Fusion model-based prediction method and related equipment

Abstract

The embodiment of the invention provides a prediction method based on a fusion model and related equipment, which are used for efficiently and stably predicting an enzyme EE value. The method provided by the embodiment of the invention comprises the following steps: acquiring a target data set aiming at an enzyme EE value from a target database; extracting to-be-screened sequence features of all protein sequences in the target data set; performing training prediction on all the to-be-screened sequence features, screening out to-be-predicted sequence features, and grouping all the to-be-predicted sequence features according to a preset grouping proportion to obtain a to-be-predicted training set and a to-be-predicted test set; inputting all the to-be-predicted sequence features into a plurality of to-be-predicted fusion sub-models, and performing cross validation on each to-be-predicted fusion sub-model to obtain a cross validation value and a sub-model prediction value of each to-be-predicted fusion sub-model; obtaining a model weight value according to the cross validation value of the to-be-predicted fusion sub-model; and obtaining a target model prediction value of the to-be-predicted fusion model according to each model weight value and the sub-model prediction value.

Description

Technical Field

[0001] This application relates to the fields of data processing and machine learning fusion algorithms, and particularly to a prediction method and related equipment based on a fusion model. Background Technology

[0002] Enantiomeric purity (EE) is an important indicator describing the ratio of two enantiomers in chiral compounds and has significant applications in pharmaceuticals, chemical engineering, and biotechnology. Enzyme-catalyzed enantioselective reactions, due to their high efficiency and selectivity, have become important methods for preparing chiral compounds. Therefore, studying the EE value of enzymes is of great significance for understanding the mechanisms of enzyme-catalyzed reactions, optimizing reaction conditions, and designing highly efficient enzyme catalysts.

[0003] Currently, there are two main methods for studying enzyme EE values: experimental determination and prediction. While accurate, experimental determination is time-consuming, costly, and requires extensive experimental procedures and optimization. For novel or poorly studied enzymes, experimental determination may be difficult. In prediction methods, molecular simulation and computational chemistry methods have high computational complexity, rely on the construction of accurate enzyme-substrate models, have long simulation times, and the accuracy of the results depends heavily on the choice of computational methods and parameters.

[0004] In conclusion, considering both cost and accuracy, the most suitable and promising approach is to use artificial intelligence for prediction. However, there are currently few suitable research algorithms available. For example, the UniRep method and random forest-machine learning (ML) are commonly used to build prediction models, but these methods have poor robustness and cannot meet the needs of most practical situations. Summary of the Invention

[0005] This application provides a prediction method and related equipment based on a fusion model for efficient and stable prediction of enzyme EE values.

[0006] The first aspect of this application provides a prediction method based on a fusion model, comprising:

[0007] Obtain a target dataset for the optical purity (EE) value of an enzyme from a target database; wherein the target dataset includes multiple protein sequences of the enzyme;

[0008] Extract the sequence features to be screened from all protein sequences in the target dataset;

[0009] All the sequence features to be screened are trained and predicted, and the sequence features to be predicted that meet the prediction conditions are selected. All the sequence features to be predicted are then grouped according to a preset grouping ratio to obtain a training set and a test set to be predicted. The sequence features to be predicted are distributed in the training set and the test set to be predicted according to the preset grouping ratio.

[0010] All the predicted sequence features in the predicted training set and the predicted test set are input into multiple predicted fusion sub-models in the predicted fusion model to perform cross-validation on each predicted fusion sub-model, so as to obtain the cross-validation value and the sub-model prediction value for each predicted fusion sub-model; wherein, the predicted fusion model includes at least two predicted fusion sub-models.

[0011] Based on the cross-validation value of each fusion sub-model to be predicted, the model weight value corresponding to each fusion sub-model to be predicted is obtained;

[0012] Based on the model weight value and the predicted value of each fusion sub-model to be predicted, the target model predicted value of the fusion model to be predicted is obtained; wherein, the target model predicted value is the predicted value for the enzyme EE value.

[0013] Optionally, extracting the sequence features to be screened from all protein sequences in the target dataset includes:

[0014] The frequency of the target amino acid type in the amino acid sequence of all protein sequences is calculated to obtain descriptor features;

[0015] Based on the large language model, feature vectors of each protein sequence are extracted to obtain the mean features corresponding to each protein sequence;

[0016] The descriptor features and the meanized features are merged to obtain the sequence features to be screened.

[0017] Optionally, the step of training and predicting all the sequence features to be screened, selecting the sequence features to be predicted that meet the prediction conditions, and grouping all the sequence features to be predicted according to a preset grouping ratio to obtain a training set and a test set to be predicted includes:

[0018] The target dataset is randomly split according to the preset grouping ratio to obtain multiple sets of training sets and test sets to be selected, which are grouped according to the preset grouping ratio; wherein, each set of training sets and test sets to be selected includes all the sequence features to be selected.

[0019] The variance of all the sequence features to be screened in the training set and test set to be screened in each group is normalized to obtain the sequence features to be predicted.

[0020] The random forest algorithm (RF) is used to predict the features of the sequence to be screened after variance normalization, and a group score is obtained for the combination of the training set and the test set to be screened corresponding to any group; wherein, the ratio of the training set to be screened to the test set in any group is the preset grouping ratio.

[0021] The training set and test set to be selected that are located before the target ranking in the group scores of all groups are determined, and the corresponding training set to be selected is used as the training set to be predicted, and the corresponding test set to be selected is used as the test set to be predicted; wherein, the target ranking is the ranking of each group from high to low according to the corresponding group scores, and the target ranking is related to the preset group ratio.

[0022] Optionally, the step of inputting all the predicted sequence features from the predicted training set and the predicted test set into multiple predicted fusion sub-models in the predicted fusion model, to perform cross-validation on each predicted fusion sub-model to obtain a cross-validation value for each predicted fusion sub-model, includes:

[0023] Determine the dataset size values ​​for the training set to be predicted and the test set to be predicted;

[0024] According to the preset grouping ratio, the training set to be predicted and the test set to be predicted are randomly grouped according to the data values ​​of the dataset, resulting in grouped training sets and grouped test sets after grouping according to the number of groups of the dataset size; wherein, the ratio of the grouped training set to the grouped test set in each group is the preset grouping ratio;

[0025] The training set and test set of each group are input into each fusion sub-model to be predicted for cross-validation to obtain the training validation score corresponding to the training set and the test validation score corresponding to the test set.

[0026] Calculate the validation ratio between the training validation score and the test validation score, and use the validation ratio as the cross-validation value.

[0027] Optionally, obtaining the sub-model prediction value includes:

[0028] Based on the features of the sequences to be predicted in the training sets of all groups, each of the fusion sub-models to be predicted is trained to obtain the sub-model parameter values ​​corresponding to each of the fusion sub-models to be predicted.

[0029] The fusion sub-model to be predicted with the sub-model parameter values ​​is determined as the target fusion sub-model, and the prediction sequence features of the group test sets in all groups are input into the target fusion sub-model, and the predicted value of the sub-model is output.

