(423b) Advancing Solubility Prediction Using an AI-Driven Tool for Resource Optimization in Pharmaceutical Research

In recent years, the application of data-driven models has been adopted by various research and development fields such as materials science, chemistry, and drug discovery and development. Machine learning techniques are now commonly used to predict physicochemical properties and process parameters. In materials science, the availability of large amounts of data has led to the development of advanced tools for materials informatics. While mechanistic models have been used to predict solubility values, machine learning methods have also been gaining attention. Deep learning techniques, such as convolutional neural networks, have emerged as a promising approach for capturing complex structure-property relationships for property prediction. Computer vision techniques have also been used to predict solubility by analyzing images of the solute and solvent [1]. By leveraging these data-driven approaches, scientists and engineers can develop more accurate and efficient models for predicting solubility, which can have important applications in fields such as drug discovery and material science.

Predicting solubility of complex pharmaceutical compounds is crucial at several stages of active pharmaceutical ingredient process development. Through the lens of green chemistry, solvent selection has become increasingly important in process design and optimization, aiming to reduce the use of the most hazardous, environmentally damaging solvents [2]. Although traditionally this challenge has been addressed by experimental solubility measurements [3], these are time- and material-consuming, limiting the opportunity of solvent screening. Hence, there is a need for computational models that can predict compound solubility efficiently, not only accelerating but optimizing the early stages of pharmaceutical development.

Hereby, a novel approach for predicting the solubility of chemical compounds is presented. The proposed method is based on a hybrid neural network architecture that combines deep learning, computer vision, and experimental input data (Figure 1). To the best of the authors’ knowledge, there is no other model in the literature that uses such a hybrid architecture for solubility prediction. The proposed model leverages deep learning to extract features from images and experimental data, and of computer vision to analyze and interpret these features. The model is trained on a large dataset of over 240 systems and associated solubility and turbidity measurements and is optimized using continuous refinement and optimization techniques. The authors demonstrate that the proposed model achieves remarkable accuracy in predicting solubility (98%), outperforming other state-of-the-art models. The model has been validated through a robust cross-validation test.

The proposed model has the potential to significantly improve the efficiency of solubility screening in early-stage research and development, and to reduce the experimental workload and resource consumption in the chemical and pharmaceutical industries. A data-driven approach for solubility screening can significantly reduce the time and resources required for experimental solvent screening. Furthermore, data-driven models have the potential to outperform traditional mechanistic models [4]. This study presents a promising direction for future research in the field of solubility prediction and highlights the potential of hybrid neural network architectures for solving complex problems in chemistry and related fields.

References

1. Singh, V., et al., Recent trends in computational tools and data-driven modeling for advanced materials. Materials Advances, 2022. 3(10): p. 4069-4087.
2. Byrne, F.P., et al., Tools and techniques for solvent selection: green solvent selection guides. Sustainable Chemical Processes, 2016. 4: p. 1-24.
3. Könczöl, Á. and G. Dargó, Brief overview of solubility methods: Recent trends in equilibrium solubility measurement and predictive models. Drug Discovery Today: Technologies, 2018. 27: p. 3-10.
4. Piccard, P.-J., et al., Organic Solvent Nanofiltration and Data-Driven Approaches. Separations, 2023. 10(9): p. 516.