A unified ML framework for solubility prediction across organic solvents

Vassileiou, Antony D. and Robertson, Murray N. and Wareham, Bruce G. and Soundaranathan, Mithushan and Ottoboni, Sara and Florence, Alastair J. and Hartwig, Thoralf and Johnston, Blair F. (2023) A unified ML framework for solubility prediction across organic solvents. Digital Discovery, 2 (2). pp. 356-367. ISSN 2635-098X (https://doi.org/10.1039/D2DD00024E)

[thumbnail of Vassileiou-etal-DD-2023-A-unified-ML-framework-for-solubility-prediction]
Preview
Text. Filename: Vassileiou_etal_DD_2023_A_unified_ML_framework_for_solubility_prediction.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo

Download (1MB)| Preview

Abstract

We report a single machine learning (ML)-based model to predict the solubility of drug/drug-like compounds across 49 organic solvents, extensible to more. By adopting a cross-solvent data structure, we enable the exploitation of valuable relational information between systems. The effect is major, with even a single experimental measurement of a solute in a different solvent being enough to significantly improve predictions on it, and successive ones improving them further. Working with a sparse dataset of only 714 experimental data points spanning 75 solutes and 49 solvents (81% sparsity), a ML-based model with a prediction RMSE of 0.75 log S (g/100 g) for unseen solutes was produced. This compares favourably with conductor-like screening model for real solvents (COSMO-RS), an industry-standard model based on thermodynamic laws, which yielded a prediction RMSE of 0.97 for the same dataset. The error for our method reduced to a mean RMSE of 0.65 when one instance of the solute (in a different solvent) was included in the training data; this iteratively reduced further to 0.60, 0.57 and 0.56 when two, three and four instances were available, respectively. This standard of performance not only meets or exceeds those of alternative ML-based solubility models insofar as they can be compared but reaches the perceived ceiling for solubility prediction models of this type. In parallel, we assess the performance of the model with and without the addition of COSMO-RS output as an additional descriptor. We find that a significant benefit is gained from its addition, indicating that mechanistic methods can bring insight that simple molecular descriptors cannot and should be incorporated into a data-driven prediction of molecular properties where possible.