[0030] Optionally, obtaining the model weight value corresponding to each fusion sub-model to be predicted based on the cross-validation value of each fusion sub-model to be predicted includes:

[0031] The cross-validation value of each fusion sub-model to be predicted is input into the model weight value calculation formula to obtain the model weight value corresponding to each fusion sub-model to be predicted; wherein, the model weight value calculation formula is:

[0032]

[0033]

[0034]

[0035] Wherein, the S RF S is the cross-validation value of the first fusion sub-model to be predicted. XGB S is the cross-validation value of the second fusion sub-model to be predicted. kNN α is the cross-validation value of the third fusion sub-model to be predicted, β is the model weight value of the first fusion sub-model to be predicted, γ is the model weight value of the second fusion sub-model to be predicted, and γ is the model weight value of the third fusion sub-model to be predicted.

[0036] Optionally, obtaining the target model prediction value of the fusion model to be predicted based on the model weight value and the predicted value of each sub-model to be predicted includes:

[0037] The model weights and predicted values ​​of each sub-model to be predicted are input into the model prediction value calculation formula to obtain the target model prediction value corresponding to the fusion model to be predicted; wherein, the model prediction value calculation formula is: y pred =α×y RF +β×y XGB +γ×y kNN ,

[0038] Wherein, the y predThe target model predicts the value, where α is the model weight value of the first fusion sub-model to be predicted in the fusion model to be predicted, β is the model weight value of the second fusion sub-model to be predicted in the fusion model to be predicted, γ is the model weight value of the third fusion sub-model to be predicted in the fusion model to be predicted, and y is the model weight value of the third fusion sub-model to be predicted in the fusion model to be predicted. RF The y is the sub-model prediction value of the first fusion sub-model to be predicted. XGB The y is the sub-model prediction value of the second fusion sub-model to be predicted. kNN The predicted value of the sub-model of the third fusion sub-model to be predicted.

[0039] A second aspect of this application provides a prediction system based on a fusion model, comprising:

[0040] An acquisition unit is used to acquire a target dataset for the optical purity (EE) value of an enzyme from a target database; wherein the target dataset includes multiple protein sequences of the enzyme;

[0041] An extraction unit is used to extract the sequence features to be screened from all protein sequences in the target dataset;

[0042] The filtering unit is used to train and predict all the sequence features to be filtered, filter out the sequence features to be predicted that meet the prediction conditions, and group all the sequence features to be predicted according to a preset grouping ratio to obtain a training set to be predicted and a test set to be predicted; wherein, the sequence features to be predicted are distributed in the training set to be predicted and the test set to be predicted according to the preset grouping ratio.

[0043] An input unit is used to input all the predicted sequence features from the training set to be predicted and the test set to be predicted into multiple predicted fusion sub-models in the predicted fusion model, so as to perform cross-validation on each predicted fusion sub-model to obtain cross-validation value and sub-model prediction value for each predicted fusion sub-model; wherein, the predicted fusion model includes at least two predicted fusion sub-models.

[0044] The acquisition unit is further configured to obtain the model weight value corresponding to each fusion sub-model to be predicted based on the cross-validation value of each fusion sub-model to be predicted;

[0045] The acquisition unit is further configured to obtain the target model prediction value of the fusion model to be predicted based on the model weight value and the predicted value of each fusion sub-model to be predicted; wherein the target model prediction value is the prediction value for the enzyme EE value.

[0046] The prediction system based on a fusion model provided in the second aspect of this application is used to perform the prediction method based on a fusion model described in the first aspect.

[0047] A third aspect of this application provides a prediction apparatus based on a fusion model, comprising:

[0048] Central processing unit, memory, input / output interfaces, wired or wireless network interfaces, and power supply;

[0049] The memory is either a short-term storage memory or a persistent storage memory;

[0050] The central processing unit is configured to communicate with the memory and execute instructions in the memory to perform the prediction method based on the fusion model described in the first aspect.

[0051] A fourth aspect of this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the prediction method based on a fusion model as described in the first aspect.

[0052] A fifth aspect of this application provides a computer program product comprising instructions that, when executed on a computer, cause the computer to perform the prediction method based on a fusion model as described in the first aspect.

[0053] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: By using the prediction method based on the fusion model disclosed in the embodiments of this application, the sequence features of protein sequences in the dataset are extracted, the sequence features of proteins are quickly predicted, the calculation strategy of the advantageous group is screened, and the advantageous group is used to cross-validate each sub-model to obtain the most stable reasonable model, thus realizing the requirements of high accuracy and strong versatility of the fusion model. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0055] Figure 1 This is a schematic diagram of the system architecture of a fusion model disclosed in an embodiment of this application;

[0056] Figure 2 This is a flowchart illustrating a prediction method based on a fusion model disclosed in an embodiment of this application;

[0057] Figure 3This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application;

[0058] Figure 4 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application;

[0059] Figure 5 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application;

[0060] Figure 6 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application;

[0061] Figure 7 This is an importance analysis diagram of a fusion model disclosed in an embodiment of this application;

[0062] Figure 8 This is a dataset distribution diagram disclosed in an embodiment of this application;

[0063] Figure 9 This is a prediction result diagram of a fusion model disclosed in an embodiment of this application;

[0064] Figure 10 This is a comparison chart of prediction results for a fusion model disclosed in an embodiment of this application;

[0065] Figure 11 This is a schematic diagram of the structure of a prediction system based on a fusion model disclosed in an embodiment of this application;

[0066] Figure 12 This is a schematic diagram of the structure of a prediction device based on a fusion model disclosed in an embodiment of this application. Detailed Implementation

[0067] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0068] It should be noted that the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed in this application.

[0069] Currently, there are two main methods for obtaining the optical purity (EE) value: experimental determination and prediction. Experimental determination: (1) The EE value of the reaction product is directly determined by techniques such as high performance liquid chromatography (HPLC), gas chromatography (GC), and nuclear magnetic resonance (NMR); (2) Directed evolution and enzyme engineering methods are used to improve the enantioselectivity of enzymes through directed evolution or structural modification, and the EE value is obtained through experimental verification. Prediction methods: (1) Molecular simulation and computational chemistry methods are used to predict the interaction between enzymes and substrates and their impact on the EE value of products by using quantum chemical calculations, molecular dynamics simulations, docking analysis, etc.; (2) Machine learning and data-driven methods are used to train machine learning models (such as neural networks, support vector machines, etc.) to predict EE values ​​by using known EE value data of enzyme-catalyzed reactions. However, with the rapid development of artificial intelligence, research on enzyme function prediction based on artificial intelligence (machine learning, deep learning) is constantly advancing, which is the main direction of development in this field. At present, there are some research examples of enzyme EE value prediction based on artificial intelligence, but overall there are few. For example, in existing schemes, the UniRep method is used to calculate the descriptor of the protein sequence, and a random forest machine learning (RF) model is used to train the collected enzyme EE value data to establish a predictive model between activity and sequence.

[0070] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0071] Please see Figure 1 , Figure 1This is a schematic diagram of the system architecture of a fusion model disclosed in an embodiment of this application.

[0072] Depend on Figure 1 As can be seen, the fusion model disclosed in this application includes at least a source data layer, a database layer, a screening layer, and a validation layer. The source data layer mainly describes the method by which the fusion model acquires data. The data sources can be various collagen literature, private experimental data, or public databases, etc., and are not specifically limited here. Then, the collected data is input into the database layer, which can parse the protein fragments and place them into the protein fragment library of an AI-assisted cosmetic new raw material innovation research and development platform (RIMP, Readline Innovative raw material platform), thereby obtaining the protein fragment sequences, efficacy, or targets. Then, the protein sequences from the database layer are input into the screening layer. After screening using Python, the frequency of custom short peptides or functional peptide fragments appearing in the protein fragments is obtained, and their importance is ranked. Finally, the frequency of the custom short peptides and functional peptide fragments ranked by importance is input into the validation layer to complete the experimental validation and analyze the practicality of the fusion model's predictive function for EE values.

