Given recent technological advances in proteomics, it is now possible to quantify plasma proteomes in large cohorts of patients to screen for biomarkers and to guide the early diagnosis and treatment of depression. Here we used CatBoost machine learning to model and discover biomarkers of depression in UK Biobank data sets (depression n = 4,479, healthy control n = 19,821). CatBoost was employed for model construction, with Shapley Additive Explanations (SHAP) being utilized to interpret the resulting model. Model performance was corroborated through 5-fold cross-validation, and its diagnostic efficacy was evaluated based on the area under the receiver operating characteristic (AUC) curve. A total of 45 depression-related proteins were screened based on the top 20 important features output by the CatBoost model in six data sets. Of the nine diagnostic models for depression, the performance of the traditional risk factor model was improved after the addition of proteomic data, with the best model having an average AUC of 0.764 in the test sets. KEGG pathway analysis of 45 screened proteins showed that the most significant pathway involved was the cytokine-cytokine receptor interaction. It is feasible to explore diagnostic biomarkers of depression using data-driven machine learning methods and large-scale data sets, although the results require validation.