Methods, apparatus and media for predicting the effect of amino acid variations on protein solubility

Abstract

This invention belongs to the field of bioinformatics and relates to a method, apparatus, and medium for predicting the impact of amino acid variations on protein solubility: The method involves acquiring mutated amino acids and their feature value sets from a dataset; inputting the feature values ​​from the feature value sets into a three-classification module, which outputs a first probability vector indicating that the mutated amino acid belongs to different predicted categories; inputting the feature values ​​from the feature value sets into three two-layer three-classification modules corresponding to different predicted categories, and based on the outputs of each two-layer three-classification module, obtaining a second, third, and fourth probability vector indicating that the mutated amino acid belongs to different predicted categories; and inputting the first, second, third, and fourth probability vectors into a meta-learner, which outputs the predicted category of the mutated amino acid. This scheme, based on a multi-layer stacking strategy, constructs a prediction model that covers all possible binary and three-class classifications, exhibiting high prediction accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method, apparatus, and medium for predicting the effect of amino acid variations on protein solubility. Background Technology

[0002] Protein solubility is closely related to a variety of diseases, especially neurodegenerative diseases (such as Alzheimer's disease, Parkinson's disease) and amyloidosis. Predicting the impact of amino acid variations on protein solubility can provide important insights for the identification of these diseases.

[0003] Existing technologies use the PON-Sol, SODA, and Protein-Sol models to study the relationship between amino acids and protein solubility. The PON-Sol model, in particular, uses a two-layer, three-class classification model based on random forests to predict the impact of amino acid variations on protein solubility. It collects high-quality, experimentally validated datasets from the literature, classifying 443 samples in this dataset into three categories: increased solubility, no effect on solubility, and decreased solubility. Because the dataset contains a large number of samples leading to increased protein solubility, the first layer of its two-layer, three-class classification model first determines the probability of each sample causing increased protein solubility, and then the second layer determines the probability of it causing decreased or no solubility. Since the first layer of this two-layer, three-class classification model focuses on learning samples that increase protein solubility, the model becomes overly biased towards learning this type of sample, making it difficult to capture the feature patterns of other categories, thus affecting the model's prediction accuracy. The SODA model uses solubility-related protein physicochemical properties for prediction. It first uses the PASTA, ESpritz, KyteDoolittle, and FELLS models to predict the influence of aggregate tendency, intrinsic disordered regions, hydrophobicity, and secondary structure preference of variant amino acids on protein solubility. Finally, a linear model combines these influence scores to output the total influence score of variant amino acids on protein solubility. Since each model requires computation and time, this significantly impacts prediction efficiency in large-scale predictions. Furthermore, because the final prediction depends on the results of multiple models, and each model introduces additional errors and uncertainties, the final prediction accuracy is low. Similar to SODA, the Protein-Sol model comprehensively considers multiple eigenvalues ​​of variant amino acids to predict their influence on protein solubility. It uses multiple models to predict the influence score of each eigenvalue on protein solubility, then determines the weight of each eigenvalue by comparing differences in solubility distribution, and finally predicts the influence score of variant amino acids on protein solubility based on the influence score and weight of each eigenvalue. This model also suffers from low prediction efficiency and accuracy.

[0004] In summary, existing methods for predicting the impact of amino acid variations on protein solubility suffer from low accuracy and efficiency. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of low accuracy and efficiency in the existing methods for predicting the effect of amino acid variations on protein solubility.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting the effect of amino acid variations on protein solubility, comprising:

[0007] Obtain the set of variant amino acids and their feature values ​​from the dataset;

[0008] Input the feature values ​​from the feature value set into the three-classification module, and output the first probability vector of the variant amino acid belonging to different predicted categories;

[0009] The feature values ​​in the feature value set are input into three two-layer three-classification modules corresponding to different prediction categories. Based on the output of each two-layer three-classification module, the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories are obtained.

[0010] Input the first probability vector, the second probability vector, the third probability vector, and the fourth probability vector into the meta-learner, and output the predicted category of the mutated amino acid.

[0011] Preferably, the feature values ​​in the feature value set are respectively input into three two-layer three-classification modules corresponding to different prediction categories. Based on the output of each two-layer three-classification module, the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories are obtained, including:

[0012] The feature values ​​in the feature set are input into the first base classifier, which outputs the probability that the mutated amino acid belongs to the first predicted category; the feature values ​​in the feature set are input into the second base classifier, which outputs the probability that the mutated amino acid belongs to the second and third predicted categories; based on the outputs of the first and second base classifiers, a second probability vector of the mutated amino acid belonging to different predicted categories is obtained.

[0013] The feature values ​​in the feature set are input into the third base classifier, which outputs the probability that the mutated amino acid belongs to the third predicted category; the feature values ​​in the feature set are input into the fourth base classifier, which outputs the probability that the mutated amino acid belongs to the first predicted category and the probability that it belongs to the second predicted category; based on the outputs of the third and fourth base classifiers, a third probability vector of the mutated amino acid belonging to different predicted categories is obtained.

[0014] The feature values ​​in the feature set are input into the fifth base classifier, which outputs the probability that the mutated amino acid belongs to the second predicted category. The feature values ​​in the feature set are input into the sixth base classifier, which outputs the probability that the mutated amino acid belongs to the first and third predicted categories. Based on the outputs of the fifth and sixth base classifiers, a fourth probability vector of the mutated amino acid belonging to different predicted categories is obtained.

[0015] Preferably, obtaining the feature value set of the mutated amino acids includes:

[0016] Obtain multiple feature values ​​and solubility values ​​for each variant amino acid in the dataset;

[0017] Discretize the solubility values ​​of all variant amino acids and each characteristic value of all variant amino acids to obtain multiple distribution intervals of solubility values ​​and multiple distribution intervals of each characteristic value;

[0018] Based on the distribution information of the solubility values ​​of all variant amino acids in each solubility value distribution interval, and the distribution information of each characteristic value of all variant amino acids in each distribution interval of that characteristic value, the mutual information value between each characteristic value and the solubility of the variant amino acid is calculated.

