Protein Subcellular Localization Prediction based on Residual Kolmogorov-Arnold Networks

Protein subcellular localization (PSL) prediction is a crucial cross-disciplinary task that provides insights into the function, regulation, and interactions of proteins within cellular systems. Understanding the precise localization of proteins in different tissue types can facilitate a deeper comprehension of cell-to-cell communication mechanisms and contribute to the identification of potential cancer biomarkers. Immunohistochemistry (IHC) images from the Human Protein Atlas (HPA) database offer a valuable resource for visualizing and analyzing the PSL of human proteins across various tissues. Recently, various deep-learning-based methods have been proposed for PSL on IHC images and have achieved some improvements on large-scale IHC datasets. However, the deep-learning-based PSL methods suffer from their high computational burden, primarily due to their reliance on Multi-layer perceptrons (MLPs) as the fundamental network units. To alleviate this issue, inspired by the Kolmogorov-Arnold Networks (KANs), we introduce the Kolmogorov-Arnold network structure into deep learning modeling for PSL and propose a novel Residual Kolmogorov-Arnold Network (ResKANet). Specifically, the enhanced residual module is employed to acquire deep features from the IHC image, which serves as input for KAN to capture global information. By relying on a spline interpolation function, KAN effectively fits the objective function with fewer parameters, extracting more comprehensive information when provided with identical deep feature inputs and enhancing model prediction performance. Consequently, the ResKANet achieves a new state-of-the-art performance on the IHC benchmark for PSL.