Assessing the impact of data harmonization on human liver microsomal stability prediction model performance

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry and Cheminformatics

Background:

• Metabolic stability, assessed via human liver microsomes (HLM), is vital for drug clearance and pharmacokinetics (PK).
• Experimental determination of metabolic stability is resource-intensive.
• Existing machine learning (ML) models often use heterogeneous ChEMBL data, potentially introducing variability.

Purpose of the Study:

• To systematically assess the influence of data heterogeneity versus homogeneity on HLM stability prediction model performance.
• To compare different ML algorithms (KNN, SVC, XGBoost, Chemprop) using standardized and combined datasets.
• To evaluate model generalizability on external test sets.

Main Methods:

• Trained ML models (KNN, SVC, XGBoost, Chemprop) on three standardized datasets (NCATS, Biogen, ChEMBL) and their combinations.
• Assessed model performance and generalizability using two external test sets (Astra-Zeneca, Polaris).
• Analyzed the impact of dataset source and composition on predictive accuracy.

Main Results:

• Model performance varied significantly based on the dataset's homogeneity and source.
• Homogeneous datasets generally led to more robust and generalizable predictive models.
• The study identified trade-offs between using large, heterogeneous datasets and smaller, homogeneous ones.

Conclusions:

• Dataset homogeneity is a critical factor for developing reliable ML models for HLM metabolic stability prediction.
• Careful dataset curation and selection are essential for building generalizable drug discovery tools.
• Findings guide the development of more accurate and cost-effective PK/PD modeling.