Scholarlink

<p>Malonylation modification of proteins is closely related to many diseases, such as diabetes and cancer. Therefore, accurate identification of malonylation modification sites is crucial for elucidating the molecular mechanisms underlying these diseases. Traditional experimental methods suffer from the problems of high cost, long cycle time, difficulty, etc. With advancements in artificial intelligence, the prediction of protein post-translational modification sites through computational methods has emerged as a vital complement to experimental approaches. In this paper, we present a malonylation site prediction model, Catsoft_Kmalsite, the core innovation of which lies in its integration of complementary information from protein three-dimensional structural features and sequence/physicochemical features, coupled with a soft voting ensemble strategy based on Bayesian-optimized base classifiers. Specifically, we utilize AlphaFold2 to acquire protein tertiary structural information and employ CTDC, EAAC, and EGAAC methods to extract protein sequence and physicochemical features. Subsequently, two base classifiers are constructed using the CatBoost algorithm based on these two distinct feature sets, respectively. Following parameter fine-tuning of the base classifiers via Bayesian optimization, they are ultimately integrated using a soft voting strategy. All ablation experimental results show that the Catsoft_Kmalsite model exhibited good robustness and generalization ability. Across six metrics, including AUC, ACC, Sen, Pre, F1, and MCC, the model achieved average performances of 94.03%, 87.91%, 89.15%, 86.91%, 88.00%, and 0.7585, respectively, in fivefold cross-validation and specific performance of 95.18%, 89.55%, 90.87%, 88.79%, 89.82%, and 0.7912 on the independent test set; Catsoft_Kmalsite also outperformed other state-of-the-art studies in all evaluated metrics. Furthermore, we have developed a website for users to use (<a href="http://1.94.102.146:8501/Catsoft_Kmalsite">http://1.94.102.146:8501/Catsoft_Kmalsite</a>). The code and dataset of Catsoft_Kmalsite are available at <a href="https://github.com/flyinsky6/Catsoft_Kmalsite">https://github.com/flyinsky6/Catsoft_Kmalsite</a>.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing the identification of malonylation sites using AlphaFold2 and ensemble learning

  • 期刊论文

  • 发表日期：

    2025年10月5日

查看全文
  • Linlin Xu,
  • Yuting Qian,
  • Jiayi Yang,
  • Xiaowei Xu,
  • Zhiqiang Li,
  • Yanhan Wang,
  • Enhui Lv,
  • Xingxing Kang,
  • Hongwei Zhang,
  • Yaping Lu,
  • Fei Wang,
  • Xin Liu

摘要

Malonylation modification of proteins is closely related to many diseases, such as diabetes and cancer. Therefore, accurate identification of malonylation modification sites is crucial for elucidating the molecular mechanisms underlying these diseases. Traditional experimental methods suffer from the problems of high cost, long cycle time, difficulty, etc. With advancements in artificial intelligence, the prediction of protein post-translational modification sites through computational methods has emerged as a vital complement to experimental approaches. In this paper, we present a malonylation site prediction model, Catsoft_Kmalsite, the core innovation of which lies in its integration of complementary information from protein three-dimensional structural features and sequence/physicochemical features, coupled with a soft voting ensemble strategy based on Bayesian-optimized base classifiers. Specifically, we utilize AlphaFold2 to acquire protein tertiary structural information and employ CTDC, EAAC, and EGAAC methods to extract protein sequence and physicochemical features. Subsequently, two base classifiers are constructed using the CatBoost algorithm based on these two distinct feature sets, respectively. Following parameter fine-tuning of the base classifiers via Bayesian optimization, they are ultimately integrated using a soft voting strategy. All ablation experimental results show that the Catsoft_Kmalsite model exhibited good robustness and generalization ability. Across six metrics, including AUC, ACC, Sen, Pre, F1, and MCC, the model achieved average performances of 94.03%, 87.91%, 89.15%, 86.91%, 88.00%, and 0.7585, respectively, in fivefold cross-validation and specific performance of 95.18%, 89.55%, 90.87%, 88.79%, 89.82%, and 0.7912 on the independent test set; Catsoft_Kmalsite also outperformed other state-of-the-art studies in all evaluated metrics. Furthermore, we have developed a website for users to use (http://1.94.102.146:8501/Catsoft_Kmalsite). The code and dataset of Catsoft_Kmalsite are available at https://github.com/flyinsky6/Catsoft_Kmalsite.

Graphical abstract