The identification of drug-target interaction (DTI) plays an important role in drug discovery. Recent research has shown the powerful capability of deep learning based on large datasets in DTI prediction. However, due to the competition, pharmaceutical institutions are reluctant to disclose the screened active compounds against specific disease targets before completing long-term drug development. Federated learning is a machine learning paradigm that allows numerous clients to train a global model without raw data exchange. For DTI prediction, federated learning allows multiple pharmaceutical institutions to jointly train a well-supervised DTI prediction model while safeguarding their intellectual property. In this paper, we propose an efficient, high-performance and privacy-preserving federated learning framework for collaborative DTI prediction. A BERT model is pre-trained on data from the PDB database and used to extract latent semantic representations of proteins. We utilize the Graph Convolutional Network to extract the features from molecule graphs. A novel molecule-specific prompt tuning method is proposed to guide the BERT model in learning the downstream DTI prediction task. As the parameters of the BERT model are frozen, our method improves the efficiency of computation and communication in federated learning. Compared with other methods, our method demonstrates state-of-the-art performance in DTI prediction. We also validate that our method can resist gradient attacks to a certain extent which protects the data security of each client.