Method for predicting impact of mutation on protein stability, and training method and apparatus

Abstract

A training method and apparatus for a neural network model which is used for predicting the impact of mutation on protein stability. The method comprises: acquiring multiple sets of training samples, each set of samples comprising a wild-type protein sequence, a mutant protein sequence, and a ground truth label for free energy change data ΔΔG; respectively extracting feature vectors of the two sequences by means of two graph neural networks having the same parameters, and inputting the feature vectors into a prediction sub-network so as to obtain a prediction result of ΔΔG, and on the basis of the prediction result and the ground truth label, synchronously adjusting the parameters of the two graph neural networks and maintaining the consistency of same. Further comprised are a method and apparatus for using the trained neural network model to perform prediction.

Description

Method for predicting the effect of mutations on protein stability, training method and device

[0001] Cross-reference to related applications

[0002] This application claims priority to Chinese patent application No. 202411022256.2, filed on July 26, 2024, the entire contents of which are incorporated by reference. TECHNICAL FIELD

[0003] The present disclosure relates to the field of biological macromolecule performance prediction, in particular to the technical field of application of artificial intelligence, biological computing, etc. in protein stability prediction, and particularly relates to a method for predicting the effect of mutations on protein stability, a training method and a device. BACKGROUND

[0004] Proteins, as important biological macromolecules in living organisms, play an important role in constituting tissues and cells of the organism, catalysis, immunity, energy supply, etc. Proteins can only exert specific functions by adopting specific three-dimensional structures, and the ability to fold into corresponding three-dimensional structures is affected by protein stability. Therefore, it is crucial to accurately predict the effect of mutations on protein stability (free energy of folding). SUMMARY

[0005] It would be advantageous to provide a mechanism that alleviates, mitigates or even eliminates one or more of the above problems.

[0006] According to an aspect of the present disclosure, a training method of a neural network model for predicting the influence of a mutation on protein stability is provided, wherein an initial neural network model comprises a first initial graph neural network and a second initial graph neural network that are isomorphic and have the same initial parameters, and comprises an initial predictor network, the method comprising: based on a wild-type protein sequence before mutation, a mutant protein sequence after mutation, and free energy change data ΔΔG, constructing a plurality of sets of training samples, each set of training samples comprising a first protein amino acid sequence, a second protein amino acid sequence, and a true value label of free energy change; and performing model training on the initial neural network model using the plurality of sets of training samples, comprising: processing the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; processing the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; inputting the first target feature vector and the second target feature vector into the initial predictor network to obtain a prediction result of the free energy change; and adjusting parameters of the initial neural network model based on the prediction result and the true value label, wherein in response to determining that at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously, so that the parameters of the adjusted first initial graph neural network and the second initial graph neural network are the same.

[0007] According to the training method of the present disclosure, the first protein amino acid sequence comprises a plurality of first protein amino acid residues, wherein processing the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence comprises: preprocessing the first protein amino acid sequence to construct first graph data comprising node features and edge features; updating the node features in the first graph data based on the node features and the edge features in the first graph data using the first initial graph neural network to obtain a first intermediate feature vector for each of the plurality of first protein amino acid residues; and fusing the first intermediate feature vectors of the plurality of first protein amino acid residues to obtain the first target feature vector.

[0008] In some embodiments, preprocessing the first protein amino acid sequence to construct first graph data comprising node features and edge features comprises:

[0009] encoding the first protein amino acid sequence using a protein pre-training model to obtain an encoded feature vector for each of the first protein amino acid residues; determining a contact probability for each of a plurality of first protein amino acid residue pairs in the first protein amino acid sequence, the plurality of first protein amino acid residue pairs including any two first protein amino acid residue pairs formed by matching combinations of the first protein amino acid residues in the first protein amino acid sequence; and constructing the first graph data by using the encoded feature vector for each of the first protein amino acid residues as a node feature and determining an edge feature based on the contact probability for each of the first protein amino acid residue pairs.

[0010] In some embodiments, constructing the first graph data by using the encoded feature vector for each of the first protein amino acid residues as a node feature and determining an edge feature based on the contact probability for each of the first protein amino acid residue pairs includes:

[0011] For each of the first protein amino acid residue pairs, in response to determining that the contact probability of the first protein amino acid residue pair is greater than a preset threshold, establishing an edge connection between the first protein amino acid residue pair and setting the edge feature as the contact probability of the first protein amino acid residue pair.

[0012] In some embodiments, the second protein amino acid sequence includes a plurality of second protein amino acid residues, and processing the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence includes:

[0013] preprocessing the second protein amino acid sequence to construct a second graph data including node features and edge features; updating the node features in the second graph data based on the node features and the edge features in the second graph data using the second initial graph neural network to obtain a second intermediate feature vector for each of the second protein amino acid residues; and fusing the second intermediate feature vectors for the second protein amino acid residues to obtain the second target feature vector.

[0014] In some embodiments, preprocessing the second protein amino acid sequence to construct a second graph data including node features and edge features includes:

[0015] encoding the second protein amino acid sequence using a protein pre-training model to obtain an encoded feature vector of each of the plurality of second protein amino acid residues; determining a contact probability of a plurality of second protein amino acid residue pairs in the second protein amino acid sequence, the plurality of second protein amino acid residue pairs comprising any two second protein amino acid residue pairs formed by matching combinations of the second protein amino acid residues in the second protein amino acid sequence; and constructing the second graph data by taking the encoded feature vector of each of the plurality of second protein amino acid residues as a node feature and determining an edge feature based on the contact probability of the plurality of second protein amino acid residue pairs.

[0016] In some embodiments, constructing the second graph data by taking the encoded feature vector of each of the plurality of second protein amino acid residues as a node feature and determining an edge feature based on the contact probability of the plurality of second protein amino acid residue pairs comprises:

[0017] For each of the plurality of second protein amino acid residue pairs, in response to determining that the contact probability of the second protein amino acid residue pair is greater than a preset threshold, establishing an edge connection between the second protein amino acid residue pair and setting the edge feature as the contact probability of the second protein amino acid residue pair.

[0018] In some embodiments, the protein pre-training model is a ProtT5-XL-UniRef50 model.

[0019] In some embodiments, inputting the first target feature vector and the second target feature vector into the initial predictor network to obtain a prediction result of the free energy change comprises:

[0020] fusing the first target feature vector and the second target feature vector to obtain a fused feature vector; and processing the fused feature vector using a fully connected layer and an output layer included in the initial predictor network to obtain the prediction result of the free energy change.

[0021] In some embodiments, the type of the mutation is a single amino acid substitution, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 2, the graph convolution output dimension is 32, the number of fully connected layers is 3, and the fully connected output dimension is 64.

[0022] In some embodiments, the type of the mutation is a plurality of amino acid substitutions, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 3, the graph convolution output dimension is 32, the number of fully connected layers is 3, and the fully connected output dimension is 64.

[0023] In some embodiments, the type of the mutation is an insertion and / or deletion mutation, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 3, the graph convolution output dimension is 64, the number of fully connected layers is 2, and the fully connected output dimension is 64.

[0024] In some embodiments, the free energy change data is ΔΔG = (ΔG1- ΔG2), where ΔG1 is the free energy change of the wild-type protein sequence before mutation, and ΔG2 is the free energy change of the mutant protein sequence after mutation.

[0025] In some embodiments, the first initial graph neural network and the second initial graph neural network are based on SAGEConv.

[0026] In some embodiments, based on the wild-type protein sequence before mutation, the mutant protein sequence after mutation, and the free energy change data ΔΔG, a plurality of training samples are constructed, including:

[0027] a first training sample is constructed by taking the wild-type protein sequence before mutation as the first protein amino acid sequence, taking the mutant protein sequence after mutation as the second protein amino acid sequence, and taking the free energy change data as the free energy change true value label; and a second training sample is constructed by taking the wild-type protein sequence before mutation as the second protein amino acid sequence, taking the mutant protein sequence after mutation as the first protein amino acid sequence, and taking the opposite number of the free energy change data as the free energy change true value label.

[0028] According to another aspect of the present disclosure, a method for predicting the influence of mutation on protein stability using a neural network is provided, wherein the neural network model comprises a trained graph neural network and a predictor network, and the neural network model is obtained by training the above method, which comprises: obtaining a wild-type protein sequence before mutation of a to-be-predicted protein as a first protein amino acid sequence and a sequence of the to-be-predicted protein as a second protein amino acid sequence; processing the first protein amino acid sequence using the trained graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; processing the second protein amino acid sequence using the trained graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; and inputting the first target feature vector and the second target feature vector into the predictor network to obtain a prediction result of the free energy change.

[0029] According to another aspect of the present disclosure, there is provided a device for training a neural network model for predicting the effect of a mutation on protein stability, wherein the initial neural network model comprises a first initial graph neural network and a second initial graph neural network that are isomorphic and have the same initial parameters, and comprises an initial predictor network, the device comprising: a sample construction unit configured to construct a plurality of sets of training samples based on a wild-type protein sequence before mutation, a mutant protein sequence after mutation, and free energy change data ΔΔG, each set of training samples comprising a first protein amino acid sequence, a second protein amino acid sequence, and a true value label of free energy change; and a training unit configured to perform model training on the initial neural network model using the plurality of sets of training samples, the training unit comprising: a first processing sub-unit configured to process the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; a second processing sub-unit configured to process the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; a predictor unit configured to input the first target feature vector and the second target feature vector into the initial predictor network to obtain a prediction result of the free energy change; and a parameter adjustment sub-unit configured to adjust the parameters of the initial neural network based on the prediction result and the true value label, wherein in response to determining that at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously, so that the adjusted parameters of the first initial graph neural network and the second initial graph neural network are the same.

[0030] According to another aspect of the present disclosure, there is provided a device for predicting the effect of a mutation on protein stability using a neural network model, wherein the neural network model comprises a trained graph neural network and a predictor network, and the neural network model is trained using the device described above, the device comprising: an acquisition unit configured to acquire a wild-type protein sequence before mutation of a protein to be predicted as a first protein amino acid sequence and a sequence of the protein to be predicted as a second protein amino acid sequence; a first processing unit configured to process the first protein amino acid sequence using the trained graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; a second processing unit configured to process the second protein amino acid sequence using the trained graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; and a prediction unit configured to input the first target feature vector and the second target feature vector into the predictor network to obtain a prediction result of the free energy change.

[0031] According to another aspect of the present disclosure, a computer device is provided, comprising: at least one processor; and a memory having a computer program stored thereon, wherein the computer program, when executed by the processor, causes the processor to perform the above method.

[0032] According to another aspect of the present disclosure, a computer readable storage medium is provided, having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the above method.

[0033] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, causes the processor to perform the above method.

[0034] According to another aspect of the present disclosure, a neural network model trained according to the training method of the neural network model for predicting the influence of mutations on protein stability is provided.

