BACKGROUND Depression is a kind of serious personal and public mental health problem nowadays. Self-report is the main method used to test whether a person is of depression and the severity of depression. However, it is not easy to discover patients with depression as they feel shame to disclose or discuss their mental health conditions with others. Moreover, self-report is time-consuming, and usually leads to miss a certain number of cases. Therefore, automatic discovering patients with depression from other sources, such as social media, attracts more and more attention. Social media as one of the most important daily communication systems connects a large quantities of people including depression patients, and provides one channel to discover depression patients. In this paper, we investigate deep learning methods on depression risk prediction for Chinese microblogs, which has potential to discover patients with depression and to trace their mental health conditions. OBJECTIVE The aim of this study is to explore the potential of some state-of-the-art deep learning methods on depression risk prediction for Chinese microblogs. METHODS Deep learning methods with pretrained language representation models, including BERT, RoBERTa and XLNET, are investigated for depression risk prediction, and are compared with previous methods on an manually annotated benchmark dataset with depression risks at four levels from 0 to 3, where 0, 1, 2 and 3 denote no inclination, mild, moderate and severe, respectively. The dataset is collected from Weibo. We also compare the different deep learning methods with pretrained language representation models in two settings: 1) publicly released pretrained language representation models, and 2) language representation models further pretrained on a large-scale unlabeled dataset collected from Weibo. Precision, recall and F1 score are used as our performance evaluation measures. RESULTS Among the three deep learning methods, BERT achieves the best performance with a micro-averaged F1 score of 0.856. RoBERTa achieves the best performance with a macro-averaged F1 score of 0.424 on depression risks at level 1, 2 and 3, which is new benchmark result on the dataset. The further pretrained language representation models bring improvement over the publicly released ones. CONCLUSIONS In this study, we use deep learning methods with pretrained language representation models to predict depression risk for Chinese microblogs automatically. The experimental results show that the deep learning methods perform better than previous methods, and have greater potential to discover patients with depression and to trace their mental health conditions. CLINICALTRIAL