Depression impacts around 280 million people worldwide. It is defined by enduringsadness and a persistent loss of interest. Limited access to treatment due to high costsand availability issues highlights the need for affordable early detection methods. Machine learning has shown promise in detecting depression, especially using text datafrom social media, where users share emotions openly. This study investigates the useof BERT, a transformer model, combined with the Grey Wolf Optimizer (GWO) todetect depression in tweets by applying a professionally re-labelled Kaggle dataset toenhance early detection. The optimized parameters include pre-trained models, batchsizes, and learning rates. This study reveals that the GWO significantly enhancesthe performance of BERT in text-based depression detection. The best performanceis achieved using BERT optimized by GWO; it is outperforming when using BERTalone. The best parameter combination, which achieves the best validation f1-score,is a model name called bert-base-cased-finetuned-mrpc, batch size of 64, and learningrate of 0.0001. The testing set results an accuracy of 0.8400 and precision, recall, andf1-score of 0.8356.