Protein stability prediction method and device, computer equipment and storage medium

Abstract

Embodiments of the present application provide a protein stability prediction method and device, computer equipment and a storage medium. The method comprises: obtaining a first to-be-tested sequence corresponding to an original protein and a second to-be-tested sequence corresponding to a mutant protein; determining a plurality of first target residue features from the first to-be-tested sequence according to the attention mechanism of the protein language model and extracting the global features thereof to obtain first virtual structure microenvironment features; similarly, obtaining second virtual structure microenvironment features corresponding to a plurality of second target residue features; determining a first difference based on the first virtual structure microenvironment features and the second virtual structure microenvironment features, and determining a second difference based on a first global sequence feature corresponding to the first to-be-tested sequence and a second global sequence feature corresponding to the second to-be-tested sequence; and predicting based on the first difference and the second difference to obtain a protein stability prediction result. In this way, the accuracy of predicting the stability of the protein can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information, and particularly relates to a protein stability prediction method and device, a computer device and a storage medium. BACKGROUND

[0002] Protein stability refers to the ability of a protein to maintain its three-dimensional structure under specific conditions. Protein stability can provide important insights into protein folding and function, and is an important part of protein evolutionary fitness. However, the substitution of a single or multiple amino acids in a protein sequence can affect the structure of the protein, and in turn affect the overall structural stability and function of the protein. For example, certain amino acids such as proline and glycine can significantly change the conformation and stability of the protein due to their special chemical properties.

[0003] In related technologies, sequence methods are generally used to predict the stability of proteins. However, sequence methods can only roughly capture the overall structure of a protein for analysis, resulting in inaccurate predictions of protein stability. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a protein stability prediction method, device, computer device and storage medium, which can improve the accuracy of predicting the stability of proteins.

[0005] To achieve the above purpose, a first aspect of the embodiments of the present application provides a protein stability prediction method, which comprises:

[0006] obtaining a first to-be-tested sequence corresponding to an original protein, and a second to-be-tested sequence of a mutant protein corresponding to the original protein;

[0007] determining a plurality of first target residue features from the first to-be-tested sequence according to an attention mechanism of a pre-trained protein language model, and extracting global features of the plurality of first target residue features to obtain first virtual structure microenvironment features;

[0008] determining a plurality of second target residue features from the second to-be-tested sequence according to the attention mechanism of the protein language model, and extracting global features of the plurality of second target residue features to obtain second virtual structure microenvironment features;

[0009] obtaining a first global sequence feature of the first to-be-tested sequence and a second global sequence feature of the second to-be-tested sequence through the protein language model;

[0010] determine a first difference value based on the first virtual structure microenvironment feature and the second virtual structure microenvironment feature, and determine a second difference value based on the first global sequence feature and the second global sequence feature;

[0011] perform protein stability prediction on the mutant protein according to the first difference value and the second difference value, to obtain a protein stability prediction result.

[0012] Correspondingly, a second aspect of the embodiment of the present application provides a protein stability prediction device, the device comprising:

[0013] a first obtaining module, configured to obtain a first to-be-predicted sequence corresponding to an original protein, and a second to-be-predicted sequence of a mutant protein corresponding to the original protein;

[0014] a first extracting module, configured to determine a plurality of first target residue features from the first to-be-predicted sequence according to an attention mechanism of a pre-trained protein language model, and extract global features from the plurality of first target residue features, to obtain a first virtual structure microenvironment feature;

[0015] a second extracting module, configured to determine a plurality of second target residue features from the second to-be-predicted sequence according to the attention mechanism of the protein language model, and extract global features from the plurality of second target residue features, to obtain a second virtual structure microenvironment feature;

[0016] a second obtaining module, configured to obtain a first global sequence feature of the first to-be-predicted sequence and a second global sequence feature of the second to-be-predicted sequence through the protein language model;

[0017] a determining module, configured to determine a first difference value based on the first virtual structure microenvironment feature and the second virtual structure microenvironment feature, and determine a second difference value based on the first global sequence feature and the second global sequence feature;

[0018] a prediction module, configured to perform protein stability prediction on the mutant protein according to the first difference value and the second difference value, to obtain a protein stability prediction result.

[0019] In some embodiments, the prediction module is further configured to:

[0020] input the first difference value corresponding to the first target residue feature and the second target residue feature into a pre-trained prediction model to perform prediction of protein folding free energy in a residue feature change dimension, to obtain a first prediction result;

[0021] inputting the second difference value corresponding to the first to-be-sequenced sequence and the second to-be-sequenced sequence into the prediction model to perform prediction of the global change dimension protein folding free energy, to obtain a second prediction result;

[0022] obtaining a protein stability prediction result based on the mean of the first prediction result and the second prediction result.

[0023] In some embodiments, the first extraction module is further configured to:

[0024] generating a first overall attention map corresponding to the first to-be-sequenced sequence according to the attention mechanism of the pre-trained protein language model;

[0025] performing global pooling processing on the first overall attention map to obtain a first target overall attention map;

[0026] determining a plurality of first positions of a plurality of mutant residue features from the first to-be-sequenced sequence;

[0027] based on matching the plurality of first positions in the first target overall attention map, obtaining a plurality of first attention score vectors corresponding to the plurality of mutant residue features;

[0028] obtaining a first residue feature selection number, and determining a same number of first target residue features as the first residue feature selection number from the plurality of mutant residue features based on the plurality of first attention score vectors.

[0029] In some embodiments, the first extraction module is further configured to:

[0030] obtaining a pre-set first residue feature initial vector, and splicing the first residue feature initial vector and the plurality of first target residue features to obtain a first target feature sequence; wherein the first residue feature initial vector is located at the first position of the first target feature sequence;

[0031] based on the first residue feature initial vector, aggregating the corresponding each first target residue feature in the first target feature sequence and the mutual relationship between the plurality of first target residue features to obtain a first virtual structure microenvironment feature for representing global information of the first target feature sequence.

