Multi-label Protein Subcellular Localization (PSCL) prediction is an important research area in bioinformatics, and multi-label classification, as a common challenge in PSCL, is particularly crucial. In this work, a multi-label PSCL prediction model is proposed based on the C-Tran architecture, with features extracted by a protein language model. For a protein, its amino acid sequence is encoded into a sequence of feature vectors, while its labels are converted into a group of learnable vectors, which are appended to the feature vectors. During training, the extended sequence is masked and fed into the model, and the model tries to predict the correct labels in the form of a binary vector. In this way, the model can learn the dependencies among amino acids, the correlation among labels, and the interconnection between the protein sequence and its labels. During inference, the features in the same form is sent to the model with all the label vectors masked, without any modification to the architecture.