IDRnet: A Novel Pixel-enlightened Neural Network for Predicting Protein Subcellular Location Based on Interactive Pointwise Attention

Abstract

Background: Traditional approaches to protein subcellular pattern analysis are primarily based on feature concatenation and classifier design. However, highly complex structures and poor performance are prominent shortcomings of these traditional approaches. In this paper, we report the development of an end-to-end pixel-enlightened neural network (IDRnet) based on Interactive Pointwise Attention (IPA) for the prediction of protein subcellular locations using immunohistochemistry (IHC) images. Patch splitting was adopted to reduce interference caused by tissue microarrays, such as bubbles, edges, and blanks. The IPA unit was constructed with a Depthwise and Pointwise convolution (DP) unit, and a pointwise pixel-enlightened algorithm was applied to modify and enrich protein subcellular location information. Methods: IDRnet was able to achieve 97.33% accuracy in single-label IHC patch images and 88.59% subset accuracy in mixed-label IHC patch images, and outperformed other mainstream deep learning models. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) was adopted to visualize the spatial information of proteins in the feature map, which helped to explain and understand the IHC image's abstract features and concrete expression form. Results: IDRnet was able to achieve 97.33% accuracy in single-label IHC patch images and 88.59% subset accuracy in mixed-label IHC patch images, and outperformed other mainstream deep learning models. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) was adopted to visualize the spatial information of proteins in the feature map, which helped to explain and understand the IHC image's abstract features and concrete expression form.