Depression is a highly prevalent and recurrent mental disorder, and early screening remains challenging in clinical practice. Existing screening methods based on physiological signals often suffer from underutilization of multimodal information, a lack of interpretability in the decision-making process, and limited accuracy and generalization. This study proposes an interpretable sleep staging method for depression screening based on cross-modal attention mechanisms. The proposed method constructs epoch- and sequence-level cross-modal Transformers, integrates EEG and EOG signals, and leverages both cross-modal and intramodal attention to achieve high-precision sleep staging and obtain visualized staging criteria. Subsequently, seven sleep architecture features are extracted from the staging results and fed into a depression screening model with feature selection capabilities to achieve depression screening and explanation of key features. On the publicly available dataset Sleep-EDF, the sleep staging model achieves an accuracy of 83.57% and a Macro F1 score of 79.07%. In a depression screening task on a self-built dataset, the screening model achieves an accuracy of 94.29%. This study constructs an end-to-end interpretable analysis path from raw PSG signals to depression screening, providing a new technical solution for objective and explainable auxiliary screening of depression.