Method, device and equipment for predicting properties of mutant protein

Abstract

The invention discloses a method, device and equipment for predicting the property of a mutant protein, and the method comprises the steps: carrying out the fusion processing of the embedding characteristics and physicochemical characteristics of an original protein, the embedding characteristics and physicochemical characteristics of the mutant protein generated by the mutation of the original protein, and the interaction characteristics of residues of the mutant protein, obtaining initial node characteristics of the mutant protein; the interaction characteristic represents the interaction between every two residues in the mutant protein; the embedded feature represents an amino acid sequence of the protein, the physicochemical feature represents physicochemical properties of residues forming the protein, and a prediction model is used for processing initial node features and structural features of the original protein so as to predict multiple properties of the mutant protein; and predicting model loss of the model, wherein the model loss is determined according to a sample property predicted value representing multiple properties and a plurality of sample property measured values of the sample, wherein the sample property predicted value and the sample property measured values are obtained by the prediction model.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method, apparatus, and device for predicting the properties of mutant proteins. Background Technology

[0002] Proteins are the main executors of life activities, and their function is closely related to their structure. Mutations in proteins can lead to changes in their structure and function, thereby affecting various physiological processes in organisms. Therefore, accurately predicting the property changes of mutated proteins (i.e., proteins that have undergone mutation) is of great significance for scientific research in biomedical fields such as understanding disease mechanisms, drug design, and protein engineering.

[0003] One current method for predicting mutant proteins involves constructing protein feature data that characterizes the structure and mutation location of the mutant protein, and then using a trained model to process this protein feature data to predict a certain property of the mutant protein (such as activity or stability). However, this method suffers from low accuracy. Summary of the Invention

[0004] Therefore, this application discloses the following technical solution:

[0005] The first aspect of this application provides a method for predicting the properties of mutant proteins, comprising:

[0006] Obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein.

[0007] The first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction feature of the mutant protein residues are fused to obtain the initial node feature of the mutant protein; wherein, the interaction feature characterizes the interaction between every two residues in the mutant protein;

[0008] The initial node features and the structural features of the original protein are processed using a prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein; wherein, the model loss of the prediction model is determined based on the sample property prediction values ​​characterizing the various properties obtained by processing the samples by the prediction model and the measured values ​​of the multiple sample properties of the samples.

[0009] Optionally, the fusion processing of the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction features of the mutant protein residues to obtain the initial node features of the mutant protein includes:

[0010] The first embedding feature and the first physicochemical feature are subjected to dimensionality reduction processing to obtain the first protein sequence feature;

[0011] The second embedding feature and the second physicochemical feature are subjected to dimensionality reduction processing to obtain the second protein sequence feature;

[0012] The initial node features of the mutant protein are obtained by splicing together the first protein sequence features, the second protein sequence features, and the interaction features of the mutant protein residues.

[0013] Optionally, the process of obtaining the structural features of the original protein includes:

[0014] The β-carbon coordinates of each residue in the original protein are determined based on the protein structure data (PDB) of the original protein.

[0015] The adjacency matrix of the original protein is determined based on the β-carbon atom coordinates of every two residues in the original protein, and the adjacency matrix includes the distance between the β-carbon atoms of every two residues in the original protein.

[0016] The structural features of the original protein are determined based on the adjacency matrix.

[0017] Optionally, the step of processing the initial node features and the structural features of the original protein using a prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein includes:

[0018] The initial node features and the structural features of the original protein are processed sequentially using a series of multi-layer neural networks in the prediction model to obtain the network output features of the mutant protein; wherein, the input of the first layer of the neural network includes the initial node features and the structural features, and the input of each layer of the neural network except the first layer includes the output of the previous layer of the neural network and the structural features;

[0019] For each prediction module of the prediction model, the network output features are processed by the prediction module to obtain a property prediction value that characterizes a property of the mutant protein; wherein, the prediction model includes multiple prediction modules, and the property prediction values ​​output by different prediction modules characterize different properties of the mutant protein.

[0020] Optional, also includes:

[0021] Based on the network output features and the mutation site information of the original protein, the mutation local features of the mutated protein are obtained;

[0022] The process of using the prediction module to process the network output features to obtain a property prediction value characterizing a property of the mutant protein includes:

[0023] The prediction module processes the network output features and the mutation local features to obtain a property prediction value that characterizes a property of the mutant protein.

[0024] Optionally, the process of obtaining the interaction characteristics of the mutant protein residues includes:

[0025] Extract the multiple sequence alignment profile of the mutant protein;

[0026] The interaction characteristics of the mutant protein residues were calculated based on the multiple sequence alignment spectrum of the mutant protein.

[0027] Optionally, the process of determining the model loss of the prediction model includes:

[0028] For each of the predicted values ​​of the sample property, the loss value corresponding to the predicted value of the sample property is determined based on the predicted value of the sample property and the measured value of the sample property corresponding to the same property.

[0029] Based on preset loss weights, the loss values ​​corresponding to the predicted values ​​of various sample properties are fused to obtain the model loss of the prediction model.

[0030] Optionally, the prediction model includes a series of multi-layer neural networks and multiple prediction modules;

[0031] The process of constructing the prediction model includes:

[0032] Construct a multi-layer neural network in series in the prediction model based on the initial dataset to obtain a pre-trained multi-layer neural network;

[0033] Determine the various properties that need to be predicted;

[0034] The initial prediction model is obtained by combining the multiple prediction modules corresponding to the various properties and the pre-trained multilayer neural network.

