Pruned Machine Learning Models to Predict Aqueous Solubility.

Area of Science:

• Drug discovery and development
• Computational chemistry
• Chemical biology

Background:

• Accurate prediction of aqueous solubility is essential for therapeutic compounds.
• Insoluble compounds can compromise the reliability of assays throughout drug discovery.
• Developing robust predictive models for solubility is a key challenge.

Purpose of the Study:

• To develop and evaluate naïve Bayesian classifier models for predicting aqueous solubility.
• To investigate the impact of data pruning on the accuracy of solubility prediction models.
• To provide curated training sets for future machine learning applications in solubility prediction.

Main Methods:

• Utilized publicly available aqueous solubility data to construct training sets.
• Created full, pruned, and fused training set variations.
• Trained naïve Bayesian classifier models on these datasets.
• Evaluated model performance using two independent, external test sets.

Main Results:

• The best pruned and fused model demonstrated significantly higher accuracy in predicting external test sets compared to full or fused models.
• Careful data pruning enhanced the predictive power of machine learning models for aqueous solubility.
• The study highlights the benefit of strategic data selection in model development.

Conclusions:

• Data pruning is an effective strategy for enhancing the accuracy of machine learning models for aqueous solubility prediction.
• Curated training sets and insights from this study can benefit future computational drug discovery efforts.
• Optimized machine learning models can improve the efficiency and reliability of solubility assessments.