COMP 345 |
| We have developed robust QSAR models of Blood-Brain Barrier (BBB) permeability using k-Nearest Neighbors and Support Vector Machines approaches and molecular topological descriptors. The modeling set of 175 compounds was divided into external evaluation set (20 compounds) and multiple training and test sets (the remaining 155 compounds). The consensus QSAR model accuracies were q2=0.96, R2=0.96, and R2=0.85 for training, test, and external evaluation sets, respectively. These models were applied to additional external evaluation sets consisting of 124 drugs (from the WOMBAT-PK dataset) and 299 organic compounds classified as permeable (BBB+) or non-permeable (BBB-), and the prediction accuracies were 70% and 59%, respectively. Noticeable prediction accuracy improvements were achieved after excluding outliers from the modeling set and applying applicability domain threshold for the prediction of evaluation sets: the accuracy for the first external evaluation set increased to R2=0.94 and for the two additional external sets to 92% and 83%, respectively. |
|
Poster Session
6:00 PM-8:00 PM, Tuesday, August 21, 2007 BCEC -- Ballroom Foyer, Poster
Sci-Mix
Division of Computers in Chemistry |