[0073] To resolve the technical issues described above, please refer to [link / reference]. Figure 2 , Figure 2 This is a flowchart illustrating a prediction method based on a fusion model disclosed in an embodiment of this application. It includes steps 201-206.

[0074] 201. Obtain the target dataset for enzyme optical purity (EE) values ​​from the target database.

[0075] Depend on Figure 1 As shown, the target dataset for the enzyme's EE values ​​can be obtained from the target database. It should be noted that the target dataset includes multiple protein sequences of the enzyme.

[0076] In one specific embodiment, experimental data on enzyme EE values ​​are first collected from the RIMP database. This data can be real experimental data measured under the same conditions or real experimental data measured under different conditions; details are not elaborated here. However, it should be noted that, to facilitate consistent feature extraction and calculation of the dataset, it is stipulated that, in this embodiment, the data in the dataset consists of experimental data measured under the same conditions.

[0077] Furthermore, using the experimental data described above, a dataset, namely the target dataset, is constructed. Since an enzyme is a protein, its protein sequence can be obtained, and subsequently, the amino acid sequence within that protein can be extracted; details will not be elaborated here.

[0078] It should also be noted that the target database proposed in this embodiment can be not only the RIMP database, but also other publicly available academic papers or databases in other models, fields, or servers. These databases store data such as amino acid sequences of various proteins under different temperature conditions, including publicly available and self-developed peptide information (basic properties, physicochemical and biological properties, product or efficacy, etc.).

[0079] Furthermore, for ease of subsequent description, the dataset in this embodiment is set to a target dataset with 236 data (protein sequences).

[0080] 202. Extract the sequence features to be screened from all protein sequences in the target dataset.

[0081] After collecting protein sequences from the target dataset, the sequence features of all protein sequences in the target dataset can be extracted, thus obtaining the sequence features to be screened.

[0082] In one specific embodiment, descriptors and features are calculated from the sequences in the dataset. Specifically, the Pyprotein toolkit and a large language model based on ESM-2 can be used to extract features from the protein sequences in the dataset, thereby obtaining the features of the sequences to be screened. Further details regarding the technical aspects can be found in [reference needed]. Figure 3 The illustrated embodiment.

[0083] It's important to note that Pyprotein is a Python library for basic computations on peptides / proteins / enzymes, such as calculating isoelectric points (similar to the "ExPASy Compute pI / Mw tool") and frequency calculations (single amino acids, dipeptides, 3-mers). It includes methods in the protein class designed for use in scripts that automate computations. Currently, Pyprotein supports computations on compounds with amino acid residues (proteins / enzymes / peptides), and can also accept amino acid abundance or amino acid sequences as input. Furthermore, it includes pKa tables from three sources for calculating isoelectric points and charge states. ESM-2 is a deep learning model specifically designed to understand the "language" of proteins—the patterns and rules governing structure and function within the protein's amino acid sequence. Its goal is masked language modeling, rather than autoregressive (causal language models), making it more suitable for certain protein-related tasks. ESM-2 outperforms all tested single-sequence protein language models in a range of structure prediction tasks. Additionally, ESM-2 can be used to generate accurate structure predictions directly from protein sequences. For simplicity, this will not be elaborated upon further.

[0084] 203. Train and predict all the sequence features to be selected, select the sequence features that meet the prediction conditions, and group all the sequence features to be predicted according to the preset grouping ratio to obtain the training set and the test set to be predicted.

[0085] Based on step 202, since there are differences between different sequence features, and some sequence features may not be suitable for subsequent model training and prediction, it is necessary to train and predict all the sequence features to be selected, thereby filtering out the sequence features that meet the prediction conditions. Then, all the sequence features to be predicted are grouped according to a preset grouping ratio to obtain the training set and the test set. It should be noted that the sequence features to be predicted are distributed in the training set and the test set according to the preset grouping ratio.

[0086] In one specific embodiment, the extracted sequence features from the target dataset are standardized to normalize the variance of the sequence features, resulting in the sequence features to be predicted. Simultaneously, the dataset is randomly split according to a preset grouping ratio, and the system's training module is used to quickly predict the sequence features, obtaining scores for each group. The groups with higher scores are then selected as the training and test sets for the dataset.

[0087] Furthermore, to facilitate rapid selection of the training and test sets while reducing the amount of data required for subsequent computation, the preset grouping ratio is set to a 4:1 ratio for the training set and the test set. Therefore, the ratio of the dataset to be predicted to the test set is also 4:1. In this embodiment, the four groups with the highest scores can be selected as the training and test sets for the target dataset. The higher the score, the more likely it is to be used as the training set.

[0088] Furthermore, for a more detailed description of the technical aspects, please refer to [reference needed]. Figure 4 The illustrated embodiment.

[0089] 204. Input all the features of the sequences to be predicted in the training set and the test set to be predicted into multiple sub-models to be predicted in the fusion model to be predicted, so as to perform cross-validation on each sub-model to be predicted, and obtain the cross-validation value and the predicted value of each sub-model.

[0090] After obtaining the training set and test set to be predicted through grouping, all the features of the sequences to be predicted can be input into multiple sub-models of the fusion model to be predicted. Thus, the features of the sequences to be predicted from the training set and test set are input into the sub-models to be predicted, allowing for cross-validation of each sub-model. This yields the cross-validation value and the predicted value for each sub-model. It should be noted that the fusion model to be predicted must include at least two sub-models.

[0091] In one specific embodiment, as described in step 203, the training set and test set to be predicted, in a 4:1 ratio, are used to cross-validate the fusion sub-model to be predicted. For example, suppose there are datasets 1, 2, 3, 4, and 5. Datasets 1-4 are the training set to be predicted, and dataset 5 is the test set. Therefore, datasets 1-4 are used as the training set to train the fusion sub-model to be predicted, and dataset 5 is used as the test set to validate the training results. At this point, the cross-validation value (in short, the average of the five-fold cross-validation results for each sub-model in the fusion model to be predicted) and the sub-model prediction value are obtained for one training and test run of the fusion sub-model. Then, datasets 2-5 are used as the training set to be predicted, and dataset 1 is used as the test set, and the same steps are repeated for training and validation. This process continues until each dataset from 1-5 is used as both the training and test sets for training and validation. Finally, the cross-validation value and sub-model prediction value for five training and validation runs are obtained.

[0092] Furthermore, the average of the cross-validation values ​​and sub-model predictions from five training and validation iterations can be used as the final cross-validation value and sub-model prediction value for the fused sub-model to be predicted.

[0093] Furthermore, for a more detailed description of the technical aspects, please refer to [reference needed]. Figure 5 The illustrated embodiment.

[0094] 205. Based on the cross-validation value of each fusion sub-model to be predicted, obtain the model weight value corresponding to each fusion sub-model to be predicted.

[0095] Then, based on step 205, the model weight value corresponding to each fusion sub-model to be predicted can be obtained according to the cross-validation value of each fusion sub-model to be predicted.

