Intelligent screening system of mitochondrial effector molecules and construction method and application thereof

Abstract

The application relates to an intelligent screening system of a mitochondrion effect molecule, a construction method and application thereof, and belongs to the technical field of molecular biology. The construction method of the intelligent screening system of the mitochondrion effect molecule comprises the following steps: 1, establishing a target protein library; 2, obtaining a data set of the mitochondrion effect molecule; 3, adopting Morgan molecular fingerprint to characterize the mitochondrion effect molecule in the data set, carrying out deduplication and decontamination processing, and then carrying out molecular similarity processing to obtain an input set of a model; and 4, taking accuracy and AUC value as evaluation indexes to construct a support vector machine model. The application utilizes the support vector machine model to carry out prediction in a large amount of molecular data set, and gives a molecule with a high probability score which may have an effective effect on mitochondria, the model is helpful for researchers in the field of mitochondria to reduce parameter adjustment time and improve work efficiency.

Description

Technical Field

[0001] This invention belongs to the field of molecular biology technology, specifically relating to an intelligent screening system for mitochondrial effector molecules, its construction method, and its application. Background Technology

[0002] Mitochondria are double-membrane organelles found in most eukaryotic cells, often referred to as the cell's powerhouse. As a key organelle for intracellular energy metabolism, mitochondria provide energy for normal cellular activities but are also susceptible to oxidative damage and metabolic disorders, leading to mitochondrial dysfunction, reduced cellular function, and ultimately, mitochondrial diseases. With in-depth research into mitochondria, it has become increasingly clear that they play crucial roles in cellular metabolism, cell growth, cell survival, and signal transduction. Mitochondrial dysfunction is closely related to various physiological processes and diseases, such as aging, immune responses, diabetes, cancer, neurodegenerative diseases, and cardiovascular diseases. Therefore, increasing research focuses on the pathogenic role of mitochondrial damage and how to better maintain and protect mitochondrial function. Currently, many drugs targeting mitochondria have been discovered that can effectively treat mitochondrial diseases. In addition to mitochondrial-targeted drugs, researchers in the field are actively exploring functional molecules that regulate mitochondrial function to help improve mitochondrial dysfunction.

[0003] However, despite the promising preventative and therapeutic effects of mitochondrial-targeted drugs and mitochondrial nutrients on mitochondrial diseases, the discovery and screening of mitochondrial effector molecules is typically time-consuming, a characteristic of traditional biological screening. Traditional drug or molecular screening is a highly complex process, involving the acquisition of protein molecular information using proteomics and microarray technology, bioinformatics analysis, and in vivo experiments. While current biological experimental techniques have rapidly advanced—RNA interference, cell microarray, protein fluorescence labeling, and nuclear magnetic resonance (NMR) have all been used to confirm target proteins and drugs—these traditional methods, constrained by enormous human and material resources, are insufficient for the large-scale, high-throughput screening of molecules acting on specific targets in modern pharmaceutical research and development. With the rapid development of information processing technology, intelligent computing technology has emerged, offering the potential to provide efficient solutions for compound molecular screening through its large-scale analytical capabilities and systematic screening mechanisms. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to design and provide an intelligent screening system for mitochondrial effector molecules, along with its construction method and applications. This invention utilizes machine learning for intelligent screening of mitochondrial effector molecules. From the establishment and input of the training set to model parameter tuning and final prediction on a large dataset, it successfully screens out molecules with potential mitochondrial effects, establishing an intelligent screening system for mitochondrial effector molecules. This invention uses a support vector machine model for prediction on a large molecular dataset and identifies molecules with high probability scores that are likely to have mitochondrial effects. This model helps researchers in the field of mitochondrial research reduce parameter tuning time and improve work efficiency.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for constructing a smart screening system for mitochondrial effector molecules, characterized by the following steps:

[0007] (1) Collect basic information on target proteins of biomolecules and establish a target protein library;

[0008] (2) Based on the target protein library established in step (1) above, IC50 and AC50 values ​​are used as screening criteria for effector molecules to assist the activity of effector molecules in the ChEMBL database, screen effector molecules targeting proteins on mitochondria, and obtain a dataset of mitochondrial effector molecules.

[0009] (3) The mitochondrial effect molecules in the dataset obtained in step (2) above are characterized by Morgan molecular fingerprinting, and after deduplication and decontamination, molecular similarity processing is performed to obtain the input set of the model.

