A recursive-partitioning model for blood-brain barrier permeation.

Area of Science:

• Computational chemistry
• Drug discovery
• Quantitative Structure-Activity Relationship (QSAR) modeling

Background:

• Predicting drug properties like log(BB) is crucial for drug discovery.
• P-glycoprotein (P-gp) mediated efflux presents a challenge in extrapolating predictive models to new chemical spaces.
• Recursive partitioning offers a flexible approach for building predictive models.

Purpose of the Study:

• To develop and evaluate bagged recursive partitioning models for predicting log(BB).
• To assess the impact of physical property descriptors on model performance.
• To investigate the challenges of model extrapolation to broader chemical spaces, particularly concerning P-gp substrates.

Main Methods:

• Bagged recursive partitioning models were employed.
• Leave Group Out Cross-Validation (LGO-CV) was used for model evaluation.
• Three sets of physical property descriptors (CPSA, Ro5x, MOE) were assessed.
• Models were tested on Pfizer chemical space, considering P-gp efflux.

Main Results:

• Q2 values for the models ranged from 0.51 to 0.53.
• Extrapolation to Pfizer chemical space was challenging due to P-gp efflux.
• Removing P-gp substrates improved correlation coefficients (R2 = 0.39).
• Simple linear models showed improved correlation within specific chemical series.

Conclusions:

• Bagged recursive partitioning models can predict log(BB) with moderate success.
• P-gp efflux significantly impacts model extrapolation accuracy.
• Focusing on specific chemical series and excluding P-gp substrates enhances predictive performance.
• Further refinement of QSAR models is needed for broader applicability.