[0019] The mutual information values ​​between various feature values ​​and the solubility of the variant amino acid are arranged from largest to smallest. The N feature values ​​corresponding to the top N mutual information values ​​are selected as the N target feature values. The feature value set is obtained based on the N target feature values ​​of the variant amino acid.

[0020] Preferably, the set of feature values ​​is obtained based on N target feature values ​​of the variant amino acids, including:

[0021] Input each target feature value of each variant amino acid into the trained variant amino acid solubility prediction model, and calculate the SHAP value of each target feature value of the variant amino acid based on the solubility prediction value and the solubility value of the variant amino acid.

[0022] Calculate the mean SHAP value for each target feature of all variant amino acids to obtain the mean SHAP value for each target feature.

[0023] Arrange the SHAP mean values ​​of various target features from largest to smallest, remove the M target features corresponding to the last M SHAP mean values, and obtain the feature set based on the remaining target feature values ​​of the mutated amino acids.

[0024] After obtaining the set of variant amino acids and their feature values ​​from the dataset, the following is also included:

[0025] The recursive feature elimination strategy is used to filter the feature values ​​in the feature value set to obtain the feature subsets corresponding to the three classification modules and the feature subsets corresponding to each base classifier.

[0026] Input the feature subset corresponding to the three-classification module into the three-classification module, and input the feature subset corresponding to each base classifier into each base classifier respectively.

[0027] Preferably, after inputting the first probability vector, the second probability vector, the third probability vector, and the fourth probability vector into the meta-learner and outputting the predicted category of the mutated amino acid, the method further includes:

[0028] The loss function is calculated based on the first probability vector, the second probability vector, the third probability vector, the fourth probability vector, and the predicted category.

[0029] Based on the variable amino acids and their feature sets in the dataset, iterative training of the three-classification module, each base classifier, and the meta-learner is performed using iterative random parameter search and cross-validation until the loss function is minimized. Based on the trained three-classification module, each base classifier, and the meta-learner, a trained prediction model is obtained.

[0030] Preferably, the steps for calculating the loss function include:

[0031] Using the Focal Loss function, the first classification loss of the three-class classification module is calculated based on the first probability vector output by the three-class classification module and the true category of the mutated amino acid.

[0032] Using the Focal Loss function, the second classification loss of each two-layer three-classification module is calculated based on the probability vector output by each module and the true category of the mutated amino acid.

[0033] Using the cross-entropy loss function, the third classification loss of the meta-learner is calculated based on the predicted category output by the meta-learner and the true category of the mutated amino acid.

[0034] The value of the loss function is obtained by weighting the first classification loss, each second classification loss, and the third classification loss.

[0035] Preferably, the process of constructing the dataset includes:

[0036] A keyword search was performed in the PubMed database, and a collection of documents was obtained based on the search results; the keywords included one or more of the following: protein, variant, and solubility.

[0037] Extract the literature set that is identical to the literature in the VariBench database to obtain a subset of literature;

[0038] Data mining tools are used to identify documents in a subset of documents, obtain documents containing amino acid variation information, and obtain the target document set.

[0039] The amino acid variation data and the corresponding protein solubility change data of each document in the target literature set are extracted to obtain the dataset.

[0040] The present invention also provides an apparatus for predicting the effect of amino acid variations on protein solubility, comprising:

[0041] The data acquisition module is used to acquire the variant amino acids and their feature value sets in the dataset;

[0042] The first three-class classification module is used to input the feature values ​​in the feature value set into the three-class classification module and output the first probability vector of the variant amino acid belonging to different predicted categories.

[0043] The second and third classification modules are used to input the feature values ​​in the feature value set into the three two-layer three classification modules corresponding to different prediction categories. Based on the output of each two-layer three classification module, the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories are obtained.

[0044] The category prediction module is used to input the first probability vector, the second probability vector, the third probability vector, and the fourth probability vector into the meta-learner and output the predicted category of the mutated amino acid.

[0045] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above for predicting the effect of amino acid variations on protein solubility.

[0046] The method for predicting the effect of amino acid variations on protein solubility provided in this application has the following beneficial effects:

[0047] Since there are three categories of influence of variant amino acids on protein solubility, this application first uses a three-class classification module to directly predict the variant amino acids and output the probability value of the different influences of the amino acid on protein solubility. Since the first layer of the two-layer three-class classification module focuses on learning the category with a larger number of samples in the dataset, this application constructs three two-layer three-class classification modules corresponding to the three categories respectively. Each two-layer three-class classification module focuses on learning the feature information of variant amino acids related to each category and outputs the probability value of the different influences of the variant amino acid on protein solubility. This allows the model to clearly capture the feature patterns of each type of data, and the accuracy of its prediction results is not affected by the data distribution characteristics in the dataset. Finally, all probability combinations are used as input to a meta-learner, which integrates the output results of all classification modules to improve the model's ability to handle imbalanced data. Since the prediction model built by this application based on the multi-layer stacking strategy covers all possibilities of binary and tri-class classification, it avoids the problem of low prediction accuracy caused by the model focusing on learning a certain type of data. It has high prediction accuracy regardless of whether the data distribution in the dataset is uniform. At the same time, it avoids the problem of low prediction accuracy and efficiency caused by using multiple models to predict different feature values ​​of mutated amino acids and then integrating them. Attached Figure Description

[0048] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0049] Figure 1 A flowchart of the method for predicting the effect of amino acid variations on protein solubility provided in this application;

[0050] Figure 2 A schematic diagram of the structure of the prediction model provided in this application;