[0010] (4) Using accuracy and AUC as evaluation indicators, a support vector machine algorithm was used for learning and training to construct a mitochondrial effect molecule screening model.

[0011] In the construction method described above, the basic information of the target protein in step (1) includes its name, protein number, mechanism, and location.

[0012] The construction method described above, wherein the target proteins in step (1) include carnitine palmitoyltransferase, long fatty acyl-CoA, kynurenase, monoamine oxidase, coenzyme Q-cytochrome c reductase, cytochrome c, NADH dehydrogenase, succinate dehydrogenase, glycerol-3-phosphate dehydrogenase, adenosine triphosphate synthase, carnitine palmitoyltransferase II, uncoupling protein, MT-ND1, MT-ND2, MT-ND3, MT-ND4, MT-ND5, MT-ND6, MT-CYB, and MT-CO1. MT-CO2, MT-CO3, dihydroorotate dehydrogenase, citrate synthase, cis-aconitase, isocitrate dehydrogenase, α-ketopentyl dihydrogenase complex, succinate-coenzyme A synthase, fumarate, malate dehydrogenase, aspartate aminotransferase, glutamate dehydrogenase, pyruvate dehydrogenase complex, carbamoyl phosphate synthase I, ornithine transcarbamoylase, N-acetylglutamate synthase, acetaldehyde dehydrogenase, creatine kinase, adenosine kinase, cholesterol side chain lyase, aldosterone synthase, T1M10, T1M23.

[0013] In the construction method described above, the database in step (2) includes the ChEMBL database.

[0014] The construction method described herein includes, in step (2), a dataset comprising mitochondrial effector molecules with IC50 and / or AC50 values ​​less than 1000 nm, mitochondrial effector molecules with IC50 and / or AC50 values ​​greater than 50000 nmol, mitochondrial effector molecules without IC50 and AC50 values ​​but clearly active against the target in the database, and mitochondrial effector molecules without IC50 and AC50 values ​​but clearly inactive against the target in the database.

[0015] The construction method described herein sets mitochondrial effector molecules with IC50 and / or AC50 values ​​less than 1000 nmol, as well as mitochondrial effector molecules without IC50 and AC50 values ​​but with clearly defined target activity in the database, as positive samples.

[0016] The construction method described herein sets mitochondrial effector molecules with IC50 and / or AC50 values ​​greater than 50,000 nmol, as well as mitochondrial effector molecules with no IC50 and AC50 values ​​and clearly inactive against the target in the database, as negative samples.

[0017] In the construction method described above, step (2) uses the Python toolkit to search for and screen effector molecules targeting proteins on mitochondria.

[0018] A smart screening system for mitochondrial effector molecules is established using any of the construction methods described herein.

[0019] The application of the intelligent screening system for mitochondrial effector molecules in the discovery and screening of mitochondrial effector molecules.

[0020] The principle of this invention is to predict mitochondrial effector molecules using machine learning. This involves leveraging machine learning's ability to learn from data, training the model using existing medical data, and enabling the model to predict new data. The research on screening mitochondrial effector molecules based on machine learning mainly consists of the following steps.

[0021] Step 1: Define the research objective and collect the corresponding data. Data collection is a crucial step, as the quality of the dataset directly determines the upper limit of the model's effectiveness. The raw data in this study mainly consist of molecules that have effects on specific proteins on mitochondria. These molecules are generally collected from databases, and the key information for effector molecules should include molecule ID, molecule name, target protein name, and specific mechanism of action.

[0022] The mechanisms by which effector molecules target mitochondria are complex. Most of these molecules affect proteins located on mitochondria. Therefore, a mitochondrial target protein library was established by searching for typical proteins localized to mitochondria, including those involved in mitochondrial biochemical processes such as the tricarboxylic acid cycle and oxidative phosphorylation. Molecules that have effector effects on target proteins can be considered to target and affect mitochondria; therefore, these molecules were included in the mitochondrial effector molecule dataset.

[0023] Step 2: Data Processing. The collected raw data may contain a lot of invalid information, including missing values, duplicate values, feature redundancy, and high-dimensional sparse features, making it unsuitable for direct model training. Preprocessing is necessary before using the raw data to train the model. Furthermore, since most of the collected raw data comes from databases, where molecular descriptions are primarily expressed as SMILES strings, preprocessing typically involves using molecular descriptors to process these SMILES strings and extract the relevant chemical information from the molecules for computer processing.