[0035] According to one or more embodiments of the present disclosure, for the task of predicting the influence of mutations on protein stability, two graph neural networks with the same structure and parameters are used to process the first protein amino acid sequence and the second protein amino acid sequence in the sample respectively during training, and the parameters of the two graph neural networks are adjusted synchronously during adjustment of the neural network, so as to ensure that the parameters of the two graph neural networks are always the same. Through the above-mentioned manner, the trained neural network model can output accurate prediction results of the change in free energy.

[0036] It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain exemplary implementations of the application. The illustrated embodiments are merely examples and do not limit the scope of the claims. In all the drawings, like reference numerals refer to like elements throughout the accompanying drawings.

[0038] FIG. 1 shows a flowchart of a training method of a neural network model according to an exemplary embodiment of the present application;

[0039] FIG. 2 shows a flowchart of a process of processing a first protein amino acid sequence using a first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence according to an exemplary embodiment of the present application;

[0040] FIG. 3 shows a flowchart of a process of preprocessing a first protein amino acid sequence and constructing first graph data containing node features and edge features, according to an example embodiment of the present application;

[0041] FIG. 4 shows a flowchart of a process of processing a second protein amino acid sequence using a second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence, according to an example embodiment of the present application;

[0042] FIG. 5 shows a flowchart of a process of preprocessing a second protein amino acid sequence and constructing second graph data containing node features and edge features, according to an example embodiment of the present application;

[0043] FIG. 6 shows a flowchart of a process of processing the first target feature vector and the second target feature vector using a predictor network to obtain a prediction result of free energy change, according to an example embodiment of the present application;

[0044] FIG. 7 shows a schematic diagram of a neural network model, according to an example embodiment of the present application;

[0045] FIG. 8 shows a flowchart of a method of predicting the effect of a mutation on protein stability, according to an example embodiment of the present application;

[0046] FIG. 9 shows a structural block diagram of a training device of a neural network model for predicting the effect of a mutation on protein stability, according to an example embodiment of the present application;

[0047] FIG. 10 shows a structural block diagram of a device for predicting the effect of a mutation on protein stability, according to an example embodiment of the present application; and

[0048] FIG. 11 is a block diagram showing an example computer device that can be applied to the example embodiments. DETAILED DESCRIPTION

[0049] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to help the understanding of the present disclosure. These should be considered in the context of the overall description and should not be considered to limit the scope of the present disclosure. Also, for the sake of brevity and clarity, descriptions of well-known functions and structures are omitted from the following description.

[0050] In the present disclosure, the terms "first", "second", and the like are used to describe various elements only for the purpose of describing particular examples, and are not intended to limit the positional relationship, the timing relationship, or the importance relationship of the elements. In some examples, the first element and the second element can refer to the same instance of the element, and in some cases, based on the context of the description, they can also refer to different instances.

[0051] The terms used in the description of various described examples in the present disclosure are only for the purpose of describing the particular examples, and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. In addition, the term "and / or" used in the present disclosure encompasses any one of the listed items and all possible combinations thereof.

[0052] In the related art, in the method for predicting the stability of a protein after a protein mutation, the way of using protein information is relatively single, and the prediction effect needs to be improved.

[0053] To solve this problem, the present disclosure uses two isomorphic and same-parameter graph neural networks to process the first protein amino acid sequence and the second protein amino acid sequence in the sample respectively during training, and synchronously adjusts the parameters of the two graph neural networks when adjusting the neural network, to ensure that the parameters of the two are always the same. Through the above-mentioned manner, the trained neural network model can output accurate prediction results of the free energy change.

[0054] The exemplary embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0055] According to an aspect of the present disclosure, a method for training a neural network model for predicting the effect of a mutation on protein stability is provided. FIG. 1 shows a flowchart of a method for training a neural network for predicting the effect of a mutation on protein stability according to an example embodiment of the present disclosure. The initial neural network model includes a first initial graph neural network and a second initial graph neural network that are isomorphic and have the same initial parameters, and includes an initial predictor network. As shown in FIG. 1, the method 100 includes: step S101, constructing a plurality of sets of training samples based on a wild-type protein sequence before mutation, a mutant protein sequence after mutation, and free energy change data ΔΔG, each set of training samples including a first protein amino acid sequence, a second protein amino acid sequence, and a true value label of free energy change; step S102, training the initial neural network model using the plurality of sets of training samples; step S1021, processing the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; step S1022, processing the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; step S1023, inputting the first target feature vector and the second target feature vector into the initial predictor network to obtain a prediction result of the free energy change; and step S1024, adjusting parameters of the initial neural network model based on the prediction result and the true value label, wherein in response to determining that at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously, so that the parameters of the first initial graph neural network and the second initial graph neural network after adjustment are the same.

[0056] Thus, by using two isomorphic and parameter-same graph neural networks to process the first protein amino acid sequence and the second protein amino acid sequence in the sample respectively at the time of training for the task of predicting the effect of a mutation on protein stability, and synchronously adjusting the parameters of the two graph neural networks when adjusting the neural network to ensure that the parameters of the two graph neural networks are always the same, the trained neural network can output an accurate prediction result of the free energy change.

[0057] In the present disclosure, although each operation is depicted in the accompanying drawings as following a specific order, this should not be understood as requiring that the operations must be performed in the specific order shown or in sequential order, for example, the order of step S1021 and step S1022 can be replaced with each other or can be performed simultaneously. In addition, it can be understood that steps S1021-S1024 can be sub-steps of step S102.

[0058] In some embodiments, in step S101, the mutation refers to a change in the genetic material (DNA) of the organism, and the types of mutations include but are not limited to: single-point mutations, multi-point mutations, mutations to mutations, insertions, and deletions. The wild-type protein sequence before mutation refers to the amino acid sequence of the protein obtained from nature, that is, the sequence translated from the gene or DNA sequence without mutation. The mutant protein sequence after mutation refers to the changed sequence translated from the gene that has undergone mutation, and the change can be single-point or multi-point substitution (replacement), insertion, or deletion of the protein amino acid sequence.

[0059] Free energy generally refers to the energy available to the system when the protein changes from one state to another under certain conditions. In some embodiments, free energy can refer to the thermodynamic folding stability value of the protein, which refers to the ability of the protein to maintain its functional three-dimensional structure under certain conditions. This stability is the result of the combined action of various interaction forces such as hydrogen bonds, hydrophobic interactions, van der Waals forces, ionic bonds, etc. By determining the thermodynamic folding stability value of the protein, we can gain a deeper understanding of the structure, function of the protein, and how they function in the organism, and further analyze whether the mutation before and after the protein will cause changes in the stability of the protein.

[0060] In some embodiments, the free energy change data ΔΔG = (ΔG1-ΔG2), ΔG1 is the free energy change of the wild-type protein sequence before mutation, and ΔG2 is the free energy change of the mutant protein sequence after mutation. Further, ΔΔG directly reflects the influence of mutation on protein stability. If the value is positive, it means that the mutation may cause the stability of the protein to decrease; if it is negative, it may enhance the stability. When designing protein variants with specific functions, ΔΔG can be calculated to predict the potential impact of mutations on protein stability, thereby guiding rational design.

[0061] In some embodiments, in step S101, based on the wild-type protein sequence before mutation, the mutant protein sequence after mutation, and the free energy change data ΔΔG, a plurality of training samples are constructed, and the plurality of training samples can include: taking the wild-type protein sequence before mutation as the first protein amino acid sequence, taking the mutant protein sequence after mutation as the second protein amino acid sequence, and taking the free energy change data as the free energy change true value label to construct the first training sample; and taking the wild-type protein sequence before mutation as the second protein amino acid sequence, taking the mutant protein sequence after mutation as the first protein amino acid sequence, and taking the opposite number of the free energy change data as the free energy change true value label to construct the second training sample.

[0062] The number of training samples is doubled by reversing the data of the wild-type protein sequence before mutation and the mutant protein sequence after mutation. Reversing, that is, the change of folding free energy (ΔΔG) is inversely symmetrical when the mutation occurs under the same environmental conditions. More precisely, if protein B is a mutant of protein A, then ΔΔG(A→B) = -ΔΔG(B→A), which is consistent with the law of conservation of energy in physics. According to the above definition, the first training sample is combined with the reversed data (the second training sample) to double the original data amount. By reversing this method, the original data set is expanded, and the data set is balanced at the same time, making the robustness of the final prediction model stronger.

[0063] In some exemplary embodiments, the wild-type protein sequence before mutation, the mutant protein sequence after mutation, and the free energy change data ΔΔG are selected from the high-throughput data of protein thermodynamic folding stability determined by cDNA proteolysis, which includes the thermodynamic folding stability determination values (ΔG) of single-point mutation, two-point mutation and insertion-deletion mutation of proteins. The data set has more than 500,000 data, which well makes up for the problem of less training data, is beneficial to train a graph neural network model with stronger generalization ability, and avoids the small data set from easily causing model overfitting and being unable to support complex graph neural network model operation.

[0064] In some exemplary embodiments, in order to better optimize the training data set, the high-throughput data of protein thermodynamic folding stability determined by cDNA proteolysis can be cleaned. After cleaning, the data set has a total of 325,908 experimental data, including single-point mutation data, multi-point mutation data, and insertion-deletion data. Among them, there are 32,314 experimental data of stable mutations and 293,594 experimental data of unstable mutations. In order to obtain a relatively balanced data set, the original data is reversed, and the specific steps include: removing data with empty ΔG / ΔΔG values; removing data of wild-type mutations; removing data of artificial synthetic proteins; removing data not based on wild-type protein sequence mutations; and reversing and merging the data. The data reversing method is the same as described above.

[0065] In some embodiments, in step S102, the initial neural network model is trained using multiple sets of training samples, which includes steps S1021, S1022, S1023 and S1024. These steps can further improve the generalization ability of the model, make the model learn more extensive features, reduce overfitting to specific samples, and thus improve the generalization ability on unseen data; improve model performance, help the model better understand the patterns and relationships in the data; enhance robustness, making the model more robust to small changes or errors in input data; and improve the interpretability of the model, better understanding the decision-making process and potential biases of the model.

[0066] In some embodiments, step S1021 uses the first initial graph neural network to process the first protein amino acid sequence to obtain a first target feature vector representing the first protein amino acid sequence. The first target feature vector is usually a high-dimensional and dense representation that encodes local or global graph structure information of the target node and related feature information.

[0067] In some embodiments, step S1022 uses the second initial graph neural network to process the second protein amino acid sequence to obtain a second target feature vector representing the second protein amino acid sequence. The second target feature vector is usually a high-dimensional and dense representation that encodes local or global graph structure information of the target node and related feature information.

[0068] A graph neural network is a deep learning model based on graph structure data, used to process structured data of nodes and edges. In some embodiments, the first initial graph neural network and the second initial graph neural network in steps S1021 and S1022 can adopt a graph convolutional network (GCN), a graph attention network (GAT), a graph isomorphism network (GIN), a diffusion convolutional neural network (DCNN), a graph sampling and aggregation (GraphSAGE), or other structured graph neural networks.

