An Effective Approach for Predicting P-value using High-dimensional SNPs data with Small Sample Size

Alzheimer's disease has seriously affected the normal life of elder people. As an important biomarker of Alzheimer's disease, Single Nucleotide Polymorphisms (SNPs) is highly important for exploring the pathogenesis of Alzheimer's disease. Genome-Wide Association Studies are widely used for extracting important SNPs from extensive genetic data. It can provide a P-value to measure the significance of the association between SNPs and Alzheimer's disease the most significantly correlated SNPs are selected as features to determine the patient's status. However, these approaches suffer from some key limitations: (1) high dimensionality of genetic data requires feature selection or dimension reduction before it can be used. (2) genome-wide association studies require sufficiently genetic samples for ensuring the production of reliable P-values. In order to overcome above limits, we propose an automated framework for modeling P-values for high-dimensional and small-sample SNPs feature selection. Specifically, we transform the feature extraction problem into a regression problem and use a neural network to fit the change of P-value on a small number of samples. Secondly, we propose a new loss function to better measure the quality of model predictions. Extensive experimental results show that the proposed method outperforms Genome-Wide Association Studies methods on ADNI-1 dataset.