[0035] The initial prediction model is fine-tuned based on samples corresponding to the various properties to obtain the prediction model.

[0036] A second aspect of this application provides an apparatus for predicting the properties of mutant proteins, comprising:

[0037] The obtaining unit is used to obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein.

[0038] A fusion unit is used to fuse the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction features of the mutant protein residues to obtain the initial node features of the mutant protein; wherein, the interaction features characterize the interaction between every two residues in the mutant protein;

[0039] The processing unit is used to process the initial node features and the structural features of the original protein using a prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein; wherein, the model loss of the prediction model is determined based on the sample property prediction values ​​characterizing the various properties obtained by processing the sample with the prediction model and the measured values ​​of the multiple sample properties of the sample.

[0040] A third aspect of this application provides an electronic device, including a memory and a processor;

[0041] The memory is used to store computer programs;

[0042] The processor is used to execute the computer program, specifically to implement the method for predicting the properties of mutant proteins provided in any of the first aspects of this application.

[0043] The beneficial effects of this scheme are as follows: Firstly, this scheme constructs a prediction model based on the difference between the predicted values ​​of multiple properties simultaneously predicted by the model and the corresponding measured values. Therefore, when this model is used to predict protein properties, it can combine the interrelationships between multiple properties for prediction. Secondly, the prediction model integrates multiple features such as the embedding characteristics, physicochemical characteristics, interaction characteristics, and structural characteristics of mutant proteins for prediction, which is more accurate than models that predict a single property of a protein based on a single feature. In summary, the prediction model of this scheme can combine the interrelationships between multiple properties of a protein and the multi-dimensional features of the protein to predict the protein properties, which has higher accuracy than models that predict a single property of a protein based on a single-dimensional feature. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 This is a flowchart of a method for predicting the properties of mutant proteins provided in an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of protein feature extraction provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram illustrating the working principle of a prediction model provided in an embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the structure of a device for predicting the properties of mutant proteins provided in an embodiment of this application;

[0049] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] This application provides a method for predicting the properties of mutant proteins. Please refer to [link to relevant documentation]. Figure 1 The method may include the following steps.

[0052] S101, obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein.

[0053] In step S101, the original protein can be any wild-type protein, or a protein obtained by mutation of a wild-type protein; the mutant protein can be a protein obtained by single-point or multi-point mutation of the original protein, and its properties need to be predicted.

[0054] As an example, the original protein can be an enzyme protein extracted from biological cells without artificial intervention, while the mutant protein can be a protein obtained by performing single-point or multi-point mutations on the enzyme protein.

[0055] The embedding and physicochemical characteristics of proteins can be obtained in a variety of ways, and this embodiment does not limit the method of obtaining them.

[0056] As some examples, the first embedding features of the original protein can be obtained as follows:

[0057] First, the protein sequence of the original protein is determined. The information contained in the protein sequence can be which amino acids are linked together in sequence to form each peptide chain that makes up the protein.

[0058] Then, the protein sequence of the original protein is encoded according to the protein language model to obtain the first embedding feature of the original protein.

[0059] This embodiment does not limit the type or number of protein language models used; any protein language model disclosed in the relevant technical field can be selected. For example, please refer to [link to example]. Figure 2 The evolutionary scale modeling (ESM) model (specifically, version 2, ESM-2) and the protein transformation (ProtTrans, ProtT) model (specifically, version 5, ProtT5) can be used to process the protein sequence of the original protein, respectively, to obtain the first embedding feature X of the original protein output by the ESM model. ESM1 And the first embedding feature X of the original protein output by the ProtT model. ProtT1 The first embedding feature of the original protein includes X. ESM1 and X ProtT1 Two characteristics.

[0060] Optional, X ESM1 and X ProtT1 All are feature matrices with N rows and 1280 columns.

[0061] The primary physicochemical characteristics of the original protein can be obtained as follows:

[0062] The physicochemical properties of each amino acid that makes up the original protein are analyzed, including but not limited to the molecular weight of each amino acid, the mass of the corresponding residue in the original protein, the acidity coefficient (e.g., pKa), and the number of carbon atoms in the amino acid. All these physicochemical properties are combined into a matrix, which represents the first physicochemical characteristic of the original protein, denoted as X. prop1 ;

[0063] The first physicochemical feature can be an N-row, M-column matrix, where N is the total number of amino acids that make up the original protein, and M is the number of physicochemical property parameters detected for each amino acid. For example, if 14 physicochemical property parameters are detected for each amino acid, then the first physicochemical feature can be an N-row, 14-column matrix, where the value in the i-th row and k-th column is the value of the k-th physicochemical property parameter of the i-th amino acid of the original protein.

[0064] The second embedding feature and second physicochemical feature corresponding to the mutant protein can be obtained in the same manner as described above, and will not be repeated here. The second embedding feature of the mutant protein value obtained in the above manner may include X. ESM2 and X ProtT2 The second physicochemical characteristic can be denoted as X. prop2 .

[0065] S102, the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature and the interaction feature of the mutant protein residues are fused to obtain the initial node features of the mutant protein; wherein, the interaction feature characterizes the interaction between every two residues in the mutant protein.

[0066] The interaction characteristics of mutant protein residues can characterize the interaction between every two residues in the mutant protein. This interaction characteristic can be an N-row, N+1-column matrix, where N is the number of amino acids that make up the mutant protein, and the values ​​in the matrix represent the strength of the interaction between the amino acid residues corresponding to the row and the amino acid residues corresponding to the column.

