LocPred-Prok: Prokaryotic protein subcellular localization prediction with a dual-branch architecture and protein language model.

The precise localization of proteins within prokaryotic cells is fundamental to understanding their function. However, existing models still struggle with challenging localization classes, such as cell wall or outer membrane proteins. We introduce LocPred-Prok, a novel deep learning framework that redefines performance standards for prokaryotic subcellular localization. LocPred-Prok employs a purpose-built dual-branch architecture that synergistically integrates global and local sequence features extracted from pLM embeddings. On a stringent, homology-partitioned benchmark, LocPred-Prok achieves a state-of-the-art accuracy of 91.2 % and a Matthews Correlation Coefficient (MCC) of 0.889. Critically, it resolves long-standing prediction challenges, demonstrating exceptional performance on notoriously difficult classes like Gram-positive cell wall and Gram-negative outer membrane proteins. It substantially outperforms recent and classic methods across all organismal subgroups, representing a significant leap forward in the field. The LocPred-Prok web server is freely accessible athttps://huggingface.co/spaces/isyslab/LocPred-Prok.