[0069] In some embodiments, the first initial graph neural network and the second initial graph neural network in steps S1021 and S1022 can be based on GraphSAGE. GraphSAGE is a kind of sampling-based graph neural network, which realizes effective learning and representation of node features through sampling and aggregation of neighbor nodes. The above-mentioned characteristics of GraphSAGE make it have significant advantages in processing large-scale graph data, especially in processing protein sequences to complete the protein stability prediction task. GraphSAGE can balance efficiency and accuracy.

[0070] Further, in some exemplary embodiments, an adaptive convolution algorithm (SAGEConv) layer can be constructed in GraphSAGE. SAGEConv is a variant of graph convolution network, derived from GraphSAGE algorithm. The SAGEConv layer updates the node features through an aggregation function, such as pooling, mean or recurrent neural network. In this way, the protein representation is further processed as the input of the subsequent fully connected network layer.

[0071] In some embodiments, in step S1024, the free energy change data ΔΔG obtained based on the prediction of step S1023 is adjusted with the free energy change data ΔΔG in the corresponding training sample in step S101 as the true value label and the preset loss function, to ensure that the gap between the prediction result output by the adjusted neural network model and the corresponding true value label is reduced, and the model is continuously adjusted until the loss function of the model is minimized, indicating that the model training is completed. Wherein, when at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network needs to be adjusted synchronously, so that the parameters of the adjusted first initial graph neural network and the second initial graph neural network are the same. Synchronous adjustment of parameters is beneficial to avoid the model from falling into local optimal solution, find global optimal solution, further improve the stability of the model, improve the resource utilization efficiency, and better adapt to different data sets.

[0072] FIG. 2 shows a flowchart of a process of processing a first protein amino acid sequence using a first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence, according to an example embodiment of the present application. In the first protein amino acid sequence, the first protein amino acid sequence comprises a plurality of first protein amino acid residues. Process 200 can be used to implement step S1021 in method 100 described above. In some embodiments, as shown in FIG. 2, process 200 can include: step S201, preprocessing the first protein amino acid sequence to construct a first graph data comprising node features and edge features; step S202, using the first initial graph neural network to update the node features in the first graph data based on the node features and the edge features in the first graph data to obtain a first intermediate feature vector for each of the plurality of first protein amino acid residues; and step S203, fusing the first intermediate feature vectors of the plurality of first protein amino acid residues to obtain the first target feature vector.

[0073] Thus, by the above-described manner, the sequence information of the protein can be more effectively utilized under the framework of the graph neural network model, so as to obtain a first target feature vector with stronger prediction capability, and the accuracy and robustness of the subsequent protein stability prediction before and after mutation are improved.

[0074] In step S201, the first protein amino acid sequence comprises a plurality of first protein amino acids, and the plurality of first protein amino acid residues included in the first protein amino acid sequence can be determined as a plurality of nodes in the first graph data, and the topological relationship of the first graph data can be determined based on the plurality of first protein amino acid residues.

[0075] In some embodiments, in the preprocessing process of step S201, the protein amino acid sequence can be parsed into individual amino acid nodes, the amino acid types and other category features can be encoded, the chemical properties of the amino acids can be obtained by computational chemistry methods, and the features can be standardized or normalized. Through the preprocessing process, the success rate of data analysis and modeling can be significantly improved.

[0076] In some embodiments, the node features can represent numerical features of the chemical properties of each amino acid residue, such as hydrophobicity, polarity, molecular weight, isoelectric point. The node features can be obtained using a protein pre-training model, or can be obtained by other means. By analyzing the node features, detailed information of each amino acid residue in the sequence can be more accurately obtained, further improving the prediction accuracy.

[0077] In some embodiments, the edge features can represent interaction information, distance information, connection strength, etc. between pairs of amino acid residues in the sequence. By analyzing the edge features, detailed information between pairs of amino acid residues in the sequence can be more accurately obtained, further improving the prediction accuracy.

[0078] In some embodiments, the first initial graph neural network can include a plurality of first initial graph convolutional layers, and the number of first initial graph convolutional layers can be selected as 2 or 3. By configuring two or three graph convolutional layers in the first initial graph neural network to process the node features and edge features in the first graph data, the best prediction performance can be achieved while keeping the computational resource consumption reasonable.

[0079] In some embodiments, in step S202, the node features in the first graph data can be updated by iterations of a plurality of first initial graph convolutional layers to obtain a plurality of first intermediate feature vectors of the first protein amino acid residues.

[0080] Thus, through the iterative update of multiple layers of graph convolution, each layer of convolution can collect information from a wider neighborhood range, thereby aggregating more and more context information layer by layer. More specifically, for a two-layer graph convolutional layer architecture: the first layer mainly focuses on the features of the directly adjacent amino acid residues of each amino acid residue, including the original features of the amino acid nodes and possible edge features. The features of the adjacent amino acid residue nodes are combined with the features of the target amino acid residue node through an aggregation function (such as summation, averaging, maximum, etc.) to generate a preliminary feature representation. The second layer further expands on the basis of the first layer to collect more extensive neighborhood information of the first layer output. Here, the "neighborhood" can include the neighbors of the neighbors of certain amino acid residue nodes in the first layer.

[0081] For a three-layer graph convolutional layer architecture: the first layer mainly captures the local features of other amino acid residues close to the amino acid residue, the second layer starts to fuse neighborhood information of a larger range, and the third layer captures more global structural information.

[0082] In some embodiments, the output result of the last first initial graph convolutional layer can be used as the first intermediate feature vectors of the plurality of first protein amino acid residues, or the output results of the plurality of first initial graph convolutional layers can be fused to obtain the first intermediate feature vectors of the plurality of first protein amino acid residues.

[0083] In some embodiments, in step S203, the first intermediate feature vectors of the plurality of first protein amino acid residues can be fused by splicing, weighted summation, mean pooling, maximum pooling, attention mechanism, direct summation, using a multilayer perceptron or other neural network for processing, or any combination of the above methods, thereby obtaining a first target feature vector representing the first protein amino acid sequence.

[0084] FIG. 3 shows a flowchart of a process of pre-processing a first protein amino acid sequence and constructing a first graph data comprising node features and edge features, according to an example embodiment of the present application. The process 300 can be used to implement the step S201 in the method 200 described above. In some embodiments, as shown in FIG. 3, the process 300 can comprise: a step S301 of encoding the first protein amino acid sequence by using a protein pre-training model to obtain an encoded feature vector for each of a plurality of first protein amino acid residues; a step S302 of determining a contact probability for a plurality of first protein amino acid residue pairs in the first protein amino acid sequence, the plurality of first protein amino acid residue pairs comprising any two first protein amino acid residue pairs formed by matching combinations of any two first protein amino acid residues in the first protein amino acid sequence; and a step S303 of taking the encoded feature vector for each of the plurality of first protein amino acid residues as a node feature, and determining an edge feature based on the contact probability for the plurality of first protein amino acid residue pairs, to construct the first graph data.

[0085] Thus, by the above-described manner, the complex patterns and biological properties in the amino acid sequence can be more effectively captured under the framework of the graph neural network model, the encoded feature vector provides a higher-level and abstract feature representation than the original sequence, the representation capability is enhanced, the spatial proximity between the residues within the protein is measured in combination with the contact probability, which helps to integrate the three-dimensional structure information of the protein, and the prediction accuracy is further improved and the model robustness is improved.

[0086] It can be understood that, in some embodiments, the order of the step S301 and the step S302 can be replaced with each other or can be performed simultaneously.

[0087] In some embodiments, in the step S301, the plurality of first protein amino acid residues comprised in the first protein amino acid sequence can be directly input into the protein pre-training model to obtain the encoding result output by the protein pre-training model, i.e., the encoded feature vector for each of the plurality of first protein amino acid residues.

[0088] The pre-training model can be obtained by pre-learning language features on a large amount of unlabeled data, and then fine-tuning on a small amount of labeled data. Such a training manner can improve the performance of downstream tasks. By training on a large corpus of protein or nucleic acid sequences, the protein pre-training model used in the step S301 can be obtained.

[0089] In some embodiments, the protein pre-training model can be a ProtT5-XL-UniRef50 model. The ProtT5-XL-UniRef50 model is pre-trained on a large-scale protein sequence dataset, capable of capturing sequence information, structure information and function information of proteins, and providing rich and accurate feature representations. By using the ProtT5-XL-UniRef50 model as the protein pre-training model, the quality of feature extraction can be significantly improved, ultimately improving the accuracy of prediction.

[0090] In one exemplary embodiment, the ProtT5-XL-UniRef50 pre-training model is a protein language model trained in the UniRef50 database based on a Transformer architecture. The Transformer architecture mainly consists of an encoder and a decoder, and the last layer embedding layer of the encoder is usually used to extract amino acid encoding information when extracting sample features for downstream tasks. In this experiment, the protein sequence is input into the ProtT5-XL-UniRef50 model to obtain the 1024-dimensional features of the protein amino acids as the node features of the graph. N*1024 is obtained, where N is the length of the protein sequence, i.e., the number of amino acids. N*1024 represents the features of the protein.

[0091] In some embodiments, the protein pre-training model can also be ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XXL, Ankh or other models, which are not limited herein.

[0092] The contact probability is the possibility of a pair of protein amino acid residues in a protein sequence forming a contact. In some embodiments, the formation of a contact by a pair of protein amino acid residues can be defined as the distance between the atoms (e.g., Cβ atoms) of the two protein amino acid residues of the protein being less than a preset distance threshold (e.g., 4.5 Å). In one specific embodiment of the present disclosure, the protein amino acid sequence can be input into the SPOT-Contact-LM model to predict a contact map, which is composed of an N*N matrix, and the value in the matrix determines the contact probability of a pair of amino acid residues. The closer the contact probability is to 1, the greater the possibility of contact between the two pairs of amino acid residues.

[0093] Observations revealed that by obtaining the contact probabilities of multiple first protein amino acid residue pairs and combining the encoding feature vectors of multiple first protein amino acid residues in the first protein amino acid sequence, and then processing them with a graph neural network, a feature vector with strong predictive power for protein stability can be obtained without obtaining or predicting the spatial structure information of the first protein amino acid sequence, thereby significantly reducing the cost of protein stability prediction.

[0094] In some embodiments, step S302, determining the contact probability of multiple first protein amino acid residue pairs in the first protein amino acid sequence, may include: identifying multiple first protein amino acid residue pairs in the first protein amino acid sequence; for each amino acid residue pair, obtaining a high-dimensional feature vector between the first and second protein amino acid residues included in that amino acid residue pair; and simultaneously inputting the high-dimensional feature vectors of the multiple first protein amino acid residue pairs into a trained neural network for predicting contact probabilities to obtain the contact probability of the multiple first protein amino acid residue pairs. In this manner, the neural network for predicting contact probabilities can consider information from other amino acids when predicting the contact probability between each amino acid residue pair, thereby improving the accuracy of the predicted contact probability.