[0032] In some embodiments, the second extraction module is further configured to:

[0033] generating a second overall attention map corresponding to the second to-be-sequenced sequence according to the attention mechanism of the protein language model;

[0034] performing global pooling processing on the second overall attention map to obtain a second target overall attention map;

[0035] determining a plurality of second positions of a plurality of mutation residue features from the second to-be-sequenced sequence;

[0036] performing matching on the second target global attention map based on the plurality of second positions, to obtain a plurality of second attention score vectors corresponding to the plurality of mutation residue features;

[0037] obtaining a second residue feature selection quantity, and determining, based on the plurality of second attention score vectors, a same number of second target residue features as the second residue feature selection quantity from the plurality of mutation residue features.

[0038] In some embodiments, the second extraction module is further configured to:

[0039] obtaining a preset second residue feature initial vector, and concatenating the second residue feature initial vector with the plurality of second target residue features to obtain a second target feature sequence; wherein the second residue feature initial vector is located at the first position of the second target feature sequence.

[0040] performing feature aggregation on each second target residue feature corresponding to the second residue feature initial vector in the second target feature sequence and the interrelationship between the plurality of second target residue features based on the second residue feature initial vector, to obtain a second virtual structure microenvironment feature for representing global information of the second target feature sequence.

[0041] In some embodiments, the second obtaining module is further configured to:

[0042] obtaining a preset first global initial vector, and performing feature aggregation on a plurality of first features in the first to-be-sequenced sequence based on the first global initial vector by using the protein language model, to obtain a first global sequence feature;

[0043] obtaining a preset second global initial vector, and performing feature aggregation on a plurality of second features in the second to-be-sequenced sequence based on the second global initial vector by using the protein language model, to obtain a second global sequence feature.

[0044] Correspondingly, a third aspect of the embodiments of the present application provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the protein stability prediction method of any one of the first aspect of the embodiments of the present application when executing the computer program.

[0045] Correspondingly, a fourth aspect of the embodiments of the present application provides a computer readable storage medium, the storage medium storing a computer program, the computer program being executed by a processor to implement the protein stability prediction method according to any one of the first aspect of the embodiments of the present application.

[0046] According to the embodiments of the present application, the first to-be-sequence of the original protein and the second to-be-sequence of the mutant protein corresponding to the original protein are obtained, a plurality of first target residue features are determined from the first to-be-sequence according to the attention mechanism of the pre-trained protein language model, and global features of the plurality of first target residue features are extracted to obtain first virtual structure microenvironment features; a plurality of second target residue features are determined from the second to-be-sequence according to the attention mechanism of the protein language model, and global features of the plurality of second target residue features are extracted to obtain second virtual structure microenvironment features; the first global sequence features of the first to-be-sequence and the second global sequence features of the second to-be-sequence are obtained through the protein language model; the first difference is determined based on the first virtual structure microenvironment features and the second virtual structure microenvironment features, and the second difference is determined based on the first global sequence features and the second global sequence features; the protein stability prediction result is obtained by predicting the protein stability of the mutant protein according to the first difference and the second difference. In this way, the local environment features around the mutant residues and the global features of the entire protein sequence are combined to predict the stability of the protein, so that the protein language model can understand the influence of the mutation on the local structure of the protein and the influence of the mutation on the global structure, thereby improving the accuracy of predicting the stability of the protein. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is an architecture schematic diagram of the protein stability prediction system provided by the embodiments of the present application;

[0048] Figure 2 is a flowchart of the protein stability prediction method provided by the embodiments of the present application;

[0049] Figure 3 is a general flowchart of the protein stability prediction method provided by the embodiments of the present application;

[0050] Figure 4 is a construction process diagram of the virtual structure microenvironment provided by the embodiments of the present application;

[0051] Figure 5 is a functional module schematic diagram of the protein stability prediction device provided by the embodiments of the present application;

[0052] Figure 6 is a hardware structure schematic diagram of the computer device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0053] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not used to limit the present application.

[0054] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a manner different from the module division in the device or the sequence in the flowchart. The terms "first", "second", and the like in the description and claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0056] Protein stability refers to the ability of a protein to maintain its three-dimensional structure under certain conditions. Protein stability can provide important insights into protein folding and function, and is an important part of protein evolutionary fitness. However, a single or multiple amino acid substitutions in a protein sequence can affect the structure of the protein, and thus affect the overall structural stability and function of the protein. For example, certain amino acids such as proline and glycine can significantly change the conformation and stability of the protein after mutation due to their special chemical properties.

[0057] In the related art, sequence methods are generally used to predict the stability of proteins. However, the sequence method can only roughly capture the overall structure of the protein and analyze it, resulting in the problem that the stability of the protein cannot be accurately predicted. Alternatively, structure-dependent methods are used to predict the stability of proteins. However, structure-dependent methods require high-resolution laboratory analysis structures, and these data are severely lacking or even missing in some protein families. In the case where data are difficult to obtain comprehensively, the accurate prediction of protein stability is also affected.

[0058] Therefore, the embodiments of the present application provide a protein stability prediction method, device, computer device and storage medium, which can improve the accuracy of predicting the stability of proteins.

[0059] The protein stability prediction method, device, computer device and storage medium provided by the embodiments of the present application are specifically described by the following embodiments. First, the protein stability prediction system in the embodiments of the present application is described.

[0060] Please refer to Figure 1In some embodiments, this application provides a protein stability prediction system, including a terminal 11 and a server 12.

[0061] For example, terminal 11 can be a device through which the user directly interacts, allowing the user to access and operate the protein stability prediction system. For instance, terminal 11 can be a client of a personal computer, smartphone or tablet, workstation, or high-performance computing cluster, etc.

[0062] Furthermore, server 12 can be a backend server or cloud service facility for providing protein stability prediction services, responsible for processing requests from terminal 11, executing prediction algorithms, and returning prediction results. For example, server 12 can be a web server, application server, database server, high-performance computing cluster, cloud computing platform, etc.

[0063] Server 12 can efficiently process requests from different terminals 11 and provide accurate and reliable protein stability prediction services.

[0064] The protein stability prediction method in this application can be illustrated by the following examples.

[0065] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user will be obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent will the necessary user-related data for the normal operation of the embodiments of this application be obtained.