[0096] In one specific embodiment, the model weight value of each sub-model to be predicted can be obtained by using the sum of the cross-validation values ​​of each sub-model to be predicted in the fusion model to be predicted as the denominator and the cross-validation value of its own sub-model to be predicted as the numerator.

[0097] 206. Based on the model weights and predicted values ​​of each sub-model to be predicted, the target model predicted value of the fusion model to be predicted is obtained.

[0098] Combining the steps shown in step 205, the target model prediction value of the fusion model can be obtained based on the model weight values ​​and sub-model prediction values ​​of each fusion sub-model to be predicted. It should be noted that the target model prediction value is a prediction value for the enzyme EE value.

[0099] In one specific embodiment, the model weight value of each sub-model to be predicted is used as a coefficient, multiplied by the corresponding sub-model prediction value, and the product of each model weight value and the sub-model prediction value is added together to obtain the target model prediction value of the fusion model to be predicted, i.e., the EE value predicted by the fusion model to be predicted.

[0100] Furthermore, for a more detailed description of the technical aspects, please refer to [reference needed]. Figure 6 The illustrated embodiment.

[0101] This embodiment discloses a prediction method based on a fusion model. By extracting sequence features from protein sequences in a dataset, it rapidly predicts protein sequence features, selects a computational strategy for advantageous grouping, and then cross-validates each sub-model using these advantageous groupings to obtain the most stable and reasonable model. This achieves the requirements of high accuracy and strong versatility for the fusion model. Furthermore, by optimizing the descriptor calculation method (Pyprotein, ESM-2 large language model) and the model fusion strategy of various effective machine learning models (RF / XGBoost / kNN), this embodiment discloses a model fusion-based method for predicting enzyme EE values.

[0102] To extract sequence features from protein sequences in a dataset, please refer to [link / reference]. Figure 3 , Figure 3 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application. It includes steps 301-303.

[0103] 301. Calculate the frequency of the target amino acid in the amino acid sequence of all proteins to obtain descriptor features.

[0104] It is not difficult to understand that steps 301-303 in this embodiment are the same as those described above. Figure 2 Step 202 is similar. However, it should be noted that in this embodiment, the target amino acid type is first determined (there are 20 common amino acid types, such as glutamic acid, lysine, or tryptophan), and then the frequency of the target amino acid type in the amino acid sequence of all protein sequences is calculated to obtain the descriptor features.

[0105] In one specific embodiment, the Pyprotein toolkit was used to calculate the frequency of amino acids (single amino acids, dipeptides, 3-mers) in a sequence, totaling 8420 different descriptor features. It should be noted that Pyprotein descriptors calculate the frequency (percentage) of occurrence, and 8420 is a fixed value. Specifically, in nature, there are 20 amino acids, 20*20 = 400 dipeptides, and 20*20*20 = 8000 3-mers, representing the proportion of each sequence present in these 8420 cases. For example, the proportion of a certain type of amino acid A in a protein sequence is A = 0.500; correspondingly, the proportion of dipeptide AA in a protein sequence is AA = 0.032, and the proportion of 3-mers (GAP, three different types of amino acids G, A, and P) in a protein sequence is GAP = 0.007. In short, the above represents the proportion of amino acids in the entire feature. It should be noted that the above is only an example of the presence of one amino acid.

[0106] 302. Based on the large language model, extract the feature vector of each protein sequence to obtain the mean value feature corresponding to each protein sequence.

[0107] Meanwhile, during step 301, feature vectors for each protein sequence can be extracted based on the large language model to obtain the mean features corresponding to each protein sequence.

[0108] In one specific embodiment, the embedding layer for each protein sequence is extracted based on the ESM-2 large language model. It should be noted that the embedding layer is a layer structure in deep learning models used to transform discrete input data (such as words, characters, categories, etc.) into continuous vector representations. Specifically, the extraction method utilizes mean representations to extract a total of 1280 mean features. Furthermore, these 1280 mean features are directly calculated using the ESM-2 large language model, with each sequence as input.

[0109] 303. Merge the descriptor features and the mean features to obtain the features of the sequence to be screened.

[0110] Then, the descriptor features obtained in step 301 and the mean features obtained in step 302 are merged to obtain the sequence features to be screened.

[0111] In one specific embodiment, all features of the 8420 descriptor features and 1280 mean-valued features are merged to obtain 9700 features, which are used as descriptor inputs, i.e., the sequence features to be screened for each protein sequence. In other words, each protein sequence can be calculated using the two methods described above, and finally merged to form 9700 features.

[0112] This embodiment discloses a prediction method based on a fusion model, which can calculate descriptors and extract features from sequences in a dataset to obtain the sequence features of each protein sequence. Simultaneously, it implements an innovative feature extraction method, thereby obtaining an efficient and stable prediction model.

[0113] To achieve the grouping of training and test sets, Figure 4 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application. It includes steps 401-404.

[0114] 401. Randomly split all the sequence features to be screened in the target dataset according to a preset grouping ratio to obtain multiple sets of training sets and test sets to be screened, grouped according to the preset grouping ratio.

[0115] It is not difficult to understand that steps 401-404 in this embodiment are the same as those described above. Figure 2 Step 203 is similar. However, it should be noted that in this embodiment, all the sequence features to be screened in the target dataset can first be randomly split according to a preset grouping ratio, thereby obtaining multiple sets of training sets and test sets to be screened, grouped according to the preset grouping ratio. It should be noted that each set of training sets and test sets to be screened includes all the sequence features to be screened.

[0116] In one specific embodiment, in conjunction with the above... Figure 2 As shown in step 203, all the sequence features to be screened in the target dataset are randomly split according to (training set: test set = 4:1), thus obtaining multiple sets of training sets and test sets to be screened, grouped according to a 4:1 ratio. It should be noted that the preset grouping ratio described above is only one feasible technical solution or example. In other feasible technical solutions, it can also be set to 3:2, etc., and there is no specific limitation here.

[0117] Furthermore, different training or test sets may contain the same sequence features to be selected. However, within the same group, the sequence features to be selected will necessarily differ between the training and test sets. For example, if there are sequence features 1-5 to be selected, in the first group, the training set has sequence features 1-4, and the test set has sequence feature 5. In the second group, the training set has sequence features 2-5, and the test set has sequence feature 1. This results in multiple sets of training and test sets to be selected.

[0118] Furthermore, the dataset contains a total of 236 datasets, resulting in 188 training sets and 48 test sets. In short, the dataset consists of 236 data points (y-values), each with 9700 features (x-values). These 236 data points are randomly divided in a 4:1 ratio into 188 training sets and 48 test sets. Each data point in either the training or test set has 9700 features.

[0119] 402. Perform variance normalization on all the sequence features to be screened in the training set and test set to be screened in each group to obtain the sequence features to be predicted.

[0120] Then, the variance of all the sequence features to be screened in the training set and test set to be screened in each group is normalized to obtain the sequence features to be predicted.

[0121] In one specific implementation, the StandardScaler module under sklearn.preprocessing in Python is used to calculate and thus normalize the variance of all the data of the sequence features to be screened in the training set and the test set to be screened.

