Method for constructing protein thermal stability point mutation prediction model and prediction method

Abstract

The protein thermal stability point mutation prediction model construction method disclosed in the application selects multi-dimensional physical quantities within a preset distance range around the target residue, can more comprehensively reflect the structural characteristics of the protein, and provides a richer information basis for predicting the protein thermal stability point mutation. When training the three-dimensional convolutional neural network, the large-size convolution kernel can capture the interaction between atoms in a large range, and reveal the complex relationship; the small-size convolution kernel can focus on local details, and the combination of the two can comprehensively describe the structural characteristics from different scales, so as to ensure that the model learns the key information. The normalization method of the transfer learning neural network emphasizes the information provided by the experimental data, and at the same time, keeps the benchmark probability of the unmutated amino acid as a reference, so that the influence of the amino acid mutation on the protein thermal stability and functional activity can be more accurately predicted, especially in the case of lacking comprehensive experimental data, reasonable inference can still be made by relying on known experimental information and benchmark values.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of protein engineering, and more particularly, to a protein thermal stability point mutation prediction model construction method and a prediction method. BACKGROUND

[0002] Protein thermal stability is a key research area in protein engineering, which is of great significance for understanding protein function, optimizing protein performance, and developing new biological drugs. In the related technical field, it involves the intersection of bioinformatics, computational biology, and protein chemistry, aiming to reveal the relationship between protein structure and function through theoretical calculation and experimental research, and on this basis, to design and modify proteins.

[0003] The existing technology mainly adopts the following methods in the research of protein thermal stability point mutation. Traditional experimental methods usually rely on site-directed mutagenesis of proteins, and then determine the changes in thermal stability of the mutants through experimental methods such as differential scanning calorimetry, dynamic light scattering, etc. Although this method is direct, it requires a large amount of time and resources, because each mutant needs to go through a tedious experimental preparation and testing process. At the same time, due to the complexity of protein structure and function, it is inefficient to screen mutation sites one by one through experiments, and it is difficult to fully explore the influence of all possible mutation combinations on thermal stability.

[0004] Computational methods to some extent make up for the shortcomings of experimental methods. Some studies use molecular dynamics simulation to predict the structural changes of proteins after mutation, and then evaluate their influence on thermal stability. This method can simulate the dynamic behavior of proteins in a virtual environment, but the calculation cost is high, and the accuracy of simulation is limited by factors such as force field parameters and simulation length. In addition, machine learning-based methods have also been gradually applied to protein thermal stability prediction, such as training classification or regression models using known protein structures and thermal stability data to predict the influence of mutations on thermal stability. However, existing machine learning models often lack comprehensive feature extraction, failing to fully capture key information related to thermal stability in protein structure, resulting in limited prediction accuracy. SUMMARY

[0005] In view of at least one defect or improvement demand of the prior art, the present application provides a protein thermal stability point mutation prediction model construction method and a prediction method to overcome some defects mentioned in the background section.

[0006] To achieve the above-mentioned purpose, in a first aspect, the present application provides a protein thermal stability point mutation prediction model construction method, comprising:

[0007] The three-dimensional structure of a preset type of protein is acquired, a preset type of physical quantity within a preset distance range around a target residue is selected for raster processing to form a plurality of three-dimensional matrices, and the three-dimensional matrices are merged into a four-dimensional matrix with the type of physical quantity as a channel;

[0008] A three-dimensional convolutional neural network is trained using the four-dimensional matrix as training data to construct an initial protein thermostability point mutation prediction model; the three-dimensional convolutional neural network includes a preset large-size convolutional kernel for extracting global features and a preset small-size convolutional kernel for extracting local features;

[0009] The output of a previous layer of the initial protein thermostability point mutation prediction model output layer is taken as input, a normalization method that adjusts the probability of amino acid mutation based on enzyme activity experimental data and / or a reference probability of keeping an unmutated amino acid when there is no enzyme activity experimental data is used, a transfer learning neural network is trained, and an optimized final protein thermostability point mutation prediction model is obtained.

[0010] Further, the preset distance range is specifically:

[0011] A region between a spherical region with a radius of 5 to 15 angstroms centered on the target residue.

[0012] Further, the distance between the target residue and the surrounding atoms is weighted by a distance weighting function w(r), and the specific formula includes:

[0013]

[0014] wherein r represents the distance between the target residue and the surrounding atoms; a represents an adjustment parameter; and w(r) represents the weighted distance from the target residue.

[0015] Further, the preset type of physical quantity includes physical quantities reflecting the structure characteristics of the protein from the atomic level, the intermolecular force level, and the chemical group level.

[0016] The preset type of physical quantity specifically includes one or more of the number of carbon atoms, the number of nitrogen atoms, the solvent accessible surface area, the average charge value, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the number of aromatic rings.

[0017] Further, the sizes of the preset large-size convolutional kernel and the preset small-size convolutional kernel of the three-dimensional convolutional layer of the three-dimensional convolutional neural network are 3×3×3 and 2×2×2, respectively, and the step size is set to 1.

[0018] Further, the three-dimensional convolutional layer performs convolution operation on the input matrix through sliding window operation, and the convolution kernel and each local region of the input perform element-level multiplication and summation operation to obtain the feature value at the corresponding position, and the specific formula includes:

[0019]

[0020] wherein, I x , I y , I z respectively represent the size of the input matrix in three dimensions; f represents the size of the convolution kernel; s represents the step size; O x , O y , O z respectively represent the size of the convolution output matrix in the corresponding dimension.

[0021] Further, the normalization method of adjusting the amino acid mutation probability in combination with enzyme activity experimental data includes:

[0022] Assume that the initial mutation probability of all amino acids is 100%;

[0023] If the experimental data shows that the mutation of a certain amino acid improves the activity of the enzyme, the mutation probability corresponding to the amino acid will be increased accordingly; otherwise, it will be reduced accordingly;

[0024] The adjusted mutation probability is standardized so that the sum of all mutation probabilities is 100%; the specific formula includes:

[0025]

[0026] wherein, a i represents the relative activity percentage of the i-th amino acid associated with the experimental data; N represents the total number of all amino acids; represents the sum of the relative activity percentages of all amino acids; represents the normalized amino acid mutation probability under the scenario of enzyme activity experimental data;

[0027] The normalization method of keeping the benchmark probability of the unmutated amino acid as a reference when there is no enzyme activity experimental data includes:

[0028] Set the mutation probability of the amino acid without enzyme activity experimental data to 0%;

[0029] Set the mutation probability of the unmutated amino acid to 100% as a benchmark value, so as to ensure that in the case of no mutation, the prediction of the model can take its initial state as a reference;

[0030] The specific formula includes:

[0031]

[0032] wherein, p i represents the mutation probability after correction by experimental data; represents the sum of all amino acid mutation probabilities; The normalized amino acid mutation probability under the scenario of no enzyme activity experimental data.