[0066] In this application embodiment, the description will focus on a protein stability prediction device, which can be integrated into a computer device. See [link to relevant documentation]. Figure 2 , Figure 2 This is a flowchart illustrating the steps of the protein stability prediction method provided in this application embodiment. Taking the integration of the protein stability prediction device into a terminal or server as an example, the specific process when the processor on the terminal or server executes the program instructions corresponding to the protein stability prediction method is as follows:

[0067] Step 101: Obtain the first test sequence corresponding to the original protein and the second test sequence of the mutant protein corresponding to the original protein.

[0068] In some embodiments, in order to accurately predict the influence of mutations on protein stability, a first to-be-tested sequence corresponding to the original protein and a second to-be-tested sequence corresponding to the mutant protein of the original protein can be obtained for comparison analysis.

[0069] The original protein can be the normal form of the protein in nature, which is usually taken as a reference benchmark for comparing the characteristics of the mutant protein.

[0070] The first to-be-tested sequence can be the amino acid sequence of the original protein, which is a benchmark sequence for protein stability prediction.

[0071] The mutant protein can be a protein with one or more mutations in the original protein, and the mutations can be one or more amino acid substitutions, insertions, or deletions.

[0072] The second to-be-tested sequence can be the amino acid sequence of the mutant protein, which is a compiled sequence for protein stability prediction.

[0073] For example, assuming that there is a natural original protein without any mutations, and it is known that a mutation occurs at a certain position of the original protein, forming a mutant protein. At this time, the amino acid sequence of the original protein can be obtained as the first to-be-tested sequence, and the first to-be-tested sequence is assumed to be MAGF (only as an example, the first to-be-tested sequence can be more complex in actual situations). The amino acid sequence of the mutant protein obtained after the mutation of the original protein is obtained as the second to-be-tested sequence, and the amino acid sequence of the second to-be-tested sequence is assumed to be MAGV.

[0074] By obtaining the sequences of the original protein and the mutant protein, subsequent accurate comparison analysis and protein stability prediction can be facilitated.

[0075] In step 102, a plurality of first target residue features are determined from the first to-be-tested sequence according to the attention mechanism of the pre-trained protein language model, and global feature extraction is performed on the plurality of first target residue features to obtain first virtual structure microenvironment features.

[0076] In some embodiments, in order to capture the local environmental information around the mutant residue, a plurality of first target residue features can be determined through the attention mechanism, and the first virtual structure microenvironment features can be obtained by integrating the plurality of first target residue features and the surrounding local environmental information, so as to improve the accuracy of protein stability prediction.

[0077] The protein language model can be a deep learning model obtained through pre-training. The initial protein language model can be a trained protein language model obtained through pre-training on a large number of protein sequences. The trained protein language model can capture the correlation rules between amino acids. Further, the protein language model is usually based on a Transformer architecture, which can process protein sequence data and extract useful features therefrom.

[0078] The attention mechanism can be a mechanism in the protein language model that determines the interaction strength between different positions in the first test sequence by learning different representations of the first test sequence, and assigns different attention scores to different positions based on the importance of the different positions.

[0079] The first target residue feature can be a position-related residue feature of the mutant protein before mutation determined from the first test sequence by the attention mechanism.

[0080] The first virtual structure microenvironment feature can be a comprehensive feature representation obtained by global feature extraction on the first target residue feature, which contains local environmental information around the mutant residue before mutation.

[0081] For example, the attention mechanism of the pre-trained protein language model can output an overall attention map corresponding to the first test sequence, and the overall attention map can be pooled to obtain a pooled overall attention map. Further, the position of the mutant residue feature in the first test sequence can be determined, and the position in the pooled overall attention map can be searched to determine the attention score vector of the position.

[0082] Further, after obtaining the attention score vector of the position of each residue feature, the number of first target residue features to be selected can be determined, and the residue feature with the highest attention score can be determined. For example, if 10 first target residue features are to be selected, the 10 residue features with the highest attention scores can be selected as the first target residue features according to the attention scores of the multiple residue features. In this way, the residue information with the greatest impact on the mutation position can be captured, so that the protein language model can focus more on important information, while removing unnecessary noise and avoiding being affected by secondary information.

[0083] Further, global feature extraction can be performed on the multiple first target residue features to obtain corresponding virtual structure microenvironment features, so as to reflect the influence of mutation on the local structure by generating a comprehensive feature representation.

[0084] The plurality of first target residue features are determined through the attention mechanism, and global feature extraction is performed on the first target residue features to obtain the first virtual structure microenvironment feature, which can capture and screen important feature information and improve the accuracy of protein language model prediction.

[0085] In some embodiments, in order to improve the accuracy of protein stability prediction, a plurality of first target residue features can be determined from the first to-be-sequenced sequence according to the attention mechanism of the pre-trained protein language model, so as to capture the residue information that has the greatest impact on the mutation position, and at the same time, through screening, improve the calculation efficiency of the protein language model. For example, "determining a plurality of first target residue features from the first to-be-sequenced sequence according to the attention mechanism of the protein language model" in step 102 can include:

[0086] (102.a1) generating a first overall attention map corresponding to the first to-be-sequenced sequence according to the attention mechanism of the pre-trained protein language model;

[0087] (102.a2) performing global pooling processing on the first overall attention map to obtain a first target overall attention map;

[0088] (102.a3) determining a plurality of first positions of a plurality of mutation residue features from the first to-be-sequenced sequence;

[0089] (102.a4) matching the plurality of first positions in the first target overall attention map to obtain a plurality of first attention score vectors corresponding to the plurality of mutation residue features;

[0090] (102.a5) obtaining a first residue feature selection quantity, and determining a same number of first target residue features as the first residue feature selection quantity from the plurality of mutation residue features based on the plurality of first attention score vectors.

[0091] The first overall attention map can be used to reflect the interaction strength between different positions in the first to-be-sequenced sequence, and the first overall attention map can be a two-dimensional matrix, and each element represents an attention score corresponding to a feature of the first to-be-sequenced sequence.

[0092] The first target overall attention map can be an attention map obtained by performing global pooling processing on the first overall attention map. The pooling method can be maximum pooling, average pooling, adaptive pooling, global pooling, etc.

