Comparing the Pfizer Central Nervous System Multiparameter Optimization Calculator and a BBB Machine Learning Model.

Area of Science:

• Computational chemistry and cheminformatics
• Neuroscience drug discovery
• Machine learning applications in pharmacology

Background:

• Accurate prediction of blood-brain barrier (BBB) penetration is crucial for neuroscience drug discovery.
• Traditional methods like Pfizer's CNS-MPO offer simplistic rules but limited interpretability and accuracy.
• Sophisticated machine learning (ML) models exist but often function as "black boxes".

Purpose of the Study:

• To implement and compare two versions of Pfizer's CNS-MPO algorithm (Pf-MPO.v1, Pf-MPO.v2).
• To evaluate a Bayesian machine learning model for predicting BBB penetration.
• To assess the predictive accuracy of these models against known CNS active drugs.

Main Methods:

• Developed a Bayesian ML model using extended connectivity fingerprint descriptors on 2296 compounds.
• Compared the Bayesian model's predictions with Pf-MPO.v1 and Pf-MPO.v2 algorithms.
• Validated performance against a set of 40 known CNS active drugs.

Main Results:

• The Bayesian model achieved 92.5% accuracy (37/40 compounds).
• Pf-MPO.v1 achieved 75% accuracy (30/40 compounds) with a score ≥4.
• Pf-MPO.v2 achieved 82.5% accuracy (33/40 compounds) with a score ≥4.

Conclusions:

• The Bayesian ML model demonstrates superior accuracy in predicting BBB penetration compared to MPO algorithms.
• Machine learning models offer greater flexibility and predictive power for BBB penetration than rule-based systems.
• ML models show promise for enhancing drug discovery alongside interpretable methods.