[0095] In some embodiments, multiple first protein amino acid residue pairs may include amino acid residue pairs composed of optional first and second protein amino acid residues in the first protein amino acid sequence. That is, the contact probability between any two amino acid residues in the target protein sequence can be determined in step S302. By determining the contact probabilities between all possible amino acid residue pairs, more comprehensive protein structure information can be provided, reducing information omissions and adapting to the needs of complex proteins. Furthermore, graph neural networks rely on the relationships between nodes (amino acid residues) and edges (amino acid residue pairs) for feature learning and prediction. By determining the contact probabilities of all amino acid residue pairs, more accurate edge topology and edge features can be provided, enabling the graph neural network to make fuller use of this information, thereby improving the performance of the neural network and ultimately contributing to improving the accuracy and reliability of protein stability prediction.

[0096] In some embodiments, the high-dimensional feature vector between the first and second protein amino acid residues (i.e., the high-dimensional feature vector of the amino acid residue pair) may include any combination of the following features: the one-hot encoding of each of the first and second protein amino acid residues, the attention scores of the first and second protein amino acid residues, and / or the structural information of the target protein sequence. In an exemplary embodiment, the attention score may come from an attention map generated by a protein pre-training model based on the target protein sequence; the structural information may include the probabilities of 3-state secondary structures (SS3) and 8-state secondary structures (SS8), solvent-accessible surface area (ASA), hemispherical exposure (HSE), dihedral angles ψ, φ, θ, τ of the protein backbone, and / or other protein structural information.

[0097] In some embodiments, the above features can be combined using splicing, element-level operations (e.g., element addition, subtraction, multiplication) or other methods to obtain a high-dimensional feature vector of amino acid residue pairs.

[0098] In some embodiments, the neural network used to predict the contact probability may employ a residual network (e.g., ResNet). This neural network may be trained using labeled data describing whether or not contact occurs based on amino acid residues.

[0099] In some embodiments, in step S303, the encoding feature vectors of each of the multiple first protein amino acid residues obtained in step S301 and the edge features determined by the contact probabilities of the multiple first protein amino acid residue pairs obtained in step S302 can be used to construct the first graph data.

[0100] In some embodiments, step S303, using the encoding feature vectors of each of the plurality of first protein amino acid residues as node features and determining edge features based on the contact probability of the plurality of first protein amino acid residue pairs to construct the first graph data may include: for each of the plurality of first protein amino acid residue pairs, in response to determining that the contact probability of the first protein amino acid residue pair is greater than a preset threshold, establishing an edge connection between the first protein amino acid residue pairs, and setting the edge feature as the contact probability of the first protein amino acid residue pair.

[0101] By using the above method, we can avoid establishing edges between nodes corresponding to amino acid residues that have no contact relationship, thereby effectively reducing the number of edges, reducing the demand for computing and storage resources, and improving the inference efficiency of the first graph neural network.

[0102] In some exemplary embodiments, the preset threshold can be 0.1. When the contact probability between any two first protein amino acid base pairs is greater than 0.1, it can be considered that the two first protein amino acid residues are in contact and have an edge connection relationship. An edge connection is established between the two first protein amino acid residues, and the edge feature corresponding to the edge connection is set as the contact probability of the first protein amino acid residue pair. If the contact probability between the first protein amino acid base pairs is less than 0.1, it is considered that there is no contact and no edge connection relationship exists.

[0103] By employing the above methods, we can further simplify the model's complexity, reduce its computational burden, minimize unnecessary noise, and decrease the space required for storing graph representations, thereby further improving the accuracy and efficiency of protein stability prediction.

[0104] In some embodiments, in addition to the above methods, the first graph data can also be constructed in other ways, or the first protein amino acid sequence can be processed by the first initial graph neural network in other ways to obtain the first target feature vector characterizing the first protein amino acid sequence.

[0105] Figure 4 illustrates a flowchart of a process for processing a second protein amino acid sequence using a second initial graph neural network according to an exemplary embodiment of this application to obtain a second target feature vector characterizing the second protein amino acid sequence. In the second protein amino acid sequence, the second protein amino acid sequence contains multiple first protein amino acid residues. Process 400 can be used to implement step S1022 in the method 100 described above. In some embodiments, as shown in Figure 4, process 400 may include: step S401, preprocessing the second protein amino acid sequence to construct second graph data containing node features and edge features; step S402, using the second initial graph neural network to update the node features in the second graph data based on the node features and edge features in the second graph data to obtain first intermediate feature vectors for each of the multiple second protein amino acid residues; and step S403, fusing the second intermediate feature vectors of the multiple second protein amino acid residues to obtain a second target feature vector.

[0106] Therefore, by using the above method, protein sequence information can be utilized more effectively within the framework of graph neural network models, thereby obtaining a second target feature vector with stronger predictive power, and improving the accuracy and robustness of subsequent protein stability prediction before and after mutation.

[0107] In step S401, the second protein amino acid sequence contains multiple second protein amino acids. The multiple second protein amino acid residues included in the second protein amino acid sequence can be identified as multiple nodes in the second graph data, and the topological relationship of the second graph data can be determined based on the multiple pairs of second protein amino acid residues.

[0108] In some embodiments, during the preprocessing step S401, the protein amino acid sequence can be parsed into individual amino acid nodes, the amino acid type and other category features can be encoded, the chemical properties of the amino acids can be obtained through computational chemistry methods, and the features can be standardized or normalized. This preprocessing can significantly improve the success rate of data analysis and modeling.

[0109] In some embodiments, node features can be numerical characteristics representing the chemical properties of each amino acid residue, such as hydrophobicity, polarity, molecular weight, and isoelectric point. Node features can be obtained using a protein pre-trained model or through other methods. By analyzing node features, detailed information about each amino acid residue in the sequence can be obtained more accurately, further improving prediction accuracy.

[0110] In some embodiments, edge features can characterize information such as interaction information, distance information, and connection strength between amino acid residue pairs in a sequence. By analyzing edge features, detailed information about the amino acid residue pairs in a sequence can be obtained more accurately, further improving the accuracy of predictions.

[0111] In some embodiments, the second initial graph neural network may include multiple second initial graph convolutional layers, the number of which can be selected as 2 or 3. By configuring two or three graph convolutional layers in the second initial graph neural network to process the node features and edge features in the second graph data, optimal prediction performance can be achieved while maintaining reasonable computational resource consumption.

[0112] It should be noted that the first initial graph neural network and the second initial graph neural network are isomorphic and have the same initial parameters. Therefore, the number, structure and parameters of the second initial graph convolutional layers included in the second initial graph neural network are the same as those of the first initial graph convolutional layers included in the first initial graph neural network.

[0113] In some embodiments, in step S402, the node features in the second graph data can be updated using multiple second initial graph convolutional layers iteratively, and the second intermediate feature vector of multiple second protein amino acid residues output by each second initial graph convolutional layer can be obtained.

[0114] Therefore, through iterative updates of multi-layer graph convolutions, each convolutional layer can collect information from a broader neighborhood, thus aggregating more and more contextual information layer by layer. More specifically, for a two-layer graph convolutional architecture: the first layer mainly focuses on the features of the directly adjacent amino acid residues of each amino acid residue, including the original features of the amino acid node and possible edge features. At the same time, it combines the features of adjacent amino acid residue nodes with the features of the target amino acid residue node through aggregation functions (such as summation, averaging, maximum, etc.) to generate a preliminary feature representation; the second layer further expands on the first layer, collecting a wider range of neighborhood information from the output of the second layer. Here, "neighborhood" may include the neighbors of the neighbors of some amino acid residue nodes in the second layer.

[0115] For the three-layer graph convolutional layer architecture: the first layer mainly captures the local features of other amino acid residues that are close to the amino acid residues; the second layer begins to fuse a larger range of neighborhood information; and the third layer captures more global structural information.

[0116] In some embodiments, the output of the last second initial graph convolutional layer can be used as the second intermediate feature vector of multiple second protein amino acid residues, or the outputs of multiple second initial graph convolutional layers can be fused to obtain the second intermediate feature vector of multiple second protein amino acid residues.

[0117] In some embodiments, in step S403, fusing the second intermediate feature vectors of multiple second protein amino acid residues can be achieved by splicing, weighted summation, mean pooling, max pooling, attention mechanism, direct summation, processing with a multilayer perceptron or other neural networks, or any combination of the above methods, thereby obtaining a second target feature vector characterizing the amino acid sequence of the second protein.

[0118] In some embodiments, the fusion in step S403 may include merging the second intermediate feature vectors of the multiple second protein amino acid residues output by the multiple second graph convolutional layers using a residual network structure.

[0119] Therefore, by using the above method, the second intermediate feature vectors at different levels can be fused more effectively, enhancing the expressive power of the second target feature vector obtained after fusion, thereby improving the performance of the neural network in protein mutation prediction.

[0120] Figure 5 illustrates a flowchart of a process for preprocessing a second protein amino acid sequence to construct second graph data including node features and edge features according to an exemplary embodiment of this application. Process 500 can be used to implement step S401 in method 400 described above. In some embodiments, as shown in Figure 5, process 500 may include: step S501, encoding the second protein amino acid sequence using a protein pre-training model to obtain encoding feature vectors for each of multiple second protein amino acid residues; step S502, determining the contact probability of multiple pairs of second protein amino acid residues in the second protein amino acid sequence, wherein the multiple pairs of second protein amino acid residues include pairs of second protein amino acid residues formed by matching and combining any two second protein amino acid residues in the second protein amino acid sequence; and step S503, using the encoding feature vectors of each of the multiple second protein amino acid residues as node features, and determining edge features based on the contact probability of the multiple pairs of second protein amino acid residues to construct second graph data.

[0121] Therefore, by means of the above method, it is possible to more effectively capture complex patterns and biological characteristics in amino acid sequences within the framework of graph neural networks. Encoding feature vectors provides a higher-level and more abstract feature representation than the original sequence, enhancing the characterization ability. Combined with contact probability, it obtains a measure of the spatial proximity between residues within the protein, which helps to integrate the three-dimensional structural information of the protein, further improving prediction accuracy and model robustness.

[0122] It is understood that in some embodiments, the order of steps S501 and S502 can be interchanged or performed simultaneously.

[0123] In some embodiments, in step S501, multiple second protein amino acid residues included in the second protein amino acid sequence can be directly input into the protein pre-training model to obtain the encoding results output by the protein pre-training model, i.e., the encoding feature vectors of each of the multiple second protein amino acid residues. The protein pre-training model used in step S501 can be the protein pre-training model used in step S301 above.

[0124] Observations revealed that by acquiring the contact probabilities of multiple second protein amino acid residue pairs and combining the encoding feature vectors of multiple second protein amino acid residues in the second protein amino acid sequence, and then processing them with a graph neural network, a feature vector with strong predictive power for protein stability can be obtained without acquiring or predicting the spatial structure information of the first protein amino acid sequence, thereby significantly reducing the cost of protein stability prediction.

[0125] In some embodiments, step S502, determining the contact probability of multiple pairs of second protein amino acid residues in the second protein amino acid sequence, may include: identifying multiple pairs of second protein amino acid residues in the second protein amino acid sequence; for each amino acid residue pair, obtaining a high-dimensional feature vector between the second protein amino acid residues included in the amino acid pair; and simultaneously inputting the high-dimensional feature vectors of the multiple second protein amino acid residue pairs into a trained neural network for predicting contact probabilities to obtain the contact probability of the multiple second protein amino acid residue pairs. In this manner, the neural network for predicting contact probabilities can consider information from other amino acids when predicting the contact probability between each pair of amino acid residues, thereby improving the accuracy of the predicted contact probability.