[0067] One way to obtain the interaction signatures of mutant protein residues is as follows:

[0068] Multiple sequence alignment (MSA) spectra of the mutant protein are extracted, and then the MSA spectra are calculated using relevant prediction models. Based on these MSA spectra, the interaction characteristics of the mutant protein residues are calculated. These interaction characteristics can be denoted as X. rri .

[0069] The model used to calculate the interaction characteristics can be a protein prediction model based on the association of correlated mutations (CCMpred), or other related models; this embodiment does not limit the specific model.

[0070] In step S102, fusion processing can be performed in various ways, and this embodiment does not limit the method of fusion processing. As an example, fusion processing can be to directly splice all the above features together, and the feature formed after splicing is used as the initial node feature of the mutant protein, denoted as X. node(0) .

[0071] S103, the prediction model is used to process the initial node features and the structural features of the original protein to obtain multiple property prediction values ​​that characterize the mutant protein; wherein, the model loss of the prediction model is determined based on the sample property prediction values ​​that characterize multiple properties obtained by processing the sample with the prediction model and the measured values ​​of multiple sample properties of the sample.

[0072] In step S103, the initial node features of the mutant protein and the structural features of the original protein can be input into the prediction model. After processing by the prediction model, multiple property prediction values ​​corresponding to the mutant protein are output, where each property prediction value represents a property of the mutant protein.

[0073] For example, the prediction model can be input with four property prediction values ​​for the mutant protein, one property prediction value characterizing the stability of the mutant protein, one property prediction value characterizing the activity of the mutant protein, one property prediction value characterizing the selectivity of the mutant protein, and one property prediction value characterizing the solubility of the mutant protein.

[0074] The structural features of the original protein can be any data that can characterize the structure of the original protein. In this embodiment, the structural features of the original protein can be denoted as G.

[0075] As examples, the three-dimensional structural data of the original protein can be detected and used as the structural features of the original protein.

[0076] As another example, a matrix characterizing the contact relationship between every two residues in the original protein can be determined based on the distance between them, and this matrix can be used as a structural feature of the original protein.

[0077] For the latter example, the corresponding structural features can be obtained as follows:

[0078] The coordinates of the β-carbon atom of each residue in the original protein are determined based on the protein structure data (PDB) of the original protein.

[0079] The adjacency matrix of the original protein is determined based on the β-carbon coordinates of every two residues in the original protein. The adjacency matrix includes the distance between the β-carbon atoms of every two residues in the original protein.

[0080] The structural features of the original protein are determined based on the adjacency matrix.

[0081] Of the methods described above, one can first obtain the protein structure database (PDB) file of the original protein, and then extract the β carbon atom (i.e., C) of each residue of the original protein from this file. β The coordinates of ).

[0082] Then, for each pair of residues in the original protein, use the C of those two residues. β The coordinates of the two residues were used to calculate the C coordinates. β The distance between the two residues is determined, thus obtaining an adjacency matrix; this adjacency matrix can be an N x N matrix, where each element is the distance between the residues in the corresponding row and the residues in the corresponding column;

[0083] Finally, for each element in the adjacency matrix, the element is compared with a preset distance threshold. If the element is greater than the distance threshold, the element can be replaced with 0, indicating that there is no contact relationship between the two residues corresponding to the row and column of the element. If the element is less than or equal to the distance threshold, the element can be replaced with 1, indicating that there is a contact relationship between the two residues corresponding to the row and column of the element. This results in an N-row N-column matrix, where each element is either 1 or 0. This matrix can be regarded as the structural features of the original protein.

[0084] The beneficial effects of this scheme are as follows: Firstly, this scheme constructs a prediction model based on the difference between the predicted values ​​of multiple properties simultaneously predicted by the model and the corresponding measured values. Therefore, when this model is used to predict protein properties, it can combine the interrelationships between multiple properties for prediction. Secondly, the prediction model integrates multiple features such as the embedding characteristics, physicochemical characteristics, interaction characteristics, and structural characteristics of mutant proteins for prediction, which is more accurate than models that predict a single property of a protein based on a single feature. In summary, the prediction model of this scheme can combine the interrelationships between multiple properties of a protein and the multi-dimensional features of the protein to predict the protein properties, which has higher accuracy than models that predict a single property of a protein based on a single-dimensional feature.

[0085] In some optional embodiments, for step S102, the multiple features involved in S102 can be fused in the following manner to obtain initial node features.

[0086] A1, dimensionality reduction is performed on the first embedding feature and the first physicochemical feature to obtain the first protein sequence feature;

[0087] A2, dimensionality reduction processing is performed on the second embedding feature and the second physicochemical feature to obtain the second protein sequence feature;

[0088] A3, by splicing together the first protein sequence features, the second protein sequence features, and the interaction features of the mutant protein residues, the initial node features of the mutant protein are obtained.

[0089] In step A1, the first embedding feature and the first physicochemical feature can be concatenated to obtain the initial sequence feature corresponding to the original protein, denoted as X. amino-1 This process can be represented by the following formula (1).

[0090] X amino-1 =concatenate(X) ESM1 X ProtT1 X prop1 ), (1).

[0091] Concatenate refers to the process of combining multiple features into a single feature. For details on how to combine multiple features into a single feature, please refer to relevant technical documentation.

