Abstract <p>Deep learning for cancer intelligent diagnosis based on multi-omics data has achieved enormous advance in the condition of sufficient samples. However, these methods can’t generalize to circumstances that certain cancer samples are few, which poses a challenge for few-shot learning to apply in it. Therefore, we propose a few-shot learning algorithm, called the multi-omics feature reconstruction (MOFR), to realize the cancer classification and the metastasis prediction. Specifically, MOFR transforms the classification problem into a feature reconstruction issue in latent space. This approach needs fewer learning parameters, which facilitates model fitting in such environment. By directly regressing the ’distance’ between the support feature set and the query feature, it achieves classification tasks. Additionally, we also design a tailored feature extractor to effectively map raw data into high-level and informative features. It draws inspiration from the Transformer architecture but overcomes the need for extensive training data to train the model. According to extensive experimental results, the proposed few-shot learning algorithm surpasses homogeneous approaches in both classification and metastasis prediction performance. We provide detailed code for the new model proposed in this paper on <a href="https://github.com/vbfeobveofn/Code.git">https://github.com/vbfeobveofn/Code.git</a>.</p> Graphical Abstract <p></p>