Prediction of passive blood-brain partitioning: straightforward and effective classification models based on in silico derived physicochemical descriptors

Abstract

The distribution of compounds between blood and brain is a very important consideration for new candidate drug molecules. In this paper, we describe the derivation of two linear discriminant analysis (LDA) models for the prediction of passive blood-brain partitioning, expressed in terms of logBB values. The models are based on computationally derived physicochemical descriptors, namely the octanol/water partition coefficient (logP), the topological polar surface area (TPSA) and the total number of acidic and basic atoms, and were obtained using a homogeneous training set of 307 compounds, for all of which the published experimental logBB data had been determined in vivo. In particular, since molecules with logBB>0.3 cross the blood-brain barrier (BBB) readily while molecules with logBB<-1 are poorly distributed to the brain, on the basis of these thresholds we derived two distinct models, both of which show a percentage of good classification of about 80%. Notably, the predictive power of our models was confirmed by the analysis of a large external dataset of compounds with reported activity on the central nervous system (CNS) or lack thereof. The calculation of straightforward physicochemical descriptors is the only requirement for the prediction of the logBB of novel compounds through our models, which can be conveniently applied in conjunction with drug design and virtual screenings.

Figure 1

Percentages of good classification and statistical parameters for model 1 and model 2. Compounds with log BB values ≥ 3 readily cross the BBB, while those with log BB values < 1 are poorly distributed to the brain.

Figure 2

Plot of the log P vs. TPSA values for compounds with log BB ≥ 0.3 (in white) and compounds with log BB < 0.3 (in black). The demarcation line identified by model 1, indicated by the black diagonal line, provides 80 % of good classification for the two categories of compounds.

Figure 3

Desirability analysis of the descriptors that compose model 2. The green color indicates areas that are predicted to favor log BB ≥ −1, while the red color indicates areas that are predicted to favor log BB < −1.

Figure 4

ROC curves relative to the analysis of an external dataset of compounds with reported CNS activity of lack thereof. The area under the curve is 0.95 and 0.97 respectively, indicating a very high predictive power for both models. A perfect model would have an area of 1, while a random model, denoted by the diagonal line, would have an area of 0.5.