[0033] Further, the full connection layer of the transfer learning neural network adopts ReLU as the activation function.

[0034] The loss function used in training the three-dimensional convolutional neural network includes:

[0035]

[0036] Wherein, L(θ) represents the total loss function; The mutation probability predicted by the network; y i The actual mutation probability; θ j The weight parameter of the network; N represents the total number of training samples; λ represents the regularization term coefficient; m represents the number of parameters in the network.

[0037] In the second aspect, the present application provides a protein thermal stability point mutation prediction method, the structure data of the protein to be studied is input into the initial protein thermal stability point mutation prediction model or the final protein thermal stability point mutation prediction model in the prediction model construction method of any of the preceding description for processing, and the target amino acid of the mutation site and the required mutation is predicted.

[0038] Further, the standard deviation of each candidate site and the mutation probability of the actual amino acid are evaluated, and the sites meeting the conditions are sorted;

[0039] If the standard deviations of multiple sites are higher than the first preset value, these sites are sorted and recommended in the order of standard deviation from large to small; the sorting criteria related to the standard deviation include:

[0040]

[0041] Wherein, σ n Indicates the standard deviation corresponding to the candidate site n; Indicates the standard deviation sorting value;

[0042] If the mutation probabilities of the actual amino acids of multiple sites are not higher than the second preset value, they are sorted and recommended in the order of mutation probability from small to large; the sorting criteria related to the mutation probability include:

[0043]

[0044] Wherein, p i Indicates the mutation probability of the candidate site i; Indicates the mutation probability sorting value;

[0045] determining a final recommended order based on the weight of the standard deviation ranking value and the mutation probability ranking value, and the specific formula includes:

[0046]

[0047] wherein w σ and w p respectively represent the weight of the standard deviation ranking value and the weight of the mutation probability ranking value, w σ +w p =1; Rank i represents a comprehensive ranking value;

[0048] According to the comprehensive ranking value, the mutation site and the target amino acid of the required mutation are recommended, and the mutation site ranked at the top based on the comprehensive ranking value is preferentially selected for experimental research.

[0049] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0050] The protein thermal stability point mutation prediction model construction method disclosed in the present application selects multi-dimensional physical quantities within a preset distance range around the target residue, can more comprehensively reflect the structural characteristics of the protein, and provides a richer information basis for predicting protein thermal stability point mutations. When training the three-dimensional convolutional neural network, the large-size convolution kernel can capture the interaction between atoms in a larger range, revealing complex relationships; the small-size convolution kernel can focus on local details, and the combination of the two can comprehensively describe the structural characteristics from different scales, ensuring that the model learns key information. Further, the normalization method of the transfer learning neural network emphasizes the information provided by the experimental data, while maintaining the benchmark probability of the unmutated amino acid as a reference, which can more accurately predict the influence of amino acid mutation on the thermal stability and functional activity of the protein, especially in the absence of comprehensive experimental data, it can still rely on known experimental information and benchmark values to make reasonable inferences. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0052] Figure 1 A core flowchart of a protein thermal stability point mutation prediction model construction method provided in the embodiments of the present application;

[0053] Figure 2A system architecture diagram of a protein thermal stability point mutation recommendation system based on transfer learning provided by an embodiment of the present application is shown in FIG. 1.

[0054] Figure 3 A feature engineering unit data processing flowchart provided by an embodiment of the present application is shown in FIG. 4.

[0055] Figure 4 A three-dimensional convolutional neural network data processing flowchart provided by an embodiment of the present application is shown in FIG. 5.

[0056] Figure 5 A transfer learning neural network operation flowchart provided by an embodiment of the present application is shown in FIG. 6.

[0057] Figure 6 A recommendation unit sorting logic diagram provided by an embodiment of the present application is shown in FIG. 7.

[0058] Figure 7 A block diagram of an electronic device suitable for implementing the protein thermal stability point mutation prediction model construction method and / or prediction method described above is shown in FIG. 8. DETAILED DESCRIPTION

[0059] In order to make the objects, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0060] The terms "first", "second" or "n-th" and the like in the specification, claims or drawings of the present application are used to distinguish different objects and are not intended to describe a specific order. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units that are not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0061] The present application selects multi-dimensional physical quantities through a feature engineering unit, and adopts a reasonable distance range and a weighting mechanism, so as to more comprehensively and accurately reflect the structural characteristics of the protein, and provide more accurate information for subsequent prediction. The prior art is relatively single in feature extraction, and cannot fully consider various key factors in the protein structure, thereby affecting the accuracy of prediction. In terms of neural network architecture, the convolutional neural network module and the transfer learning neural network module of the present application improve the learning ability and prediction performance of the model through optimized structure design (such as convolution layer parameter setting, full connection layer and pooling layer cooperation, use of a pre-trained model and a ReLU activation function, etc.). In contrast, the calculation model in the prior art may have unreasonable places in structure design, high calculation cost and unsatisfactory prediction effect. The normalization method of the present application dynamically adjusts the amino acid mutation probability in combination with experimental data, and reasonably processes special cases, so that the prediction is more in line with the actual situation. However, the prior art has deficiencies in optimizing prediction in combination with experimental data, and it is difficult to accurately reflect the comprehensive influence of mutation on the function and thermal stability of the protein. The recommendation unit of the present application optimizes the recommendation order based on multi-criteria sorting, and can provide more targeted and prioritized mutation sites for experiments. However, the prior art lacks systematicness and accuracy in mutation site recommendation, and it is difficult to efficiently guide experimental research. In summary, the prior art has many limitations in the study of protein thermal stability point mutation, and the present application aims to overcome these defects and provide a more effective prediction and recommendation technology.

[0062] One embodiment of the present application provides a recommendation scheme for protein thermal stability point mutation based on transfer learning. The recommendation system mainly includes four module units: a feature engineering unit, a convolutional neural network module, a transfer learning neural network module and a recommendation unit. The first three module units are essentially the construction process of the protein thermal stability point mutation prediction model, which can be referred to in Figure 1 ; and the last module unit is the specific application of the constructed prediction model, which can be referred to in Figure 2 . The specific technical details of the four parts are introduced one by one as follows.

[0063] (1) Feature engineering unit

[0064] The three-dimensional structure data of various proteins, including but not limited to enzymes, antibodies, structural proteins, and other types of proteins, are obtained from the RCSB Protein Structure Database. Physical quantities within a specific distance range around the target residue are selected for gridding processing to form multiple three-dimensional matrices, which are then merged into a four-dimensional matrix using the type of physical quantity as the channel to serve as the input data for the convolutional neural network module. The physical quantities include the number of carbon atoms, the number of nitrogen atoms, the number of other types of atoms, the solvent accessible surface area, the average charge, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the number of aromatic rings. These physical quantities reflect the structural characteristics of proteins from multiple dimensions, such as atomic level, intermolecular force level, and chemical group level, providing comprehensive structural information input for subsequent neural network training.