[0024] Step 3: Machine Learning Algorithm Selection. As can be seen from the current state of research on drug target interaction prediction both domestically and internationally, the selection of machine learning algorithms is particularly important. To date, researchers have proposed numerous excellent machine learning algorithms. Different application scenarios may require different applicable algorithms. Commonly used machine learning algorithms include decision trees, logistic regression, support vector machines, etc. Therefore, it is necessary to select typical machine learning algorithms and deep neural network frameworks to model the mitochondrial effector molecular dataset. The grid parametric tuning method should be used to optimize the parameters involved in each model to achieve the best performance. Then, the various algorithm models should be compared using several performance evaluation metrics to select the model that is more suitable for this study and has better predictive performance.

[0025] Step 4: Model Evaluation. After the machine learning model is built, appropriate validation methods and evaluation metrics need to be selected to evaluate the model's performance. Commonly used model validation methods include five-fold cross-validation, and common classification evaluation metrics include accuracy, precision, and AUC. Based on the evaluation metrics, the parameters of various algorithms can be adjusted to select the optimal parameters best suited for each group of models.

[0026] Step 5: Molecular Prediction. Having already selected the most suitable model for screening mitochondrial effect molecules in the previous steps, we can use this optimal model to predict the performance of a large set of unknown molecules. Generally, classification algorithms will ultimately produce probability scores for each molecule, which are then sorted from highest to lowest. Molecules with higher scores are then selected for relevant biological validation. The results of this validation can also verify the model's predictive performance. Furthermore, molecules with good experimental results can be included in the dataset to expand its scope.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] This invention establishes a dataset of mitochondrial effector molecules and compares and optimizes classic algorithms for screening mitochondrial effector molecules. The models are validated, and the probabilities of a large number of compounds affecting mitochondria are predicted and evaluated. However, evaluating individual models using model evaluation metrics is insufficient to verify their performance; this invention also uses molecules proven to be mitochondrial nutrients as a test set to verify the correctness and predictive effectiveness of the models using known results. Furthermore, effector molecules targeting mitochondria are screened from a massive molecular library, and biological experiments are used to verify whether the screened molecules have mitochondrial-targeting effects. The overall effectiveness of each model in screening mitochondrial effector molecules is evaluated, thus obtaining the optimal model, which is the model of this invention. By applying the model of this invention, researchers can save a significant amount of time on parameter tuning and data input / output processing, making it available to researchers in the field of mitochondrial biology. Attached Figure Description

[0029] Figure 1 This is a technical roadmap for the present invention;

[0030] Figure 2 The parameter tuning process for the support vector machine model includes, Figure 2 The horizontal axis represents the choice of kernel function parameters, which can be linear, sigmoid, poly, or rbf. Figure 2 The horizontal axis b represents the adjustment of parameter C, with a value range of [0, 100]. Figure 2 The horizontal axis 'c' represents the adjustment of the gamma parameter, with a value range of [0, 2]. Figure 2 The ordinate of ac represents the AUC value of the support vector machine model for the corresponding parameter values. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Example 1: Establishing a target protein library

[0033] The technical roadmap of this invention is as follows: Figure 1 .

[0034] This invention collects a series of classic mitochondrial-targeting molecules discovered early on, as well as target proteins with well-defined molecular functions, such as aldehyde dehydrogenase, malate dehydrogenase, and mitochondrial complexes. Information on these proteins was retrieved from the aforementioned databases and compiled into a target protein table. Simultaneously, to ensure data reliability and comprehensiveness, classic mitochondrial proteins were also retrieved from protein databases and integrated into the target protein table. The table includes key information for each protein, including its English and Chinese names, protein number, main effects and mechanisms, and its location in the mitochondria. The integrated mitochondrial target protein table is shown in Table 1 below. Since many functionally distinct complexes exist in mitochondria, if a complex is mainly composed of several enzymes, the protein number of that complex primarily indicates the enzyme protein number. If the complex is composed of many subunits, the protein numbers of each subunit are uniformly grouped under the complex number. For target molecules whose protein location is not clearly defined in the database or literature, the location is defined as "other."

