|
Activity Number:
|
343
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, July 31, 2007 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #308776 |
|
Title:
|
Development of Nearest-Neighbor Classifiers Identifying Dermal Sensitizers Based on a Local Lymph Node Assay Database
|
|
Author(s):
|
Shengqiao Li*+ and Adam Fedorowicz
|
|
Companies:
|
Centers for Disease Control and Prevention and Centers for Disease Control and Prevention
|
|
Address:
|
1095 Willowdale Road MS L4050, Morgantown, WV, 26505,
|
|
Keywords:
|
Nearest Neighbor Classification ; KNN ; QSAR ; Skin Sensitization ; Sensitivity ; Mahalanobis Distance
|
|
Abstract:
|
K Nearest Neighbor classifiers were developed to predict skin sensitization of a new chemical based on a murine local lymph node assay database of 178 organic chemicals. Two filters were compared for pre-selection of molecular descriptors. The Fisher's Discriminant Ratio filter picked a subset of descriptors which turn out to be more discriminatory than those picked by the t-test filter. Then, a step forward search method was implemented to screen out extra descriptors and simplify the classifiers based on leave-one-out accuracy. Euclidean and Mahalanobis distance metrics were also examined and the results showed the Mahalanobis distance was appropriate for this study. The 3-nearest neighbor classifier of 13 descriptors singled out by the above methods has an especially balanced performance with sensitivity of 92% and specificity of 81% for this unbalanced dataset.
|