[0093] The mutation residue feature can be the position of the feature that occurs mutation in the first to-be-sequenced sequence. The mutation residue feature can be determined based on the comparison between the first to-be-sequenced sequence and the second to-be-sequenced sequence, or can be manually labeled by a user or automatically identified by an algorithm.

[0094] The first position can be the specific location of the mutated residue feature in the first sequence to be tested.

[0095] The first attention score vector can refer to the attention score vector at the position corresponding to the mutation position in the overall attention map of the first target.

[0096] The number of first residue features selected can be the number of mutated residue features to be selected from multiple mutated residue features. For example, the number of first residue features selected can be 10, 20, etc. The specific number of selections can be flexibly determined according to the actual situation.

[0097] For example, suppose the first test sequence is MAGF (this is just an example; in reality, the sequence structure may have multiple representations), and it is known that a mutation has occurred at the fourth amino acid position (F has become V). The first test sequence is input into a pre-trained protein language model to obtain the first test sequence feature Z1 corresponding to the first test sequence. seq The protein language model can generate a first global attention map corresponding to the first test sequence based on the features of the first test sequence through the final Transformer layer. Where H refers to the number of attention heads in the Transformer layer, and L1 refers to the sequence length of the first test sequence.

[0098] Furthermore, global pooling can be performed on the first overall attention map to obtain the first target overall attention map, in order to retain the most critical information while reducing data dimensionality. The formula is as follows:

[0099]

[0100] in, Represents the first overall attention map A last1 The first sub-attention graph of the i-th attention head in the Transformer layer, where H represents the number of attention heads in the Transformer layer. This represents the overall attention map for the primary objective.

[0101] Furthermore, the first position (pos) of each mutated residue can be determined through automatic system labeling (or by comparing the sequences before and after the protein mutation). mut1 And through the first position in the pooled first target overall attention map The search is performed to obtain the first attention score vector associated with the mutant residue features. The formula is as follows:

[0102]

[0103] Among them, A mut1a first attention score vector.

[0104] Further, the number of first residue feature selected is set to k (the value of k can be set according to actual situation), and the features of several residues with the highest first attention scores are selected from the mutation residue features:

[0105] (pos1, pos2, …, pos k1 )=argmax k1 (A mut1 )

[0106]

[0107] Wherein, argmax k1 (·) represents a function of retrieving the top K mutation residue features with the highest attention scores as the first target residue features according to the first attention score vector, represents the top K residues with the highest first attention scores of the mutation residue features.

[0108] In the above manner, the local environmental information around the mutation residue is effectively captured, the accuracy and robustness of the protein stability prediction are improved, and the computational efficiency of the protein language model is improved by reducing redundant information through feature selection.

[0109] In some embodiments, in order to capture the local environmental information around the mutation residue to determine the interaction strength between the mutation residue and other amino acids, global feature extraction can be performed on the plurality of first target residue features to obtain first virtual structure microenvironment features, so as to more accurately evaluate the influence of mutation on protein stability. For example, the "global feature extraction on the plurality of first target residue features to obtain first virtual structure microenvironment features" in step 102 can include:

[0110] (102.b1) obtaining a preset first residue feature initial vector, and splicing the first residue feature initial vector and the plurality of first target residue features to obtain a first target feature sequence; wherein the first residue feature initial vector is located at the beginning of the first target feature sequence;

[0111] (102.b2) based on the first residue feature initial vector, performing feature aggregation on each first target residue feature in the first target feature sequence and the interrelationship between the plurality of first target residue features, to obtain a first virtual structure microenvironment feature for representing the global information of the first target feature sequence.

[0112] Wherein, the first residue feature initial vector can be obtained by a pre-trained protein language model, and is a vector for capturing information of the entire sequence.

[0113] wherein the first target feature sequence can be a feature sequence obtained by concatenating the first residue feature initial vector with the plurality of first target residue features.

[0114] For example, the first residue feature initial vector Z CLS1 may be a learnable vector, After obtaining the plurality of first target residue features , the K first target residue features can be input into a transformer layer to aggregate the K first target residue features into a first virtual structure microenvironment feature, as follows:

[0115]

[0116] wherein, represents the final first virtual structure microenvironment feature, ExtractCLS represents an operation of extracting a CLS token (global vector) in the protein language model, represents a transformer layer with trainable parameters θ, represents the concatenation of the first residue feature initial vector and the plurality of first target residue features.

[0117] Through the first residue feature initial vector, the protein language model can pass additional information when processing sequence data, and better understand the interaction between the mutation residue feature and its adjacent mutation residue feature, so as to achieve the optimal feature integration effect of the first virtual structure microenvironment feature, and more accurately evaluate the influence of the mutation on the protein stability.

[0118] Step 103, determining a plurality of second target residue features from the second to-be-measured sequence according to the attention mechanism of the protein language model, and extracting global features from the plurality of second target residue features to obtain a second virtual structure microenvironment feature.

[0119] In some embodiments, in order to capture the local environment information around the mutation residue after the mutation of the protein, the plurality of second target residue features can be determined through the attention mechanism, and the second virtual structure microenvironment feature can be obtained by integrating the plurality of second target residue features and the surrounding local environment information, so as to improve the accuracy of the protein stability prediction.

[0120] ​The protein language model can be a deep learning model obtained through pre-training. The initial protein language model can be a trained protein language model obtained through pre-training on a large number of protein sequences. The trained protein language model can capture the correlation rules between amino acids. Further, the protein language model is usually based on a Transformer architecture, which can process protein sequence data and extract useful features therefrom.

[0121] The attention mechanism can be a mechanism in the protein language model that determines the interaction strength between different positions in the second test sequence by learning different representations of the second test sequence, and assigns different attention scores to different positions based on the importance of the different positions.

[0122] The second target residue feature can be a residue feature related to the mutation position of the mutant protein after mutation determined from the second test sequence by the attention mechanism.

[0123] The second virtual structure microenvironment feature can be a comprehensive feature representation obtained by global feature extraction on the second target residue feature, which contains local environmental information around the mutation residue.

