|
Abstract:
|
The availability of high throughput molecular assays allows us to build multiplatform sample classification models. Molecular features measured from different platforms vary in number (10^4-10^8) and distribution, and may also have different classification abilities. Consequently, a multiplatform classifier built by fitting a single penalized regression model that penalizes features from all platforms equally, may yield a suboptimal classifier. On the other hand, a classifier built by fitting separate models for each platform that are later combined into a single classifier is likely to miss joint cross-platform effects. We propose to construct multiplatform classifiers by fitting a lasso model that performs variable selection on all features at one time, using separate penalty parameters for each individual data type to allow for different contributions from each. Simulation studies reveal that the multi-tuning parameter lasso is robust, and it outperforms other classifiers when individual platforms contain varied number of informative features with similar effect sizes. Finally, we apply our model to a variety of real cancer data sets.
|