[0051] Figure 3 A flowchart illustrating feature selection using a recursive feature elimination strategy provided in this application;

[0052] Figure 4 A flowchart illustrating the iterative training of the learner using iterative random parameter search and cross-validation methods provided in this application;

[0053] Figure 5 A schematic diagram illustrating the predicted effects of amino acid variations on Bruton's tyrosine kinase (BTK) on protein solubility, provided in this application; wherein, Figure 5 In the diagram, (a) represents the number of variant amino acids on Bruton's tyrosine kinase (BTK) that increase protein solubility. Figure 5 In the diagram, (b) represents the number of variant amino acids on Bruton's tyrosine kinase (BTK) that have no effect on protein solubility. Figure 5 In the diagram, (c) represents the number of variant amino acids on Bruton's tyrosine kinase (BTK) that lead to reduced protein solubility. Figure 5 (d) in the figure represents the association between amino acid variations that cause changes in protein solubility and pathogenic amino acid variations;

[0054] Figure 6 This is a schematic diagram of the device provided in this application for predicting the effect of amino acid variations on protein solubility. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0056] Please see Figure 1 , Figure 1 The diagram shown is a flowchart of the method for predicting the effect of amino acid variations on protein solubility provided in this application. The method specifically includes:

[0057] S10: Obtain the set of variant amino acids and their feature values ​​from the dataset.

[0058] Furthermore, this application constructs a dataset based on literature mining, and the specific construction process includes:

[0059] S100: Perform keyword search in the PubMed database and obtain a collection of documents based on the search results; the keywords include one or more of protein, variant, and solubility.

[0060] S101: Obtain the documents in the document collection that are identical to those in the VariBench database, thus obtaining a subset of documents.

[0061] S102: Use data mining tools to identify documents in the document subset, obtain documents containing amino acid variation information, and obtain the target document set. Specifically, the data mining tools are MutationFinder, tmVar, or EMU.

[0062] S103: Extract amino acid variation data and corresponding protein solubility change data from each document in the target document set to obtain the dataset.

[0063] For example, one can manually read through each document in the target literature collection and filter out amino acid variation data and protein solubility data. One can also collect information such as protein sequences and diseases related to these amino acid variation data and protein solubility data from NCBI, so as to explore the association between the effect of amino acid variation on protein solubility and diseases based on this information.

[0064] The existing datasets used to predict the impact of amino acid variations on protein solubility contain only 443 samples. However, the dataset constructed based on literature mining in this application increases the number of samples to 6328, greatly expanding the dataset and providing a training basis for the prediction model. This avoids the problem that insufficient samples prevent the model from accurately capturing the features of various types of data, thus affecting the model's prediction accuracy.

[0065] S20: Input the feature values ​​in the feature value set into the three-classification module, and output the first probability vector of the variant amino acid belonging to different predicted categories.

[0066] S30: Input the feature values ​​in the feature value set into the three two-layer three-classification modules corresponding to different prediction categories. Based on the output of each two-layer three-classification module, obtain the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories.

[0067] S40: Input the first probability vector, the second probability vector, the third probability vector, and the fourth probability vector into the meta-learner and output the predicted category of the mutated amino acid.

[0068] Considering the impact of variant amino acids on protein solubility is categorized into three types: increase, decrease, and no effect, this application first uses a three-class classification module to directly predict the variant amino acids, outputting the probability value of the amino acid's different effects on protein solubility. Since the first layer of the two-layer three-class classification module focuses on learning the category with a larger number of samples in the dataset, this application constructs three two-layer three-class classification modules corresponding to the three categories respectively. Each two-layer three-class classification module focuses on learning the feature information of variant amino acids related to each category and outputs the probability value of the variant amino acid's different effects on protein solubility. This allows the model to clearly capture the feature patterns of each data category, and the accuracy of its prediction results is not affected by the data distribution characteristics in the dataset. Finally, all probability combinations are used as input to a meta-learner, which integrates the output results of all classification modules to improve the model's ability to handle imbalanced data. Since the prediction model built by this application based on the multi-layer stacking strategy covers all possibilities of binary and tri-class classification, it avoids the problem of low prediction accuracy caused by the model focusing on learning a certain type of data. It has high prediction accuracy regardless of whether the data distribution in the dataset is uniform. At the same time, it avoids the problem of low prediction accuracy and efficiency caused by using multiple models to predict different feature values ​​of mutated amino acids and then integrating them.

[0069] Further, in step S30, the feature values ​​in the feature value set are input into three two-layer three-classification modules corresponding to different prediction categories. Based on the output of each two-layer three-classification module, the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories are obtained, including:

[0070] S300: Input the feature values ​​in the feature value set into the first base classifier and output the probability that the mutated amino acid belongs to the first predicted category; input the feature values ​​in the feature value set into the second base classifier and output the probability that the mutated amino acid belongs to the second predicted category and the probability that it belongs to the third predicted category; based on the outputs of the first base classifier and the second base classifier, obtain the second probability vector that the mutated amino acid belongs to different predicted categories.

[0071] S301: Input the feature values ​​in the feature value set into the third base classifier and output the probability that the mutated amino acid belongs to the third predicted category; input the feature values ​​in the feature value set into the fourth base classifier and output the probability that the mutated amino acid belongs to the first predicted category and the second predicted category; based on the outputs of the third base classifier and the fourth base classifier, obtain the third probability vector of the mutated amino acid belonging to different predicted categories.

[0072] S302: Input the feature values ​​in the feature value set into the fifth base classifier and output the probability that the mutated amino acid belongs to the second predicted category; input the feature values ​​in the feature value set into the sixth base classifier and output the probability that the mutated amino acid belongs to the first predicted category and the third predicted category; based on the outputs of the fifth and sixth base classifiers, obtain the fourth probability vector of the mutated amino acid belonging to different predicted categories.