[0124] For example, the attention mechanism of the pre-trained protein language model can output an overall attention map corresponding to the second test sequence, and the overall attention map can be pooled to obtain a pooled overall attention map. Further, the position of the mutated residue feature in the second test sequence can be determined, and the position in the pooled overall attention map can be searched to determine the attention score vector of the position.

[0125] Further, after obtaining the attention score vector of the position of each residue feature, the number of second target residue features to be selected can be determined, and the residue feature with the highest attention score can be determined. For example, if 10 second target residue features are to be selected, the 10 residue features with the highest attention scores can be selected as the second target residue features according to the attention scores of the multiple residue features. In this way, the residue information that has the greatest impact on the mutation position can be captured, so that the protein language model can focus more on important information, while removing unnecessary noise and avoiding being affected by secondary information.

[0126] Further, global feature extraction can be performed on the multiple second target residue features to obtain corresponding virtual structure microenvironment features, so as to reflect the influence of mutation on the local structure by generating a comprehensive feature representation.

[0127] The second target residue features are determined through the attention mechanism, and global feature extraction is performed on the second target residue features to obtain the second virtual structure microenvironment features, which can capture and screen important feature information and improve the accuracy of protein language model prediction.

[0128] In some embodiments, in order to improve the accuracy of protein stability prediction, a plurality of second target residue features can be determined from the second to-be-sequenced sequence according to the attention mechanism of the pre-trained protein language model, so as to capture the residue information that has the greatest influence on the mutation position, and at the same time, through screening, the calculation efficiency of the protein language model is improved. For example, the "determining a plurality of second target residue features from the second to-be-sequenced sequence according to the attention mechanism of the protein language model" in step 103 can include:

[0129] (103.a1) generating a second overall attention map corresponding to the second to-be-sequenced sequence according to the attention mechanism of the protein language model;

[0130] (103.a2) performing global pooling processing on the second overall attention map to obtain a second target overall attention map;

[0131] (103.a3) determining a plurality of second positions of a plurality of mutation residue features from the second to-be-sequenced sequence;

[0132] (103.a4) matching the plurality of second positions in the second target overall attention map to obtain a plurality of second attention score vectors corresponding to the plurality of mutation residue features;

[0133] (103.a5) obtaining a second residue feature selection quantity, and determining a same number of second target residue features as the second residue feature selection quantity from the plurality of mutation residue features based on the plurality of second attention score vectors.

[0134] The second overall attention map can be used to reflect the interaction strength between different positions in the second to-be-sequenced sequence, and the second overall attention map can be a two-dimensional matrix, and each element represents an attention score corresponding to a feature of the second to-be-sequenced sequence.

[0135] The second target overall attention map can be an attention map obtained by performing global pooling processing on the second overall attention map. The pooling method can be maximum pooling, average pooling, adaptive pooling, global pooling, etc.

[0136] The mutation residue feature can be the position of the feature that occurs mutation in the second to-be-sequenced sequence. The mutation residue feature can be determined based on the alignment of the second to-be-sequenced sequence and the second to-be-sequenced sequence, or can be manually labeled by the user or automatically identified by the algorithm.

[0137] The second position can be the specific location of the mutated residue feature in the second sequence to be tested.

[0138] The second attention score vector can refer to the attention score vector at the position corresponding to the mutation position in the overall attention map of the second target.

[0139] The number of second residue features selected can be the number of mutated residue features to be selected from multiple mutated residue features. For example, the number of second residue features selected can be 10, 20, etc. The specific number of selections can be flexibly determined according to the actual situation.

[0140] For example, suppose the second test sequence is MAGF (this is just an example; in reality, the sequence structure may have multiple representations), and it is known that a mutation has occurred at the fourth amino acid position (F has become V). The second test sequence is input into a pre-trained protein language model to obtain the second test sequence feature Z2. seq The protein language model can generate a second overall attention map corresponding to the second test sequence based on the features of the second test sequence through the final Transformer layer. Here, H refers to the number of attention heads in the Transformer layer, and L2 refers to the sequence length of the second test sequence.

[0141] Furthermore, global pooling can be applied to the second overall attention map to obtain the second target overall attention map, thereby reducing data dimensionality while retaining the most critical information. The formula is as follows:

[0142]

[0143] in, Represents the second overall attention map A 1ast2 The second sub-attention graph of the j-th attention head, where H refers to the number of attention heads in the Transformer layer. This represents the overall attention map for the second objective.

[0144] Furthermore, the second position (pos) of each mutated residue can be determined through automatic system labeling (or by comparing the sequences before and after the protein mutation). mut2 And through the second position in the pooled second target overall attention map The search is performed to obtain a second attention score vector associated with the mutant residue features. The formula is as follows:

[0145]

[0146] Among them, A mut2a second attention score vector.

[0147] Further, the number of second residue feature selected is set to k (the value of k can be set according to actual situation), and the features of several residues with the highest second attention scores are selected from the mutant residue features:

[0148] (pos1, pos2, …, pos k2 )=argmax k2 (A mut2 )

[0149]

[0150] Wherein, argmax k2 (·) represents a function of retrieving the top K mutant residue features with the highest attention scores as the second target residue features according to the second attention score vector, represents the top K residues with the highest second attention scores of the mutant residue features.

[0151] In the above manner, the local environmental information around the mutant residue is effectively captured, the accuracy and robustness of the protein stability prediction are improved, and the computational efficiency of the protein language model is improved by reducing redundant information through feature selection.

[0152] In some embodiments, in order to capture the local environmental information around the mutant residue to determine the interaction strength between the mutant residue and other amino acids, global feature extraction can be performed on the plurality of second target residue features to obtain a second virtual structure microenvironment feature, so as to more accurately evaluate the influence of the mutation on the protein stability. For example, the "global feature extraction on the plurality of second target residue features to obtain a second virtual structure microenvironment feature" in step 103 can include:

[0153] (103.b1) obtaining a preset second residue feature initial vector, and splicing the second residue feature initial vector and the plurality of second target residue features to obtain a second target feature sequence; wherein the second residue feature initial vector is located at the beginning of the second target feature sequence;

[0154] (103.b2) based on the second residue feature initial vector, performing feature aggregation on each second target residue feature in the second target feature sequence and the interrelationship between the plurality of second target residue features, to obtain a second virtual structure microenvironment feature for representing the global information of the second target feature sequence.

