Protein Subcellular Localization Prediction Based on Attention Mechanism and Multi-Channel Spatial Modeling for Imbalanced Data

Determining the subcellular localization of proteins accurately is crucial for understanding protein functions and molecular processes. Immunofluorescence (IF) images, which involve targeting specific proteins and observing their distribution patterns, have been widely used for inferring their locations. In this study, a novel computational protocol was developed to predict protein subcellular locations from IF images. Firstly, class-imbalance issues were addressed through the utilization of resampling and data augmentation tactics. Additionally, we enhanced the DenseNet121 model with the Convolutional Block Attention Module (CBAM). The Focal loss function was also employed to focus on minority class samples. Our method achieved an overall accuracy of 90.67% on the Human Protein Atlas (HPA) benchmark dataset, providing reliable support for protein subcellular localization.