[0073] Furthermore, in some embodiments of this application, the first base classifier, the second base classifier, the third base classifier, the fourth base classifier, the fifth base classifier, the sixth base classifier, and the seventh base classifier are all LightGBM base classifiers.

[0074] Because the LightGBM base classifier employs a leaf-wise tree growth strategy, it can select the leaf with the highest gain for classification in each iteration, allowing it to capture relationships between data using fewer leaf nodes. Furthermore, since multi-layer stacked models have a large number of base classifiers, the LightGBM base classifier's histogram algorithm and parallel learning capabilities can accelerate the training process of multi-layer stacked models while reducing memory usage.

[0075] Furthermore, in some embodiments of this application, the meta-learner is a logistic regression meta-learner.

[0076] In multi-layer stacked models, base classifiers may overfit the training data, leading to a decline in the model's generalization ability. Logistic regressor learners can suppress overfitting base classifiers and alleviate the overfitting phenomenon of the entire model by reweighting and combining the outputs of multiple base classifiers. In addition, since multi-layer stacked models have a large number of learners, their computational load and cost are relatively high. Using logistic regressor learners can effectively fuse the output results of multiple base classifiers without increasing the computational cost too much, further improving the model's prediction efficiency.

[0077] like Figure 2 The diagram shown is a schematic of the prediction model structure provided in this application. The prediction model is designed using a stacked generalization strategy and consists of three layers of learners. The first two layers include seven base classifiers, and the third layer includes a meta-learner. The meta-learner uses the output results of the first two layers of learners as input and outputs the final three-class classification result.

[0078] Specifically, the first layer of the prediction model is a base classifier that performs a three-class classification task, and the second layer includes six base classifiers that perform a two-class classification task. These six base classifiers are divided into three pairs, each pair being responsible for a specific two-layer three-class classification task, specifically including: Task 1: first determine whether the mutated amino acid causes a decrease in protein solubility, and then determine whether it causes an increase in protein solubility or has no effect on protein solubility; Task 2: first determine whether the mutated amino acid has an effect on protein solubility, and then determine whether it causes an increase in protein solubility or a decrease in protein solubility; Task 3: first determine whether the mutated amino acid causes an increase in protein solubility, and then determine whether it causes a decrease in protein solubility or has no effect.

[0079] This model utilizes six base classifiers to form three two-layer three-classification modules. These modules, along with a three-classification learner, provide probability outputs covering all possibilities of binary and three-class classification. The probability outputs of all base classifiers are then combined to form new input data, which serves as the input to a meta-learner. The prediction of the impact of variant amino acids on protein solubility is obtained based on the output of the meta-learner. Since the first layer of each two-layer three-classification module focuses on learning the features of one type of variant amino acid, the model can clearly capture the feature patterns of each data class. The accuracy of its predictions is unaffected by the data distribution characteristics, avoiding the problem of low prediction accuracy caused by the model focusing on learning a particular data class. It exhibits high prediction accuracy regardless of whether the data distribution in the dataset is uniform.

[0080] Furthermore, the set of eigenvalues ​​of the variant amino acids is obtained, including:

[0081] Step 1: Obtain multiple feature values ​​and solubility values ​​for each variant amino acid in the dataset.

[0082] Among them, various feature values ​​include amino acid feature values ​​of variant amino acids, conservation feature values, variant type feature values, sequence context feature values, protein type feature values, etc.

[0083] Specifically, amino acid feature values ​​are calculated based on the physicochemical and biological properties of amino acids in the AAindex database, such as size, polarity, hydrophobicity, and charge-dihedral bias. The AAindex database is divided into three categories: AAIndex1, AAIndex2, and AAIndex3, containing 544, 94, and 47 features respectively. After removing invalid data, the AAindex database contains a total of 617 features. Conservation features include hit count and SIFT score, used to assess the degree of conservation of amino acids at a specific position among different protein family members. Specifically, protein sequences homologous to variant amino acids are searched in the NCBI non-redundant database, hit counts are counted, and SIFT scores are calculated through multiple sequence alignment. Variance type features include two substitution matrices: a 20*20 substitution matrix representing the substitution relationships of 20 amino acids, and a 6*6 substitution matrix describing the substitution relationships of different categories of amino acids, including hydrophobicity, negative charge, positive charge, conformation, polarity, and others. Sequence context features are related to the sequence information adjacent to the amino acid mutation site. The frequency of different amino acids is calculated within a fixed window of size 23 centered on the mutation site. Classification criteria include the polarity of the amino acid, whether it carries a charge, and whether it is positively or negatively charged. The protein type feature is the length of the amino acid.

[0084] Step 2: Discretize the solubility values ​​of all variant amino acids and each characteristic value of all variant amino acids to obtain multiple solubility value distribution intervals and multiple distribution intervals for each characteristic value.

[0085] Specifically, since the eigenvalues ​​and solubility values ​​are continuous variables, they need to be discretized first using binning or kernel density estimation to obtain multiple intervals. This allows us to obtain the solubility values ​​of each variant amino acid and the probability distribution of the eigenvalues ​​in each interval. For example, the solubility value distribution intervals can include easily soluble and poorly soluble intervals. If the eigenvalue is the hydrophobic index, then its multiple distribution intervals can be three intervals: low, medium, and high.

[0086] Step 3: Based on the distribution information of the solubility values ​​of all variant amino acids in each solubility value distribution interval, and the distribution information of each characteristic value of all variant amino acids in each distribution interval of that characteristic value, calculate the mutual information value between each characteristic value and the solubility of the variant amino acid.