[0126] In some embodiments, multiple second protein amino acid residue pairs may include amino acid residue pairs composed of optional first and second protein amino acid residues in the second protein amino acid sequence. That is, the contact probability between any two amino acid residues in the target protein sequence can be determined in step S502. By determining the contact probabilities between all possible amino acid residue pairs, more comprehensive protein structure information can be provided, reducing information omissions and adapting to the needs of complex proteins. Furthermore, graph neural networks rely on the relationships between nodes (amino acid residues) and edges (amino acid residue pairs) for feature learning and prediction. By determining the contact probabilities of all amino acid residue pairs, more accurate edge topology and edge features can be provided, enabling the graph neural network to make fuller use of this information, thereby improving the performance of the neural network and ultimately contributing to improving the accuracy and reliability of protein stability prediction.

[0127] In some embodiments, the high-dimensional feature vector between the first and second protein amino acid residues (i.e., the high-dimensional feature vector of the amino acid residue pair) may include any combination of the following features: the one-hot encoding of each of the first and second protein amino acid residues, the attention scores of the first and second protein amino acid residues, and / or the structural information of the target protein sequence. In an exemplary embodiment, the attention score may come from an attention map generated by a protein pre-training model based on the target protein sequence; the structural information may include the probabilities of 3-state secondary structures (SS3) and 8-state secondary structures (SS8), solvent-accessible surface area (ASA), hemispherical exposure (HSE), dihedral angles ψ, φ, θ, τ of the protein backbone, and / or other protein structural information.

[0128] In some embodiments, the above features can be combined using splicing, element-level operations (e.g., element addition, subtraction, multiplication) or other methods to obtain a high-dimensional feature vector of amino acid residue pairs.

[0129] In some embodiments, the neural network used to predict the contact probability may employ a residual network (e.g., ResNet). This neural network may be trained using labeled data describing whether or not contact occurs based on amino acid residues.

[0130] In some embodiments, in step S503, the encoding feature vectors of each of the multiple second protein amino acid residues obtained in step S501 and the edge features determined by the contact probabilities of the multiple second protein amino acid residue pairs obtained in step S502 can be used to construct the second graph data.

[0131] In some embodiments, step S503, using the encoding feature vectors of each of the multiple second protein amino acid residues as node features and determining edge features based on the contact probability of the multiple second protein amino acid residue pairs to construct the second graph data, may include: for each of the multiple second protein amino acid residue pairs, in response to determining that the contact probability of the second protein amino acid residue pair is greater than a preset threshold, establishing an edge connection between the second protein amino acid residue pairs, and setting the edge feature as the contact probability of the second protein amino acid residue pair.

[0132] By using the above method, we can avoid establishing edges between nodes corresponding to amino acid residues that have no contact relationship, thereby effectively reducing the number of edges, reducing the demand for computing and storage resources, and improving the inference efficiency of the second graph neural network.

[0133] In some exemplary embodiments, the preset threshold can be 0.1. When the contact probability between any two second protein amino acid base pairs is greater than 0.1, it can be considered that the two second protein amino acid residues are in contact and have an edge connection relationship. An edge connection is established between the two second protein amino acid residues, and the edge feature corresponding to the edge connection is set as the contact probability of the second protein amino acid residue pair. If the contact probability between the second protein amino acid base pairs is less than 0.1, it is considered that there is no contact and no edge connection relationship exists.

[0134] By employing the above methods, we can further simplify the model's complexity, reduce its computational burden, minimize unnecessary noise, and decrease the space required for storing graph representations, thereby further improving the accuracy and efficiency of protein stability prediction.

[0135] In some embodiments, in addition to the above methods, the second graph data can also be constructed in other ways, or the second protein amino acid sequence can be processed by the second initial graph neural network in other ways to obtain the second target feature vector characterizing the second protein amino acid sequence.

[0136] Figure 6 illustrates a flowchart of a process for obtaining a prediction result of the change in free energy by processing a first target feature vector and a second target feature vector using a prediction subnetwork according to an exemplary embodiment of this application. Process 600 can be used to implement step S1023 in the method 100 described above. In some embodiments, as shown in Figure 6, process 600 may include: step S601, fusing the first target feature vector and the second target feature vector to obtain a fused feature vector; and step S602, processing the fused feature vector using the fully connected layers and output layer included in the initial prediction subnetwork to obtain a prediction result of the change in free energy.

[0137] Therefore, by using the above method, the first target feature vector and the second target feature vector can be simplified into a fixed-length feature vector representing the target protein sequence, making subsequent processing more efficient. Furthermore, the first target feature vector and the second target feature vector, which represent different sequence sources, can be effectively fused, allowing the neural network to make full use of this information to obtain more accurate protein stability prediction results.

[0138] In some embodiments, in step S601, the first target feature vector obtained in steps S1021 and S1022 can be fused with the second target feature vector by concatenation, weighted summation, mean pooling, max pooling, attention mechanism, direct summation, processing with a multilayer perceptron or other neural network, or any combination of the above methods, to obtain the desired fused feature vector.

[0139] In an exemplary embodiment, in step S601, the first target feature vector and the second target feature vector can be merged using a residual network to obtain a fused feature vector, which serves as a comprehensive representation of the entire first protein amino acid sequence and the second protein amino acid sequence.

[0140] Therefore, by using the above method, the first target feature vector and the second target feature vector based on different protein information sources can be more effectively fused, enhancing the expressive power of the target feature vector obtained after fusion, thereby improving the performance of the neural network in protein stability prediction.

[0141] In some embodiments, in step S602, the target feature vector can be processed sequentially using a fully connected layer and an output layer to obtain a prediction result of the free energy change from the first protein amino acid model to the second protein amino acid model.

[0142] In an exemplary embodiment, in step S602, the target feature vector obtained in step S601 can be input into two fully connected layers, and the obtained feature vector can be sent into the output layer to obtain the final prediction result.

[0143] The training method for the neural network model disclosed herein can be used to train different neural network models that predict the impact of different mutation types on protein stability.

[0144] In some embodiments, the mutation type is a single amino acid substitution. For this mutation type, the following parameters can be set in the first and second initial graph neural networks: 2 graph convolutional layers, 32 graph convolutional output dimensions, 3 fully connected layers, and 64 fully connected output dimensions. By setting these parameters, the neural network model can have a stronger predictive ability for the "single amino acid substitution" mutation type.

[0145] In one exemplary embodiment, for a single amino acid substitution, the following parameters may be further employed in the first initial graph neural network and the second initial graph neural network, as well as during training: activation function 'relu', feature probability (feat_prob rate) 0.25, dropout rate 0.25, aggregation function 'pool', and learning rate 0.001.

[0146] In some embodiments, the mutation type is multiple amino acid substitutions. For this mutation type, the following parameters can be set in the first and second initial graph neural networks: 3 graph convolutional layers, 32 graph convolutional output dimensions, 3 fully connected layers, and 64 fully connected output dimensions. By setting these parameters, the neural network model can have a stronger predictive ability for the "multiple amino acid substitutions" mutation type.

[0147] In one exemplary embodiment, for multiple amino acid substitutions, the following parameters can be further set in the first initial graph neural network and the second initial graph neural network, as well as during training: activation function is 'relu', feature probability (feat_prob) is 0.25, dropout rate is 0.25, aggregation function is 'mean', and learning rate is 0.001.

[0148] In some embodiments, the mutation type is an insertion and / or deletion mutation. For this mutation type, the following parameters can be set in the first and second initial graph neural networks: 3 graph convolutional layers, 64 graph convolutional output dimensions, 2 fully connected layers, and 64 fully connected output dimensions. By setting the above parameters, the neural network model can have a stronger predictive ability for the "insertion and / or deletion mutation" mutation type.

[0149] In one exemplary embodiment, for insertion and / or deletion mutations, the following parameters may be further set in the first initial graph neural network and the second initial graph neural network, as well as during training: activation function is 'relu', feature probability (feat_prob) is 0.25, dropout rate is 0.25, aggregation function is 'mean', and learning rate is 0.0001.

[0150] By adjusting different model structure and training parameters for different types of mutations, more refined model optimization can be achieved, thereby improving the model's performance and accuracy in specific application scenarios. This approach ensures that the model starts learning from the state best suited to the characteristics of the data it processes, accelerates convergence, reduces training time, and helps to obtain neural network models that are more targeted to different mutation types.

[0151] Figure 7 illustrates a schematic diagram of a neural network model according to an exemplary embodiment of the present disclosure. As shown in Figure 7, the neural network 700 may include: a first initial graph neural network 720, a second initial graph neural network 730, and an initial prediction subnetwork 740. The first initial graph neural network 720 is configured to receive encoding feature vectors 702 for each of a plurality of first protein amino acid residues in a first protein amino acid sequence and contact probabilities 704 for pairs of first protein amino acid residues in the first protein amino acid sequence, and outputs a first target feature vector characterizing the first protein amino acid sequence. The second initial graph neural network 730 is configured to receive encoding feature vectors 708 for each of a plurality of second protein amino acid residues in a second protein amino acid sequence and contact probabilities 710 for pairs of second protein amino acid residues in a second protein amino acid sequence, and outputs a second target feature vector characterizing the second protein amino acid sequence. The prediction subnetwork 740 is configured to receive the first target feature vector and the second target feature vector, and output a prediction result 714 of the free energy change data.

[0152] In some embodiments, first graph data 706 can be constructed based on the encoding feature vectors 702 of each of the multiple first protein amino acid residues and the contact probabilities 704 of the multiple first protein amino acid residue pairs in the first protein amino acid sequence. The first initial graph neural network 720 may include multiple first graph convolutional layers 722, configured to iteratively update the node features in the first graph data and obtain the first intermediate feature vectors of each of the multiple first protein amino acid residues output by the last first graph convolutional layer 722. Furthermore, the first initial graph neural network 720 can fuse the first intermediate feature vectors of the multiple first protein amino acid residues to obtain a first target feature vector. In some embodiments, the number of first graph convolutional layers can be 2 or 3, depending on the mutation type. In some embodiments, other numbers of first graph convolutional layers can also be set as needed, and are not limited here.

[0153] In some embodiments, second graph data 712 can be constructed based on the encoding feature vectors 708 of each of the multiple second protein amino acid residues and the contact probabilities 710 of the multiple second protein amino acid residue pairs in the second protein amino acid sequence. The second initial graph neural network 730 may include multiple second graph convolutional layers 732, configured to iteratively update the node features in the second graph data and obtain the second intermediate feature vectors of each of the multiple second protein amino acid residues output by the last second graph convolutional layer 732. Furthermore, the second initial graph neural network 730 can fuse the second intermediate feature vectors of the multiple second protein amino acid residues to obtain a second target feature vector. In some embodiments, the number of layers configured for the second graph convolutional layers can be 2 or 3, depending on the mutation type. In some embodiments, other numbers of second graph convolutional layers can also be set as needed, and are not limited here.