[0035] Table 1. Integrated Mitochondrial Target Protein Table

[0036]

[0037]

[0038]

[0039]

[0040] Example 2: Selection of mitochondrial effector molecules

[0041] In the ChEMBL database, the IC50 value is the most frequent indicator of effector activity for specific targets, followed by the AC50 value. The IC50 value represents a compound's ability to inhibit mitochondrial activity; its effect is inversely proportional to the IC50 value. Conversely, the AC50 value represents the positive effect of a molecule on a specific target; the lower the AC50 value, the stronger the promoting effect. This invention will also use these two activity indicators as screening criteria for effector molecules, selecting those with IC50 and AC50 values ​​less than 1000 nm as positive samples and those with IC50 and AC50 values ​​greater than 50000 nm as negative samples. Furthermore, for effector molecules without IC50 and AC50 indicators, if their activity is clearly described in the database, they will also be included in the molecule set as positive samples.

[0042] To facilitate faster retrieval and screening of desired molecules, the official database provides a Python toolkit called chembl_webresource_client (https: / / github.com / chembl / chembl_webresource_client). This toolkit can automatically access the ChEMBL database, search for target-related compounds for specific targets, and batch output the search results to a file.

[0043] Example 3: Molecular Pretreatment

[0044] 1. Characterization of molecular information

[0045] After obtaining information about the effector molecules, it is necessary to extract the characteristic information contained in the molecules so that computers can recognize them; this is the molecular descriptor. Molecular fingerprints are a type of molecular descriptor, and they have a good feature extraction capability for characterizing molecules and obtaining the biochemical information contained in the molecules.

[0046] This invention utilizes Morgan molecular fingerprints to characterize effect molecules. Morgan molecular fingerprints can effectively describe compound molecules; by identifying the presence of a specific structure described by the Morgan fingerprint within the molecule's structure, the molecule can be converted into a binary string for computer recognition. In this invention, each molecule's Morgan molecular fingerprint is represented as a 1024-bit binary number. The molecules in the file are represented as SMIELS strings, which need to be converted into molecular structure diagrams. This allows the Morgan molecular fingerprint to analyze the molecule's chemical bonds and atomic information from the molecular diagram.

[0047] 2. Molecular similarity processing

[0048] Molecules often exhibit high similarity due to the large number of identical functional groups or atomic bonds, and highly similar molecules tend to share similar effects. However, a large number of highly similar molecules in a dataset can lead to uneven sample distribution, resulting in poor model fitting after input into the algorithm. Therefore, even after deduplication and cleansing, similarity processing is still necessary to remove highly similar molecules and ensure a uniform distribution of samples in the final dataset. Morgan fingerprinting includes a method for calculating molecular similarity, using a function to process the dataset, as shown in Table 2-5.

[0049] Table 2 shows the code processing for deduplication and cleanup.

[0050]

[0051]

[0052] Table 3 Extraction of Morgan's fingerprint molecular information

[0053]

[0054] Table 4 Molecular Similarity Processing

[0055]

[0056] Table 5. Comparison of Molecular Similarity

[0057]

[0058]

[0059] Note: N in the table represents the number of molecules that are similar to a single molecule.

[0060] The AUC and accuracy values ​​on both the validation and test sets generally increase sequentially with similarity values ​​of 0.8, 0.9, and 0.95 (i.e., 80%, 90%, and 95% similarity), demonstrating that the support vector machine model performs better when the similarity is 0.95. Within each similarity metric, the N value ranges from 0, 2, and 4, with the AUC and accuracy values ​​on both sets also generally increasing sequentially. At a similarity of 0.95, the model with the number of similar molecules per molecule controlled within the range of 4 is optimal, based on the N value. Therefore, the final dataset was selected with a similarity of 0.95 and a similarity range of 4 molecules. Subsequent research will use this dataset to input into various algorithms and evaluate the models.

[0061] Example 4: Classic Algorithm Modeling and Evaluation

[0062] This study primarily utilizes classical machine learning methods to build models for processed effector molecules and analyzes and compares the performance of different models. Common evaluation metrics exist for models; for classification tasks, accuracy and AUC are typically used. The parameters of each model are adjusted based on these metrics. Bayesian optimization is employed to fine-tune the performance of each model, aiming for optimal results. The mitochondrial effector molecule set involved in this study may not be suitable for some algorithms. Through multi-dimensional comparative analysis of model metrics, the advantages and disadvantages of each model are summarized, and a suitable algorithm model is ultimately selected for the prediction of new molecules.

[0063] As a classic machine learning algorithm, Support Vector Machine (SVM) is first used in this study by inputting the effect set into the algorithm. The parameters used to adjust the SVM are shown in Table 6.