[0087] Specifically, mutual information value The calculation formula is:

[0088] ,

[0089] Where X represents the distribution interval of the eigenvalues; This represents the number of eigenvalues ​​distributed within X; Indicates the range of solubility values; This represents the number of solubility values ​​distributed within Y; The marginal probability distribution of eigenvalues ​​represents the probability of eigenvalues ​​distributed within the distribution intervals of each eigenvalue. This represents the marginal probability distribution of solubility values, i.e., the probability of solubility values ​​falling within each solubility value distribution interval. This represents the joint probability density of eigenvalues ​​and solubility values.

[0090] Step 4: Arrange the mutual information values ​​between various feature values ​​and the solubility of the variant amino acid in descending order, select the N feature values ​​corresponding to the top N mutual information values ​​as N target feature values; obtain the feature value set based on the N target feature values ​​of the variant amino acid.

[0091] Mutual information (MI) is a metric that measures the correlation between two variables. It can capture the non-linear relationship between features and the target variable, making it particularly effective when processing high-dimensional biological data. For protein solubility prediction, the target variable is protein solubility, while the feature variables include multiple biophysical properties of amino acids (such as hydrophobicity, charge distribution, and molecular size). By calculating the mutual information between each feature and solubility, features significantly correlated with the target variable can be preliminarily screened, while noisy features unrelated to solubility are eliminated, reducing data dimensionality and improving prediction accuracy.

[0092] Alternatively, hierarchical clustering or correlation coefficient analysis can be used to further refine the N selected feature values. By aggregating similar feature values ​​into one class, feature values ​​with low correlation can be removed. For example, features such as hydrophobicity and molecular size have a strong correlation, while conservative features have a low correlation. Therefore, conservative features can be regarded as redundant features and removed.

[0093] Furthermore, based on the N target feature values ​​of the variant amino acids, its feature value set is obtained, including:

[0094] Step 4-1: Input each target feature value of each variant amino acid into the trained variant amino acid solubility prediction model, and calculate the SHAP value of each target feature value of the variant amino acid based on the solubility prediction value and the solubility value of the variant amino acid.

[0095] Specifically, the prediction model for the solubility of variant amino acids can be the XGBoost model.

[0096] Step 4-2: Calculate the mean SHAP value of each target feature for all variant amino acids to obtain the mean SHAP value of each target feature.

[0097] Step 4-3: Arrange the SHAP mean values ​​of various target features from largest to smallest, remove the M target features corresponding to the last M SHAP mean values, and obtain the feature set based on the remaining target feature values ​​of the mutated amino acids.

[0098] Optionally, in some embodiments, after each removal of some target feature values, the process can return to step 4-1 and gradually remove target feature values ​​with lower importance through multiple iterations.

[0099] Specifically, based on feature filtering using mutual information, this application also introduces recursive feature elimination driven by SHAP values. The SHAP value is used to quantify the contribution of each target feature value to the solubility prediction model, thereby eliminating feature values ​​that have little impact on protein solubility values, thus retaining important features and improving the final prediction accuracy.

[0100] In high-dimensional biological data, feature selection is a key step in improving model performance and efficiency. In the task of predicting the impact of protein solubility, due to the large number of feature values ​​and high feature dimensionality, it is difficult to improve the prediction accuracy by selecting the most informative and relevant features from a variety of features. This application designs a two-stage mitigating feature selection method based on mutual information and SHAP value to remove redundant features and retain important features, thereby improving the prediction accuracy and the prediction efficiency of the model.

[0101] Furthermore, since each base classifier has a different classification task and focuses on different feature values, this application also introduces a recursive feature elimination strategy for feature selection, thereby optimizing the training of the base classifiers in the model. As for the meta-learner, its input is the output of the base classifier after feature selection, the purpose of which is to comprehensively consider the characteristics of each base classifier, so there is no need to perform feature selection again.

[0102] Specifically, after obtaining the set of variant amino acids and their feature values ​​in the dataset, the process also includes:

[0103] By using a recursive feature elimination strategy to filter the feature values ​​in the feature value set, we can obtain the feature subsets corresponding to the three-classification module and the feature subsets corresponding to each base classifier.

[0104] Input the feature subset corresponding to the three-classification module into the three-classification module, and input the feature subset corresponding to each base classifier into each base classifier respectively.

[0105] For example, such as Figure 3 The flowchart shown is a feature selection process using a recursive feature elimination strategy. The feature selection process for each base classifier specifically includes: inputting the feature values ​​in the feature value set of the variant amino acids individually or in random combinations into the base classifier; evaluating the importance of the input feature values ​​or feature value combinations to the base classifier based on the output of each base classifier; removing the feature values ​​or feature value combinations with the lowest importance; until the number of remaining feature values ​​is equal to a preset value or the accuracy of the output result of the base classifier reaches a preset threshold; and combining the remaining feature values ​​to obtain the feature value subset of the base classifier.

[0106] Specifically, since different learners have different criteria for judging the importance of feature values, when the base classifier is a LightGBM base classifier, the frequency of use of a feature value or feature value combination or the information gain can be used as the criterion for judging the importance of that feature value or feature value combination to the base classifier.

[0107] This application uses a recursive feature elimination method in a multi-layer stacked model, which can not only effectively reduce the number of features and speed up model training and prediction, but also remove invalid and feature values ​​that affect the prediction accuracy of the base classifiers based on the characteristics of each base classifier, and retain the most informative feature values, thereby improving the prediction accuracy of the prediction model.

[0108] Furthermore, considering the interaction between multiple learners in a multi-layer stacked model, this application uses iterative random parameter search and cross-validation to iteratively train the three-classification module, each base classifier, and the meta-learner during model training.

[0109] Specifically, after inputting the first probability vector, the second probability vector, the third probability vector, and the fourth probability vector into the meta-learner and outputting the predicted category of the mutated amino acid, the process also includes:

[0110] The loss function is calculated based on the first probability vector, the second probability vector, the third probability vector, the fourth probability vector, and the predicted category.