[0092] After obtaining the initial sequence features of the original protein, a fully connected network can be used to reduce the dimensionality of these initial sequence features to obtain the first protein sequence feature corresponding to the original protein. This first protein sequence feature is denoted as X. amino-1 '.

[0093] The fully connected network used for dimensionality reduction (denoted as the dimensionality reduction module) can be a multilayer perceptron (MLP) network or other neural network structures with dimensionality reduction capabilities, without limitation.

[0094] Taking MLP networks as an example, the above dimensionality reduction process can be represented by the following formula (2).

[0095] X amino-1 =MLP(X) amino-1 ), (2).

[0096] The fully connected network used for dimensionality reduction can be part of the prediction model constructed in this embodiment, or it can be an external, independent module.

[0097] In step A2, the second embedding feature and the second physicochemical feature can be processed in the same way as in step A1 to obtain the second protein sequence feature corresponding to the mutant protein, denoted as X. amino-2 '.

[0098] In step A3, the first protein sequence features, the second protein sequence features, and the interaction features of the mutant protein residues can be spliced ​​together, and the result can be used as the initial node features of the mutant protein.

[0099] This process can be represented by the following formula (3).

[0100] X node(0) =concatenate(X) amino-1 ' , X amino-2 ' , X rri ), (3).

[0101] The advantage of obtaining the initial node features of mutant proteins in the above manner is that by first performing dimensionality reduction on the large amount of embedded features and physicochemical features, and then fusing the dimensionality-reduced features to form the initial node features, the amount of data required to obtain the initial node features can be reduced, thereby improving the processing efficiency of subsequent processes.

[0102] The prediction model in this embodiment can be a prediction model of any structure. For an example, please refer to [link to example]. Figure 3 The prediction model may include a series of multi-layer neural networks, with multiple prediction modules (e.g., connected at the end of the multi-layer neural network) linked together. Figure 3 (Two prediction modules in the text).

[0103] Each prediction module can include a global pooling layer and a prediction head. The input of the multi-layer neural network is pooled by the global pooling layer and then enters the prediction head. The prediction head processes the input data to obtain a property prediction value of the mutant protein. The property prediction values ​​output by different prediction modules characterize different properties of the mutant protein.

[0104] by Figure 3 For example, a cascaded multilayer neural network can be a multilayer graph isomorphism network (GIN). Each layer of the graph isomorphism network can include a GIN network body, a batch normalization layer (denoted as BatchNorm) connected after the GIN module, and an activation function (denoted as LeakyReLU) connected after the batch normalization layer.

[0105] Other structures can also be used in cascaded multilayer neural networks. For example, when the structural features are three-dimensional structural data of the original protein values, each layer in the multilayer neural network can be a three-dimensional graph neural network such as an equivariant graph neural network (EGNN) for processing three-dimensional data. Alternatively, each layer of the cascaded neural network can also be a graph convolutional neural network (GCN), a graph attention network (GAT), or a graph transformer network.

[0106] Based on the structure of multi-layer graph isomorphic networks, the aforementioned step S103 can be implemented as follows:

[0107] The initial node features and the structural features of the original protein are processed sequentially by a series of multi-layer neural networks in the prediction model to obtain the network output features of the mutant protein. The input of the first layer of the neural network includes the initial node features and the structural features, and the input of each layer of the neural network except the first layer includes the output and structural features of the previous layer.

[0108] For each prediction module of the prediction model, the network output features are processed by the prediction module to obtain a property prediction value that characterizes a property of the mutant protein; the prediction model includes multiple prediction modules, and the property prediction values ​​output by different prediction modules characterize different properties of the mutant protein.

[0109] Specifically, assuming the prediction model includes a series of K-layer graph isomorphic neural networks, the processing of the k-th layer (k is greater than or equal to 1 and less than or equal to K) can be represented by the following formulas (4) to (6).

[0110] X node(k)-1 =GIN k (G, X) node(k-1) ), (4).

[0111] X node(k)-2 =BatchNorm k (X node(k)-1 ), (5).

[0112] X node(k) =LeakyReLU k (X node(k)-2 ), (6).

[0113] Among them, X node(k-1) This represents the output of the (k-1)th layer of the graph isomorphic neural network. When k equals 1, X node(0) This refers to the initial node characteristics of the aforementioned mutant protein. GIN k X represents the main body of the GIN network in the k-th layer of the graph isomorphic neural network. node(k)-1 This represents the output of the main body of the GIN network in the k-th layer of the graph isomorphic neural network. BatchNorm k X represents the batch normalization layer of the k-th layer of a graph isomorphic neural network. node(k)-2 LeakyReLU represents the output of the batch normalized layer of the k-th layer of a graph isomorphic neural network. k Let X represent the activation function of the k-th layer graph isomorphic neural network. node(k) This represents the output of the k-th layer graph isomorphic neural network.

[0114] When k equals K, the output X of the Kth layer graph isomorphic neural network node(K) The aforementioned network output feature, denoted as X. L .

[0115] After obtaining the network output features, the network output features can be input into the global pooling layer of each prediction module for pooling processing to obtain the global pooling features of each prediction module. Then, the output head of the prediction module is used to process the global pooling features of that prediction module, and the output of the prediction head is the property prediction value of the prediction module.

[0116] For example, suppose the prediction model includes two prediction modules, denoted as task1 and task2, where task1 obtains the globally pooled feature X. task1 The global pooling feature X obtained by task2 can be expressed by the following formula (7). task2 It can be represented by the following formula (8).

