In the modern medical field, big data analytics play a crucial role in disease diagnosis. Taking cancer as an example, with its multiple subtypes, relying solely on a single omics dataset often proves insufficient for accurate subtype identification. Many existing methods focus on local feature extraction followed by clustering, but they often fail to capture global dependencies across different data omics, limiting their effectiveness. To address these limitations, we introduce a multi-omics clustering approach that integrates convolutional autoencoders with prototype alignment (CAPA). This approach employs convolutional autoencoders to extract local features from each omics, while leveraging a multi-head attention mechanism to capture global dependencies. In the pre-training stage, feature reconstruction is performed, followed by a prototype alignment module that addresses prototype deviation across omics, reducing differences and ensuring cross-omics consistency. Our approach employs a joint optimization strategy that integrates reconstruction loss, prototype alignment loss, and KL divergence loss to improve the stability and accuracy of clustering results. According to the experiment, CAPA performs well across multiple cancer multi-omics datasets, significantly enhancing the accuracy of cancer subtype identification, and providing a comprehensive and multidimensional approach to personalized healthcare.