Bayesian optimization-based zinc binding protein action site mixing depth prediction method and program product

Abstract

The invention relates to the field of artificial intelligence, in particular to a Bayesian optimization-based zinc binding protein action site mixing depth prediction method and device. The method comprises the following steps: based on protein sequence information, constructing an LSTM depth prediction sub-model LSTMIAM (Long Short Term Memory with Integrating Attention Multimedia) integrated with a self-attention mechanism and a machine learning component learner (Mach Learning Component Learners), and then dynamically setting a weight and adjusting parameters by using a Bayesian weighted optimization algorithm so as to enable a combined model to be optimal. The method provided by the invention is high in prediction accuracy and has good generalization and robustness.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and program product for predicting the mixed depth of zinc-binding protein action sites based on Bayesian optimization. Background Art

[0002] The binding of proteins to metal ions plays a crucial role in various physiological and biochemical processes in organisms. Among the numerous metal ions, zinc ions are particularly important. Zinc is the transition metal second only to iron in abundance in the human body. It is involved in the catalytic processes of over 300 enzymes and plays an important role in processes such as DNA synthesis, cell division, and immune function. Zinc ions bind to specific amino acid residues to form stable metalloprotein complexes, thereby regulating the biological functions of these proteins. Accurately identifying these binding sites not only helps understand the functions and mechanisms of proteins but also has important implications for drug design, disease treatment, and biotechnology applications.

[0003] Over the years, scientists have achieved numerous important results in studying protein-metal ion binding sites using biophysical methods, including X-ray crystallography, nuclear magnetic resonance (NMR), and spectroscopy. These techniques provide detailed atomic-level information on the binding patterns and structural features of metal ions in proteins.

[0004] However, although biophysical methods have powerful capabilities in analyzing the interactions between proteins and metal ions, these methods often require complex experimental conditions, expensive equipment, and a lot of time. In addition, the acquisition and interpretation of experimental data often face great challenges, especially when studying complex multi-metal binding proteins. Therefore, computational prediction methods have emerged as an efficient and economical alternative for studying protein-metal ion binding sites. In recent years, with the rapid development of machine learning and deep learning technologies, researchers have developed a variety of computational prediction models. These models use a large amount of known protein-metal ion binding data and learn the rules through training algorithms to predict unknown binding sites.

[0005] Despite progress in studying protein-zinc binding, many challenges remain. The diversity and complexity of proteins, coupled with the high specificity of zinc-binding sites, complicate prediction. A hybrid deep prediction method for zinc-binding protein binding sites based on Bayesian optimization offers new insights and approaches for research in this area. Summary of the Invention

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0007] The present invention provides a method for predicting the mixed depth of zinc-binding protein interaction sites based on Bayesian optimization, comprising the following steps:

[0008] S1: Obtain protein sequence feature information;

[0009] S2: Use the protein sequence feature information to train the LSTM deep prediction sub-model LSTM_IAM integrated with the self-attention mechanism to obtain the trained LSTM_IAM;

[0010] S3: using the sequence feature information of the protein to train all machine learning models in the machine learning component learner; the machine learning component learner includes more than one machine learning model;

[0011] S4: Multiply the output of the trained LSTM_IAM and the output of each machine learning model in the trained machine learning component learner by a weight, and add them up as the output result;

[0012] Among them, the weight of each model is calculated through the Bayesian optimization algorithm.

[0013] Preferably, the sequence feature information of the protein includes PSSM, WOP and RW-GRMTP attribute feature values.

[0014] Preferably, it is characterized in that LSTM_IAM includes an input layer, a hidden layer, a self-attention module, and an FC layer connected in sequence, wherein the FC layer is a fully connected layer.

[0015] Preferably, the machine learning models in the machine learning component learner include SVM, random forest and gradient boosting algorithms.

[0016] Preferably, when using the Bayesian optimization algorithm to calculate the weights of each model, the optimization goal is to maximize the F1-score value.

[0017] The present invention also provides a computer program product, comprising a computer program, which implements the above method when executed by a processor.

[0018] Beneficial effects:

[0019] The present invention first extracts the characteristic attributes PSSM, WOP, and RW-GRMTP from protein sequences. These attributes can effectively capture the important characteristic information of zinc binding sites in protein sequences. Next, these characteristic data are preprocessed, normalized, and denoised to ensure their quality and consistency. Next, these processed characteristic data are used as input to an LSTM deep prediction sub-model with an integrated self-attention mechanism and a machine learning component learner. The sub-models are trained independently and predictions are then performed. Finally, a Bayesian optimization module is used to perform weighted voting on the prediction results, outputting the final prediction result.

