Neural networks (NN) have been proposed as a method for detecting and characterizing gene-gene and gene-environment interactions. However, most applications of NN use a fixed set of genetic loci as input and a fixed number of hidden layer nodes and node connections. With these NN, only the weights of the network connections are optimized. This greatly limits the flexibility of the NN which in turn limits the power of the NN to identify functional loci. The goal of this study was to develop a NN approach that uses machine learning for optimizing both the selection of genetic loci as input and the NN architecture in addition to the NN weights. Essentially, this method evolves the structure of the neural network and identifies the functional loci out of a pool of many candidates. The best models identified by the NN are then evaluated using 10-fold cross validation and permutation testing. Using simulated data, we have demonstrated that this approach improves the identification of gene-gene interactions and facilitates interpretation of multilocus results. This study supports the idea that NN will be an important statistical tool for the study of complex multifactorial diseases.
|Original language||English (US)|
|Number of pages||1|
|Journal||American Journal of Medical Genetics - Neuropsychiatric Genetics|
|State||Published - Oct 8 2001|
All Science Journal Classification (ASJC) codes
- Psychiatry and Mental health
- Cellular and Molecular Neuroscience