounds obtained from several of the EU-OS proprietary screening compound collections. Leveraging on the interest of SLAS for screening scientific development it was decided to launch a joint EUOS-SLAS competition within the chemoinformatics and machine learning (ML) communities. The competition was open to real world computation experts, for the best, most predictive, classification model of compound solubility. The aim of the competition was multiple: from a practical side, the winning model should then serve as a cornerstone for future solubility predictions having used the largest training set so far publicly available. From a higher project perspective, the intent was to focus the energies and experiences, even if professionally not precisely coming from Pharma R&D; to address the issue of how to predict compound solubility. Here we report how the competition was ideated and the practical aspects of conducting it within the Kaggle framework, leveraging of the versatility and the open-source nature of this data science platform. Consideration on results and challenges encountered have been also examined.

Transparency in Modeling through Careful Application of OECD's QSAR/QSPR Principles via a Curated Water Solubility Data Set

The need for careful assembly, training, and validation of quantitative structure-activity/property models (QSAR/QSPR) is more significant than ever as data sets become larger and sophisticated machine learning tools become increasingly ubiquitous and accessible to the scientific community. Regulatory agencies such as the United States Environmental Protection Agency must carefully scrutinize each aspect of a resulting QSAR/QSPR model to determine its potential use in environmental exposure and hazard assessment. Herein, we revisit the goals of the Organisation for Economic Cooperation and Development (OECD) in our application and discuss the validation principles for structure-activity models. We apply these principles to a model for predicting water solubility of organic compounds derived using random forest regression, a common machine learning approach in the QSA/PR literature. Using public sources, we carefully assembled and curated a data set consisting of 10,200 unique chemical structures with associated water solubility measurements. This data set was then used as a focal narrative to methodically consider the OECD's QSA/PR principles and how they can be applied to random forests. Despite some expert, mechanistically informed supervision of descriptor selection to enhance model interpretability, we achieved a model of water solubility with comparable performance to previously published models (5-fold cross validated performance 0.81 R2 and 0.98 RMSE). We hope this work will catalyze a necessary conversation around the importance of cautiously modernizing and explicitly leveraging OECD principles while pursuing state-of-the-art machine learning approaches to derive QSA/PR models suitable for regulatory consideration.

Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction

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.