Depression presents a significant public health challenge, with recent advancements highlighting the need to investigate potential neuroimaging biomarkers underlying its complex neurobiological effects. This paper introduces Diffusion with Transformers for Depression (DiT-Dep), an innovative framework that utilizes DiT for effective depression detection and biomarkers identification. By modeling neuroimaging data as graphs, DiT-Dep optimizes the classification performance through a dual-objective training strategy. Moreover, a novel entropy-based attention refinement mechanism is introduced to enhance the model’s ability to learn discriminative features, coupled with perturbation-based post hoc explanation methods that clarify the relationships between functional brain networks and depression. Evaluations across multiple datasets reveal that DiT-Dep significantly outperforms leading baselines, achieving superior detection performance while also providing meaningful insights into the replicable and verifiable neuroimaging biomarkers associated with depression, thereby underscoring the promise of using advanced AI methodologies for scientific research in psychiatry. The code is available at: https://github.com/RosalindFok/DiT-Dep.git .