Deep learning and large language models (LLMs) are being investigated to detect depression from social media data. However, traditional deep learning methods such as Long Short-Term Memory (LSTM) are limited by their reliance on feature engineering, resulting in time consuming and hard to capture complex patterns arising from media data. To address this issue, we propose a novel deep-learning approach that is based on Bidirectional Encoder Representations from Transformers (BERT) framework with Transformer Regression (called BERT-TR) for depression severity prediction from media data. In addition, we employ different techniques such as under sampling, oversampling and weighted loss function to ensure robust and generalizable predictions for depression. The data from X, derived from clinical assessments aligned with the fifth edition of the (DSM-5-TR) criteria, is used for investigations. We theoretically and empirically show that BERT-TR can predict depression severity from media data effectively and efficiently.