In the digital age, there is a growing demand for accurate online mental health support, yet current platforms struggle with text-based interaction analysis. This research introduces MeTAN, an innovative approach for automatic depression detection in text, offering a private and convenient method for individuals to assess their mental health early on, before professional engagement. Unlike traditional black-box deep learning methods focused solely on classification, MeTAN prioritizes explainability in health research, which is critical for high-stakes decisions in mental health. It leverages a novel encoder that integrates hierarchical attention mechanisms, metaphorical interpretation, and temporal features to detect depression and identify key textual indicators in tweets. MeTAN aims to assist psychologists by detecting and interpreting emotional patterns in social media text, thereby enhancing diagnosis in virtual settings where anonymity is paramount. Experimental results demonstrate that MeTAN outperforms existing approaches with fewer parameters, showcasing its efficacy in depression detection.