BioSeq_Ksite: Multi-perspective feature-driven prediction of protein succinylation based on an adaptive attention module with SSBCE loss strategy

Abstract

Succinylation is a post-translational modification in which a succinyl group is transferred to the lysine residue of a protein, playing a crucial role in regulating both protein structure and cellular function. This paper introduces a novel sequential model, BioSeq_Ksite, designed to enhance succinylation prediction accuracy by integrating an adaptive attention mechanism and a joint loss function. This study first presents a new hybrid feature, ProtFusion, which combines the physicochemical properties of amino acids with pretrained models. Next, this paper introduces an adaptive attention module that enables the model to autonomously identify important features during training. Additionally, a gated network architecture is adopted to create a dual-branch sequential model. Finally, by combining sensitivity, specificity, and cross-entropy loss, a new joint loss function is proposed, which is used for succinylation prediction for the first time and significantly enhances the model's ability to handle class-imbalanced data. Evaluation on the test dataset shows that BioSeq_Ksite outperforms other models in MCC, Sn, AUC, and F1-Score, with a 7.68 % improvement in MCC over the second-best model. It provides an efficient and reliable tool for succinylation research and application. BioSeq_Ksite can be accessed at https://github.com/zzq1124ZHZ/BioSeq_Ksite.