[0117] X task1 =GlobalAttentionPooling task1 (X L ), (7).

[0118] X task2 =GlobalAttentionPooling task2 (X L ), (8).

[0119] Each prediction head can include a fully connected neural network structure. The output of the fully connected neural network structure can be a real number.

[0120] Multiple prediction heads of a prediction model can be used to output predicted values ​​for different types of properties, and can have different structures.

[0121] For example, the property prediction value output by a prediction head can be any real number within a certain range. In this case, the prediction head can be considered to be used to perform a regression task. Such a prediction head may only include a fully connected neural network structure, and the value output by this fully connected neural network structure is directly used as the property prediction value obtained by the prediction head.

[0122] A prediction head outputs a property prediction value that can be a binary classification result, such as 0 or 1. This property prediction value can characterize whether a mutant protein has a certain property. For example, 1 indicates that the protein has a certain property, and 0 indicates that the protein does not have that property. In this case, the prediction head can be considered to be used to perform a binary classification task. Such a prediction head can include a fully connected neural network structure and an activation function (such as the sigmoid function). The activation function is used to convert the real number output by the fully connected neural network structure into 0 or 1, which represent the classification result.

[0123] A prediction head outputs a property prediction value that can be a multi-class classification result, such as any one of class 1, class 2, class 3, or class 4. This property prediction value can characterize the category of a certain property of a mutant protein. For example, the prediction value corresponding to the functional property of a mutant protein can be any of the multi-class classification results. For instance, class 1 indicates that the mutant protein has function A, class 2 indicates that the mutant protein has function B, and class 3 indicates that the mutant protein has function C. In this case, the prediction head can be considered to be used to perform a multi-class classification task. Such a prediction head can include a fully connected neural network structure and a normalization function (such as the Softmax function). The normalization function is used to convert the real number output by the fully connected neural network structure into the corresponding category in the multi-class classification result.

[0124] The beneficial effect of this embodiment is that by sequentially processing the initial node features of the mutant protein through a series of multi-layer neural networks, multi-level features of the mutant protein can be extracted, thereby helping to obtain more accurate prediction results.

[0125] In some alternative embodiments, such as Figure 3 As shown, the prediction model in this embodiment may further include a mutation extraction module. Using this module, the prediction model in this embodiment may further include the following steps:

[0126] Based on the network output features and the mutation site information of the original protein, the mutation local features of the mutant protein are obtained.

[0127] Correspondingly, after obtaining the local features of the mutation, the network output features can be processed as follows to obtain the property prediction values:

[0128] By using the prediction module to process network output features and mutation local features, a property prediction value characterizing a property of the mutant protein is obtained.

[0129] The mutation extraction module can be an average pooling layer. This mutation extraction module can obtain the local mutation feature X according to the following formula (9). local .

[0130] X local =1 / m*(X L T ·I), (9).

[0131] In formula (9), I represents mutation site information. L T This indicates the output feature X of the network. L The transposition process is performed, where m is the maximum number of mutation sites in the original protein, indicating that there may be a maximum of m mutation sites in the original protein. The sites in the protein can be matched one-to-one with the protein residues.

[0132] Mutation site information can be any information that indicates the location of a mutation site in the original protein, and its specific form is not limited. As an example, a mutation site can be a one-dimensional vector containing N elements, where each element corresponds to a site in the original protein. The value of each element can be 1 or 0. If a site belongs to a mutation site, the value of its corresponding element is 1; if a site does not belong to a mutation site, the value of its corresponding element is 0. That is, I can be defined as I = {0, 1}. N .

[0133] For any prediction module, the way it processes network output features and abrupt local features can be:

[0134] The global pooling layer of the prediction module processes the network output features to obtain the global pooling features of the prediction module. Then, the mutation local features and the global pooling features of the prediction module are input into the prediction head of the prediction module. After processing by the prediction head, the property prediction value of the prediction module is obtained.

[0135] Based on the foregoing example, for instance, suppose the prediction model includes two prediction modules, denoted as task1 and task2 respectively. After the two prediction modules obtain the corresponding global pooling features, their respective prediction heads can process the mutation local features and their respective global pooling features in the manner shown in formulas (10) and (11) to obtain the property prediction value Y corresponding to the task1 prediction module. task1 The property prediction value Y corresponding to the task2 prediction module task2 .

[0136] Y task1 =MLP task1 (X task1 ||X local ), (10).

[0137] Y task2 =ML task2 (X task2 ||X local ), (11).

[0138] Among them, MLP task1 This represents the prediction header of the task1 prediction module, MLP. task2 This represents the prediction header of the task2 prediction module, and || indicates that the mutation local features and the corresponding global pooling features are concatenated to form a single feature.

[0139] The beneficial effect of this embodiment is that by introducing mutation site information that characterizes the location of the mutation site, the prediction head can focus more on the features related to the mutation site when outputting the property prediction value, thereby improving the prediction efficiency and the accuracy of the prediction results.

[0140] Optionally, the process of determining the model loss of the prediction model includes:

[0141] For each sample property prediction value, the loss value corresponding to the sample property prediction value is determined based on the sample property prediction value and the measured value of the sample property corresponding to the same property.

[0142] Based on the preset loss weights, the loss values ​​corresponding to the predicted values ​​of various sample properties are fused to obtain the model loss of the prediction model.

