ImPLoc

ImPLoc predicts protein subcellular localization from immunohistochemistry (IHC) images to support spatial proteomics analyses using Human Protein Atlas (HPA) data.

Key Features:

• Deep Learning Architecture: ImPLoc employs a deep convolutional neural network (CNN) to extract features from IHC images.
• Multi-Instance Multi-Label Model: It uses a multi-instance multi-label framework with a multi-head self-attention encoder to aggregate multiple CNN-derived feature vectors for simultaneous prediction of multiple subcellular locations.
• Benchmark Dataset: A benchmark dataset of 1,186 proteins and 7,855 HPA images covering six subcellular locations was constructed for training and validation.
• Enhanced Prediction Accuracy: Experimental results show that ImPLoc significantly outperforms existing computational methods in prediction accuracy on the benchmark.
• Differential Localization Analysis: The method was applied to a test set of 889 proteins from normal and cancer tissues and identified eight proteins with differential localization at a significance level of 0.05.

Scientific Applications:

• Spatial proteomics: Predicting subcellular localization from IHC images to inform studies of protein function and spatial distribution within cells and tissues.
• Biomarker discovery in cancer: Comparative localization analysis across normal and cancer tissues to identify differentially localized proteins as candidate cancer biomarkers.

Methodology:

Multi-instance deep learning using a CNN for feature extraction and a multi-head self-attention encoder for aggregating feature vectors; trained and validated on the 1,186-protein/7,855-image HPA benchmark and applied differential localization analysis on a test set of 889 proteins with significance threshold 0.05.