BACKGROUND: There is a surprising lack of neuroimaging studies of depression that not only identify the whole brain causal connectivity features but also explore whether these features have neurobiological correlates. METHODS: Three graph neural networks (GNN) models were applied to three types of causal connectomes (CCs): granger causality, regression DCM (rDCM), and TwoStep, obtained from a total of 1296 young adult participants in three large-scale datasets. RESULTS: GNN models showed better performance for predicting depression when using causal connectomes such as TwoStep (average precision score, 0.882), granger causality (0.878), or rDCM (0.853) compared with using functional connectomes like Pearson's (0.850) and partial (0.823) correlation. Notably, nodal features derived only from rDCM and TwoStep showed spatial associations with positron emission tomography measures of receptors for neurotransmitters such as dopamine and serotonin. Further analysis revealed the shared directed edges among the subject's edge features, which included cortical causal connections in networks such as the default mode, control, dorsal attention, peripheral visual, and parietofrontal networks. LIMITATIONS: The classification performance of leave-one-site-out cross-validation did not achieve a similar level with that of 10-fold cross-validation. CONCLUSIONS: Our findings suggest that the connectomes derived from CCs using GNN, rather than functional connectomes, provide more accurate and neurobiologically relevant information for depression. Moreover, the observed spatial heterogeneity of this relevance and subject-specific edge features emphasizes the complexity of depression. These results have the potential to advance our understanding of depression's nature and potentially contribute to precision psychiatry by aiding in its diagnosis and treatment.