[0143] In this embodiment, the prediction model can be pre-constructed based on multiple samples. When constructing the model, initial values ​​can be assigned to each model parameter in the model through random initialization to obtain an initial prediction model. Then, the initial prediction model is used to process each sample to obtain the predicted value of the sample property corresponding to each sample. Based on the difference between the predicted value of the sample property of each sample and the measured value of the sample property corresponding to the same property, the model loss is determined. The parameters of the initial prediction model are adjusted according to the model loss. After the adjustment is completed, this process is repeated until the obtained model loss converges. When the model loss converges, the initial prediction model can be regarded as the constructed prediction model.

[0144] In this embodiment, any sample used to construct the model may include the following data:

[0145] The embedding and physicochemical characteristics of wild-type proteins, the embedding and physicochemical characteristics of mutant proteins, the interaction characteristics of residues in mutant proteins, the structural characteristics of wild-type proteins, the mutation site information of wild-type proteins, and the measured values ​​of sample properties obtained by detecting the properties of mutant proteins. Mutant proteins are obtained by mutating wild-type proteins.

[0146] The measured values ​​of the sample properties contained in the sample can correspond to the several properties that the prediction model is to predict.

[0147] For example, if the prediction model is to be used to predict the stability and solubility of mutant proteins, then a sample can contain the measured values ​​of the sample stability and the measured values ​​of the sample solubility of the mutant proteins detected.

[0148] Different samples correspond to different mutant proteins. For example, sample 1 contains the above data for mutant protein A, and sample 2 contains the above data for mutant protein B.

[0149] When processing a sample using the initial prediction model, the data contained in the sample other than the measured values ​​of the sample properties can be processed using the prediction method of the aforementioned embodiment, thereby obtaining the predicted values ​​of the sample properties corresponding to the sample, such as the predicted values ​​of the sample stability and the predicted values ​​of the sample solubility.

[0150] Then, for each property that the model wants to predict, the measured value and predicted value of the sample property of each sample corresponding to that property are input into the calculation formula of the loss function to obtain the loss value corresponding to that property.

[0151] This embodiment does not limit the loss function used. As an example, the loss function can be any one of the mean squared error loss function (MSE Loss), smooth L1 loss function (Smooth L1 Loss), and L1 loss function (L1 Loss).

[0152] Combination Figure 3 For example, suppose we need to build a prediction model for predicting property 1 (e.g., stability) and property 2 (e.g., solubility), where the task1 prediction module is used to output the property prediction value for property 1, and the task2 prediction model is used to output the property prediction value for property 2.

[0153] When calculating the loss value, assuming the L1 loss function is used, the loss value Loss1 corresponding to property 1 can be calculated by the following formula (12), and the loss value Loss2 corresponding to property 2 can be calculated by the following formula (13).

[0154]

[0155]

[0156] Where H represents the total number of samples used to build the model, Y (1)，i Y represents the predicted value of property 1 corresponding to the i-th sample. (1)，i ' represents the measured value of property 1 corresponding to the i-th sample, Y (2)，i Y represents the predicted value of property 2 corresponding to the i-th sample. (2)，i ' represents the measured value of property 2 corresponding to the i-th sample.

[0157] After obtaining the loss value corresponding to each property, the loss values ​​corresponding to multiple properties can be weighted and averaged according to the preset loss weights corresponding to each property, and the result is used as the model loss.

[0158] Based on the previous example, assuming that the loss weight corresponding to property 1 is W1 and the loss weight corresponding to property 2 is W2, the model loss Loss can be calculated using the following formula (14).

[0159] Loss=W1*Loss1+W2*Loss2, (14).

[0160] Optionally, if the predicted value of a certain property belongs to the aforementioned binary or multi-class classification, then the loss value corresponding to that property can be calculated using loss functions such as cross-entropy loss function or Focal loss function. For specific calculation formulas, please refer to relevant technical literature, which will not be elaborated here.

[0161] The prediction model in this embodiment may require different parameters during construction (or training), without limitation. As some examples, the optional parameter settings are as follows.

[0162] In this embodiment, the output dimension of the dimensionality reduction module is set to 256, the output dimension of each layer of the graph isomorphic neural network in the series can be set to 256, the number of layers in the series K can be set to 3, that is, the prediction model includes 3 layers of graph isomorphic neural networks in the series, the loss weights W1 and W2 can both be set to 0.5, the maximum number of iterations is set to 200, that is, the model parameters are adjusted repeatedly 200 times according to the model loss, the batch size (i.e. the number of samples used for training) is set to 32, and the optimizer is Adam.

[0163] Optional, for Figure 3 The structure of the prediction model shown can also be constructed using transfer learning. In this case, the process of constructing the prediction model may include:

[0164] Construct a multi-layer neural network in series in the prediction model based on the initial dataset to obtain the pre-trained multi-layer neural network;

[0165] Identify the various properties that need to be predicted;

[0166] The initial prediction model is obtained by combining multiple prediction modules corresponding to various properties with a pre-trained multilayer neural network.

[0167] The initial prediction model is fine-tuned based on samples corresponding to various properties to obtain the prediction model.

[0168] The initial dataset can include a large number of samples, while a smaller number of samples can be used when fine-tuning the initial prediction model.

[0169] For specific methods of pre-training, please refer to relevant technical literature, which will not be elaborated here.

[0170] The user can select from a variety of properties that need to be predicted, such as stability and solubility.

