# Integrative Profiling for BBB Permeability Using Capillary Electrochromatography, Experimental Physicochemical Parameters, and Ensemble Machine Learning

## TL;DR

This paper introduces a new in vitro method combining capillary electrochromatography and machine learning to predict how well drugs can cross the blood-brain barrier.

## Contribution

The novel contribution is an integrative high-throughput method for predicting BBB permeability using CEC data and machine learning models.

## Key findings

- A regression model with R2 = 0.64 was developed to predict log BB values using CEC retention factor, pKa, and log D7.4.
- Machine learning classification of CEC electropherograms achieved an accuracy of 0.81 and F1weighted score of 0.8.

## Abstract

Profiling the blood–brain barrier (BBB) permeability of bioactive molecules during early drug development is critical for optimizing their pharmacokinetic profile. The in vivo ability of a compound to cross the BBB is measured by the log BB parameter; however, its determination requires costly and time-consuming animal experiments. This study aimed to develop a novel in vitro method for high-throughput prediction of log BB values. The approach combines experimental data from open-tubular capillary electrochromatography (CEC) and automated potentiometric titrations, including the CEC retention factor (k′), electropherograms, and physicochemical parameters pKa and log D7.4. The k′ parameter reflects BBB permeability using a capillary internally coated with liposomes that mimic a biological membrane. Preliminary CEC analyses were conducted for 25 neutral drugs at pH 7.4, revealing a promising correlation between the permeability parameters log k and log BB. The validation was extended to 57 ionized drugs, with additional determination of pKa and log D7.4. A regression model was developed: log BB = −2.45 + 0.1k′ + 0.3logD7.4 + 0.27pKa (R2 = 0.64). Furthermore, the analysis of CEC electropherograms enabled the machine learning-based rapid classification of compounds using Dynamic Time Warping, k-Nearest Neighbors, and the Bag-of-SFA-Symbols in Vector Space model, yielding an accuracy of 0.81 and an F1weighted score of 0.8.