This paper presents a study on the application of self-supervised learning to the problem of predicting phenotypes from gene expression data. Existing supervised learning-based machine learning and deep learning methods require a large amount of labeled data, but obtaining such data is costly and time-consuming in the case of gene expression data. To overcome these limitations, this study selected three state-of-the-art self-supervised learning methods and applied them to bulk gene expression data, and evaluated whether they improve the accuracy of phenotype prediction. Using several public gene expression datasets, we demonstrate that self-supervised learning methods can effectively capture complex information and improve prediction accuracy. We analyze the strengths and limitations of each method, make recommendations for method selection according to application cases, and suggest future research directions. This is the first study to combine bulk RNA-Seq data with self-supervised learning.