[0122] Furthermore, please refer to Figure 7 Figure 7 This is an importance analysis graph of a fusion model disclosed in an embodiment of this application. The vertical axis represents the feature name, with higher axes indicating more important features. The horizontal axis represents the model prediction result (SHAP, Shapley Additive exPlanations) for each training sample under the corresponding feature. The color represents the feature value of each training sample; larger values ​​are redder, and smaller values ​​are bluer. This value affects the model's accuracy and is only displayed by color on the graph; the specific numerical value is not shown. Figure 7 It mainly describes the impact of which features on the model, and how each training sample affects the model.

[0123] 403. Using the Random Forest (RF) algorithm, the variance-normalized features of the sequences to be screened are predicted to obtain the grouping score corresponding to the combination of the training set and the test set to be screened for any group.

[0124] After normalizing the variance of the features of the sequences to be screened, the Random Forest (RF) algorithm can be used to predict the features of the sequences after the variance normalization, thereby obtaining the grouping score corresponding to the combination of the training set and the test set to be screened for any group. It should be noted that the ratio of the training set to the test set in any group is a preset grouping ratio.

[0125] In one specific embodiment, a pre-trained module (modeled using default parameters of the random forest algorithm) is used for fast prediction, where a value between 1000 and 5000 is selected cyclically. In this embodiment, 2000 can be selected. Specifically, the 236 data points are divided into two parts (training set and test set) in a 4:1 ratio, resulting in N different possibilities. Each loop represents one combination, and in this case, the loop is repeated 2000 times, resulting in 2000 different combinations. Thus, the group score R of the training set and the test set to be selected is quickly predicted in each combination (training set:test set = 4:1). 2 Furthermore, it is also possible to obtain the score R for each dataset in the current combination, including all training and test sets to be selected. 2 It should be noted that R 2It is a scoring standard, specifically a scoring standard using the RF algorithm, with a value range of 0 to 1.

[0126] 404. Determine the training set and test set to be selected that are ranked before the target ranking in the group scores of all groups, and use the corresponding training set to be selected as the training set to be predicted, and the corresponding test set to be selected as the test set to be predicted.

[0127] Based on step 403, the training set and test set to be selected that are ranked higher than the target ranking in the group scores of all groups can be used as the training set to be predicted and the test set to be predicted. It should be noted that the target ranking is the ranking of each group from high to low according to the corresponding group scores, and the target ranking is related to the preset group ratio.

[0128] In one specific embodiment, after obtaining the group scores for each group, the top 5 groups with the highest scores can be selected as the groups for the training set and test set in the target dataset. This determines the training set and the test set to be predicted.

[0129] Furthermore, as described above, the dataset with the highest individual scores is used as the training set, but the scores on the test set are also evaluated simultaneously. Specifically, the scores on the training set and the test set are compared in R... 2 If the score difference is less than 5%, and the test set is used as the training set in the next round, and the training set is used as the test set in the next round, and the test set score is higher than the training set score (i.e., a second scoring), then the TOP5 group is selected. It should be noted that this selection method ensures the general applicability of the training and test sets.

[0130] The prediction method based on a fusion model disclosed in this embodiment can screen out advantageous groups through multiple rapid predictions, thereby providing a stable data source for the model algorithm and improving the feasibility of the solution.

[0131] To automate predictions using fusion models, a suitable fusion model needs to be constructed. Please refer to... Figure 5 Figure 5 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application. It includes steps 501-506.

[0132] 501. Determine the number of datasets for the training set and the test set to be predicted.

[0133] It is not difficult to understand that steps 501-506 in this embodiment are the same as those described above. Figure 2Step 204 is similar. However, it should be noted that in this embodiment, the number of datasets for the training set and the test set to be predicted can be determined first, i.e., the number of datasets. Combining the above embodiment, it can be understood that the number of datasets for the training set and the test set to be predicted is 5, of which the number of training sets to be predicted is 4 and the number of test sets to be predicted is 1.

[0134] 502. According to the preset grouping ratio, the training set and test set to be predicted are randomly grouped according to the data values ​​of the dataset, so as to obtain the grouped training set and grouped test set after grouping according to the number of groups of the dataset.

[0135] Then, according to a preset grouping ratio, the training set and test set to be predicted are randomly grouped based on the data values ​​in the dataset, resulting in grouped training sets and test sets with a set number of groups based on the dataset size. It should be noted that the ratio of the training set to the test set in each group is the preset grouping ratio.

[0136] In one specific embodiment, as described in the preceding embodiments, the preset grouping ratio is 4:1 for the training set and the test set. Therefore, the training set and test set to be predicted, consisting of five datasets, are randomly grouped. This results in five different groups of training sets and test sets to be predicted, and simultaneously, the grouped training set and test set for each group are determined.

[0137] For example, let's define the datasets as datasets 1, 2, 3, 4, and 5. The first group would be the training set 2, 3, 4, and 5, with test set 1. The second group would be the training set 1, 3, 4, and 5, with test set 2. And so on.

[0138] 503. Input the group training set and group test set of all groups into each fusion sub-model to be predicted for cross-validation to obtain the training validation score corresponding to the group training set and the test validation score corresponding to the group test set.

[0139] Based on step 502, the group training set and group test set of all groups can be input into each fusion sub-model to be predicted for cross-validation, thereby obtaining the training validation score value corresponding to the group training set and the test validation score value corresponding to the group test set.

[0140] In one specific embodiment, Python is used to automate the prediction of the fusion model on the selected dataset. In one feasible technical solution, the fusion sub-models to be predicted include three types: RF, XGBoost, and kNN (using scikit-learn). The datasets are trained using these three models with parameter tuning (parameters and ranges are shown in Table 1), and 5-fold cross-validation scores are obtained. It should be noted that 5-fold cross-validation is a method used to verify the selection of the optimal model from different models. The dataset is divided into 5 parts, and 4 parts are used as training data and 1 part as validation data in turn for trials. Each trial yields a corresponding accuracy rate. In this model, RF, XGBoost, and kNN will each output a value as a score for the training set and a corresponding value as a score for the test set. Thus, training validation scores and test validation scores are obtained respectively. Specific model parameters are shown in Table 1.

[0141]

[0142] It should be noted that the parentheses represent example parameters selected in this embodiment.

[0143] 504. Calculate the validation ratio between the training validation score and the test validation score, and use the validation ratio as the cross-validation value.

[0144] Then, by calculating the validation ratio between the training validation score and the test validation score, the validation ratio can be used as the cross-validation value.

[0145] In one specific embodiment, the average score of the 5-fold cross-validation is used to calculate the ratio, and the average of the ratios can be used as the cross-validation value.

[0146] Furthermore, in step 503, the scores for each group of training sets and test sets are obtained. The average score corresponding to the training set and test set scores in each group is calculated to obtain the score value for the training set and the average score for the test set. Therefore, the final cross-validation value R is determined by calculating the ratio of the training set score value to the test set average score. 2 .

[0147] Furthermore, in this embodiment, the general scoring is 5CV_R. 2 >0.75, RMSE_5CV<0.08, TEST_R 2 >0.75, RMSE_TEST <0.05. Among them, 5CV_R 2 Ri for the training set obtained by 5-fold cross-validation 2The value of RMSE_5CV is an alternative scoring method under 5-fold cross-validation, which is related to the R value of the training set. 2 Conversely. TEST_R 2 R for the test set obtained by 5-fold cross-validation 2 The value of RMSE_TEST. RMSE_TEST is another scoring criterion method under 5-fold cross-validation, and is related to the R value of the test set. 2 on the contrary.