[0065] The public protein structure database uses the RCSB Protein Structure Database, which includes three-dimensional structure data from proteins of various biological sources and with different functions. By using these data, the convolutional neural network module can capture the universal structural features between different proteins during training, while also identifying the structural differences of proteins with specific functions or sources. The structural data provided by the database not only includes the spatial arrangement of proteins and the interaction between atoms, but also covers multiple dimensions of physical and chemical information such as solvent accessible surface area, charge distribution, number of hydrogen bond donors and acceptors. These data provide rich input for the convolutional neural network, ensuring that the network can learn features at multiple levels such as protein conformation changes, interactions between amino acid residues, and protein stability.

[0066] In the feature engineering unit, the specific distance range for selecting physical quantities around the target residue is a spherical region centered on the target residue with a radius of 5 to 15 angstroms. As shown in Figure 3 The data processing flowchart of the feature engineering unit is shown in the figure. The selection of this distance range is based on a deep understanding of the relationship between protein thermal stability and the surrounding environment. By selecting this range, both key atoms and chemical groups interacting with the target residue can be covered, and excessive unrelated structural information can be avoided. Generally, a smaller distance range may result in insufficient local information, which cannot fully reflect the structural environment of the target residue, while a larger range may include too much background noise unrelated to the target residue, affecting the training effect of the subsequent model. Therefore, by selecting physical quantities within the range of 5 to 15 angstroms, the integrity of the information and the suppression of noise can be effectively balanced, providing accurate and representative input features for the neural network.

[0067] In the process of implementation, the feature engineering unit first determines the position of the target residue and collects physical quantities within a spherical region centered at the residue with a radius of 5 to 15 angstroms. These physical quantities include the number of carbon atoms, the number of nitrogen atoms, the charge distribution, the number of hydrogen bond donors and acceptors, the number of aromatic rings, and the like, which to some extent reflect the local chemical environment of the target residue and its interaction with surrounding atoms and groups. Subsequently, these physical quantities are converted into multiple three-dimensional matrices through gridding processing, and finally merged into a four-dimensional matrix with the type of physical quantity as the channel, which is sent as input data into the neural network for subsequent training.

[0068] To further ensure the rationality of the selection of these physical quantities, the distance between the target residue and the surrounding atoms can be weighted by a distance weighting function w(r) as follows, which represents the contribution of atoms in different distance ranges to the thermal stability of the target residue. The weighting function can be expressed as:

[0069]

[0070] where r is the distance between the target residue and the surrounding atoms; and a is an adjustment parameter that controls the degree of influence of distance on physical quantities. The weighting function shows that atoms closer to the target residue have a greater impact on thermal stability, so they have higher weights, while atoms farther away from the target residue have a gradually decreasing contribution to thermal stability.

[0071] In practical applications, this weighting method can be adjusted according to the force of atoms and chemical groups in different distance ranges on thermal stability. By assigning reasonable weights to the selected physical quantities, the input data can more accurately reflect the local environmental features of the target residue. This process not only helps the neural network learn the impact of local changes in protein structure on thermal stability, but also enhances the generalization ability of the model, ensuring that it can accurately predict the thermal stability of different protein structures.

[0072] Therefore, the feature engineering unit effectively improves the representativeness of the input data through a reasonable distance selection and weighting mechanism, providing accurate and comprehensive structural features for the subsequent convolutional neural network module and transfer learning neural network module, thereby enhancing the system's performance in predicting protein thermal stability.

[0073] The present application selects a plurality of physical quantities such as the number of carbon atoms, the number of nitrogen atoms, the number of other types of atoms, the solvent accessible surface area, the average value of the charge, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the number of aromatic rings. These physical quantities cover multiple dimensions such as the atomic level, the intermolecular force level, and the chemical group level, and can more comprehensively reflect the structural characteristics of the protein than traditional methods that rely on only a single or a few characteristics, thereby providing a richer information basis for accurately predicting the thermal stability of the protein point mutation.

[0074] The present application sets the selection range of the physical quantity around the target residue in a spherical region of 5 to 15 angstroms, and introduces a distance weighting function. This distance range optimally balances information integrity and noise suppression, and can cover key interaction atoms and groups while avoiding the introduction of too much irrelevant information. The weighting function assigns weights according to the distance between the atoms and the target residue, highlighting the important influence of nearby atoms on thermal stability, so that the input data can more accurately reflect the local environmental characteristics, which helps the neural network to learn the influence of local changes on thermal stability and improves the generalization ability of the model.

[0075] (2) Convolutional neural network module

[0076] The multiple sets of three-dimensional convolutional layers-three-dimensional pooling layers are sequentially stacked, and then multiple fully connected layers are connected, and finally 20 values corresponding to 20 standard amino acids are output. The module is trained using the full set or subset of data from a public protein structure database (such as the RCSB protein structure database). The three-dimensional convolutional layer extracts features from the input four-dimensional matrix, the three-dimensional pooling layer performs dimensionality reduction, and the fully connected layer integrates feature information to realize the prediction output of the amino acid.

[0077] The convolutional neural network can extract effective feature representations by processing multi-dimensional data and use them to predict the influence of amino acid residue mutations on thermal stability. In the convolutional neural network module, the three-dimensional convolutional layer processes the input four-dimensional matrix data and can identify different structural patterns. To further improve the generalization ability of the network, normalization methods such as batch normalization and layer normalization are used during training to ensure that the input data remains balanced during training, thereby avoiding problems such as gradient disappearance or gradient explosion.

[0078] During the training of the convolutional neural network, the optimized loss function is as follows:

[0079]

[0080] where L(θ) represents the total loss function, represents the predicted mutation probability of the network, y i represents the actual mutation probability, θ jwhere W represents the weight parameters of the network, N represents the total number of training samples, λ represents the coefficient of the regularization term, and m represents the number of parameters in the network. This loss function not only calculates the difference between the predicted value and the actual value through the mean square error (MSE), but also controls the complexity of the model through the regularization term to prevent overfitting.

[0081] In this way, the convolutional neural network can learn a wide range of protein structure features from the RCSB database, thereby achieving higher accuracy in mutation prediction of specific proteins.

[0082] In the convolutional neural network module, the filter (convolution kernel) size of the three-dimensional convolutional layer is 3x3x3 or 2x2x2, and the step size is set to 1. As shown in Figure 4 The structure diagram of the convolutional neural network module is shown, which effectively controls the computational complexity while ensuring that local features in the protein three-dimensional structure can be effectively extracted, ensuring the training and prediction efficiency of the network. The filter size and step size of the three-dimensional convolutional layer are crucial for extracting protein structure features, especially when dealing with high-dimensional and complex structures. Proper filter design can help the network better learn local and global structure information.

