Revisiting the Application of Machine Learning Approaches in Predicting Aqueous Solubility.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Environmental Science

Background:

• Accurate prediction of aqueous solubility is vital across pharmaceutical, environmental, and agrochemical sectors.
• Despite its importance, predicting chemical solubility accurately remains a significant scientific challenge.
• Machine learning and molecular descriptors offer promising avenues for improving solubility predictions.

Purpose of the Study:

• To evaluate and compare popular machine learning (ML) methods for predicting aqueous solubility.
• To assess the effectiveness of various molecular featurization techniques in ML models.
• To identify key molecular descriptors contributing to accurate solubility predictions.

Main Methods:

• Comparative analysis of diverse machine learning algorithms.
• Implementation of various molecular featurization techniques, including graph convolution and attention mechanisms.
• Evaluation of over 4000 molecular descriptors for their predictive contribution.

Main Results:

• Graph-based ML methods showed exceptional predictive power on high-quality datasets.
• Models using molecular descriptors demonstrated superior interpretability and resilience to data noise.
• Approximately 800 out of 4000 analyzed molecular descriptors were found to be significant for solubility prediction.

Conclusions:

• Machine learning models offer significant improvements in aqueous solubility prediction.
• The choice of ML method and featurization technique impacts performance and interpretability.
• Future research should focus on robust descriptor selection and noise-robust modeling for enhanced solubility prediction.