[0148] For easier understanding, please refer to Figure 10 , Figure 10 This image shows a comparison of prediction results for a fusion model disclosed in an embodiment of this application. The results show that ESM-2 (best model) has a prediction accuracy of 5CV_R. 2 =0.54, TEST_R 2 =0.67; ESM-2+Pyprotein+ fusion model (best model): 5CV_R 2 =0.77, TEST_R 2 =0.77. This shows that it scores relatively high in the fusion model.

[0149] 505. Based on the features of the sequences to be predicted in the training sets of all groups, train each fusion sub-model to be predicted separately to obtain the sub-model parameter values ​​corresponding to each fusion sub-model to be predicted.

[0150] Then, based on the features of the sequences to be predicted in the training sets of all groups, each fusion sub-model to be predicted is trained to obtain the sub-model parameter values ​​corresponding to each fusion sub-model to be predicted.

[0151] In one specific embodiment, the features of the sequences to be predicted from the grouped training sets of all groups can be input into each sub-model to be predicted, thereby training each sub-model to be predicted. This allows for the training of a converged sub-model to be predicted, from which the sub-model parameter values ​​are obtained. For example, please refer to... Figure 8 , Figure 8 This is a dataset distribution map disclosed in an embodiment of this application. It is not difficult to understand that... Figure 8 Main description of the fusion model to be predicted Figure 8 (a) and the kNN sub-model Figure 8 The dataset in (b). Figure 8 It can be seen that, after training and validation, the datasets corresponding to the fusion model and kNN are similar, thus demonstrating significant comparability and applicability.

[0152] Therefore, once the training of the corresponding fusion sub-model to be predicted converges, the sub-model parameter values ​​of the fusion sub-model to be predicted can be determined.

[0153] 506. Identify the fusion sub-model to be predicted with sub-model parameter values ​​as the target fusion sub-model, and input the to-be-predicted sequence features of the group test sets in all groups into the target fusion sub-model, and output the sub-model prediction value.

[0154] Based on step 505, the fusion sub-model to be predicted with sub-model parameter values ​​is determined as the target fusion sub-model, and the prediction sequence features of the group test sets in all groups are input into the target fusion sub-model, and the sub-model prediction value is output.

[0155] In one specific embodiment, all trained and converged fusion sub-models to be predicted can be identified as the target fusion sub-model. Then, the features of the sequences to be predicted from different group test sets are input into the target fusion sub-model to obtain the sub-model prediction value corresponding to the target fusion sub-model. Furthermore, the results include the optimal parameters of each model (output in a CSV file), the best score result (output in a CSV file), the dataset EE value vs. the normal distribution data distribution plot (PNG), importance analysis (PNG), and the predicted result (output in a CSV file). See details in [link to documentation]. Figures 7 to 10 The description.

[0156] This embodiment discloses a prediction method based on a fusion model, which uses three efficient models for prediction and merges the prediction results by updating the weights to obtain the most stable and reasonable model. Simultaneously, the program can directly output optimal parameters, predicted values, statistical graphs, and importance analysis, and achieve CPU-accelerated computation. In short, it can automatically process data through data import, perform model fusion, obtain prediction results, and monitor the running process, optimal parameters, and results in real time. Finally, it can also provide corresponding statistical graphs of the results, making it convenient to use.

[0157] pass Figure 5 The target fusion model obtained in the illustrated embodiment can accurately output the model's predicted values. Please refer to... Figure 6 , Figure 6 This is a flowchart illustrating another prediction method based on a fusion model disclosed in an embodiment of this application. It includes steps 601-602.

[0158] 601. Input the cross-validation value of each fusion sub-model to be predicted into the model weight value calculation formula to obtain the model weight value corresponding to each fusion sub-model to be predicted.

[0159] 602. Input the model weight value and sub-model prediction value of each fusion sub-model to be predicted into the model prediction value calculation formula to obtain the target model prediction value corresponding to the fusion model to be predicted.

[0160] It is not difficult to understand that steps 601-602 in this embodiment are the same as those described above. Figure 2 Steps 205-206 are similar, but it should be noted that in this embodiment, the model weight value is calculated as follows:

[0161]

[0162]

[0163]

[0164] Among them, S RF Five-fold cross-validation RF for the RF of the fusion sub-model to be predicted 2 Value, S XGB Cross-validation R for the XGBoost fusion sub-model to be predicted 2 Value, S kNN Cross-validation R for the kNN fusion sub-model to be predicted 2 The values ​​are: α is the model weight value of the fusion sub-model RF to be predicted, β is the model weight value of the fusion sub-model XGBoost to be predicted, and γ is the model weight value of the fusion sub-model to be predicted.

[0165] Furthermore, the formula for calculating the model prediction is:

[0166] y pred =α×y RF +β×y XGB +γ×y kNN ,

[0167] Among them, y pred This represents the target model's predicted value, i.e., the EE value predicted by model fusion. RF Let y be the sub-model prediction value of the fusion sub-model RF to be predicted. XGB Let y be the sub-model prediction value of the XGBoost fusion sub-model to be predicted. kNN The predicted values ​​of the sub-models in the kNN fusion sub-model to be predicted are given.

[0168] Furthermore, for ease of understanding, please refer to Figure 9 , Figure 9 This is a prediction result diagram of a fusion model disclosed in an embodiment of this application. Wherein, Figure 9 The upper part of the graph shows the EE values ​​of the dataset and the distribution of the normally distributed data. Therefore, it can be seen that the predicted values ​​of the target model of the fusion model conform to the corresponding normal distribution. Furthermore, in the lower part, after fitting the EE values ​​of the dataset to a curve, the theoretical values ​​(x-axis) and actual values ​​(y-axis) can be attributed to the same curve. Therefore, this demonstrates that the fusion model in this embodiment achieves both high accuracy and strong versatility.

[0169] This embodiment discloses a prediction method based on a fusion model, and combines it with... Figures 2 to 6 The embodiments shown in this application demonstrate that the technical solution can effectively predict enzyme function based on the enzyme EE value of the fusion model by optimizing the descriptor calculation method (Pyprotein, ESM-2 large language model) and the model fusion strategy of multiple effective machine learning models (RF / XGBoost / kNN).

[0170] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0171] Please see Figure 11 , Figure 11 This is a schematic diagram of the structure of a prediction system based on a fusion model disclosed in an embodiment of this application.

[0172] The acquisition unit 1101 is used to acquire a target dataset for the optical purity (EE) value of an enzyme from a target database; wherein the target dataset includes multiple protein sequences of the enzyme.

[0173] Extraction unit 1102 is used to extract the sequence features to be screened from all protein sequences in the target dataset;

[0174] The filtering unit 1103 is used to train and predict all the sequence features to be filtered, filter out the sequence features to be predicted that meet the prediction conditions, and group all the sequence features to be predicted according to a preset grouping ratio to obtain a training set to be predicted and a test set to be predicted; wherein, the sequence features to be predicted are distributed in the training set to be predicted and the test set to be predicted according to the preset grouping ratio.

[0175] The input unit 1104 is used to input all the predictable sequence features in the training set and the test set to be predicted into multiple predictable fusion sub-models in the predictable fusion model, so as to perform cross-validation on each predictable fusion sub-model and obtain the cross-validation value and the sub-model prediction value for each predictable fusion sub-model; wherein, the predictable fusion model includes at least two predictable fusion sub-models.