[0171] During the above construction process, various prediction modules with different structures can be pre-set according to different properties. For example, prediction modules corresponding to regression tasks, binary classification tasks, and multi-class classification tasks can be set up separately. In this way, when building the model in the manner described in this embodiment, the appropriate prediction module can be selected and assembled according to the task corresponding to the property to be predicted.

[0172] The structure of the initial prediction model obtained by combining multiple prediction modules corresponding to various properties and a pre-trained multilayer neural network can be: Figure 3 The structure of the prediction model is shown.

[0173] For specific methods of fine-tuning, please refer to relevant technical literature, which will not be elaborated here.

[0174] The beneficial effect of this embodiment is that after pre-training a multi-layer neural network using a large initial dataset, the multi-layer neural network can be combined with the corresponding prediction module as needed. After combination, fine-tuning based on a small number of samples can obtain a model that can predict the predicted values ​​of the corresponding properties. On the one hand, it can expand the processing capability of the model, enabling the prediction model to predict the predicted values ​​of various different properties as needed. On the other hand, it can reduce the time and computational resource overhead required to retrain the model each time it is used to predict different properties.

[0175] This application also provides an apparatus for predicting the properties of mutant proteins; please refer to [link to relevant documentation]. Figure 4 The device may include the following units.

[0176] The obtaining unit 401 is used to obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein.

[0177] The fusion unit 402 is used to fuse the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction feature of the mutant protein residues to obtain the initial node features of the mutant protein; wherein, the interaction feature characterizes the interaction between every two residues in the mutant protein.

[0178] The processing unit 403 is used to process the initial node features and the structural features of the original protein using a prediction model to obtain multiple property prediction values ​​that characterize the mutant protein. The model loss of the prediction model is determined based on the sample property prediction values ​​that characterize the multiple properties obtained by processing the sample with the prediction model and the measured values ​​of the multiple sample properties of the sample.

[0179] Optionally, when the fusion unit 402 fuses the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction features of the mutant protein residues to obtain the initial node features of the mutant protein, it can be used for:

[0180] The first embedding feature and the first physicochemical feature are subjected to dimensionality reduction processing to obtain the first protein sequence feature;

[0181] The second embedding feature and the second physicochemical feature are subjected to dimensionality reduction processing to obtain the second protein sequence feature;

[0182] By splicing together the first protein sequence features, the second protein sequence features, and the interaction features of the mutant protein residues, the initial node features of the mutant protein are obtained.

[0183] Optionally, the process of obtaining structural features of the original protein by obtaining unit 401 may include:

[0184] The coordinates of the β-carbon atom of each residue in the original protein are determined based on the protein structure data (PDB) of the original protein.

[0185] The adjacency matrix of the original protein is determined based on the β-carbon coordinates of every two residues in the original protein. The adjacency matrix includes the distance between the β-carbon atoms of every two residues in the original protein.

[0186] The structural features of the original protein are determined based on the adjacency matrix.

[0187] Optionally, when processing unit 403 processes the initial node features and the structural features of the original protein using the prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein, it can be used for:

[0188] The initial node features and the structural features of the original protein are processed sequentially by a series of multi-layer neural networks in the prediction model to obtain the network output features of the mutant protein. The input of the first layer of the neural network includes the initial node features and the structural features, and the input of each layer of the neural network except the first layer includes the output and structural features of the previous layer.

[0189] For each prediction module of the prediction model, the network output features are processed by the prediction module to obtain a property prediction value that characterizes a property of the mutant protein; the prediction model includes multiple prediction modules, and the property prediction values ​​output by different prediction modules characterize different properties of the mutant protein.

[0190] Optionally, the processing unit 403 is also used for:

[0191] Based on the network output features and the mutation site information of the original protein, the mutation local features of the mutant protein are obtained.

[0192] When processing unit 403 uses the prediction module to process the network output features and obtains a property prediction value characterizing a property of the mutant protein, it can be used for:

[0193] By using the prediction module to process network output features and mutation local features, a property prediction value characterizing a property of the mutant protein is obtained.

[0194] Optionally, the process of obtaining the interaction characteristics of mutant protein residues by unit 401 may include:

[0195] Extracting multiple sequence alignment profiles of mutant proteins;

[0196] The interaction characteristics of mutant protein residues were calculated based on the multiple sequence alignment spectrum of the mutant protein.

[0197] Optionally, the device further includes a construction unit 404 for determining the model loss of the prediction model based on the samples and constructing the prediction model based on the model loss.

[0198] The process by which building unit 404 determines the model loss of the prediction model may include:

[0199] For each sample property prediction value, the loss value corresponding to the sample property prediction value is determined based on the sample property prediction value and the measured value of the sample property corresponding to the same property.

[0200] Based on the preset loss weights, the loss values ​​corresponding to the predicted values ​​of various sample properties are fused to obtain the model loss of the prediction model.

[0201] Optionally, the prediction model includes a cascaded multilayer neural network and multiple prediction modules;

[0202] The process of building a predictive model using building unit 404 may include:

[0203] Construct a multi-layer neural network in series in the prediction model based on the initial dataset to obtain the pre-trained multi-layer neural network;

[0204] Identify the various properties that need to be predicted;

[0205] The initial prediction model is obtained by combining multiple prediction modules corresponding to various properties with a pre-trained multilayer neural network.

[0206] The initial prediction model is fine-tuned based on samples corresponding to various properties to obtain the prediction model.

