Identifying abnormally expressed genes is a critical step in cancer diagnosis and has attracted significant attention within the biomedical research community. Gene expression datasets typically involve high-dimensional data, which poses major challenges during the pre-processing stage, particularly in maintaining the biological relevance and interpretability of selected genes. Traditional gene selection techniques often struggle with high computational demands and fail to preserve the intrinsic biological meaning of genes. In this study, we present an effective two-phase framework for gene selection and classification tailored for cancer diagnosis. The first phase employs a Variational Autoencoder (VAE), a deep learning-based technique, to reduce data dimensionality while capturing essential gene expression patterns. In the second phase, we utilize an Adjusted Genetic Algorithm (AGA) to search for a subset of informative genes. To further enhance classification performance, we integrate a wrapper-based approach within the AGA to individually classify genes relevant to different cancer types. Our method was evaluated on two publicly available microarray datasets. The experimental results reveal that the proposed framework outperforms several existing approaches in terms of classification accuracy, while maintaining reasonable computational efficiency. The integration of VAE and AGA offers a robust and biologically interpretable approach to gene selection, making it a promising tool for advancing precision oncology. These findings underscore the potential of combining deep learning and evolutionary algorithms for effective biomarker discovery in high-dimensional genomic data.