Microarray technology offers insights into gene expression levels, facilitating the classification of various diseases, particularly cancers. However, the high-dimensional nature of microarray data, coupled with small sample sizes, poses a challenge for effective classification. This paper introduces a novel feature selection and classification approach using Graph Neural Networks (GNNs), tailored to handle complex gene interdependencies. The proposed method constructs a graph where nodes represent genes and edges represent gene interactions, using both biological knowledge and statistical similarity measures. The GNN framework integrates feature selection with classification, achieving state-of-the-art performance on multiple microarray datasets.