[0064] Table 6. List of Support Vector Machine Parameters

[0065]

[0066] Bayesian optimization is used to tune model parameters. The general steps of Bayesian hyperparameter tuning are: defining the objective function to be minimized, defining the parameter search space, and storing all point combinations and their effects during the search process. Bayesian hyperparameter tuning can automatically search for the most suitable parameter combination for the model within the defined parameter space and display the results.

[0067] To visually represent the numerical changes of each parameter during the adjustment process, the specific adjustment process for each parameter is visualized. Figure 2 . Figure 2 The horizontal axis represents the parameter values, which are derived from the parameter space definition of the Support Vector Machine (SVM). The vertical axis represents the AUC value. The points in the graph show the distribution of parameter values ​​during the parameter tuning process. Bayesian optimization can automatically search for the optimal point in the parameter space. The visualized image corresponds to the specific parameter points selected in the algorithm. By comparing the vertical coordinates (AUC values) of each value in the parameter space, the parameter value corresponding to the maximum AUC value is selected, which is represented by the horizontal axis value in the graph.

[0068] from Figure 2As can be seen, for the selection of kernel function parameters, the optimal kernel function is the poly function, with the highest AUC value achieved when the C parameter is set to 100 and the gamma parameter to 0.5. Therefore, the optimal parameter combination is: C parameter with C=100, gamma value of 0.5, and the poly kernel function. When the support vector machine model applies the above parameters, the obtained model has the largest AUC value, indicating that the model with this set of parameters has the best performance. Since the parameter tuning methods used in this paper are all Bayesian optimization parameter tuning methods, the parameter tuning process of each algorithm model in the following text is basically the same as that of the support vector machine model. Therefore, the selection of optimal parameters will not be elaborated in detail later, only a visualization of the parameter tuning process and the corresponding optimal parameter values ​​are given.

[0069] The support vector machine (SVM) model was evaluated using the following parameter combination: C = 100, gamma = 0.5, and poly kernel function. Cross-validation was employed to evaluate the model, calculating the accuracy and AUC values ​​for the training, validation, and test sets. The final SVM model achieved the following accuracy: training set accuracy = 0.997 ± 0.001, AUC = 0.999 ± 0.001; validation set accuracy = 0.895 ± 0.013, AUC = 0.951 ± 0.006; and test set accuracy = 0.898 ± 0.015, AUC = 0.955 ± 0.009.

[0070] This invention first models classic machine learning algorithms used for classification, then applies Bayesian optimization to each model to select the optimal parameter combination, and finally compares the performance of each model. The comparison and evaluation of the five models are shown in Table 7 below.

[0071] Table 7 Comparison and Evaluation of Five Models

[0072]

[0073] Through comparison of various models, both Support Vector Machine (SVM) and XgBoost algorithms achieved high scores across all performance metrics in modeling mitochondrial effector molecular data. While the AUC values ​​of the SVM and XgBoost models were largely similar across datasets, the accuracy of the SVM model was higher than that of the XgBoost model. Therefore, the SVM model was chosen as the primary model for subsequent validation and prediction of new molecules.

[0074] Example 5: Verification of Known Molecules

[0075] Mitochondrial effector molecules validated in the literature were used as verification factors. Based on the group's previous research on mitochondrial nutrients, 21 nutrients that passed activity validation were used to verify the applicability of the support vector machine model for screening mitochondrial effector molecules. Information on the 21 molecules is shown in Table 8.

[0076] Table 8. Nutrient molecules in the validation model

[0077]

[0078]

[0079] These 21 classic nutrients were used to validate the model. Typically, for classification models, a prediction score greater than 0.5 is considered a positive sample, and less than 0.5 is considered a negative sample. In this study, a prediction score greater than 0.5 indicates that the molecule has an effect on mitochondria, and less than 0.5 indicates that the molecule has no effect on mitochondria. The optimal model, the Support Vector Machine (SVM), scored greater than 0.5 for all molecules in these 21 categories, indicating that the learning machine believes all molecules in this set target mitochondria and have an effect on them. The validation results for known nutrients show that the SVM model has high accuracy. Considering the model's evaluation metrics, the algorithm is relatively reliable and can be applied to predict large sets of molecules and screen for new mitochondrial effector molecules.