[0083] Specifically, the 3x3x3 convolution kernel can capture the interaction information between atoms within the protein molecule in a larger spatial range, as it can consider the spatial relationship and structure pattern around the target residue in a larger range, which helps to reveal the complex interactions between different atoms and amino acids. For example, this convolution kernel helps to capture long-range interactions or large-scale environmental influences, which are crucial for understanding the thermal stability of proteins. In contrast, the 2x2x2 convolution kernel focuses on local detailed features, which captures the finer atomic arrangement and local chemical environment around the target residue by sliding in a smaller range, thereby providing the network with higher-resolution local information. The combination of these two convolution kernels can comprehensively describe the three-dimensional structure features of proteins from global and local perspectives, ensuring that the model can accurately learn the key structure features that affect thermal stability and function.

[0084] In the implementation process, the convolutional layer performs convolution operation on the input matrix through sliding window operation, and the convolution kernel performs element-level multiplication and summation operation with each local region of the input to obtain the feature value at the corresponding position. Assuming that the convolution kernel size is fxfxf, the step size is s, the size of the input matrix is I x ×I y ×I z , and the size of the convolution output matrix is O x ×O y ×O z It can be calculated by the following formula:

[0085]

[0086] where I x ,I y ,I z represent the dimensions of the input matrix in three dimensions, f is the size of the convolution kernel, s is the stride, and O x ,O y ,O z is the size of the convolution output matrix in the corresponding dimension. By adjusting the size f of the convolution kernel and the stride s, the receptive field of the convolution operation (i.e., the size of the region on which the convolution kernel slides over the input matrix) can be controlled, effectively capturing structural features at different scales.

[0087] This convolution kernel design not only captures the microscopic features of local atomic interactions of proteins, but also learns the global interatomic relationships at a larger scale. By introducing multiple convolution layers in the neural network and using convolution kernels of different sizes for hierarchical feature extraction, the network can gradually aggregate and refine structural information at different levels from local to global, thus better reflecting the thermal stability and function of proteins in the final prediction.

[0088] The three-dimensional convolution layer, with reasonable convolution kernel size and stride settings, not only effectively reduces the computational complexity, but also efficiently extracts multi-scale features of protein structures, providing strong data support for subsequent feature integration and thermal stability prediction.

[0089] The filter size of the three-dimensional convolution layer is 3x3x3 or 2x2x2, and the stride is 1. This setting effectively extracts local features of protein three-dimensional structures while controlling computational complexity. The 3x3x3 convolution kernel can capture large-scale atomic interactions and reveal complex relationships, while the 2x2x2 convolution kernel focuses on local details. The combination of the two fully describes structural features from different scales, ensuring that the model learns key information, and through reasonable convolution operation receptive field settings, it can effectively reduce computational complexity and improve training and prediction efficiency.

[0090] After a series of three-dimensional convolution layers-three-dimensional pooling layers are stacked and connected to multiple layers of fully connected layers, the three-dimensional pooling layer can reduce the data dimension, reduce the computational complexity and preserve important features, and the fully connected layer further integrates feature information to achieve accurate prediction output of amino acids. This structure design enables the network to gradually aggregate and refine information at different levels of structure, enhancing the learning ability of protein structure features.

[0091] (3) Transfer learning neural network module

[0092] The output of the previous layer of the convolutional neural network module is taken as the input of the module, which is composed of several layers of full connection, one-dimensional convolution-one-dimensional pooling, or normalization-full connection, and finally outputs 20 values corresponding to 20 standard amino acids. The four-dimensional matrix obtained by feature engineering on the unmutated structure of the experimental protein is taken as the input, and the normalized experimental observable physical quantities such as protein melting temperature, dynamic light scattering particle size, or Michaelis constant of enzymatic reaction, which characterize the thermal stability of the protein, are taken as the mutation probability output of the corresponding amino acid. This module uses the characteristics of transfer learning to further optimize the prediction of amino acid mutation probability combined with experimental data, effectively improving the prediction accuracy of the model for specific protein thermal stability related mutations.

[0093] The aforementioned convolutional neural network module can be trained to obtain an initial protein thermal stability point mutation prediction model. To further optimize the learning efficiency and accuracy of the initial protein thermal stability point mutation prediction model, in the transfer learning neural network module, the diversity of protein data from the RCSB database helps the network quickly adapt to the characteristics of different proteins. In the process of transfer learning, the pre-trained model learns a large amount of standard protein data, enabling the model to quickly capture potential structural patterns and further refine the prediction based on these patterns. In this way, the transfer learning neural network module can improve the prediction accuracy of specific mutations, especially for protein structures that have not been seen before, achieving good generalization.

[0094] In the transfer learning neural network module, the normalization method adjusts the amino acid mutation probability by combining enzyme activity experimental data to ensure that the model's prediction results better reflect the actual functional activity of the protein.

[0095] Specifically, the normalization process is carried out through the following steps:

[0096] First, without any external correction, the initial mutation probability of all amino acids is assumed to be 100%, i.e., the potential of each amino acid mutation is equal in the initial stage. This initial setting provides a unified starting point for subsequent correction. Next, the model adjusts the mutation probability of each amino acid based on experimental data, especially the relative activity percentage of enzyme activity experiments. For example, if experimental data shows that the mutation of a certain amino acid significantly improves the activity of the enzyme, the mutation probability of that amino acid will be increased accordingly. Conversely, if a mutation leads to a decrease in enzyme activity, the mutation probability of that amino acid will be reduced. Finally, to ensure that the sum of all probabilities is 100%, all adjusted mutation probabilities will be standardized, i.e., the final mutation probability of each amino acid is calculated by dividing its corrected probability value by the sum of all amino acid mutation probabilities.

[0097] This normalization method combines enzyme activity experimental data, making the model's predictions not only focus on changes in protein thermal stability, but also consider the impact of mutations on enzyme activity and its function. Specifically, let p i be the initial mutation probability of the i-th amino acid, a i is the relative activity percentage of the corresponding amino acid in the experimental data, and the normalized amino acid mutation probability The calculation formula is:

[0098]

[0099] where a i represents the relative activity percentage of the i-th amino acid associated with experimental data, N is the total number of all amino acids, is the sum of the relative activity percentages of all amino acids, denotes the normalized amino acid mutation probability under the scenario of enzyme activity experimental data. Through this calculation, the normalized ensures that the probability sum of all amino acids is 100%, while reasonably allocating the mutation probability of different amino acids according to enzyme activity data.

[0100] This normalization method enables the transfer learning neural network module to optimize the probability prediction of amino acid mutations based on enzyme activity experimental data, more accurately reflecting the comprehensive impact of different mutations on protein thermal stability and functional activity. In practical applications, using this normalization method, the recommendation system can not only optimize the thermal stability of proteins, but also maintain the enzyme activity and function after mutation, thereby providing more practical mutation suggestions in protein engineering.