[0155] Wherein, the second residue feature initial vector can be obtained by a pre-trained protein language model, and is a vector for capturing information of the entire sequence.

[0156] wherein the second target feature sequence can be a feature sequence obtained by concatenating the second residue feature initial vector with the plurality of second target residue features.

[0157] For example, the second residue feature initial vector Z CLS2 may be a learnable vector, After obtaining the plurality of second target residue features , the K second target residue features can be aggregated into a second virtual structure microenvironment feature by a transformer layer.

[0158]

[0159] wherein, represents the final second virtual structure microenvironment feature, ExtractCLS represents an operation of extracting a CLS token (global vector) in the protein language model, represents a transformer layer with trainable parameters θ, represents the concatenation of the second residue feature initial vector and the plurality of second target residue features.

[0160] Through the second residue feature initial vector, the protein language model can pass additional information when processing sequence data, and better understand the interaction between the mutation residue feature and its adjacent mutation residue feature, so that the second virtual structure microenvironment feature achieves the optimal feature integration effect, thereby more accurately evaluating the influence of the mutation on the protein stability.

[0161] Step 104, obtaining a first global sequence feature of the first test sequence and a second global sequence feature of the second test sequence by the protein language model.

[0162] In some embodiments, in order to better understand the overall structure and function of the protein, integrate the information of the mutation residue and its surrounding environment, the first global sequence feature of the first test sequence and the second global sequence feature of the second test sequence can be obtained to compare and analyze the influence of the mutation on the protein stability at the global level.

[0163] wherein the first global sequence feature can be a global feature extracted from the first test sequence by the protein language model. The first global sequence feature can be used to understand the overall structure and function of the original protein.

[0164] wherein the second global sequence feature can be a global feature extracted from the second test sequence by the protein language model. The second global sequence feature can be used to understand the overall structure and function of the mutant protein.​

[0165] For example, the first global sequence feature can be obtained by concatenating the first global initial vector with the first sequence to be tested, and aggregating each first sequence feature in the first sequence to be tested and the mutual relationship between the first sequence features according to the first global initial vector.

[0166] Similarly, the second global sequence feature can be obtained by concatenating the second global initial vector with the second sequence to be tested, and aggregating each second sequence feature in the second sequence to be tested and the mutual relationship between the second sequence features according to the second global initial vector.

[0167] By obtaining the first global sequence feature and the second global sequence feature, a global perspective can be provided to better understand the overall structure and function of the original protein and the mutant protein, which helps to integrate the information of the mutation site and its surrounding environment, and to compare and analyze the influence of the mutation on the protein stability at the global level, thereby improving the accuracy and robustness of the prediction.

[0168] In some embodiments, in order to comprehensively evaluate the influence of the mutation on the protein stability, the first global sequence feature corresponding to the original protein can be obtained to understand the state of the protein before the mutation, and the second global sequence feature corresponding to the mutant protein can be obtained to understand the state of the protein after the mutation, so as to accurately predict the stability of the protein. For example, step 104 can include:

[0169] (104.1) obtaining a preset first global initial vector, and aggregating a plurality of first features in the first sequence to be tested based on the first global initial vector by the protein language model to obtain the first global sequence feature;

[0170] (104.2) obtaining a preset second global initial vector, and aggregating a plurality of second features in the second sequence to be tested based on the second global initial vector by the protein language model to obtain the second global sequence feature.

[0171] The first global initial vector can be a first vector used to initialize the feature sequence before the feature aggregation. The first global initial vector is used to help the protein language model better understand the overall structure and function of the original protein.

[0172] The second global initial vector can be a first vector used to initialize the feature sequence before the feature aggregation. The second global initial vector is used to help the protein language model better understand the overall structure and function of the mutant protein.

[0173] Exemplarily, a preset first global initial vector Z q1 CLS is obtained by concatenating the first global initial vector with the first to-be-sequenced sequence (Z 11 ,…,Z 1END ). seq1 q1 CLS 11 ,…,Z 1END ) is obtained. Further, based on the first global learning vector, a plurality of first features in the first to-be-sequenced sequence can be aggregated by a Transformer layer or other feature aggregation method in the protein language model to obtain a first global sequence feature Z Q1 CLS , so as to effectively capture the overall structure and function information of the original protein and better understand the state of the protein before mutation.

[0174] In some embodiments, a pre-trained time sequence model such as a long short-term memory network can be obtained, and the first to-be-sequenced sequence is input into the time sequence model to obtain a first to-be-sequenced sequence with enhanced time sequence information. Then, the first to-be-sequenced sequence with enhanced time sequence information is further processed to generate a corresponding first global sequence feature. In this way, the time sequence evolution of the protein structure and function after mutation can be better captured, the understanding of the model in the time dimension is further enhanced, and more accurate prediction results can be provided.

[0175] Exemplarily, a preset second global initial vector Z q2 CLS is obtained. Then, a second global initial vector is concatenated with a second to-be-sequenced sequence (Z 21 ,…,Z 2END ) to obtain Z seq2 q2 CLS 21 ,…,Z 2END ). Further, based on the second global learning vector, a plurality of second features in the second to-be-sequenced sequence can be aggregated by a Transformer layer or other feature aggregation method in the protein language model to obtain a second global sequence feature Z Q2 CLS . So as to effectively capture the overall structure and function information of the mutated protein and better understand the state of the protein after mutation.

[0176] ​​​​In some embodiments, a pre-trained time sequence model, such as a long short-term memory network, can be obtained, and the second to-be-tested sequence is input into the time sequence model to obtain a second to-be-tested sequence with enhanced time sequence information. Then, the second to-be-tested sequence with enhanced time sequence information is further processed to generate corresponding second global sequence features. In this way, the time sequence evolution of the protein structure and function after mutation can be better captured, the understanding of the model in the time dimension is further enhanced, and more accurate prediction results can be provided.