[0176] The acquisition unit 1101 is also used to obtain the model weight value corresponding to each fusion sub-model to be predicted based on the cross-validation value of each fusion sub-model to be predicted;

[0177] The acquisition unit 1101 is further configured to obtain the target model prediction value of the fusion model to be predicted based on the model weight value and the sub-model prediction value of each fusion sub-model to be predicted; wherein the target model prediction value is the prediction value for the enzyme EE value.

[0178] For example, the system further includes: a computing unit 1105 and a merging unit 1106;

[0179] The calculation unit 1105 is used to calculate the frequency of the target amino acid in the amino acid sequence of all protein sequences to obtain descriptor features;

[0180] Extraction unit 1102 is specifically used to extract the feature vector of each protein sequence based on the large language model, so as to obtain the mean value feature corresponding to each protein sequence;

[0181] The merging unit 1106 is used to merge the descriptor features and the mean features to obtain the features of the sequence to be screened.

[0182] For example, the system further includes: a splitting unit 1107, a processing unit 1108, a prediction unit 1109, and a determination unit 1110;

[0183] The splitting unit 1107 is used to randomly split all the sequence features to be screened in the target dataset according to a preset grouping ratio to obtain multiple sets of training sets and test sets to be screened according to the preset grouping ratio; wherein, each set of training sets and test sets to be screened includes all the sequence features to be screened.

[0184] Processing unit 1108 is used to perform variance normalization on all the sequence features to be screened in the training set and test set to be screened in each group to obtain the sequence features to be predicted.

[0185] The prediction unit 1109 is used to predict the features of the sequence to be screened after variance normalization using the Random Forest algorithm (RF) to obtain the grouping score of the combination of the training set and the test set to be screened for any group; wherein, the ratio of the training set to be screened to the test set in any group is a preset grouping ratio.

[0186] The determining unit 1110 is used to determine the training set and test set to be selected that are located before the target ranking in the group scores of all groups, and to use the corresponding training set to be selected as the training set to be predicted and the corresponding test set to be selected as the test set to be predicted; wherein, the target ranking is the ranking of each group from high to low according to the corresponding group scores, and the target ranking is related to the preset group ratio.

[0187] For example, the system further includes: grouping unit 1111;

[0188] The determination unit 1110 is specifically used to determine the number of datasets in the training set and the test set to be predicted;

[0189] Grouping unit 1111 is used to randomly group the training set and test set to be predicted according to the dataset data value according to a preset grouping ratio, so as to obtain the grouped training set and grouped test set after grouping according to the number of groups of the dataset size value; wherein, the ratio of the grouped training set and grouped test set in each group is the preset grouping ratio.

[0190] The input unit 1104 is specifically used to input the group training set and group test set of all groups into each fusion sub-model to be predicted for cross-validation, so as to obtain the training validation score value corresponding to the group training set and the test validation score value corresponding to the group test set.

[0191] The calculation unit 1105 is specifically used to calculate the validation ratio between the training validation score and the test validation score, and to use the validation ratio as the cross-validation value.

[0192] For example, the system further includes: a training unit 1112;

[0193] Training unit 1112 is used to train each fusion sub-model to be predicted based on the features of the sequence to be predicted in the group training set of all groups, so as to obtain the sub-model parameter values ​​corresponding to each fusion sub-model to be predicted.

[0194] The determining unit 1110 is specifically used to determine the fusion sub-model to be predicted with sub-model parameter values ​​as the target fusion sub-model, and input the to-be-predicted sequence features of the group test sets in all groups into the target fusion sub-model, and output the sub-model prediction value.

[0195] For example, the system includes:

[0196] Input unit 1104 is specifically used to input the cross-validation value of each fusion sub-model to be predicted into the model weight value calculation formula to obtain the model weight value corresponding to each fusion sub-model to be predicted; wherein, the model weight value calculation formula is:

[0197]

[0198]

[0199]

[0200] Among them, S RF S is the cross-validation value of the first fusion sub-model to be predicted. XGB S is the cross-validation value of the second fusion sub-model to be predicted. kNN α is the cross-validation value of the third fusion sub-model to be predicted, β is the model weight value of the first fusion sub-model to be predicted, γ is the model weight value of the second fusion sub-model to be predicted, and γ is the model weight value of the third fusion sub-model to be predicted.

[0201] For example, the system includes:

[0202] Input unit 1104 is specifically used to input the model weight values ​​and predicted values ​​of each fusion sub-model to be predicted into the model prediction value calculation formula to obtain the target model prediction value corresponding to the fusion model to be predicted; wherein, the model prediction value calculation formula is:

[0203] y pred =α×y RF +β×y XGB +γ×y kNN , where y pred Let α be the predicted value of the target model, β be the model weight of the first sub-model to be predicted in the fusion model, γ be the model weight of the second sub-model to be predicted in the fusion model, and y be the model weight of the third sub-model to be predicted in the fusion model. RF y is the sub-model prediction value of the first fusion sub-model to be predicted. XGB y is the sub-model prediction value of the second fusion sub-model to be predicted. kNN This is the predicted value of the sub-model of the third fusion sub-model to be predicted.

[0204] Please refer to the following: Figure 12 The schematic diagram of a prediction device based on a fusion model disclosed in this application includes:

[0205] Central processing unit 1201, memory 1205, input / output interface 1204, wired or wireless network interface 1203, and power supply 1202;

[0206] Memory 1205 is either a short-term storage memory or a persistent storage memory;

[0207] The central processing unit 1201 is configured to communicate with the memory 1205 and execute instructions stored in the memory 1205 to perform the aforementioned operations. Figures 2 to 6The prediction method based on the fusion model in any of the illustrated embodiments.

[0208] This application also provides a chip system, characterized in that the chip system includes at least one processor and a communication interface, the communication interface and the at least one processor are interconnected via a circuit, and the at least one processor is used to run computer programs or instructions to perform the aforementioned... Figures 2 to 6 The prediction method based on the fusion model in any of the illustrated embodiments.

[0209] This application also provides a computer-readable storage medium, which includes instructions that, when executed on a computer, cause the computer to perform the aforementioned actions. Figures 2 to 6 The prediction method based on the fusion model in any of the illustrated embodiments.

[0210] This application also provides a computer program product containing instructions, which, when run on a computer, causes the computer to perform the aforementioned... Figures 2 to 6 The prediction method based on the fusion model in any of the illustrated embodiments.

[0211] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0212] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0213] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0214] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0215] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A prediction method based on a fusion model, characterized in that, The method includes: Obtain a target dataset for the optical purity (EE) value of an enzyme from a target database; wherein the target dataset includes multiple protein sequences of the enzyme; Extract the sequence features to be screened from all protein sequences in the target dataset; All the sequence features to be screened are trained and predicted, and the sequence features to be predicted that meet the prediction conditions are selected. All the sequence features to be predicted are then grouped according to a preset grouping ratio to obtain a training set and a test set to be predicted. The sequence features to be predicted are distributed in the training set and the test set to be predicted according to the preset grouping ratio. All the predicted sequence features in the predicted training set and the predicted test set are input into multiple predicted fusion sub-models in the predicted fusion model to perform cross-validation on each predicted fusion sub-model, so as to obtain the cross-validation value and the sub-model prediction value for each predicted fusion sub-model; wherein, the predicted fusion model includes at least two predicted fusion sub-models. Based on the cross-validation value of each fusion sub-model to be predicted, the model weight value corresponding to each fusion sub-model to be predicted is obtained; Based on the model weight value and the predicted value of each fusion sub-model to be predicted, the target model predicted value of the fusion model to be predicted is obtained; wherein, the target model predicted value is the predicted value for the enzyme EE value.

