Abstract
BACKGROUND: Traditional biological experiments for protein subcellular localization are costly and inefficient, while sequence-based methods fail to capture spatial dynamics of protein translocation. Existing deep learning models primarily rely on convolutions and lack global image integration, particularly in small-sample scenarios. METHODS: We propose ProteinFormer, a novel model integrating biological images with an enhanced pre-trained transformer architecture. It combines ResNet for local feature extraction and a modified transformer for global information fusion. To address data scarcity, we further develop GL-ProteinFormer, which incorporates residual learning, inductive bias, and a ConvFFN. RESULTS: ProteinFormer achieves state-of-the-art performance on the Cyto_2017 dataset for both single-label (91% [Formula: see text]-score) and multi-label (81% [Formula: see text]-score) tasks. GL-ProteinFormer demonstrates superior generalization on the limited-sample IHC_2021 dataset (81% [Formula: see text]-score), with ConvFFN improving Accuracy by 4% while reducing computational costs. CONCLUSION: ProteinFormer and its GL-ProteinFormer variant show superior performance over existing convolution-based methods. By fusing biological images with transformer-based global feature modeling, the proposed approach offers a robust and efficient solution for protein subcellular localization, especially in data-limited settings.