[0177] In this way, the overall structure and function information of the original protein and the mutant protein can be effectively captured, so that the state of the protein before and after mutation can be better understood, and the accuracy and robustness of the subsequent prediction of the stability of the protein can be improved.

[0178] In step 105, a first difference value is determined based on the first virtual structure microenvironment feature and the second virtual structure microenvironment feature, and a second difference value is determined based on the first global sequence feature and the second global sequence feature.

[0179] In some embodiments, in order to effectively capture the influence of mutation in local and global scales, a first difference value can be determined based on the first virtual structure microenvironment feature and the second virtual structure microenvironment feature, and a second difference value can be determined based on the first global sequence feature and the second global sequence feature, so as to evaluate whether the mutation has destroyed important local domains and the function of the protein, thereby improving the accuracy and robustness of the prediction.

[0180] The first difference value can be a numerical value determined by comparing the difference between the first virtual structure microenvironment feature and the second virtual structure microenvironment feature. The first difference value can be used to represent the change of the protein in the local microenvironment before and after mutation, so as to intuitively reflect the influence degree of mutation on the local microenvironment feature.

[0181] The second difference value can be a numerical value determined by comparing the difference between the first global sequence feature and the second global sequence feature. The second difference value can be used to represent the change of the protein in the overall structure before and after mutation, so as to comprehensively evaluate the influence of mutation on the global feature of the protein.

[0182] For example, if the first virtual structure microenvironment feature is The second virtual structure microenvironment feature is The first difference value is the difference between the first virtual structure microenvironment feature and the second virtual structure microenvironment feature, i.e.

[0183] For example, if the first global sequence feature is The second global sequence feature is The second difference value is the difference between the first global sequence feature and the second global sequence feature, i.e.

[0184] In this way, multi-scale information can be integrated to effectively capture the influence of protein mutations on local and global scales, thereby improving the accuracy and robustness of the prediction.

[0185] At step 106, the protein stability of the mutant protein is predicted according to the first difference value and the second difference value, to obtain a protein stability prediction result.

[0186] In some embodiments, the first difference value provides the change in the local microenvironment of the protein before and after mutation, and the second difference value provides the change in the overall structure of the protein before and after mutation. Therefore, the protein stability of the mutant protein can be predicted according to the first difference value and the second difference value, to ensure that the prediction not only considers local changes, but also considers global changes, making the prediction result more reliable and accurate.

[0187] The protein stability prediction result can be a result of predicting the protein stability by analyzing the difference in the change of the local microenvironment and the overall structure of the protein before and after mutation. The protein stability prediction result can be used to evaluate the degree of influence of mutation on protein stability.

[0188] For example, the first difference value and the second difference value can be input into a pre-trained prediction model to obtain a first prediction result corresponding to the first difference value and a second prediction result corresponding to the second difference value. Further, the first prediction result and the second prediction result can be added and averaged to obtain the protein stability prediction result. Specifically, the prediction model can be a multilayer perceptron, a deep residual network, etc.

[0189] For example, when the prediction model is a multilayer perceptron, the data processing process is as follows:

[0190]

[0191] ΔΔG = (ΔΔG CLS + ΔΔG ENV ) / 2

[0192] wherein, is the first virtual structure microenvironment feature, is the second virtual structure microenvironment feature, ΔΔG ENV is the first prediction result predicted according to the first difference value; is the first global sequence feature, is the second global sequence feature, ΔΔG CLS is the second prediction result predicted according to the second difference value.

[0193] Further, the first prediction result and the second prediction result can be adjusted and averaged to obtain the protein stability prediction result according to their importance. Specifically, a first weight can be set for the first prediction result, and a second weight can be set for the second prediction result. For example, the protein stability prediction result = (a*first prediction result + b*second prediction result) / 2. Wherein a represents the first weight, b represents the second weight, a and b are both values in the interval of 0 and 1, the sum of a and b is 1, and the values of a and b can be adjusted according to actual conditions.

[0194] By analyzing the differences in local microenvironment and overall structure changes of the protein before and after mutation to predict the protein stability, and inputting the first difference and the second difference into the pre-trained prediction model, the protein stability prediction result is obtained. This way can effectively capture the influence of mutation on local and global scales, thereby improving the accuracy and robustness of the prediction.

[0195] In some embodiments, in order to effectively capture the influence of mutation on local and global scales, the protein stability of the mutant protein can be predicted according to the first difference and the second difference respectively, thereby improving the accuracy and robustness of the prediction. For example, step 106 can include:

[0196] (106.1) input the first difference corresponding to the first target residue feature and the second target residue feature into the pre-trained prediction model to predict the protein folding free energy of the residue feature change dimension, and obtain the first prediction result;

[0197] (106.2) input the second difference corresponding to the first test sequence and the second test sequence into the prediction model to predict the protein folding free energy of the global change dimension, and obtain the second prediction result;

[0198] (106.3) based on the mean of the first prediction result and the second prediction result, obtain the protein stability prediction result.

[0199] Wherein, the prediction model can be a machine learning model or a deep learning model, which can predict the stability change of the protein after mutation according to the input first difference after training. For example, the prediction model can be a multilayer perception, a convolutional neural network, a deep residual network, a recurrent neural network, etc. The present application does not make specific limitations, as long as the selected prediction model can accurately predict the first prediction result according to the first difference and the second prediction result according to the second difference after pre-training.

[0200] The protein folding free energy can refer to the free energy released or absorbed during the folding of a protein from its linear amino acid sequence into a three-dimensional structure. It can be understood that, since the mutation of a protein will change the folding free energy of the protein, the stability of the protein before and after mutation can be determined by predicting the folding free energy of the protein.

[0201] The first prediction result can be a predicted value of the change in the folding free energy of the protein in the local microenvironment before and after mutation, obtained by inputting the first difference into a pre-trained prediction model.

[0202] The second prediction result can be a predicted value of the change in the folding free energy of the protein in the global structure before and after mutation, obtained by inputting the second difference into a pre-trained prediction model.

[0203] Specifically, the first difference corresponding to the first target residue feature and the second target residue feature, i.e., the first difference determined based on the first virtual structure microenvironment feature and the second virtual structure microenvironment feature. After obtaining the first difference, the first difference can be directly input into the prediction model to obtain the first prediction result.