[0154] In some embodiments, the prediction subnetwork 740 may include: a fusion module 742 configured to fuse a first target feature vector with a second target feature vector to obtain a comprehensive target feature vector; and a fully connected layer and an output layer 744 configured to process the comprehensive target feature vector to obtain a prediction result 714.

[0155] In this disclosure, the terms neural network, neural network model, deep learning model, model or other similar terms may be used interchangeably and, unless otherwise specified, may refer to the initial neural network model used in method 100 or the trained neural network model in method 800, which will be described below.

[0156] According to another aspect of this disclosure, a method for predicting the effect of mutations on protein stability using a neural network is provided. Figure 8 shows a flowchart of a method for predicting the effect of mutations on protein stability using a neural network according to an exemplary embodiment of this disclosure. The neural network model includes a trained graph neural network and a prediction sub-network, and the neural network model is obtained by training using method 100. Method 800 includes: step S801, obtaining the wild-type protein sequence before mutation as a first protein amino acid sequence and the sequence of the protein to be predicted as a second protein amino acid sequence; step S802, processing the first protein amino acid sequence using a graph neural network to obtain a first target feature vector characterizing the first protein amino acid sequence; step S803, processing the second protein amino acid sequence using a graph neural network to obtain a second target feature vector characterizing the second protein amino acid sequence; and step S804, inputting the first target feature vector and the second target feature vector into the prediction sub-network to obtain a prediction result for the change in free energy.

[0157] Therefore, within the framework of a graph neural network model, for the task of predicting the impact of mutations on protein stability, two isomorphic graph neural networks with identical parameters are used during training to process the first and second protein amino acid sequences in the sample, respectively. The parameters of the two graph neural networks are simultaneously adjusted during adjustment to ensure they remain identical. This approach enables the trained neural network to output accurate predictions of free energy changes, thereby accurately predicting protein stability before and after mutation.

[0158] It is understandable that the operations and effects of steps S801-S804 in method 800, namely the wild-type protein sequence before mutation of the protein to be predicted and the sequence of the protein to be predicted, can be referred to the descriptions of steps S101, S1021, S1022 and S1023 in method 100 above (only the first protein amino acid sequence in each training sample needs to be replaced with the wild-type protein sequence before mutation of the protein to be predicted, and the second protein amino acid sequence in each training sample needs to be replaced with the sequence of the protein to be predicted), and will not be elaborated here.

[0159] Furthermore, the order of steps S802 and S803 can be interchanged or performed simultaneously.

[0160] In some embodiments, a neural network model specific to a particular mutation type can be selected for prediction based on that mutation type.

[0161] When the mutation type of the protein to be predicted is a single amino acid substitution (single-point mutation), a neural network model specifically designed for this mutation type can be used. In this neural network model, the number of graph convolutional layers can be 2, the graph convolutional output dimension can be 32, the number of fully connected layers can be 3, and the fully connected output dimension can be 64. After training, the neural network model with these parameter settings has a stronger predictive ability for the "single amino acid substitution" mutation type.

[0162] When the mutation type of the protein to be predicted is multiple amino acid substitutions (multi-point mutation), a neural network model specifically designed for this mutation type can be used. In this neural network model, the number of graph convolutional layers can be 3, the graph convolutional output dimension can be 32, the number of fully connected layers can be 3, and the fully connected output dimension can be 64. After training, the neural network model with these parameter settings has a stronger predictive ability for the "multi-amino acid substitution" mutation type.

[0163] When the mutation type of the protein to be predicted is insertion and / or deletion mutation, a neural network model specifically designed for this mutation type can be used. In this neural network model, the number of graph convolutional layers can be 3, the graph convolutional output dimension can be 64, the number of fully connected layers can be 2, and the fully connected output dimension can be 64. After training, the neural network model with the above parameter settings has a stronger predictive ability for the "insertion and / or deletion mutation" mutation type.

[0164] After training the initial neural network model using the training method proposed above, the parameters of the first initial graph neural network and the second initial graph neural network are exactly the same. Therefore, in the inference and prediction stage, only one of them needs to be used as a graph neural network to process the first protein amino acid sequence and the second protein amino acid sequence to obtain the first target feature vector representing the first protein amino acid sequence and the second target feature vector representing the second protein amino acid sequence, respectively.

[0165] In some embodiments, in order to test the generalization performance of the trained neural network, the independent test set shown in Table 1 can be selected to validate the model, and data with a protein amino acid length greater than 1000 in all test sets can be deleted.

[0166] Table 1 Summary of Test Sets

[0167] As shown in Table 1, the test sets can be divided into four independent sets based on the type of mutation. These test sets are as follows:

[0168] 1. Single-point mutation test data set: including S669 (Pancotti, C.; Benevenuta, S.; Birolo, G.; Alberini, V.; Repetto, V.; Sanavia, T.; Capriotti, E.; Fariselli, P. Predicting Protein Stability Changes upon Single-Point Mutation: A Thorough Comparison of the Available Tools on a New Dataset.Brief.Bioinform.2022,23(2),bbab555.https: / / doi.org / 10.1093 / bib / bbab555.) and S96 (Montanucci, L.; Capriotti, E.; Birolo, G.; Benevenuta, S.; Pancotti, C.; Lal, D.; Fariselli, P.DDGun: An Untrained Predictor of Protein Stability Changes upon Amino Acid Variants.Nucleic Acids Res.2022,50(W1),W222–W227.https: / / doi.org / 10.1093 / nar / gkac325.), by removing data from each test dataset that have a protein amino acid length greater than 1000 and a similarity to the training sequence greater than 25%, we obtained S462 and S72 respectively.

[0169] 2. Multi-point mutation test set: including M28 (Montanucci, L.; Capriotti, E.; Birolo, G.; Benevenuta, S.; Pancotti, C.; Lal, D.; Fariselli, P. DDGun: An Untrained Predictor of Protein Stability Changes upon Amino Acid Variants. Nucleic Acids Res. 2022, 50(W1), W222–W227. https: / / doi.org / 10.1093 / nar / gkac325.) and M218 (from the databases ProThermDB (https: / / web.iitm.ac.in / bioinfo2 / prothermdb / index.html) and ThermoMutDB (http: / / biosig.unimelb.edu.au / thermomutdb / )). The sequence similarity of these two datasets to the training set is no greater than 25%.

[0170] 3. Mutant Protein to Mutation Test Dataset: Mutated proteins with the same Uniprot number in the M38 dataset are randomly paired to generate a new dataset containing 316 mutation data points from 6 proteins. Specifically, this includes 98 single-point mutation data points and 218 multi-point mutation data points, denoted as S98 and M218, respectively.

[0171] 4. Insertion / Deletion Mutation Dataset: A new dataset was acquired as an independent test set, including 12 insertions / deletions from P04925, 22 from P00720, and 4 from P00644.

[0172] To better verify the accuracy of protein stability data, the following performance metrics were selected to validate the predictive performance of the neural network model for protein stability:

[0173] The Pearson product-moment correlation coefficient is a commonly used indicator to measure the linear correlation between two sets of values. The closer it is to 1, the stronger the linear correlation between the two sets of values. A value of 0 indicates that there is no linear correlation between the two sets of values, and a negative value indicates that there is a negative correlation between the two sets of values. In this experiment, this indicator is denoted as R.

[0174] RMSE: Root Mean Square Error, used to measure the deviation between two sets of variables. The larger the value, the greater the difference between the predicted value and the true value.

[0175] To better demonstrate the predictive accuracy of the proposed neural network model for predicting protein stability, the following experimental results compare the proposed method with other protein stability prediction methods. It is important to note that this experiment proposes a general stability model framework capable of accurate predictions across various mutation tasks, and further proposes three sub-models for different tasks: Single_model, Multi_model, and Indel_model, corresponding to single-point mutation, multi-point mutation, and insertion / deletion mutation tasks, respectively. Specifically, the S98 mutation to single-point mutation data will be tested in the single-point model, and the M218 mutation to multi-point mutation data will be tested in the multi-point model. To evaluate the differences in results between the proposed method and other prediction methods on the single-point mutation dataset, the performance of the proposed method is compared with existing methods on test data, including INPS, ACDC-NN, PremPS, DDGun, ThermoNet, DynaMut, PoPMuSiC, DUET, mCSM, I-Mutant3.0, FoldX, SDM, SAAFEC-SEQ, DynaMut2, MUpro, and MAESTRO. The above methods are known in the art and can be found in the literature: Savojardo, C.; Fariselli, P.; Martelli, PL; Casadio, R. INPS-MD: A Web Server to Predict Stability of Protein Variants from Sequence and Structure. Bioinformatics 2016,32(16),2542–2544;Pancotti,C.;Benevenuta,S.;Repetto,V.;Birolo,G.;Capriotti,E.;Sanavia,T.;Fariselli,PA Deep-Learning Sequence-Based Method to Predict Protein Stability Changes Upon Genetic Variations.Genes 2021,12(6),911;Chen,Y.;Lu,H.;Zhang,N.;Zhu,Z.;Wang,S.;Li,M. PremPS: Predicting the Impact of Missense Mutations on Protein Stability.PLoS Comput.Biol.2020,16(12),e1008543;Montanucci,L.;Capriotti,E.;Birolo,G.;Benevenuta,S.;Pancotti,C.;Lal,D.;Fariselli,P.DDGun:An Untrained Predictor of Protein Stability Changes upon Amino Acid Variants.Nucleic Acids Res.2022,50(W1),W222–W227;Li,B.;Yang,Y.T.;Capra,J.A.;Gerstein,M.B.Predicting Changes in Protein Thermodynamic Stability upon Point Mutation with Deep 3D Convolutional Neural Networks.PLOS Comput.Biol.2020,16(11),e1008291;Rodrigues,C.H.;Pires,D.E.;Ascher,D.B.DynaMut:Predicting the Impact of Mutations on Protein Conformation,Flexibility and Stability.Nucleic Acids Res.2018,46(W1),W350–W355;Dehouck,Y.;Kwasigroch,J.M.;Gilis,D.;Rooman,M.PoPMuSiC 2.1:A Web Server for the Estimation of Protein Stability Changes upon Mutation and Sequence Optimality.BMC Bioinformatics 2011,12(1),151;Pires,D.E.V.;Ascher,D.B.;Blundell,T.L.DUET:A Server for Predicting Effects of Mutations on Protein Stability Using an Integrated Computational Approach.Nucleic Acids Res.2014,42(W1),W314–W319;Pires,D.E.V.;Ascher,D.B.;Blundell,T.L.mCSM:Predicting the Effects of Mutations in Proteins Using Graph-Based Signatures.Bioinformatics 2014,30(3),335–342;Capriotti,E.;Fariselli,P.;Rossi,I.;Casadio,R.A Three-State Prediction of Single Point Mutations on Protein Stability Changes.BMC Bioinformatics 2008,9(S2),S6;Schymkowitz,J.;Borg,J.;Stricher,F.;Nys,R.;Rousseau,F.;Serrano,L.The FoldX Web Server:An Online Force Field.Nucleic Acids Res.2005,33(Web Server),W382–W388;Worth,C.L.;Preissner,R.;Blundell,T.L.SDM--a Server for Predicting Effects of Mutations on Protein Stability and Malfunction.Nucleic Acids Res.2011,39(suppl),W215–W222;Li,G.;Panday,S.K.;Alexov,E.SAAFEC-SEQ:A Sequence-Based Method for Predicting the Effect of Single Point Mutations on Protein Thermodynamic Stability.Int.J.Mol.Sci.2021,22(2);Rodrigues,C.H.M.;Pires,D.E.V.;Ascher,D.B.DynaMut2:Assessing Changes in Stability and Flexibility upon Single and Multiple Point Missense Mutations.Protein Sci.2021,30(1),60–69;Cheng,J.;Randall,A.;Baldi,P.Prediction of Protein Stability Changes for Single-site Mutations Using Support Vector Machines.Proteins Struct.Funct.Bioinforma.2006,62(4),1125–1132;Laimer,J.;Hofer,H.;Fritz,M.;Wegenkittl,S.;Lackner,P.MAESTRO-Multi Agent Stability Prediction upon Point Mutations.BMC Bioinformatics 2015,16(1),116;Montanucci,L.;Capriotti,E.;Birolo,G.;Benevenuta,S.;Pancotti,C.;Lal,D.;Fariselli,DDGun: An Untrained Predictor of Protein Stability Changes upon Amino Acid Variants.Nucleic Acids Res.2022,50(W1),W222–W227;Li,G.;Panday,SK;Alexov,E.SAAFEC-SEQ:A Sequence-Based Method for Predicting the Effect of Single Point Mutations on Protein Thermodynamic Stability.Int.J.Mol.Sci.2021,22(2),E606. .