[0101] In the normalization method in the transfer learning neural network module, in addition to adjusting the amino acid mutation probability based on enzyme activity experimental data, amino acids without experimental data are also considered. For those amino acids without experimental data, their mutation probability is set to 0% to avoid unnecessary interference with the final prediction results. At the same time, for unmutated amino acids, their mutation probability is set to 100% as a reference value, ensuring that in the absence of mutations, the model's prediction can be referenced to its initial state. This processing method highlights the impact of amino acids with experimental data while combining the reference probability of unmutated amino acids, ensuring that the normalized probability distribution can more reasonably reflect the mutation situation.

[0102] Specifically, assume that the relative activity percentage of amino acid i in the experimental data is a i , if there is no experimental value (i.e., a i = 0), the mutation probability p iis set to 0%. For non-mutated amino acids, their mutation probability is set to 100%. Then, the amino acid mutation probabilities of all experimental data will be normalized according to the following formula:

[0103]

[0104] wherein, represents the normalized amino acid mutation probability under the scenario of no enzymatic activity experimental data, p i is the mutation probability after experimental data correction (0% if there is no experimental data), is the sum of all amino acid mutation probabilities. The normalized probability value ensures that the probability of each amino acid is reasonable, and in the absence of experimental data, its mutation probability is limited to 0%, while the non-mutated amino acid remains at 100%.

[0105] The purpose of this normalization method is to emphasize the information provided by experimental data while maintaining the benchmark probability of non-mutated amino acids as a reference. Through this method, the transfer learning neural network module can more accurately predict the impact of amino acid mutations on protein thermal stability and functional activity, especially in the absence of comprehensive experimental data, it can still rely on known experimental information and benchmark values to make reasonable inferences. This not only enhances the practical application value of the model, but also makes the distribution of mutation probabilities more consistent with biological significance, especially in functional prediction and protein engineering, it can provide more targeted and practical results.

[0106] In the transfer learning neural network module, the ReLU (Rectified Linear Unit) is used as the activation function in the fully connected layer. This choice plays an important role in the training and performance optimization of the model. The mathematical expression of ReLU is:

[0107] ReLU(x) = max(0, x);

[0108] where x is the input signal. The main advantage of ReLU function is its simplicity in calculation and non-linear characteristics, which makes it effective in handling large-scale data while avoiding the problem of gradient vanishing. Gradient vanishing problem usually occurs when using traditional activation functions such as Sigmoid or Tanh, when the gradient of these functions is small, the training speed of the network will be significantly slowed down, or even unable to effectively update the parameters. In contrast, the gradient of ReLU function is constant when the input is positive, and does not depend on the size of the input, so it can speed up the convergence of the network.

[0109] Specifically, the output of the ReLU function is non-negative, which makes it suitable for probability-related tasks. In the transfer learning neural network module, each amino acid mutation probability should be a non-negative value, and the sum should be 1 (or close to 1) to meet the basic requirements of probability. The non-negativity of the ReLU function ensures that the output of each node meets this requirement, helping to ensure the reasonableness and effectiveness of the output of the transfer learning neural network.

[0110] By introducing the ReLU activation function in the fully connected layer, the network can quickly adjust the parameters during training, increasing its learning ability for specific protein thermal stability-related mutations. The non-linear nature of ReLU allows the network to capture more complex patterns in the input data, especially when dealing with high-dimensional data with multiple hierarchical structures. ReLU can effectively enhance the model's expressive power and complexity, thereby improving the final prediction accuracy.

[0111] To further optimize the performance of the model, the ReLU activation function is also combined with other techniques such as batch normalization and regularization methods to ensure that the network can train stably when faced with large amounts of data, avoiding overfitting while ensuring the model has good generalization ability. During training, the gradient of ReLU is effectively propagated to the previous layer, allowing the network to quickly and effectively learn the complex relationship between the three-dimensional structure of the protein and its thermal stability.

[0112] ReLU activation function not only has obvious advantages in reducing computational complexity and accelerating convergence speed, but also has non-negative output characteristics and good gradient propagation characteristics, making it play a key role in protein thermal stability prediction models, especially in transfer learning neural network modules. Through the ReLU activation function, the network can maintain high efficiency and accuracy in large-scale data processing and complex pattern learning, ensuring that the final output meets the biological and probabilistic rationality.

[0113] The transfer learning neural network module takes the output of the previous layer of the convolutional neural network module as input, and uses structures such as fully connected layer stacking, one-dimensional convolution layer-one-dimensional pooling layer stacking, or normalization layer-fully connected layer stacking to pre-train using public database data, and then fine-tune with experimental protein data (such as protein melting temperature, dynamic light scattering particle size, or enzyme reaction Michaelis constant, etc.), effectively improving the prediction accuracy of specific protein thermal stability-related mutations, especially when data is limited, the transfer learning feature can quickly adapt and optimize the prediction. Figure 5 The transfer learning neural network module operation flowchart is shown.

[0114] The full connection layer of the transfer learning neural network module adopts a ReLU activation function, which is simple to calculate, has strong non-linear characteristics, avoids the problem of gradient disappearance, and accelerates network convergence. At the same time, the ReLU function outputs non-negative, which meets the mutation probability requirement of amino acids, helps to ensure the rationality of the output results, and can stabilize the training, enhance the model expression ability and generalization ability, so that the network can better learn the relationship between protein structure and thermal stability.

[0115] In the transfer learning neural network module, the normalization method adjusts the mutation probability of amino acids by combining enzyme activity experimental data. Initially, it is assumed that the mutation probability of all amino acids is 100%, and then it is adjusted according to the change in enzyme activity in the experimental data. If the mutation increases enzyme activity, the probability is increased, and vice versa. Finally, normalization ensures that the sum of the probabilities is 100%. This way, the prediction takes into account the impact of mutations on enzyme activity and function, optimizes mutation probability prediction, and more accurately reflects the overall impact.

[0116] The mutation probability of amino acids without experimental data is set to 0% to avoid interference, and the mutation probability of amino acids without mutation is set to 100% as a reference value. By highlighting experimental data information and combining reference values, the normalized probability distribution is more reasonable, and in the absence of comprehensive experimental data, reasonable inferences can still be made based on known information, thereby enhancing the value of the model in practical applications and providing more targeted and practical mutation suggestions.

[0117] (4) Recommendation unit

[0118] According to the output results of the convolutional neural network module or the transfer learning neural network module, the mutation site and the target amino acid required by the mutation are recommended according to specific criteria. The criteria include sites where the standard deviation of the probability of several standard amino acids predicted by the model is higher than the preset value, i.e., sites with high uncertainty; and sites where the probability corresponding to the actual amino acid in the model predicted structure is not higher than the set value, i.e., sites that do not match the experimental results. In this way, sites with high mutation potential and corresponding target amino acids are selected, providing targeted guidance for the modification of protein thermal stability.

