Discovering the nuclear localization signal universe through a deep learning model with interpretable attention units

Abstract

Abstract Nuclear localization signals (NLSs) are essential peptide fragments within proteins that play a decisive role in guiding proteins into the cell nucleus. Determining the existence and precise locations of NLSs experimentally is time-consuming and complicated, resulting in a scarcity of experimentally validated NLS fragments. Consequently, annotated NLS datasets are relatively limited, presenting challenges for data-driven approaches. In this study, we propose an innovative interpretable approach, NLSExplorer, which leverages large-scale protein language models to capture crucial biological information with a novel attention-based deep network for NLS identification. By enhancing the knowledge retrieved from protein language models with a novel attention to key area module, NLSExplorer achieves superior predictive performance compared to existing methods on two NLS benchmark datasets. Additionally, NLSExplorer is able to detect various kinds of segments highly correlated with nuclear transport, such as nuclear export signals. We employ NLSExplorer to investigate potential NLSs and other domains that are important for nuclear transport in nucleus-localized proteins in the Swiss-Prot database. Furthermore, the comprehensive pattern analysis for all these segments uncovers a potential NLS space and internal relationship of important nuclear transport segments for 416 species. This study not only introduces a powerful tool for predicting and exploring NLS space, but also offers a versatile network that is powerful for detecting characteristic domains and motifs of NLSs.