The invention belongs to the field of gene expression spectrum classification, discloses a gene expression spectrum distance measurement method based on deep learning, and belongs to the mining and application of deep learning on biological big data. First, a convolutional neural network model suitable for gene feature metric learning is designed to extract the characteristics of the data, and then the distance between the data is calculated by using the improved cosine distance, and finally the classification effect of the classification algorithm is used to measure the performance of the method excellent. This method can quickly and efficiently measure the similarity between different gene expression profiles, and provide data for subsequent studies such as gene classification, clustering, differential expression analysis, and compound screening. Compared with the traditional gene enrichment method, this method significantly improves the distance measurement effect between data, and can effectively reduce the manual intervention in gene expression profile analysis, avoiding the over-fitting phenomenon that is easy to occur in conventional deep networks. This method has strong transferability.