[0020] The present invention can fully utilize the advantages of different sub-models to improve the accuracy and stability of predictions. In the LSTM deep prediction sub-model integrated with the self-attention mechanism, the LSTM network can effectively process long dependencies in sequence data, and the self-attention mechanism can further enhance the model's ability to capture global information. The integration of the two can better understand the complex relationships and patterns in protein sequences. The machine learning component learner achieves high accuracy and generalization performance that is significantly superior to that of a single learner by utilizing the unique advantages of different base classifiers. The results of the LSTM deep prediction sub-model integrated with the self-attention mechanism and the machine learning component learner are then weighted and optimized using a Bayesian optimization algorithm to capture the common advantages of both the deep model and the base learner. The optimized optimal weight combination is used for weighted voting to obtain the optimal prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0022] Figure 1 It is the overall framework diagram of the method of the present invention.

[0023] Figure 2 This is the structural diagram of the LSTM deep prediction sub-model that integrates the self-attention mechanism.

[0024] Figure 3 These are the ROC curves and recall precision curves of different methods based on the Passerini dataset. DETAILED DESCRIPTION

[0025] In order to make the purposes, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] like Figures 1 to 3 As shown, a hybrid depth prediction method for zinc-binding protein interaction sites based on Bayesian optimization specifically includes the following steps:

[0027] S1: Obtaining sequence feature information of a protein; the sequence feature information includes PSSM, WOP, and RW-GRMTP attribute feature values;

[0028] In this example, the protein sequences in the fasta file were batch processed using Python tools, and the Swissprot database was searched using PSI-BLAST software to perform sequence alignment to obtain conservation information, thereby obtaining the PSSM, WOP, and RW-GRMTP property feature values.

[0029] S2: Use the protein sequence feature information to train the LSTM deep prediction sub-model LSTM_IAM integrated with the self-attention mechanism to obtain the trained LSTM_IAM;

[0030] In this embodiment, LSTM_IAM includes an input layer, a hidden layer, a self-attention module, and an FC layer connected in sequence, where the FC layer is a fully connected layer;

[0031] S3: using the sequence feature information of the protein to train all machine learning models in the machine learning component learner; the machine learning component learner includes more than one machine learning model;

[0032] In this embodiment, the machine learning models in the machine learning component learner include SVM, random forest, and gradient boosting algorithms;

[0033] S4: Multiply the output of the trained LSTM_IAM and the output of each machine learning model in the trained machine learning component learner by a weight, and add them up as the output result;

[0034] Among them, the weight of each model is calculated by Bayesian optimization algorithm;

[0035] In this embodiment, when the Bayesian optimization algorithm is used to calculate the weights of each model, the optimization goal is to maximize the F1-score value.

[0036] This method takes into account that the LSTM network is responsible for short-term and local dependency processing, and the self-attention mechanism enhances the model's understanding of global and long-distance dependencies. This paper adds a self-attention mechanism module between the hidden state output layer (LSTM_out) and the fully connected layer (FC) of the LSTM, integrates the two, and proposes an LSTM deep prediction sub-model LSTM_IAM that integrates the self-attention mechanism to achieve their complementarity and enhancement. This sub-model can not only capture the correlation of sequence information, but also effectively process global feature dependencies, improve the processing ability of the protein sequence attribute feature matrix, and make the model more robust. The structure of the LSTM deep prediction sub-model with integrated self-attention mechanism is as follows: Figure 2 shown.

[0037] In order to make full use of the strong learning ability and unique advantages of some classic base classifiers, this paper selects SVM, random forest and gradient boosting algorithm as component learners.

[0038] In the machine learning component Learner, the SVM classifier effectively handles the high-dimensional features of protein properties by selecting an appropriate kernel function, avoiding the curse of dimensionality. Furthermore, by optimizing the decision boundary, the SVM can find the optimal hyperplane in the feature space, achieving high-precision classification. Random Forest automatically selects the most discriminative features in the protein feature matrix by ensembling multiple decision trees, helping the model optimize without manual feature selection. Random Forest calculates a feature importance score to quantify the contribution of each feature to the prediction result. This feature importance metric calculates the importance of each feature at the split point across all trees, providing a deep understanding of the protein feature matrix. In the gradient boosting algorithm, each model training cycle focuses on samples that performed poorly in the previous model iteration. By adjusting sample weights, misclassified protein samples are given more weight, thereby giving greater attention to these difficult samples in subsequent model training. This is particularly helpful for handling complex and unbalanced data in the protein feature matrix, enabling the model to better distinguish difficult-to-predict protein sequence information. By using a weighted loss function, the gradient boosting algorithm assigns higher weights to minority class samples, ensuring that the model does not overlook them. This mechanism improves the model's sensitivity to minority classes, thereby improving the overall performance of the protein binding site prediction task.

