多源特征结合机器学习算法预测钾离子(K⁺)与钠离子(Na⁺)的结合位点

Prediction of Potassium (K⁺) and Sodium (Na⁺) Ion Binding Sites Using Multi-Source Features and Machine Learning Algorithms

摘要: 钾离子(K⁺)与钠离子(Na⁺)是生物体内重要的电解质，在维持细胞渗透压平衡、调节神经信号传导以及参与酶促反应调控中发挥关键作用。准确识别蛋白质中的金属离子结合位点，对于深入理解离子调控机制及相关疾病的分子基础具有重要意义。本文基于BioLiP数据库获取K⁺和Na⁺结合蛋白序列，利用CD-HIT进行序列去冗余处理，并按5:1的比例划分为训练集和测试集。采用SMOTEENN算法对训练集进行类别平衡处理，从序列、结构与能量三个层面共提取9类特征(PSSM、氨基酸组分、密码子频率、相对可及表面积、SASA-RASA、疏水性、二级结构、结合能和图能量)，并分别使用7种机器学习算法(Logistic Regression, SVM, KNN, Random Forest, Gradient Boosting, XGBoost, LightGBM)进行模型构建与性能评估。结果表明，单特征PSSM在K⁺和Na⁺结合位点的预测中均表现最优，其中K⁺结合位点预测的敏感性Sn = 100%，特异性Sp = 85.3%，总精度Acc = 85.6%，AUC值达到0.984；Na⁺结合位点预测的敏感性Sn = 100%，特异性Sp = 86.5%，总精度Acc = 86.6%，AUC值达到0.978。鉴于梯度提升算法在处理非线性关系的能力较强，同时对特征交互的捕捉更高效等优点，随后在LightGBM算法下，采用最优特征PSSM与其他8种特征作逐一融合，结果发现：特征融合后K⁺和Na⁺结合位点的预测精度的各项指标都有一定的提高；同时也发现特征融合不是越多越好，部分特征间存在一定信息冗余，故合理的特征选择与融合策略对模型优化至关重要。本研究对于离子通道蛋白功能解析，靶向药物研发等方面具有一定的生物学意义。

Abstract: K⁺ and Na⁺ are important electrolytes in organisms, which play a key role in maintaining cell osmotic pressure balance, regulating nerve signaling, and participating in the regulation of enzymatic reactions. Accurate identification of K⁺ and Na⁺ binding sites in proteins is of great significance for in-depth understanding of ion regulation mechanisms and the molecular basis of related diseases. In this paper, the sequences of K⁺ and Na⁺ binding proteins were selected from the BioLiP database, and the sequence redundancy was removed by CD-HIT. The sequence is divided into training and test sets according to a 5:1 ratio. Balance training data was the SMOTEENN algorithm, nine types of features (PSSM, amino acid components, codon frequency, relative accessible surface area, SASA-RASA, hydrophobicity, secondary structure, binding energy and graph energy) from three levels (sequence, structure and energy information) were extracted, and seven machine learning algorithms (Logistic Regression, SVM, KNN, Random Forest, Gradient Boosting, XGBoost, and LightGBM) were used to build models and evaluate performance. The results showed that the single-feature PSSM performed the best in the prediction of K⁺ and Na⁺ binding sites, among which the Sn = 100%, Sp = 85.3%, Acc = 85.6%, and AUC value reached 0.984, and the Sn = 100%, Sp = 86.5%, Acc = 86.6%, and AUC value of Na⁺ binding site prediction reached 0.978. In view of the advantages of the gradient algorithm in processing nonlinear relationships and more efficient capture of feature interactions, the optimal feature PSSM was used to fuse with 8 other features one by one under the LightGBM algorithm, and the results showed that the prediction accuracy of K⁺ and Na⁺ binding sites was improved to a certain extent after feature fusion. At the same time, it is also found that using more features does not yield better results, due to information redundancy among some features. So reasonable feature selection and fusion strategies are very important for model optimization. This study has certain biological significance for the functional elucidation of ion channel proteins and the development of targeted drugs.