DeepSP: A Deep Learning Framework for Spatial Proteomics

The study of protein subcellular localization (PSL) isa fundamentalstep toward understanding the mechanism of protein function. The recentdevelopment of mass spectrometry (MS)-based spatial proteomics toquantify the distribution of proteins across subcellular fractionsprovides us a high-throughput approach to predict unknown PSLs basedon known PSLs. However, the accuracy of PSL annotations in spatialproteomics is limited by the performance of existing PSL predictorsbased on traditional machine learning algorithms. In this study, wepresent a novel deep learning framework named DeepSP for PSL predictionof an MS-based spatial proteomics data set. DeepSP constructs thenew feature map of a difference matrix by capturing detailed changesbetween different subcellular fractions of protein occupancy profilesand uses the convolutional block attention module to improve the predictionperformance of PSL. DeepSP achieved significant improvement in accuracyand robustness for PSL prediction in independent test sets and unknownPSL prediction compared to current state-of-the-art machine learningpredictors. As an efficient and robust framework for PSL prediction,DeepSP is expected to facilitate spatial proteomics studies and contributesto the elucidation of protein functions and the regulation of biologicalprocesses.

Automated summary

The study introduces a new deep learning framework, DeepSP, for protein subcellular localization (PSL) prediction using mass spectrometry-based spatial proteomics data. By capturing detailed changes in protein occupancy profiles across subcellular fractions, DeepSP constructs a new feature map and utilizes a convolutional block attention module to enhance PSL prediction performance. DeepSP demonstrates significant improvement in accuracy and robustness compared to current machine learning predictors, making it a valuable tool for spatial proteomics studies and the understanding of protein function.