[0111] Based on the variable amino acids and their feature sets in the dataset, iterative training of the three-classification module, each base classifier, and the meta-learner is performed using iterative random parameter search and cross-validation until the loss function is minimized. Based on the trained three-classification module, each base classifier, and the meta-learner, a trained prediction model is obtained.

[0112] like Figure 4The diagram shows a flowchart of iterative training of the three-class classification module, each base classifier, and the meta-learner using iterative random parameter search and cross-validation. Specifically, the iterative training steps for each learner (three-class classification module, each base classifier, and meta-learner) include:

[0113] 1. Obtain all parameters to be optimized for the learner, set multiple candidate values ​​for each parameter to be optimized, and obtain a list of candidate values ​​based on all candidate values ​​of all parameters to be optimized;

[0114] 2. Randomly select a candidate value for each parameter to be optimized from the candidate value list as the parameter of the learner, obtain the set of variant amino acids and their feature values ​​from the database, input the feature values ​​in the feature value set into the learner, and calculate the average performance score of the learner using the cross-validation method.

[0115] 3. Randomly select candidate values ​​for each parameter to be optimized from the candidate value list to update the parameters of the learner, and re-obtain the variant amino acids and their feature value set from the database. Input the feature values ​​in the feature value set into the updated learner, and calculate the average performance score of the learner again using the cross-validation method.

[0116] 4. Repeat step 3 until the preset number of executions is reached, and select the combination of candidate parameters with the best average performance score of the learner as the optimal parameters of the learner.

[0117] Optionally, such as Figure 4 As shown, after iteratively training each learner to obtain its optimal parameters, we can also obtain the set of mutated amino acids and their feature values ​​from the database. The feature values ​​in the feature value set are then input into each learner, and the result of predicting the effect of the mutated amino acid on protein solubility is output. The average performance score of the entire prediction model is calculated based on the effect result using cross-validation, and then the process returns to step 2 until the preset number of iterations is reached. Thus, not only can the performance of each learner be optimized through parameter optimization, but the overall performance of the entire prediction model can also be optimized.

[0118] Since multi-layer stacked models contain multiple learners, iterating through all parameter selections is costly and leads to low training efficiency. Excessive parameter search can also cause overfitting. This application uses an iterative random parameter search method to update the parameters of each learner, which is more efficient and makes it easier to escape local optima, thereby improving the training speed and prediction accuracy of the model.

[0119] Furthermore, the steps for calculating the loss function include:

[0120] Using the Focal Loss function, the first classification loss of the three-classification module is calculated based on the first probability vector output by the three-classification module and the true category of the mutated amino acid.

[0121] Using the Focal Loss function, the second classification loss of each two-layer three-classification module is calculated based on the probability vectors output by each module and the true category of the mutated amino acid.

[0122] Using the cross-entropy loss function, the third classification loss of the meta-learner is calculated based on the predicted category output by the meta-learner and the true category of the mutated amino acid.

[0123] The value of the loss function is obtained by weighting the first classification loss, each second classification loss, and the third classification loss.

[0124] The formulas for calculating the first classification loss and the second classification loss are as follows:

[0125] ,

[0126] in, The adjustment weights, representing class imbalance, are increased for classes with less training data. To increase the contribution of this category to the loss; This represents the probability vector output by a three-class classification module or a two-layer three-class classification module. , , , These represent the predicted probabilities that the mutated amino acid belongs to one of the three categories, respectively. Adjustment weights representing the difficulty of samples reduce the loss contribution of easily classified samples, making the model pay more attention to difficult samples. Indicates the true category of the variant amino acid. , , , These represent the true probabilities of the mutated amino acid belonging to the three categories, respectively. When the mutated amino acid belongs to the first category... When the mutated amino acid belongs to the second category, When the mutated amino acid belongs to the third category, .

[0127] The formula for calculating the third category loss is:

[0128] ,

[0129] in, This represents the predicted class probability vector of the mutated amino acids output by the meta-learner. , , , These represent the predicted probabilities of the mutated amino acids output by the meta-learner belonging to the three categories, respectively.

[0130] The formula for calculating the loss function is:

[0131] ,

[0132] in, Indicates the first category loss; , , These represent the second classification losses of the three two-layer three-classification modules, respectively.

[0133] The method proposed in this application for predicting the impact of amino acid variations on protein solubility expands the dataset through literature mining, improving the training foundation of the model. Simultaneously, the prediction model constructed using a multi-layer stacking strategy addresses the problem of decreased prediction accuracy caused by uneven sample distribution in the dataset, improving the model's prediction accuracy across various datasets. Compared to the traditional two-layer three-classification strategy, this prediction model does not require adjustment based on data distribution characteristics, exhibiting greater versatility. Furthermore, combining feature selection, recursive feature elimination strategies, and iterative random parameter search for feature selection and parameter tuning further enhances the performance of the prediction model. Moreover, this prediction model demonstrates excellent prediction performance and range capability on the test set, achieving a Matthews Correlation Coefficient (MCC) of 0.501. The predictive model provided in this application can not only accurately predict the impact of amino acid variations on protein solubility, but also predict the pathogenicity of amino acid variations in diseases related to changes in protein solubility. This helps to identify amino acid variations associated with diseases such as Alzheimer's disease and cataracts, and can be further applied to fields such as protein engineering and protein crystallization. It provides support for the mechanism analysis of related variations and the design of potential treatment strategies, and provides a reference for biomedical research and application.

