DLS-SUC：タンパク質言語モデル（ESM-2）と二重の不均衡戦略を統合したリジン修飾部位の精密予測フレームワーク

Lysine succinylation (Ksucc), a negatively charged post-translational modification, plays a crucial role in regulating protein function, cellular signaling, and disease pathogenesis, including cancer, neurodegenerative disorders, and metabolic syndromes. Accurate identification of Ksucc sites is therefore vital for elucidating their molecular mechanisms and guiding novel diagnostic and therapeutic developments. However, current deep learning models still face challenges in achieving high predictive accuracy and generalizability. To address these issues, we propose DLS-SUC, a novel deep learning framework for precise Ksucc site prediction. DLS-SUC integrates One-hot encoding with ESM-2 pretrained protein language model features and combines Dense Convolutional Network (DenseNet) and Bidirectional Long Short-Term Memory (BiLSTM) architectures to capture both local sequence motifs and long-range dependencies. Furthermore, a Squeeze-and-Excitation Network (SENet) attention mechanism is incorporated to adaptively recalibrate feature channel importance, enhancing feature discrimination. To mitigate class imbalance, DLS-SUC employs a dual "algorithm-system" strategy, utilizing a weighted cross-entropy loss function to improve learning on the minority class and a cross-validation-based homogeneous ensemble to reduce variance and bolster model stability. The experimental results demonstrate that on an independent test set, DLS-SUC achieved an Sn of 75.96 %, Sp of 77.66 %, BAcc of 76.81 %, MCC of 44.9 %, and AUC of 84.73 %. The model outperforms current state-of-the-art methods, achieving a 7.14 percentage point gain in MCC. In conclusion, DLS-SUC stands as a reliable computational tool for the high-precision prediction of Ksucc sites, offering a new perspective for in-depth investigations of their biological mechanisms and roles in associated diseases.