[0119] In the recommendation unit, in order to recommend mutation sites and target amino acids according to the criteria, the system first evaluates the standard deviation of each candidate site and the mutation probability of the actual amino acid, and then sorts the sites that meet the conditions. For example, Figure 6The recommended unit ranking logic diagram is shown. Specifically, if the standard deviation of multiple sites is higher than the preset value, the system will rank and recommend these sites in order of standard deviation from large to small; if the mutation probability of the actual amino acid of multiple sites is not higher than the set value, it will be ranked and recommended in order of mutation probability from small to large. The core goal of this ranking method is to prioritize sites with higher mutation potential and sites with experimental results inconsistent with predicted amino acid probabilities, in order to provide more reasonable and prioritized mutation selection order for experimental research, thereby improving experimental efficiency and success rate.

[0120] To further quantify this ranking method, the standard deviation and mutation probability can be ranked and prioritized by the following formula. Assuming that for the ith candidate site, the standard deviation is σ i , and the probability of actual amino acid mutation is p i , the system will process the ranking according to the following two criteria:

[0121] Standard deviation ranking: For sites with a larger standard deviation, it reflects that the site has a higher uncertainty and may have greater potential in structure prediction. The ranking criteria are set as follows:

[0122]

[0123] Where σ n represents the standard deviation of the candidate site n, is the standard deviation ranking value. Sites with a larger standard deviation will be given a higher priority.

[0124] Mutation probability ranking: For sites with a smaller mutation probability, it indicates that the actual amino acid of the site deviates greatly from the model predicted amino acid, and may have higher potential for optimization. The ranking criteria are set as follows:

[0125]

[0126] Where p i is the mutation probability of the ith site, is the mutation probability ranking value. Sites with a smaller mutation probability will be given a higher priority.

[0127] Through these two ranking methods, the final recommended order can be obtained by combining the standard deviation ranking value and the mutation probability ranking value. The comprehensive ranking value is set as follows:

[0128]

[0129] Where w σ and w p are the weights of the standard deviation ranking value and the mutation probability ranking value, respectively, w σ+w p = 1. This weighted comprehensive ranking method can adjust the weights according to the actual situation to balance the influence of standard deviation and mutation probability on the ranking results.

[0130] The ranked recommended sites will be arranged according to the comprehensive ranking value, and the sites with larger standard deviation and smaller mutation probability will be recommended first, so as to ensure that the recommended results give priority to those sites with higher mutation potential and inconsistent with experimental results. Through this ranking method, more prioritized and targeted mutation sites can be provided for experimental research, thereby improving the rationality of experimental design, optimizing the research process and improving the success rate.

[0131] The recommendation unit sets criteria for recommending mutation sites and target amino acids according to the model-predicted amino acid probability standard deviation and actual amino acid mutation probability. Sites with standard deviation higher than the preset value (high uncertainty) and actual amino acid probability not higher than the set value (inconsistent with experimental results) are ranked and recommended. Through the quantitative ranking formula, the sites are ranked according to the size of standard deviation and mutation probability, respectively, and the final recommendation order is obtained by setting weights on both, giving priority to sites with higher mutation potential and inconsistent with experimental results, thereby improving the accuracy and rationality of the recommendation, providing more prioritized and targeted mutation options for experimental research, and helping to improve experimental efficiency and success rate.

[0132] The above aspects improve the synergistic effect, from providing accurate input from feature engineering, optimizing neural network architecture and enhancing learning ability through training, optimizing probability prediction through normalization method, to accurate recommendation by the recommendation unit, which comprehensively improves the prediction accuracy of the system for protein thermostability point mutation, can more effectively guide the thermostability modification work in protein engineering, and reduce the blindness of experiments, saving time and resources.

[0133] The use of public databases (such as RCSB Protein Structure Database) and the flexibility of neural network architecture make this scheme not only suitable for various types of proteins, but also easily extended to other related protein property research or engineering applications, with strong universality and potential application expansion.

[0134] The above is a separate introduction to each module unit. The following introduces the connection of each module unit in combination with two more specific embodiments to reveal the situation of the model in specific application.

[0135] One more specific embodiment of the present application

[0136] 1. Preparation of data

[0137] Three-dimensional structure data of various proteins, including but not limited to enzymes, antibodies and structural proteins, etc. of different types of proteins, are obtained from the RCSB Protein Structure Database.

[0138] For each protein, determine the target residue, and in the spherical region with a radius of 5 to 15 angstroms centered on the target residue, select physical quantities such as the number of carbon atoms, the number of nitrogen atoms, the number of other types of atoms, the solvent accessible surface area, the average value of the charge, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the number of aromatic rings.

[0139] The selected physical quantities are rasterized to form a plurality of three-dimensional matrices, and are merged into a four-dimensional matrix as a channel of the physical quantity type, serving as input data of the convolutional neural network module.

[0140] 2. Training of the convolutional neural network module

[0141] The convolutional neural network module is constructed, the filter size of the three-dimensional convolutional layer is set to 3x3x3 and 2x2x2, the step size is 1, a plurality of groups of three-dimensional convolutional layers-three-dimensional pooling layers are sequentially stacked, and a plurality of fully connected layers are connected.

[0142] The full set or subset of the public protein structure database data is used for training, batch normalization and layer normalization methods are used in the training process, the loss function (including mean square error calculation and regularization term control model complexity) is optimized, the model is ensured to learn the protein structure characteristics, overfitting is avoided, and the generalization ability is improved.

[0143] 3. Training and optimization of the transfer learning neural network module

[0144] The output of the layer before the output layer of the convolutional neural network module is used as the input of the transfer learning neural network module, and a network structure composed of a plurality of layers of fully connected layers, one-dimensional convolutional layers-one-dimensional pooling layers, or normalization layers-fully connected layers is constructed.

[0145] The four-dimensional matrix obtained by feature engineering on the unmutated structure of the experimental protein is used as the input, and the experimental observable physical quantities representing the thermal stability of the protein, such as the protein melting temperature, dynamic light scattering particle size, or Michaelis constant of enzymatic reaction, are normalized and used as the mutation probability of the corresponding amino acid as the output.

[0146] The fully connected layer adopts the ReLU activation function, which is combined with batch normalization and regularization methods to accelerate network convergence, enhance the expression ability and generalization ability of the model, and enable the model to better learn the relationship between protein structure and thermal stability, thereby improving the prediction accuracy of the thermal stability-related mutations of specific proteins.

[0147] 4. Application of the recommendation unit

[0148] The structure data of the protein to be studied is input into the trained system, and the convolutional neural network module and the transfer learning neural network module are used for processing to obtain the amino acid mutation probability of each site.