[0134] To further verify the effectiveness of the prediction model provided in this application, this application used the trained prediction model to predict the effect of all possible amino acid variations on Bruton's tyrosine kinase (BTK) on solubility. Specifically, for each site, there are 19 amino acid variations and their corresponding prediction results, such as... Figure 5 The image shown is a schematic diagram of the prediction results, in which... Figure 5 In the figure, (a) represents the number of variant amino acids on BTK that increase protein solubility. Figure 5 In the figure, (b) represents the number of variant amino acids on BTK that have no effect on protein solubility. Figure 5In the figure, (c) represents the number of variant amino acids on BTK that lead to reduced protein solubility. Furthermore, to investigate the relationship between changes in protein solubility and disease, this application also predicts the pathogenic effects of all possible amino acid variants, such as... Figure 5 (d) in the equation represents the association between variant amino acids that cause changes in protein solubility and pathogenic variant amino acids.

[0135] Studies have shown that low protein solubility is associated with many human diseases. Figure 5 (c) and Figure 5 As can be seen from (d) in the figure, there is an overlap between the variant amino acids that lead to reduced protein solubility and the variant amino acids that cause pathogenicity. This indicates that there is a certain correlation between the amino acid variants that lead to reduced protein solubility and the amino acid variants that cause pathogenicity. Therefore, the prediction results of this application are consistent with the theoretical study, which verifies that the prediction model provided by this application has high prediction accuracy and that its prediction results have a significant auxiliary effect on pathogenic amino acid variants.

[0136] Based on the method for predicting the effect of amino acid variations on protein solubility provided in the above embodiments, this application also provides an apparatus for predicting the effect of amino acid variations on protein solubility, such as... Figure 6 As shown, the device specifically includes:

[0137] The data acquisition module 10 is used to acquire the variable amino acids and their characteristic value sets in the dataset.

[0138] The first three-classification module 20 is used to input the feature values ​​in the feature value set into the three-classification module and output the first probability vector of the variant amino acid belonging to different predicted categories.

[0139] The second and third classification module 30 is used to input the feature values ​​in the feature value set into three two-layer three-class classification modules corresponding to different prediction categories. Based on the output of each two-layer three-class classification module, the second probability vector, the third probability vector, and the fourth probability vector of the mutated amino acid belonging to different prediction categories are obtained.

[0140] The category prediction module 40 is used to input the first probability vector, the second probability vector, the third probability vector and the fourth probability vector into the meta-learner and output the predicted category of the mutated amino acid.

[0141] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above for predicting the effect of amino acid variations on protein solubility.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0146] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method of predicting the effect of an amino acid variation on protein solubility, characterized by, The method comprises the following steps: obtaining a set of characteristic values of a variant amino acid in a data set; Specifically, the method comprises the following steps: obtaining a plurality of characteristic values and solubility values of each variant amino acid in the data set; discretizing the solubility values of all variant amino acids and each characteristic value of all variant amino acids to obtain a plurality of solubility value distribution intervals and a plurality of distribution intervals of each characteristic value; calculating mutual information values between each characteristic value and the solubility of the variant amino acid based on the distribution information of the solubility values of all variant amino acids in each solubility value distribution interval and the distribution information of each characteristic value of all variant amino acids in each distribution interval of the characteristic value; arranging the mutual information values between the plurality of characteristic values and the solubility of the variant amino acid from large to small, selecting N characteristic values corresponding to the first N mutual information values as N target characteristic values, and obtaining a set of characteristic values of the variant amino acid based on the N target characteristic values of the variant amino acid; wherein the set of characteristic values of the variant amino acid based on the N target characteristic values of the variant amino acid comprises the following steps: inputting each target characteristic value of each variant amino acid into a trained variant amino acid solubility prediction model, and calculating the SHAP value of each target characteristic value of the variant amino acid based on the solubility prediction value and the solubility value of the variant amino acid; calculating the mean value of the SHAP value of each target characteristic value of all variant amino acids to obtain the SHAP mean value of each target characteristic value; arranging the SHAP mean values of the target characteristic values from large to small, removing M target characteristic values corresponding to the last M SHAP mean values, and obtaining a set of characteristic values of the variant amino acid based on the remaining target characteristic values of the variant amino acid; inputting the characteristic values in the set of characteristic values into a three-classification module to output a first probability vector of the variant amino acid belonging to different prediction categories; inputting the characteristic values in the set of characteristic values into three two-layer three-classification modules corresponding to different prediction categories, and obtaining a second probability vector, a third probability vector and a fourth probability vector of the variant amino acid belonging to different prediction categories based on the outputs of the two-layer three-classification modules; specifically, the method comprises the following steps: inputting the characteristic values in the set of characteristic values into a first base classifier to output the probability of the variant amino acid belonging to a first prediction category; inputting the characteristic values in the set of characteristic values into a second base classifier to output the probability of the variant amino acid belonging to a second prediction category and a third prediction category; and obtaining the second probability vector of the variant amino acid belonging to different prediction categories based on the outputs of the first base classifier and the second base classifier; inputting the characteristic values in the set of characteristic values into a third base classifier to output the probability of the variant amino acid belonging to the third prediction category; inputting the characteristic values in the set of characteristic values into a fourth base classifier to output the probability of the variant amino acid belonging to the first prediction category and the second prediction category; and obtaining the third probability vector of the variant amino acid belonging to different prediction categories based on the outputs of the third base classifier and the fourth base classifier. inputting the feature values in the feature value set into the fifth base classifier to output probabilities of the variant amino acid belonging to a second predicted category; inputting the feature values in the feature value set into the sixth base classifier to output probabilities of the variant amino acid belonging to a first predicted category and a third predicted category; obtaining a fourth probability vector of the variant amino acid belonging to different predicted categories based on outputs of the fifth base classifier and the sixth base classifier; inputting the first probability vector, the second probability vector, the third probability vector and the fourth probability vector into the meta-learner to output a predicted category of the variant amino acid; calculating a value of a loss function based on the first probability vector, the second probability vector, the third probability vector, the fourth probability vector and the predicted category; iteratively training the three-classification module, the base classifiers and the meta-learner based on the variant amino acids in the data set and their feature value sets by using an iterative random parameter search and a cross-validation method until the value of the loss function is minimum, and obtaining a trained prediction model based on the trained three-classification module, the base classifiers and the meta-learner.

