Effect of Data Quality and Data Quantity on the Estimation of Intrinsic Solubility: Analysis Based on a Single-Source...

Area of Science:

• Drug discovery and development
• Computational chemistry
• Pharmacokinetics

Background:

• Aqueous solubility is a critical physicochemical property influencing oral drug absorption.
• Current in silico models struggle with predicting solubility for novel chemical spaces.
• A large, high-throughput aqueous solubility dataset from Johnson & Johnson offers a unique resource for model improvement.

Purpose of the Study:

• To investigate the impact of data quality and quantity on in silico solubility prediction models.
• To identify reasons for limited model performance in novel chemical spaces.
• To develop improved methods for predicting intrinsic aqueous solubility.

Main Methods:

• Generated six intrinsic solubility datasets with varying sizes and noise levels using different data processing approaches.
• Trained random forest regressors using descriptors from RDKit, ADMET predictor, and Mordred.
• Evaluated model performance using nested cross-validation and refined test sets.

Main Results:

• High-quality data consistently leads to better model performance for a given dataset size.
• Larger datasets with analytical variability can achieve accuracy comparable to smaller, cleaner datasets.
• Noise from amorphous solid residue significantly biases predictions and cannot be overcome by increasing data size, underscoring the need for data quality control.

Conclusions:

• Critical data review and curation are paramount for reliable in silico solubility predictions.
• A single-source, curated dataset significantly enhances the prediction of intrinsic solubility.
• Optimized models achieved superior performance on a benchmark solubility challenge dataset.