[0149] The recommendation unit sets criteria according to the standard deviation of the amino acid probability predicted by the model and the actual amino acid mutation probability, and filters out sites with a standard deviation higher than a preset value (such as 0.5) and sites with an actual amino acid probability not higher than a set value (such as 0.2).

[0150] The filtered sites are sorted according to the quantitative sorting formula, sorted according to the size of the standard deviation and the size of the mutation probability respectively, and the weights of the two are set comprehensively (such as the standard deviation sorting weight is 0.6, and the mutation probability sorting weight is 0.4) to obtain the final recommended order.

[0151] According to the final recommended order, the mutation sites with high ranking are preferentially selected for experimental research, for example, the top 5 sites are subjected to site-directed mutagenesis experiment to measure the change of thermal stability of the mutant, and the accuracy and effectiveness of the recommendation system are verified.

[0152] Second, another more specific embodiment of the present application

[0153] 1. Research on a specific protein (taking a protein with a certain function as an example)

[0154] Select a protein with a specific function (such as a certain oxidoreductase), and obtain its three-dimensional structure data from the RCSB protein structure database.

[0155] The residues in the key functional region of the protein are selected as target residues, and the physical quantities in the spherical region around the target residues within 5 to 15 angstroms are selected and processed according to the feature engineering unit method to obtain four-dimensional matrix input data.

[0156] Using existing experimental data related to the protein (such as enzyme activity experimental data), adjust the amino acid mutation probability in the transfer learning neural network module by normalization method combined with experimental data, and process the mutation probability of amino acids without experimental data and unmutated amino acids according to the set rules, so that the prediction is more in line with the actual situation.

[0157] After the processing of the convolutional neural network module and the transfer learning neural network module, the recommendation unit recommends the mutation sites and target amino acids according to the multi-criteria sorting, and provides targeted mutation suggestions for the thermal stability modification of the specific protein, such as recommending specific site mutations to specific amino acids to improve its thermal stability while maintaining enzyme activity.

[0158] 2. Model verification and optimization

[0159] Prepare mutants by experiment according to the recommended sites, and measure the change of thermal stability of the mutants by differential scanning calorimetry, dynamic light scattering and other experimental methods, and measure the change of enzyme activity by enzyme activity detection experiment.

[0160] The experimental results are compared with the prediction results of the recommendation system to evaluate the prediction accuracy. For example, if it is predicted that the thermal stability of a site mutation is significantly improved, and the experimental verification result is consistent, it means that the model prediction at the site is effective; if there is a deviation, analyze the reasons, which may be that the selected physical quantities in the feature engineering are not comprehensive enough, the neural network model structure or parameters need to be adjusted, etc.

[0161] According to the verification result, the model is optimized, such as adjusting the selection weight of the physical quantity in the feature engineering unit, optimizing the convolution layer parameters of the convolutional neural network module, improving the combination structure of the full connection layer and other layers in the transfer learning neural network module, etc., and then training and verifying again, continuously improving the prediction performance of the model for protein thermal stability site mutation, so as to more accurately guide the experimental research and practical application in protein engineering.

[0162] Figure 7 A block diagram of an electronic device suitable for implementing the protein thermal stability site mutation prediction model construction method and / or prediction method described above according to embodiments of the present application is schematically shown. Figure 7 The electronic device shown is only an example and should not bring any limitation to the function and use range of embodiments of the present application.

[0163] As shown in Figure 7 The electronic device 1000 described in the present embodiment includes a processor 1001 which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 1002 or loaded from a storage portion 1008 into a random access memory (RAM) 1003. The processor 1001 can include, for example, a general-purpose microprocessor (such as a CPU), an instruction set processor, and / or a related chipset, and / or a special-purpose microprocessor (such as an application-specific integrated circuit (ASIC)), etc. The processor 1001 can also include an on-board memory for cache use. The processor 1001 can include a single processing unit or a plurality of processing units for performing different actions of the protein thermal stability site mutation prediction model construction method and / or prediction method according to embodiments of the present application.

[0164] In the RAM 1003, various programs and data required by the electronic device 1000 to operate are stored. The processor 1001, the ROM 1002, and the RAM 1003 are connected to each other via the bus 1004. The processor 1001 performs various operations according to the protein thermal stability point mutation prediction model construction method and / or prediction method procedure of the embodiments of the present application by executing the programs in the ROM 1002 and / or the RAM 1003. It is noted that the programs can also be stored in one or more memories other than the ROM 1002 and the RAM 1003. The processor 1001 can also perform various operations according to the protein thermal stability point mutation prediction model construction method and / or prediction method procedure of the embodiments of the present application by executing the programs stored in the one or more memories.

[0165] According to the embodiments of the present application, the electronic device 1000 can further include an input / output (I / O) interface 1005, which is also connected to the bus 1004. The electronic device 1000 can further include one or more of the following components connected to the I / O interface 1005: an input part 1006 including a keyboard, a mouse, and the like; an output part 1007 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker, and the like; a storage part 1008 including a hard disk, and the like; and a communication part 1009 including a network interface card such as a LAN card, a modem, and the like. The communication part 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as necessary. A removable recording medium 1011 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is attached to the drive 1010 as necessary, so that a computer program read therefrom is installed in the storage part 1008 as necessary.

[0166] The protein thermal stability point mutation prediction model construction method and / or prediction method procedure according to the embodiments of the present application can be implemented as a computer software program. For example, the embodiments of the present application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program codes for executing the protein thermal stability point mutation prediction model construction method and / or prediction method shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication part 1009, and / or installed from the removable recording medium 1011. When the computer program is executed by the processor 1001, the above-described functions defined in the system of the embodiments of the present application are performed. According to the embodiments of the present application, the systems, devices, apparatuses, modules, and / or units described above can be implemented by computer program modules.

[0167] The embodiments of the present application further provide a computer readable storage medium, which can be included in the device / apparatus / system described in the above embodiments, or can exist independently without being assembled into the device / apparatus / system. The computer readable storage medium carries one or more programs, when the one or more programs are executed, the steps of the protein thermal stability point mutation prediction model construction method and / or the prediction method according to the embodiments of the present application can be implemented.

[0168] According to the embodiments of the present application, the computer readable storage medium can be a non-volatile computer readable storage medium, for example, can include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the embodiments of the present application, the computer readable storage medium can be any tangible medium containing or storing a program, which can be used by or in conjunction with an instruction execution system, apparatus or device. For example, according to the embodiments of the present application, the computer readable storage medium can include one or more memories other than the ROM 1002 and / or the RAM 1003 described above.

[0169] It should be noted that each functional module in the various embodiments of the present application can be integrated in one processing module, or each module can exist physically independently, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. When the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, essentially or the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product.

[0170] The computer program product of the present application can be a computer program product comprising a computer readable storage medium and a computer program mechanism embedded in the computer readable storage medium. Such computer program product can further include a computer readable storage medium and program instructions. These program instructions commonly display the operations described herein as cognitive computing. According to the application, these program instructions might be transferred to a computer from computer readable storage medium rather than using a computer program mechanism.