2. The prediction method based on a fusion model according to claim 1, characterized in that, The step of extracting the sequence features to be screened from all protein sequences in the target dataset includes: The frequency of the target amino acid type in the amino acid sequence of all protein sequences is calculated to obtain descriptor features; Based on the large language model, feature vectors of each protein sequence are extracted to obtain the mean features corresponding to each protein sequence; The descriptor features and the meanized features are merged to obtain the sequence features to be screened.

3. The prediction method based on a fusion model according to claim 1, characterized in that, The process involves training and predicting all the sequence features to be selected, filtering out the sequence features that meet the prediction conditions, and grouping all the sequence features to be predicted according to a preset grouping ratio to obtain a training set and a test set to be predicted, including: The target dataset is randomly split according to the preset grouping ratio to obtain multiple sets of training sets and test sets to be selected, which are grouped according to the preset grouping ratio; wherein, each set of training sets and test sets to be selected includes all the sequence features to be selected. The variance of all the sequence features to be screened in the training set and test set to be screened in each group is normalized to obtain the sequence features to be predicted. The random forest algorithm (RF) is used to predict the features of the sequence to be screened after variance normalization, and a group score is obtained for the combination of the training set and the test set to be screened corresponding to any group; wherein, the ratio of the training set to be screened to the test set in any group is the preset grouping ratio. The training set and test set to be selected that are located before the target ranking in the group scores of all groups are determined, and the corresponding training set to be selected is used as the training set to be predicted, and the corresponding test set to be selected is used as the test set to be predicted; wherein, the target ranking is the ranking of each group from high to low according to the corresponding group scores, and the target ranking is related to the preset group ratio.

4. The prediction method based on a fusion model according to claim 1, characterized in that, The step of inputting all the predicted sequence features from the predicted training set and the predicted test set into multiple predicted fusion sub-models in the predicted fusion model, and performing cross-validation on each predicted fusion sub-model to obtain a cross-validation value for each predicted fusion sub-model, includes: Determine the dataset size values ​​for the training set to be predicted and the test set to be predicted; According to the preset grouping ratio, the training set to be predicted and the test set to be predicted are randomly grouped according to the data values ​​of the dataset, resulting in grouped training sets and grouped test sets after grouping according to the number of groups of the dataset size; wherein, the ratio of the grouped training set to the grouped test set in each group is the preset grouping ratio; The training set and test set of each group are input into each fusion sub-model to be predicted for cross-validation to obtain the training validation score corresponding to the training set and the test validation score corresponding to the test set. Calculate the validation ratio between the training validation score and the test validation score, and use the validation ratio as the cross-validation value.

5. The prediction method based on a fusion model according to claim 4, characterized in that, The obtained sub-model prediction values ​​include: Based on the features of the sequences to be predicted in the training sets of all groups, each of the fusion sub-models to be predicted is trained to obtain the sub-model parameter values ​​corresponding to each of the fusion sub-models to be predicted. The fusion sub-model to be predicted with the sub-model parameter values ​​is determined as the target fusion sub-model, and the prediction sequence features of the group test sets in all groups are input into the target fusion sub-model, and the predicted value of the sub-model is output.

6. The prediction method based on a fusion model according to claim 1, characterized in that, The step of obtaining the model weight value corresponding to each fusion sub-model to be predicted based on the cross-validation value of each fusion sub-model to be predicted includes: The cross-validation value of each fusion sub-model to be predicted is input into the model weight value calculation formula to obtain the model weight value corresponding to each fusion sub-model to be predicted; wherein, the model weight value calculation formula is: Wherein, the S RF S is the cross-validation value of the first fusion sub-model to be predicted. XGB S is the cross-validation value of the second fusion sub-model to be predicted. kNN α is the cross-validation value of the third fusion sub-model to be predicted, β is the model weight value of the first fusion sub-model to be predicted, and γ is the model weight value of the third fusion sub-model to be predicted.

7. The prediction method based on a fusion model according to claim 1, characterized in that, The step of obtaining the target model prediction value of the fusion model to be predicted based on the model weight value and the predicted value of each sub-model to be predicted includes: The model weight value and the predicted value of each sub-model to be predicted are input into the model prediction value calculation formula to obtain the target model prediction value corresponding to the fusion model to be predicted; wherein, the model prediction value calculation formula is: and pred =α×y RF +β×y XGB +γ×y kNN , Wherein, the y pred The target model predicts the value, where α is the model weight value of the first fusion sub-model to be predicted in the fusion model to be predicted, β is the model weight value of the second fusion sub-model to be predicted in the fusion model to be predicted, γ is the model weight value of the third fusion sub-model to be predicted in the fusion model to be predicted, and y is the model weight value of the third fusion sub-model to be predicted in the fusion model to be predicted. RF The y is the sub-model prediction value of the first fusion sub-model to be predicted. XGB The y is the sub-model prediction value of the second fusion sub-model to be predicted. kNN The predicted value of the sub-model of the third fusion sub-model to be predicted.

8. A prediction system based on a fusion model, characterized in that, The system includes: An acquisition unit is used to acquire a target dataset for the optical purity (EE) value of an enzyme from a target database; wherein the target dataset includes multiple protein sequences of the enzyme; An extraction unit is used to extract the sequence features to be screened from all protein sequences in the target dataset; The filtering unit is used to train and predict all the sequence features to be filtered, filter out the sequence features to be predicted that meet the prediction conditions, and group all the sequence features to be predicted according to a preset grouping ratio to obtain a training set to be predicted and a test set to be predicted; wherein, the sequence features to be predicted are distributed in the training set to be predicted and the test set to be predicted according to the preset grouping ratio. An input unit is used to input all the predicted sequence features from the training set to be predicted and the test set to be predicted into multiple predicted fusion sub-models in the predicted fusion model, so as to perform cross-validation on each predicted fusion sub-model to obtain cross-validation value and sub-model prediction value for each predicted fusion sub-model; wherein, the predicted fusion model includes at least two predicted fusion sub-models. The acquisition unit is further configured to obtain the model weight value corresponding to each fusion sub-model to be predicted based on the cross-validation value of each fusion sub-model to be predicted; The acquisition unit is further configured to obtain the target model prediction value of the fusion model to be predicted based on the model weight value and the predicted value of each fusion sub-model to be predicted; wherein the target model prediction value is the prediction value for the enzyme EE value.

9. A prediction device based on a fusion model, characterized in that, The device includes: Central processing unit, memory, input / output interfaces, wired or wireless network interfaces, and power supply; The memory is either a short-term storage memory or a persistent storage memory; The central processing unit is configured to communicate with the memory and execute instructions in the memory to perform the prediction method based on the fusion model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes instructions that, when executed on a computer, cause the computer to perform the prediction method based on a fusion model as described in any one of claims 1 to 7.

Images