[0080] Example 6: Model Predicts New Molecules

[0081] This example uses a set of active molecules from the ZINC database. All molecules in this set have undergone in vitro experiments, and most exhibit specific effects. The set contains 140,000 molecules. The following section will perform predictions on this large dataset. After predicting the large dataset using a support vector machine model, the molecules are sorted from highest to lowest probability score, and the top 100 molecules are selected. Molecules appearing in the training set and those with similar structures are removed. The top 10 molecules are then selected from highest to lowest probability score. The predicted molecule information is shown in Table 9 below.

[0082] Table 9 Predicted Molecular Information

[0083]

[0084]

[0085] In summary, this invention first validates the Support Vector Machine (SVM) model. Twenty-one experimentally validated mitochondrial nutrients were selected as inputs to the SVM model. The SVM model achieved prediction scores higher than 0.5 for all 21 molecules, with a distribution range of [0.7, 1], indicating that the model demonstrated good prediction performance for all 21 molecules. This demonstrates the applicability of SVM to mitochondrial effect molecule datasets. Next, a large dataset containing 140,000 molecules from the ZINC database was used as the prediction set. The SVM model was applied to predict this large dataset, selecting molecules ranked highest based on their probability scores from highest to lowest.

Claims

1. A method for constructing a smart screening system of mitochondrial effector molecules, characterized by It comprises the following steps: (1) Collecting basic information of target proteins affected by biomolecules, and establishing a target protein library; (2) Based on the target protein library established in the above step (1), taking IC50 and AC50 values as screening criteria of effect molecules, and assisted by the activity of effect molecules in ChEMBL database, screening effect molecules targeting proteins on mitochondria, and obtaining a data set of mitochondrial effect molecules; The data set comprises mitochondrial effect molecules with IC50 and / or AC50 values less than 1000 nmol, mitochondrial effect molecules with IC50 and / or AC50 values greater than 50000 nmol, mitochondrial effect molecules without IC50 and AC50 values but with clear activity to the target in the database, and mitochondrial effect molecules without IC50 and AC50 values and with no activity to the target in the database; the mitochondrial effect molecules with IC50 and / or AC50 values less than 1000 nmol and the mitochondrial effect molecules without IC50 and AC50 values but with clear activity to the target in the database are set as positive samples; the mitochondrial effect molecules with IC50 and / or AC50 values greater than 50000 nmol and the mitochondrial effect molecules without IC50 and AC50 values and with no activity to the target in the database are set as negative samples; (3) Morgan molecular fingerprint is used to characterize the mitochondrial effect molecules in the data set obtained in the above step (2), and after de-duplication and decontamination processing, molecular similarity processing is performed to obtain an input set of the model; (4) Taking accuracy and AUC value as evaluation indexes, support vector machine algorithm is used for learning and training to construct a mitochondrial effect molecule screening model.

2. The construction method of claim 1, wherein The basic information of the target proteins in the step (1) comprises name, protein number, mechanism and location.

3. The construction method of claim 1, wherein The target proteins in the step (1) comprise carnitine palmitoyltransferase, long fatty acyl coenzyme A, kynureninease, monoamine oxidase, coenzyme Q-cytochrome c reductase, cytochrome c, NADH dehydrogenase, succinate dehydrogenase, glycerol-3-phosphate dehydrogenase, adenosine triphosphate synthase, carnitine palmitoyltransferase II, uncoupling protein, MT-ND1, MT-ND2, MT-ND3, MT-ND4, MT-ND5, MT-ND6, MT-CYB, MT-CO1, MT-CO2, MT-CO3, dihydroorotate dehydrogenase, citrate synthase, cis-aconitate, isocitrate dehydrogenase, a-ketoglutarate dehydrogenase complex, succinyl-CoA synthetase, fumarase, malate dehydrogenase, glutamate transaminase, glutamate dehydrogenase, pyruvate dehydrogenase complex, carbamoyl phosphate synthase I, ornithine transcarbamoylase, N-acetylglutamate synthase, acetaldehyde dehydrogenase, creatine kinase, adenylate kinase, cholesterol side-chain cleavage enzyme, aldosterone synthase, T1M10 and T1M23.

4. The construction method of claim 1, wherein In the step (2), the effect molecules targeting proteins on mitochondria are searched and screened by using a python toolkit.

5. An intelligent screening system for mitochondrial effector molecules, characterized in that It is obtained by the construction method in any one of claims 1-4.

6. Use of the smart screening system of claim 5 for the discovery and screening of mitochondrially affecting molecules.

Images