[0171] Those skilled in the art can understand that the features described in various embodiments and / or claims of the present application can be combined and / or integrated in various combinations and / or integrations, even if such combinations or integrations are not explicitly described in the present application. In particular, the technical features described in various embodiments and / or claims of the present application can be combined and / or integrated in various combinations and / or integrations without departing from the spirit and teachings of the present application, and all such combinations and / or integrations fall within the scope of the present application.

[0172] Although the present application has been shown and described with respect to certain exemplary embodiments thereof, it should be understood by those skilled in the art that various alterations, modifications and / or additions can be made thereto, and it is intended to cover all such alterations, modifications and / or additions as fall within the scope of the present application. Therefore, the scope of the present application should not be limited to the above described embodiments, but should be defined only by the appended claims and equivalents thereof.

Claims

1. A method for constructing a protein thermostability point mutation prediction model, characterized by, The method comprises the following steps: obtaining a three-dimensional structure of a preset type of protein, selecting a preset type of physical quantity within a preset distance range around a target residue for raster processing to form a plurality of three-dimensional matrices, and merging the three-dimensional matrices into a four-dimensional matrix with the type of physical quantity as a channel; training a three-dimensional convolutional neural network using the four-dimensional matrix as training data to construct an initial protein thermal stability point mutation prediction model; the three-dimensional convolutional neural network comprises a preset large-size convolutional kernel for extracting global features and a preset small-size convolutional kernel for extracting local features; using the output of a previous layer of the output layer of the initial protein thermal stability point mutation prediction model as input, adjusting the amino acid mutation probability based on enzyme activity experimental data and / or using a normalization method taking the baseline probability of an unmutated amino acid as a reference when there is no enzyme activity experimental data to train a transfer learning neural network, and obtaining an optimized final protein thermal stability point mutation prediction model.

2. The prediction model construction method of claim 1, wherein, The preset distance range is specifically: a region between spherical regions with a radius of 5 to 15 angstroms centered on the target residue.

3. The prediction model building method of claim 2, wherein, By a distance weighting function The distance from the target residue is weighted, and the specific formula includes: ； where r represents the distance between the target residue and the surrounding atoms; a represents the adjustment parameter; represents the weighted distance from the target residue.

4. The prediction model building method of claim 1, wherein, The preset type of physical quantity includes physical quantities reflecting the structure characteristics of the protein from the atomic level, the intermolecular force level, and the chemical group level; The preset type of physical quantity specifically includes one or more of the number of carbon atoms, the number of nitrogen atoms, the solvent accessible surface area, the average charge, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the number of aromatic rings.

5. The prediction model building method of claim 1, wherein, The sizes of the preset large-size convolutional kernel and the preset small-size convolutional kernel of the three-dimensional convolutional layer of the three-dimensional convolutional neural network are 3×3×3 and 2×2×2, respectively, and the step size is set to 1.

6. The prediction model building method of claim 5, wherein, The three-dimensional convolutional layer performs convolution operation on the input matrix through sliding window operation, and the convolution kernel performs element-level multiplication and summation operation on each local region of the input to obtain the feature value at the corresponding position. The specific formula includes: ； wherein, denote the dimensions of the input matrix in three dimensions, respectively; denotes the size of the convolution kernel; s denotes the stride; denote the dimensions of the convolution output matrix in the corresponding dimensions, respectively.

7. The prediction model building method of claim 1, wherein, The normalization method for adjusting the amino acid mutation probability based on enzyme activity experimental data includes: assuming that the initial mutation probability of all amino acids is 100%; if the experimental data shows that the mutation of a certain amino acid improves the activity of the enzyme, the mutation probability corresponding to the amino acid will be increased accordingly; otherwise, it will be decreased accordingly; standardizing the adjusted mutation probability so that the sum of all mutation probabilities is 100%; the specific formula includes: ； wherein, represents the relative percentage of the i-th amino acid associated with the experimental data; N represents the total number of all amino acids; represents the sum of the relative percentages of the relative activity of all amino acids; represents the normalized amino acid mutation probability under the scenario of experimental data with enzyme activity. The normalization method for taking the baseline probability of an unmutated amino acid as a reference when there is no enzyme activity experimental data includes: setting the mutation probability of the amino acid in the no-enzyme activity experimental data to 0%; setting the mutation probability of the unmutated amino acid to 100% as the baseline value, so as to ensure that the prediction of the model can take its initial state as a reference in the case of no mutation; The specific formula includes: ； wherein, represents the mutation probability after experimental data correction; represents the sum of all amino acid mutation probabilities; represents the normalized amino acid mutation probability under the scenario of no enzymatic activity experimental data.

8. The predictive model building method of claim 1, wherein, The activation function of the fully connected layer of the transfer learning neural network is ReLU; The loss function used when training the three-dimensional convolutional neural network includes: ； wherein, represents the total loss function; represents the network predicted probability of a mutation; represents the actual probability of a mutation; represents the weight parameters of the network; N represents the total number of training samples; λ represents the regularization term coefficient; and m represents the number of parameters in the network.

9. A method for predicting point mutations in protein thermal stability, characterized in that, Inputting the structure data of the protein to be studied into the initial protein thermostability point mutation prediction model or the final protein thermostability point mutation prediction model in the prediction model construction method of any one of claims 1-8 for processing to obtain an output result, and then predicting the mutation site and the target amino acid to be mutated according to a specific criterion; The specific criterion includes sites with a standard deviation of the probability of several standard amino acids predicted by the model higher than a preset value, and sites with a probability corresponding to the actual amino acid in the model-predicted structure not higher than a set value.

10. The prediction method of claim 9, wherein, The standard deviation and the mutation probability of the actual amino acid of each candidate site are evaluated, and the sites meeting the conditions are sorted; If there are multiple sites with a standard deviation higher than the first preset value, these sites are sorted and recommended in descending order of the standard deviation; The sorting criteria related to the standard deviation include: ； wherein, represents the standard deviation of the candidate site n; represents the standard deviation ranking value; If there are multiple sites with a mutation probability of the actual amino acid not higher than the second preset value, they are sorted and recommended in ascending order of the mutation probability; the sorting criteria related to the mutation probability include: ； wherein, represents the probability of mutation for candidate site i; represents the probability of mutation ranking value; Based on the weight of the standard deviation sorting value and the mutation probability sorting value, the final recommended order is determined, and the specific formula includes: ； wherein, and respectively represent the weight of the standard deviation ranking value and the weight of the mutation probability ranking value, ; represents the comprehensive ranking value; According to the comprehensive sorting value, the mutation site and the target amino acid to be mutated are recommended, and the mutation site ranked higher based on the comprehensive sorting value is selected for experimental study.

Images