[0207] The working principle of the device for predicting the properties of mutant proteins provided in this embodiment can be found in the relevant steps of the method for predicting the properties of mutant proteins provided in any embodiment of this application, and will not be repeated here.

[0208] This application also provides an electronic device; please refer to [link to relevant documentation]. Figure 5 The device may include a memory 501 and a processor 502.

[0209] Memory 501 is used to store computer programs;

[0210] The processor 502 is used to execute a computer program, specifically to implement the method for predicting the properties of mutant proteins provided in any embodiment of this application.

[0211] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0212] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0213] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0214] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0215] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting the properties of mutant proteins, characterized in that, include: Obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein. The first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction feature of the mutant protein residues are fused to obtain the initial node feature of the mutant protein; wherein, the interaction feature characterizes the interaction between every two residues in the mutant protein; The initial node features and the structural features of the original protein are processed using a prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein; wherein, the model loss of the prediction model is determined based on the sample property prediction values ​​characterizing the various properties obtained by processing the samples by the prediction model and the measured values ​​of the multiple sample properties of the samples.

2. The method according to claim 1, characterized in that, The process of fusing the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction features of the mutant protein residues to obtain the initial node features of the mutant protein includes: The first embedding feature and the first physicochemical feature are subjected to dimensionality reduction processing to obtain the first protein sequence feature; The second embedding feature and the second physicochemical feature are subjected to dimensionality reduction processing to obtain the second protein sequence feature; The initial node features of the mutant protein are obtained by splicing together the first protein sequence features, the second protein sequence features, and the interaction features of the mutant protein residues.

3. The method according to claim 1, characterized in that, The process of obtaining the structural features of the original protein includes: The β-carbon coordinates of each residue in the original protein are determined based on the protein structure data (PDB) of the original protein. The adjacency matrix of the original protein is determined based on the β-carbon atom coordinates of every two residues in the original protein, and the adjacency matrix includes the distance between the β-carbon atoms of every two residues in the original protein. The structural features of the original protein are determined based on the adjacency matrix.

4. The method according to claim 1, characterized in that, The process of using a prediction model to process the initial node features and the structural features of the original protein yields multiple property prediction values ​​characterizing various properties of the mutant protein, including: The initial node features and the structural features of the original protein are processed sequentially using a series of multi-layer neural networks in the prediction model to obtain the network output features of the mutant protein; wherein, the input of the first layer of the neural network includes the initial node features and the structural features, and the input of each layer of the neural network except the first layer includes the output of the previous layer of the neural network and the structural features; For each prediction module of the prediction model, the network output features are processed by the prediction module to obtain a property prediction value that characterizes a property of the mutant protein; wherein, the prediction model includes multiple prediction modules, and the property prediction values ​​output by different prediction modules characterize different properties of the mutant protein.

5. The method according to claim 4, characterized in that, Also includes: Based on the network output features and the mutation site information of the original protein, the mutation local features of the mutated protein are obtained; The process of using the prediction module to process the network output features to obtain a property prediction value characterizing a property of the mutant protein includes: The prediction module processes the network output features and the mutation local features to obtain a property prediction value that characterizes a property of the mutant protein.

6. The method according to claim 1, characterized in that, The process of obtaining the interaction characteristics of the mutant protein residues includes: Extract the multiple sequence alignment profile of the mutant protein; The interaction characteristics of the mutant protein residues were calculated based on the multiple sequence alignment spectrum of the mutant protein.

7. The method according to claim 1, characterized in that, The process of determining the model loss of the prediction model includes: For each of the predicted values ​​of the sample property, the loss value corresponding to the predicted value of the sample property is determined based on the predicted value of the sample property and the measured value of the sample property corresponding to the same property. Based on preset loss weights, the loss values ​​corresponding to the predicted values ​​of various sample properties are fused to obtain the model loss of the prediction model.

8. The method according to claim 1, characterized in that, The prediction model includes a cascaded multilayer neural network and multiple prediction modules; The process of constructing the prediction model includes: Construct a multi-layer neural network in series in the prediction model based on the initial dataset to obtain a pre-trained multi-layer neural network; Determine the various properties that need to be predicted; The initial prediction model is obtained by combining the multiple prediction modules corresponding to the various properties and the pre-trained multilayer neural network. The initial prediction model is fine-tuned based on samples corresponding to the various properties to obtain the prediction model.

9. An apparatus for predicting the properties of mutant proteins, characterized in that, include: The obtaining unit is used to obtain the first embedding feature and the first physicochemical feature corresponding to the original protein, and the second embedding feature and the second physicochemical feature corresponding to the mutant protein generated by the mutation of the original protein; wherein, the embedding feature characterizes the amino acid sequence of the corresponding protein, and the physicochemical feature characterizes the physicochemical properties of the residues constituting the corresponding protein. A fusion unit is used to fuse the first embedding feature, the first physicochemical feature, the second embedding feature, the second physicochemical feature, and the interaction features of the mutant protein residues to obtain the initial node features of the mutant protein; wherein, the interaction features characterize the interaction between every two residues in the mutant protein; The processing unit is used to process the initial node features and the structural features of the original protein using a prediction model to obtain multiple property prediction values ​​characterizing various properties of the mutant protein; wherein, the model loss of the prediction model is determined based on the sample property prediction values ​​characterizing the various properties obtained by processing the sample with the prediction model and the measured values ​​of the multiple sample properties of the sample.

10. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program, specifically to implement the method for predicting the properties of mutant proteins as described in any one of claims 1 to 8.

Images