MoRFs_TransFuse: a MoRFs predictor based on multimodal feature fusion and the lightweight Transformer network

Molecular recognition features (MoRFs) can facilitate specific protein-protein interactions by undergoing disorder-to-order transitions when binding to their protein partners. Thus, it is essential to accurately predict MoRFs. In this paper, we propose an innovative MoRFs prediction method, named MoRFs_TransFuse, based on multimodal feature fusion and a lightweight Transformer network. To construct high-quality biological features, MoRFs_TransFuse innovatively integrates physicochemical properties, evolutionary features, and pre-trained model embeddings, while retaining optimal feature combinations through multi-window extraction and Random Forest secondary screening. In terms of architecture, MoRFs_TransFuse overcomes the limitations of modeling long-range dependencies by using a self-attention mechanism to accurately capture long-range residue associations in protein sequences. Comparative experiments on benchmark datasets show that MoRFs_TransFuse significantly outperforms existing single component and combined component predictors. Additionally, the lightweight design greatly improves computational efficiency while ensuring prediction accuracy.