Multi-Label Protein Subcellular Localization Using Graph Attention and Self-Recalibrated Feature Representations

Abstract

Accurately predicting protein subcellular localization is essential for understanding biological function and informing medical research. To address the limitations of traditional laboratory techniques, this study introduces two deep learning frameworks—ML-FGAT and ML-GRat—for multi-label protein subcellular localization (ML-PSL). ML-FGAT integrates seven diverse feature encoding schemes—DC, PsePSSM, CTD, GO, CT, DDE, and EBGW [5]—followed by Differential Evolution (DE)-based feature fusion and entropy-guided selection. To enhance representation quality, a self-attention-based feature recalibration (SAFR) module is introduced to emphasize biologically relevant features. A FeatureGenerative Adversarial Network (F-GAN) then balances class distributions, and classification is performed using a Graph Attention Network (GAT). ML-FGAT achieved OAA scores ranging from 93.5% to 98.8% across five test datasets. ML-GRat uses DE for feature weighting and Canonical Correlation Analysis (CCA) for dimensionality reduction, followed by SAFR and a hybrid GAT-ResNet classifier. This model achieved OAA scores between 94.0% and 98.9% on six independent datasets, including SARS-CoV-2 and human proteins [6]. The proposed models demonstrate robust generalization, high predictive performance, and improved interpretability for ML-PSL tasks in computational biology.