Subcellular localization enables and determines protein function. Protein shuttling between the nucleus and the cytoplasm is fundamental for intracellular processes as well as responses to external signals and is usually considered to be coded by specific sequences, including nuclear localization signals and nuclear export signals. However, the weak pattern of these sequence signals presents a challenge in accurate decoding for protein nuclear localization. In this study, we constructed a deep learning model named pNuLoC based on a hybrid convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) architecture with an attention mechanism to infer protein sequence determinants essential for nuclear localization. pNuLoC could ab initio identify nucleus-cytoplasm shuttling-related protein regions, including classical nuclear localization/export signal (NLS/NES) and unconventional sequences critical for nuclear localization, with state-of-the-art performance. With this powerful tool, we systematically analyzed the dysregulated localization driven by genetic variations in cancer, and then validated some of the prediction results through experiments, among which the disrupted nuclear localization of KAT8 by the R144C mutation was found to inhibit its function of promoting cancer progression. Thus, pNuLoC could greatly contribute to dissecting the molecular details of nuclear localization and reveal the protein variants mediating dysregulated localization and dysfunction in cancers and diseases.