[0176] The following are the test results for different mutation types:

[0177] 1. Performance comparison of single-point mutation stability prediction results:

[0178] Table 2 compares the test results of different existing models and the model disclosed herein on the single point mutation dataset S462.

[0179] Table 3 compares the test results of different existing models and the model disclosed herein on the single point mutation dataset S72.

[0180] Table 4 compares the test results of different existing models and the model disclosed herein on the single-point mutation to mutation dataset S98.

[0181] As can be seen from Tables 2, 3, and 4, the Single_model model performs significantly better than other models in the three sets of single-point mutation test datasets, demonstrating good generalization performance.

[0182] 2. Performance comparison of multi-point mutation stability prediction results

[0183] To demonstrate the model's performance on multipoint mutation datasets, tests were conducted on three multipoint mutation test sets: M28, M38, and M218, and comparisons were made with other methods, including DDGun, DDGun3D, MAESTRO, and FoldX.

[0184] Table 5 compares the test results of different existing models with the model disclosed herein on the multipoint mutation dataset M28.

[0185] Table 6 compares the test results of different existing models with the model disclosed herein on the multipoint mutation dataset M38.

[0186] Table 7 compares the test performance of different existing models on the multipoint mutation to mutation dataset M218.

[0187] As shown in Tables 5, 6, and 7, the Multi_model model significantly outperforms other models in the test results across the three multi-point mutation datasets. In previous tests, the predicted starting sequence was typically the wild-type. However, in actual industrial production, it is highly likely that further mutations will be added to an already mutated protein. This means that the starting sequence is a protein sequence that already contains a mutation, necessitating the prediction of the stability of the protein sequence undergoing further mutation. This method validates the performance of predicting the stability of superimposed mutated proteins, and the results remain stable.

[0188] 3. Performance comparison of insertion / deletion mutation stability prediction results

[0189] To demonstrate the model's performance in inserting missing data, tests were conducted on three proteins: P04925, P00720, and P00644, and the results were compared with the FoldX method.

[0190] Table 8 compares the results of the three protein insertion / deletion datasets, P04925, P00720, and P00644 (Indel represents the set of three protein datasets).

[0191] As can be seen from Table 8, the Indel_model model performed well in the prediction results of the three protein insertion and deletion datasets, and also showed good performance in the merged dataset.

[0192] The above comparison results further demonstrate that, within the framework of a graph neural network model, for the task of predicting the impact of mutations on protein stability, two isomorphic graph neural networks with identical parameters are used during training to process the first and second protein amino acid sequences in the sample, respectively. The parameters of the two graph neural networks are simultaneously adjusted during adjustment to ensure they remain identical. This approach enables the trained neural network model to output accurate predictions of free energy changes, thereby accurately predicting protein stability before and after mutation.

[0193] According to another aspect of this disclosure, a training apparatus for a neural network model predicting the effect of mutations on protein stability is provided. FIG9 shows a structural block diagram of the training apparatus for a neural network model predicting the effect of mutations on protein stability according to an exemplary embodiment of this disclosure. The initial neural network model includes a first initial graph neural network and a second initial graph neural network that are isomorphic and have the same initial parameters, and includes an initial prediction subnetwork. As shown in FIG9, the apparatus 900 includes: a sample construction unit 910 configured to construct multiple sets of training samples based on the wild-type protein sequence before mutation, the mutant protein sequence after mutation, and the free energy change data ΔΔG, each set of training samples including a first protein amino acid sequence, a second protein amino acid sequence, and a truth label of the free energy change; and a training unit 920 configured to train the initial neural network model using multiple sets of training samples. The training unit 920 includes: a first processing subunit 921 configured to process the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector characterizing the first protein amino acid sequence; a second processing subunit 922 configured to... The method is configured to process the second protein amino acid sequence using a second initial graph neural network to obtain a second target feature vector characterizing the second protein amino acid sequence; a prediction sub-unit 923 is configured to input the first target feature vector and the second target feature vector into the initial prediction sub-network to obtain a prediction result of the change in free energy; and a parameter tuning sub-unit 924 is configured to adjust the parameters of the initial neural network based on the prediction result and the truth label, wherein, in response to determining that at least one parameter of the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously so that the adjusted parameters of the first initial graph neural network and the second initial graph neural network are the same.

[0194] It is understood that the operation of units 910-920 in device 900 and sub-units 921-924 in unit 920 can refer to the description of steps S101-S102 and steps S1021-S1024 in method 100 above, and will not be repeated here.

[0195] According to another aspect of this disclosure, an apparatus for predicting the effect of mutations on protein stability using a neural network model is provided. Figure 10 shows a structural block diagram of a neural network model apparatus according to an exemplary embodiment of this disclosure. The neural network model includes a trained graph neural network and a prediction subnetwork, and the neural network model is trained using an apparatus 900. As shown in Figure 10, the apparatus 1000 includes: an acquisition unit 1010 configured to acquire the wild-type protein sequence before mutation of the protein to be predicted as a first protein amino acid sequence and the sequence of the protein to be predicted as a second protein amino acid sequence; a first processing unit 1020 configured to process the first protein amino acid sequence using the trained graph neural network to obtain a first target feature vector characterizing the first protein amino acid sequence; a second processing unit 1030 configured to process the second protein amino acid sequence using the trained graph neural network to obtain a second target feature vector characterizing the second protein amino acid sequence; and a prediction unit 1040 configured to input the first target feature vector and the second target feature vector into the prediction subnetwork to obtain a prediction result for the change in free energy.

[0196] It is understood that the operation of units 1010-1040 in device 1000 can refer to the description of steps S801-S804 in method 800 above, and will not be repeated here.

[0197] According to one aspect of this disclosure, a computer device is provided, including a memory, a processor, and a computer program stored in the memory. The processor is configured to execute the computer program to implement the steps of any of the method embodiments described above.

[0198] According to one aspect of this disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the method embodiments described above.

[0199] According to one aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of any of the method embodiments described above.

[0200] The following description, in conjunction with Figure 11, provides illustrative examples of such computer devices, non-transitory computer-readable storage media, and computer program products.

[0201] Figure 11 shows an example configuration of computer device 1100 that can be used to implement the methods described herein. The aforementioned devices 900 and 1000 can also be implemented wholly or at least partially by computer device 1100 or similar devices or systems.

[0202] Computer device 1100 can be a variety of different types of devices. Examples of computer device 1100 include, but are not limited to: desktop computers, server computers, laptop or netbook computers, mobile devices (e.g., tablets, cellular or other wireless phones (e.g., smartphones), notebook computers, mobile stations), wearable devices (e.g., glasses, watches), entertainment devices (e.g., entertainment appliances, set-top boxes communicatively coupled to a display device, game consoles), televisions or other display devices, automotive computers, and so on.

[0203] Computer device 1100 may include at least one processor 1102, memory 1104, multiple communication interfaces 1106, display device 1108, other input / output (I / O) devices 1110, and one or more mass storage devices 1112, capable of communicating with each other, such as via system bus 1114 or other suitable connections.

[0204] Processor 1102 may be a single processing unit or multiple processing units, and all processing units may include single or multiple computing units or multiple cores. Processor 1102 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operating instructions. Among other capabilities, processor 1102 may be configured to acquire and execute computer-readable instructions stored in memory 1104, mass storage device 1112, or other computer-readable media, such as program code of operating system 1116, program code of application program 1118, program code of other program 1120, etc.

[0205] Memory 1104 and mass storage device 1112 are examples of computer-readable storage media for storing instructions executed by processor 1102 to perform the various functions described above. For example, memory 1104 may generally include both volatile and non-volatile memory (e.g., RAM, ROM, etc.). Furthermore, mass storage device 1112 may generally include hard disk drives, solid-state drives, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network-attached storage, storage area networks, etc. Both memory 1104 and mass storage device 1112 may be collectively referred to herein as memory or computer-readable storage media, and may be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code, which may be executed by processor 1102 as a specific machine configured to perform the operations and functions described in the examples herein.

[0206] Multiple programs can be stored on mass storage device 1112. These programs include operating system 1116, one or more applications 1118, other programs 1120, and program data 1122, and they can be loaded into memory 1104 for execution.

[0207] Although illustrated in Figure 11 as being stored in memory 1104 of computer device 1100, modules 1116, 1118, 1120, and 1122, or portions thereof, may be implemented using any form of computer-readable medium accessible by computer device 1100. As used herein, “computer-readable medium” includes at least two types of computer-readable media: computer-readable storage media and communication media.

[0208] Computer-readable storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, DVD, or other optical storage devices, magnetic cassettes, magnetic tapes, disk storage devices or other magnetic storage devices, or any other non-transmission medium that can be used to store information for access by a computer device. In contrast, communication media can embody computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms. Computer-readable storage media as defined herein do not include communication media.

[0209] One or more communication interfaces 1106 are used for exchanging data with other devices, such as via a network, direct connection, etc. Such communication interfaces can be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), wired or wireless (such as IEEE 802.11 Wireless LAN (WLAN)) interface, Wi-MAX interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth. TM Interfaces, near field communication (NFC) interfaces, etc. Communication interface 1106 can facilitate communication across various network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, etc. Communication interface 1106 can also provide communication with external storage devices (not shown) such as storage arrays, network-attached storage, storage area networks, etc.

