DeepMLP: A Proteomics-Driven Deep Learning Framework for Identifying Mis-Localized Proteins across Pan-Cancer.

Area of Science:

• Proteomics and Bioinformatics
• Cancer Biology
• Computational Biology

Background:

• Accurate protein subcellular localization (PSL) is crucial for protein function; mis-localization is linked to diseases like cancer.
• Existing spatial proteomics methods are costly and complex, limiting large-scale analysis of cancer-related mis-localized proteins (MLPs).

Purpose of the Study:

• To develop a deep learning framework, DeepMLP, for identifying MLPs in cancer using mass spectrometry-based proteomics data.
• To enhance the accuracy and stability of PSL prediction in both normal and tumor conditions.

Main Methods:

• Developed DeepMLP, a proteomics-driven deep learning framework integrating pathway-aware protein representations with dynamic protein-protein interaction (PPI) networks.
• Utilized cross-attention mechanisms for representation construction and graph attention networks for integrating PPI networks.
• Benchmarked DeepMLP against state-of-the-art methods for PSL prediction.

Main Results:

• DeepMLP demonstrated superior accuracy and stability in PSL prediction compared to existing methods.
• Systematically identified potential MLPs across various cancer types, including mis-localized protein kinases.
• Functional enrichment analyses revealed significant involvement of identified MLPs in cancer-related metabolic and signaling pathways.

Conclusions:

• DeepMLP offers a powerful, cost-effective approach for identifying MLPs in cancer.
• The identified MLPs and kinases hold potential as novel biomarkers or therapeutic targets in tumorigenesis.
• This framework advances the understanding of protein mis-localization's role in cancer development.