TIGP (BIO)—DualLoc: Advancing Multi-Compartment Protein Localization with Dual PLMs

Abstract

Accurate prediction of protein subcellular localization is critical for understanding cellular function and disease mechanisms. Existing approaches, such as DeepLoc 2.0, rely on lightweight fine-tuning of protein language models (PLMs) but struggle with multi-compartment localization. We present DualLoc, a novel framework that leverages full-parameter fine-tuning of dual PLMs—ProtBERT, ESM-2, and ProtT5—enhanced with attention and dropout layers to improve multi-label localization across ten cellular compartments. Evaluated via cross-validation on Swiss-Prot and external testing on the Human Protein Atlas, DualLoc consistently outperforms baseline methods. Notably, DualLoc-ProtT5 achieves 0.5872 accuracy, 0.8271 micro-F₁, and 0.7811 macro-F₁, with marked improvements in Matthews correlation for nucleus (+0.13), extracellular space (+0.07), and cell membrane (+0.13). Pointwise mutual information analysis further uncovers biologically meaningful compartment couplings, such as Golgi–endoplasmic reticulum (PMI = 0.25, P < 10⁻⁶), highlighting coordination within the secretory pathway. DualLoc offers a robust foundation for exploring protein multi-localization dynamics.

附件下載