A deep learning framework for high-definition prediction and interpretation of protein localization

Proteins are molecules that perform diverse functions in various cell compartments and subcellular organelles. Aberrant localization of proteins may lead to harmful effects on cells, including poorly functional traits in plants, or disease in humans and animals. Protein localization is a complicated biological process controlled by many factors.

Thus, for most proteins, their localization mechanisms are not well understood. Moreover, experimental methods for measuring the degree of protein localization are time- and labor-consuming. Therefore, it is of great significance to develop protein localization analysis methods.

Current computational methods are often lacking in accuracy to quantify protein localization at the suborganelle level. In addition, most methods lack capacities to predict the effects of mutations on protein localization, or reveal target signals and provide information important for elucidating the mechanism of this process. This project will help address this gap in methos by developing an interpretable deep-learning approach and related informatics infrastructure for protein localization studies.

The outcome will not only improve the protein localization prediction accuracy and resolution, but also shed light on localization mechanisms. Furthermore, the deep-learning framework can be applied to several other bioinformatics problems, such as mRNA localization prediction, enzyme classification and active site prediction, and protein function prediction.

The project will also provide training and research experience in machine learning and real-world software development for various students, especially underrepresented minorities. Virtual workshops and community-wide competitions will be hosted annually to provide machine learning training for students with different backgrounds.

The project will develop a deep-learning framework that incorporates a sequence-based neural network and graph neural network for suborganellar protein localization prediction, together with applications of machine-learning attention mechanisms. The framework’s interpretability will enable studies of high-definition localization mechanisms, including potential novel targeting signal identification and prediction of mislocalization driven by mutation or regulatory alteration.

In addition, the framework will be extended to study tissue-specific or cell-type-specific localization by incorporating single-cell data. An all-in-one web portal for protein localization prediction will be developed to provide a code-free environment for protein localization analysis. All the functionalities, and related data to be used and generated in this project will be provided on the platform.

The web resource will also be an educational tool for AI learning and practices at various levels, such as high-school biology, together with many visualization and playground features. The prediction web services, as well as project progress and training materials will be provided at https://www.mu-loc.org/.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.