[0039] This machine learning component, the learner, can effectively reduce the risk of overfitting and enhance generalization capabilities. It can also smooth out fluctuations in individual model predictions and, after being weighted by the Bayesian optimization algorithm, improve the stability and consistency of the overall model.

[0040] The basic idea of ​​Bayesian optimization is to select the next optimal parameter point by constructing a probabilistic model of the objective function to maximize or minimize the objective function value. This paper uses the Bayesian optimization algorithm to find the weights of the optimal component learners by calculating the maximum F1-score value.

[0041] Verification example:

[0042] To verify the classification performance of different classification algorithms, we compared the performance of our proposed method, ZnSite_HDPM_Bayes, with four other methods: LSTM_IAM, random forest, SVM, and gradient boosting. The experimental results are shown in Table 1. As shown in Table 1, the accuracy, recall, Mcc, F1-score, and G-mean evaluation metrics of our proposed algorithm are all higher than those of the other four algorithms, indicating that our proposed method is significantly superior to the other four methods.

[0043] Table 1 Performance index values ​​based on the Passerini dataset

[0044]

[0045] Compared to the component learner, the ZnSite_HDPM_Bayes method achieved superior performance in metrics such as Accuracy, Recall, MCC, F1-score, and G-mean, reaching 0.882, 0.863, 0.765, 0.882, and 0.882, respectively. This represents improvements of at least 1.97%, 1.77%, 4.79%, 1.97%, and 1.97% over the component learner. Although the Precision and Specificity of the ZnSite_HDPM_Bayes method were slightly lower than those of the SVM, they were only lower by 0.001 and 0.01. This result indicates that the ZnSite_HDPM_Bayes method exhibits greater reliability in distinguishing positive and negative samples, making it suitable for a wide range of protein-zinc binding site prediction tasks.

[0046] In the performance comparison experiment based on the passerini dataset, the AUC values ​​of LSTM+SA, Random Forest, SVM, GDBoosting and ZnSite_HDPM_Bayes were 0.932, 0.926, 0.929, 0.937 and 0.939 respectively, and the AUPRC values ​​were 0.938, 0.936, 0.932, 0.939 and 0.945 respectively. ZnSite_HDPM_Bayes achieved the maximum value in both AUC and AUPRC values, which were at least 0.21% and 0.64% higher than the base classifier respectively. This result shows that the proposed method performs better in distinguishing positive and negative samples and handling imbalance problems, and has higher practical value. The ROC curve and recall accuracy curve of ZnSite_HDPM_Bayes and its component learners are shown in the figure below. Figure 3 (a) and 3(b).

[0047] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for predicting the mixed depth of zinc-binding protein interaction sites based on Bayesian optimization, characterized in that: The following steps are involved: S1: Obtain protein sequence feature information; S2: Use the protein sequence feature information to train the LSTM deep prediction sub-model LSTM_IAM integrated with the self-attention mechanism to obtain the trained LSTM_IAM; S3: using the sequence feature information of the protein to train all machine learning models in the machine learning component learner; the machine learning component learner includes more than one machine learning model; S4: Multiply the output of the trained LSTM_IAM and the output of each machine learning model in the trained machine learning component learner by a weight, and add them up as the output result; Among them, the weight of each model is calculated through the Bayesian optimization algorithm.

2. The method for predicting the hybrid depth of zinc-binding protein interaction sites based on Bayesian optimization according to claim 1, characterized in that: The sequence feature information of the protein includes PSSM, WOP and RW-GRMTP attribute feature values.

3. The method for predicting the mixed depth of zinc-binding protein interaction sites based on Bayesian optimization according to claim 1, characterized in that: LSTM_IAM includes an input layer, a hidden layer, a self-attention module, and an FC layer, which are connected in sequence. The FC layer is a fully connected layer.

4. The method for predicting the mixed depth of zinc-binding protein interaction sites based on Bayesian optimization according to claim 1, characterized in that: The machine learning models in the machine learning component learner include SVM, random forest and gradient boosting algorithms.

5. The method for predicting the mixed depth of zinc-binding protein interaction sites based on Bayesian optimization according to claim 1, characterized in that: When using the Bayesian optimization algorithm to calculate the weights of each model, the optimization goal is to maximize the F1-score value.

6. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images