With the continued exploration of biological resources in extreme environments, functional proteins such as acidophilic proteins have attracted increasing attention. These proteins can maintain structural stability and biological functionality under highly acidic conditions (pH < 3), demonstrating significant application potential. However, the current identification of acidophilic proteins still relies on labor-intensive, time-consuming, and costly wet-lab experiments. Existing shallow machine learning methods (e.g., SVM, RF) are limited by their constrained model capacity─characterized by fewer parameters and shallow architectures─which restricts their ability to capture complex sequence-function relationships in acidophilic proteins. To address this issue, we propose an integrated computational framework, AcidProNet, which incorporates three key components: a CNN-based generative adversarial module, DCGAN-GP, for data augmentation; a sparsely and discretely activated, parameter-sharing Mixture-of-Experts Transformer, PS-MoE, designed to effectively optimize encoded features; and the utilization of a pretrained protein language model, ESM C, for extracting biologically meaningful protein embeddings. On an independent test set, our method outperforms existing models. Further expert-grouped and ablation experiments confirm its advantages in model stability and representational power. This study is the first to integrate embedding generation, data augmentation, and expert modeling within a unified framework, providing an efficient and scalable approach for functional protein prediction and laying a methodological foundation for advances in protein engineering applications. Additionally, we have developed a Web site for direct identification of acidophilic proteins, which can be accessed at http://www.bioai-lab.com/AcidNetPro.