2. The method of predicting the effect of an amino acid variation on protein solubility according to claim 1, wherein, After obtaining the variant amino acids in the data set and their feature value sets, the method further includes: filtering the feature values in the feature value set by using a recursive feature elimination strategy to obtain a feature subset corresponding to the three-classification module and feature subsets corresponding to the base classifiers; inputting the feature subset corresponding to the three-classification module into the three-classification module and inputting the feature subsets corresponding to the base classifiers into the base classifiers respectively.

3. The method of predicting the effect of an amino acid variation on protein solubility according to claim 1, wherein, The calculation steps of the loss function include: calculating a first classification loss of the three-classification module based on the first probability vector output by the three-classification module and the true category of the variant amino acid by using a Focal Loss loss function; calculating a second classification loss of each two-layer three-classification module based on the probability vector output by the two-layer three-classification module and the true category of the variant amino acid by using a Focal Loss loss function; calculating a third classification loss of the meta-learner based on the predicted category output by the meta-learner and the true category of the variant amino acid by using a cross-entropy loss function; obtaining the value of the loss function based on a weighted sum of the first classification loss, the second classification losses and the third classification loss.

4. The method of predicting the effect of an amino acid variation on protein solubility according to claim 1, wherein, The construction process of the data set includes: performing keyword search in a PubMed database to obtain a literature set based on the search results; wherein the keywords include one or more of protein, variation and solubility; obtaining a literature subset from the literature set by selecting the same literature as in the VariBench database; identifying the literature in the literature subset by using a data mining tool to obtain literature containing amino acid variation information, and obtaining a target literature set; extracting the amino acid variation data and the corresponding protein solubility change data in each literature in the target literature set to obtain the data set.

5. An apparatus for predicting the effect of an amino acid variation on protein solubility, comprising: The method includes: a data acquisition module configured to obtain variant amino acids in a data set and their feature value sets; specifically including: obtaining multiple feature values and solubility values of each variant amino acid in the data set; discretizing the solubility values of all variant amino acids and each feature value of all variant amino acids to obtain multiple solubility value distribution intervals and multiple distribution intervals of each feature value; The mutual information value between each feature value and the solubility of the variant amino acid is calculated based on distribution information of solubility values of all variant amino acids in each solubility value distribution interval and distribution information of each feature value of all variant amino acids in each distribution interval of the feature value; The mutual information values between multiple feature values and the solubility of the variant amino acid are arranged from large to small, and N feature values corresponding to the first N mutual information values are selected as N target feature values; a feature value set of the variant amino acid is obtained based on the N target feature values of the variant amino acid; wherein the feature value set of the variant amino acid is obtained based on the N target feature values of the variant amino acid, comprising: The SHAP value of each target feature value of each variant amino acid is calculated by inputting each target feature value of each variant amino acid into the trained variant amino acid solubility prediction model and based on the solubility prediction value and the solubility value of the variant amino acid; The SHAP mean value of each target feature value is obtained by calculating the mean value of the SHAP values of each target feature value of all variant amino acids; The SHAP mean values of various target feature values are arranged from large to small, and the last M target feature values corresponding to the last M SHAP mean values are removed, and a feature value set of the variant amino acid is obtained based on the remaining target feature values of the variant amino acid; The first three-classification module is configured to input the feature values in the feature value set into the three-classification module and output a first probability vector of the variant amino acid belonging to different prediction categories; The second three-classification module is configured to input the feature values in the feature value set into three two-layer three-classification modules corresponding to different prediction categories, and obtain a second probability vector, a third probability vector, and a fourth probability vector of the variant amino acid belonging to different prediction categories based on outputs of the three two-layer three-classification modules; and the second three-classification module specifically comprises: The first base classifier is configured to input the feature values in the feature value set and output a probability of the variant amino acid belonging to a first prediction category; the second base classifier is configured to input the feature values in the feature value set and output probabilities of the variant amino acid belonging to a second prediction category and a third prediction category; and the second probability vector of the variant amino acid belonging to different prediction categories is obtained based on outputs of the first base classifier and the second base classifier; The third base classifier is configured to input the feature values in the feature value set and output a probability of the variant amino acid belonging to the third prediction category; the fourth base classifier is configured to input the feature values in the feature value set and output probabilities of the variant amino acid belonging to the first prediction category and the second prediction category; and the third probability vector of the variant amino acid belonging to different prediction categories is obtained based on outputs of the third base classifier and the fourth base classifier; The fifth base classifier is configured to input the feature values in the feature value set and output a probability of the variant amino acid belonging to the second prediction category; the sixth base classifier is configured to input the feature values in the feature value set and output probabilities of the variant amino acid belonging to the first prediction category and the third prediction category; and the fourth probability vector of the variant amino acid belonging to different prediction categories is obtained based on outputs of the fifth base classifier and the sixth base classifier. The category prediction module is configured to input the first probability vector, the second probability vector, the third probability vector and the fourth probability vector into the meta-learner, and output a predicted category of the variant amino acid; calculate a value of a loss function based on the first probability vector, the second probability vector, the third probability vector, the fourth probability vector and the predicted category; and perform iterative training on the three-classification module, each base classifier and the meta-learner based on the variant amino acid and the set of characteristic values thereof in the data set by using an iterative random parameter search and a cross-validation method until the value of the loss function is minimum, and obtain a trained prediction model based on the trained three-classification module, each base classifier and the meta-learner.

6. A computer readable storage medium characterized by, The computer program is stored on the computer readable storage medium and is executed by the processor to implement the steps of the method for predicting the influence of the amino acid variation on the protein solubility according to any one of claims 1 to 4.

Images