SimLite‐MAPE: A Lightweight and Efficient PPI Prediction Model With SimAM Attention and VQ‐VAE Architecture

Protein–protein interactions (PPIs) are essential for numerous biological processes and are closely associated with disease mechanisms. Accurate and efficient prediction of PPIs is critical for understanding complex biological systems and for guiding drug discovery. In this study, we propose SimLite‐microenvironment‐aware protein embedding (MAPE), a lightweight PPI prediction model built upon the VQ‐VAE architecture and optimized for both accuracy and computational efficiency. The model introduces a heterogeneous decoder that reduces parameters and floating‐point operations without compromising predictive capability. Furthermore, the parameter‐free SimAM is integrated into the encoder, enabling more effective extraction of protein sequence and structural features. The classification module is enhanced by incorporating additive attention and residual connections, thereby improving stability and robustness. Extensive experiments conducted on multiple benchmark datasets, including STRING, SHS27k, and SHS148k, demonstrate that SimLite‐MAPE consistently outperforms state‐of‐the‐art baseline methods in terms of predictive accuracy, generalization, and efficiency. These results confirm that SimLite‐MAPE provides a scalable and practical solution for large‐scale PPI prediction.