[0210] In some examples, a display device 1108, such as a monitor, may be included for displaying information and images to the user. Other I / O devices 1110 may be devices that receive various inputs from the user and provide various outputs to the user, and may include touch input devices, gesture input devices, cameras, keyboards, remote controls, mice, printers, audio input / output devices, and so on.

[0211] The technologies described herein can be supported by these various configurations of computer device 1100, and are not limited to specific examples of the technologies described herein. For example, the functionality can also be implemented wholly or partially on a “cloud” using a distributed system. A cloud includes and / or represents a platform for resources. The platform abstracts the underlying functionality of the cloud’s hardware (e.g., servers) and software resources. Resources may include applications and / or data that can be used when performing computational processing on a server remote from computer device 1100. Resources may also include services provided via the Internet and / or via subscriber networks such as cellular or Wi-Fi networks. The platform can abstract resources and functionality to connect computer device 1100 to other computer devices. Therefore, the implementation of the functionality described herein can be distributed throughout the cloud. For example, the functionality can be implemented partly on computer device 1100 and partly through a platform that abstracts the functionality of the cloud.

[0212] Although this disclosure has been described and illustrated in detail in the accompanying drawings and the foregoing description, such description and illustration should be considered illustrative and suggestive, not restrictive; this disclosure is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the appended claims, those skilled in the art will be able to understand and implement variations of the disclosed embodiments in practice with respect to the claimed subject matter. In the claims, the word "comprising" does not exclude other elements or steps not listed, the indefinite article "a" or "an" does not exclude a plurality, the term "a plurality" means two or more, and the term "based on" should be interpreted as "at least partially based on". The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be beneficial.

Claims

1. A method for training a neural network model for predicting the effect of a mutation on protein stability, wherein, The initial neural network model includes a first initial graph neural network and a second initial graph neural network that are isomorphic and have the same initial parameters, and includes an initial predictor network, and the method includes: Based on the wild-type protein sequence before mutation, the mutant protein sequence after mutation and the free energy change data ΔΔG, a plurality of sets of training samples are constructed, each set of the training samples including a first protein amino acid sequence, a second protein amino acid sequence, and a true value label of free energy change; and Model training is performed on the initial neural network model using the plurality of sets of training samples, including: processing the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; processing the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; inputting the first target feature vector and the second target feature vector into the initial predictor network to obtain a prediction result of the free energy change; and adjusting parameters of the initial neural network model based on the prediction result and the true value label, wherein in response to determining that at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously, so that the parameters of the adjusted first initial graph neural network and the second initial graph neural network are the same.

2. The method of claim 1, wherein, The first protein amino acid sequence includes a plurality of first protein amino acid residues, and processing the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence includes: preprocessing the first protein amino acid sequence to construct first graph data including node features and edge features; updating the node features in the first graph data based on the node features and the edge features in the first graph data using the first initial graph neural network to obtain a first intermediate feature vector of each of the plurality of first protein amino acid residues; and fusing the first intermediate feature vectors of the plurality of first protein amino acid residues to obtain the first target feature vector.

3. The method of claim 2, wherein, The preprocessing of the first protein amino acid sequence to construct first graph data including node features and edge features includes: encoding the first protein amino acid sequence using a protein pre-training model to obtain an encoding feature vector of each of the plurality of first protein amino acid residues; determining contact probabilities of a plurality of first protein amino acid residue pairs in the first protein amino acid sequence, the plurality of first protein amino acid residue pairs including first protein amino acid residue pairs formed by any two first protein amino acid residue matching combinations in the first protein amino acid sequence; and using the encoding feature vector of each of the plurality of first protein amino acid residues as a node feature, and determining an edge feature based on the contact probabilities of the plurality of first protein amino acid residue pairs to construct the first graph data.

4. The method of claim 3, wherein, The method further includes: constructing, by the processor, second graph data comprising node features and edge features, wherein the node features are the encoding feature vectors of the second protein amino acid residues, and the edge features are determined based on the contact probabilities of the second protein amino acid residue pairs; and updating, by the processor, the node features in the second graph data based on the node features and the edge features in the second graph data, to obtain second intermediate feature vectors of the second protein amino acid residues. The method further includes: for each of the first protein amino acid residue pairs, in response to determining that the contact probability of the first protein amino acid residue pair is greater than the preset threshold, establishing an edge connection between the first protein amino acid residue pair, and setting the edge feature as the contact probability of the first protein amino acid residue pair.

5. The method of claim 1, wherein, The second protein amino acid sequence comprises a plurality of second protein amino acid residues, and the method further includes: processing the second protein amino acid sequence by using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence. The method further includes: preprocessing the second protein amino acid sequence to construct second graph data comprising node features and edge features. The method further includes: updating, by the processor, the node features in the second graph data based on the node features and the edge features in the second graph data, to obtain second intermediate feature vectors of the second protein amino acid residues; and fusing the second intermediate feature vectors of the second protein amino acid residues to obtain the second target feature vector. The method further includes: preprocessing the second protein amino acid sequence to construct second graph data comprising node features and edge features.

6. The method of claim 5, wherein, The method further includes: encoding the second protein amino acid sequence by using a protein pre-training model to obtain encoding feature vectors of the second protein amino acid residues; determining contact probabilities of second protein amino acid residue pairs in the second protein amino acid sequence, the second protein amino acid residue pairs comprising any two second protein amino acid residue combinations in the second protein amino acid sequence; and taking the encoding feature vectors of the second protein amino acid residues as the node features, and determining the edge features based on the contact probabilities of the second protein amino acid residue pairs to construct the second graph data. The method further includes: for each of the second protein amino acid residue pairs, in response to determining that the contact probability of the second protein amino acid residue pair is greater than the preset threshold, establishing an edge connection between the second protein amino acid residue pair, and setting the edge feature as the contact probability of the second protein amino acid residue pair. The protein pre-training model is a ProtT5-XL-UniRef50 model. The method further includes: inputting the first target feature vector and the second target feature vector into the initial prediction sub-network to obtain a prediction result of the free energy change, wherein the prediction result of the free energy change is used as the prediction result of the protein-protein interaction. The method further includes: fusing the first target feature vector and the second target feature vector to obtain a fused feature vector; and inputting the fused feature vector into the initial prediction sub-network to obtain the prediction result of the protein-protein interaction. ​ 8. The method of claim 3 or 6, wherein, ​ 9. The method of any one of claims 1-7, wherein, ​ ​ processing the fusion feature vector by using a fully connected layer and an output layer comprised in the initial predictor network to obtain a prediction result of the free energy change.

10. The method of claim 9, wherein, The type of the mutation is single amino acid substitution, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 2, the graph convolution output dimension is 32, the number of fully connected layers is 3, and the fully connected output dimension is 64.

11. The method of claim 9, wherein, The type of the mutation is multiple amino acid substitution, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 3, the graph convolution output dimension is 32, the number of fully connected layers is 3, and the fully connected output dimension is 64.

12. The method of claim 9, wherein, The type of the mutation is insertion and / or deletion mutation, and in the first initial graph neural network and the second initial graph neural network, the number of graph convolution layers is 3, the graph convolution output dimension is 64, the number of fully connected layers is 2, and the fully connected output dimension is 64.

13. The method of any one of claims 1-7, wherein, The free energy change data ΔΔG = (ΔG1-ΔG2), ΔG1 is the free energy change of the wild type protein sequence before mutation, and ΔG2 is the free energy change of the mutant protein sequence after mutation.

14. The method of any one of claims 1-7, wherein, The first initial graph neural network and the second initial graph neural network are based on SAGEConv.

15. The method of any one of claims 1-7, wherein, Based on the wild type protein sequence before mutation, the mutant protein sequence after mutation, and the free energy change data ΔΔG, a plurality of training samples are constructed, including: a first training sample is constructed by taking the wild type protein sequence before mutation as the first protein amino acid sequence, taking the mutant protein sequence after mutation as the second protein amino acid sequence, and taking the free energy change data as the free energy change true value label; and a second training sample is constructed by taking the wild type protein sequence before mutation as the second protein amino acid sequence, taking the mutant protein sequence after mutation as the first protein amino acid sequence, and taking the opposite number of the free energy change data as the free energy change true value label.

16. A method of predicting the effect of a mutation on protein stability using a neural network model, wherein, The neural network model includes a trained graph neural network and a predictor network, and the neural network model is trained by the method according to any one of claims 1-15, the method comprising: obtaining a wild type protein sequence before mutation of a to-be-predicted protein as a first protein amino acid sequence and a sequence of the to-be-predicted protein as a second protein amino acid sequence; processing the first protein amino acid sequence by using the trained graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; processing the second protein amino acid sequence by using the trained graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; and inputting the first target feature vector and the second target feature vector into the predictor network to obtain a prediction result of the free energy change.

17. An apparatus for training a neural network model that predicts the effect of a mutation on protein stability, wherein, The initial neural network model includes a first initial graph neural network and a second initial graph neural network which are isomorphic and have the same initial parameters, and includes an initial predictor network, and the device comprises: a sample construction unit configured to construct a plurality of groups of training samples based on a wild-type protein sequence before mutation, a mutant protein sequence after mutation, and free energy change data ΔΔG, each group of the training samples comprising a first protein amino acid sequence, a second protein amino acid sequence, and a true value label of free energy change; and a training unit configured to perform model training on an initial neural network model using the plurality of groups of training samples, the training unit comprising: a first processing sub-unit configured to process the first protein amino acid sequence using the first initial graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; a second processing sub-unit configured to process the second protein amino acid sequence using the second initial graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; a prediction sub-unit configured to input the first target feature vector and the second target feature vector into the initial prediction sub-network to obtain a prediction result of free energy change; and a parameter adjustment sub-unit configured to adjust parameters of the initial neural network model based on the prediction result and the true value label, wherein in response to determining that at least one parameter in the first initial graph neural network and / or the second initial graph neural network needs to be adjusted, the corresponding parameter in the other initial graph neural network is adjusted synchronously, so that the parameters of the adjusted first initial graph neural network and the second initial graph neural network are the same.

18. An apparatus for predicting the effect of a mutation on protein stability using a neural network model, wherein, The neural network model comprises a trained graph neural network and a prediction sub-network, and the neural network model is trained by the apparatus according to claim 17, the apparatus comprising: an acquisition unit configured to acquire a wild-type protein sequence before mutation of a to-be-predicted protein as a first protein amino acid sequence and a sequence of the to-be-predicted protein as a second protein amino acid sequence; a first processing unit configured to process the first protein amino acid sequence using the trained graph neural network to obtain a first target feature vector representing the first protein amino acid sequence; a second processing unit configured to process the second protein amino acid sequence using the trained graph neural network to obtain a second target feature vector representing the second protein amino acid sequence; and a prediction unit configured to input the first target feature vector and the second target feature vector into the prediction sub-network to obtain a prediction result of free energy change.

19. An electronic device comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-16.

20. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1-16.

21. A computer program product comprising a computer program, wherein, The computer program, when executed by a processor, implements the method of any one of claims 1-16. The computer program, when executed by a processor, implements the method of any one of claims 1-16.

22. A neural network model trained according to the method of any one of claims 1-15.

Images