[0204] Specifically, the second difference corresponding to the first to-be-measured sequence and the second to-be-measured sequence, i.e., the second difference determined based on the first global sequence feature and the second global sequence feature. After obtaining the second difference, the second difference can be directly input into the prediction model to obtain the second prediction result.

[0205] For example, if the first difference is calculated to be [-0.1, -0.1, 0.2, 0.1, 0.1], and the first difference is input into the prediction model, the prediction result of the protein folding free energy under the dimension of the virtual structure microenvironment feature can be -1.5 kcal / mol. The second difference is calculated to be [0.1, 0.1, 0.1, 0.1, 0.1], and the second difference is input into the prediction model, the prediction result of the protein folding free energy under the dimension of the global sequence feature can be -1.0 kcal / mol. Then, the final protein stability prediction result is (-1.5-1.0) / 2=-1.25 kcal / mol.

[0206] Further, in order to determine the degree of influence of the mutation on the protein stability, the protein stability prediction result can be compared with a preset threshold, and a classification decision is made accordingly. For example, when the protein stability prediction result is less than the preset threshold (e.g., 1.0 kcal / mol), it is considered that the mutation has little influence on the protein stability; when the protein stability prediction result is greater than the positive threshold (e.g., 1.0 kcal / mol), it is considered that the mutation reduces the protein stability, etc. The preset threshold can be flexibly set according to the actual situation, which is not specifically limited here.

[0207] In some embodiments, the first virtual structure microenvironment, the second virtual structure microenvironment, the first global sequence feature, and the second global sequence feature can be mapped into a graph structure, and the topological structure is modeled by a graph neural network to further capture the relationship between the protein residues. Further, the difference between the first virtual graph structure corresponding to the first virtual structure microenvironment and the second virtual graph structure corresponding to the second virtual structure microenvironment can be identified by the graph neural network to obtain a first difference value, and the difference between the first global graph structure corresponding to the first global sequence feature and the second global graph structure corresponding to the second global sequence feature can be identified by the graph neural network to obtain a second difference value, and the first difference value and the second difference value are predicted by a prediction model (which can be a multilayer perceptron or a graph neural network) respectively, and finally a protein stability prediction result is obtained. In this way, the structural relationship between the protein residues can be more accurately captured.

[0208] Please refer to Figure 3 , Figure 3 The embodiments introduced in the overall flow of the present application. For example, the original protein and the mutant protein corresponding to the original protein can be respectively input into the pre-trained protein language model to extract sequence features, and the first virtual structure microenvironment feature and the first global sequence feature corresponding to the original protein, and the second virtual structure microenvironment feature and the second global sequence feature corresponding to the mutant protein are obtained. Further, the first virtual structure microenvironment feature corresponding to the original protein and the second virtual structure microenvironment feature corresponding to the mutant protein can be subtracted to obtain a first difference value, and the first global sequence feature corresponding to the original protein and the second global sequence feature corresponding to the mutant protein can be subtracted to obtain a second difference value. The first difference value and the second difference value are respectively input into the pre-trained prediction model to obtain a first prediction result corresponding to the first difference value and a second prediction result corresponding to the second difference value. Further, the first prediction result and the second prediction result can be averaged to obtain a protein stability prediction result.

[0209] Further, in combination with Figure 4For the construction of the first virtual structural microenvironment feature and the second virtual structural microenvironment feature, it should be noted that the construction processes of the first virtual structural microenvironment feature and the second virtual structural microenvironment feature are the same, and in order to facilitate the introduction, the construction of the virtual structural microenvironment feature is introduced below. The following process is suitable for the construction of the first virtual structural microenvironment feature and the second virtual structural microenvironment feature.

[0210] As shown in Figure 4 , the sequence to be tested S is input into the protein language model to obtain the sequence feature Z corresponding to the sequence to be tested seq , and the protein language model can output the overall attention map of the last Transformer layer , where H represents the number of attention heads of the transfomer, and L represents the length of the sequence to be tested. First, the output overall attention map can be pooled to obtain the target overall attention map:

[0211]

[0212] , where represents the overall attention map of the i-th attention head in the overall attention map A last .

[0213] Subsequently, given the position pos of the mutation residue in the sequence to be tested mut , the position is searched in the pooled target overall attention map to obtain the attention score vector related to the mutation residue

[0214]

[0215] K (the value of K can be set according to actual conditions) features of a number of residues with the highest attention scores are selected:

[0216] (pos1, pos2, …, pos k ) = argmax k (A mut )

[0217]

[0218] , where argmax k (·) represents a function of retrieving the first K residues with the largest attention scores according to the attention score vector, represents the first K residues with the largest attention scores of the mutation residue.

[0219] Further, a learnable residue feature initial vector is set, and the a layer of transformer is fed to integrate the features of the K residues into a virtual structural microenvironment feature:

[0220]

[0221] wherein, represents the final virtual structural microenvironment feature, ExtractCLS represents the operation of extracting the CLS token in the protein language model, represents a layer of transformer with trainable parameters θ, represents the concatenation of features.

[0222] The application obtains a first to-be-sequenced sequence of an original protein and a second to-be-sequenced sequence of a mutant protein corresponding to the original protein; determines a plurality of first target residue features from the first to-be-sequenced sequence according to an attention mechanism of a pre-trained protein language model, and extracts global features of the plurality of first target residue features to obtain a first virtual structural microenvironment feature; determines a plurality of second target residue features from the second to-be-sequenced sequence according to the attention mechanism of the protein language model, and extracts global features of the plurality of second target residue features to obtain a second virtual structural microenvironment feature; obtains a first global sequence feature of the first to-be-sequenced sequence and a second global sequence feature of the second to-be-sequenced sequence through the protein language model; determines a first difference based on the first virtual structural microenvironment feature and the second virtual structural microenvironment feature, and determines a second difference based on the first global sequence feature and the second global sequence feature; and predicts the protein stability of the mutant protein according to the first difference and the second difference to obtain a protein stability prediction result. In this way, the stability of the protein can be predicted by combining the local environmental features around the mutant residues and the global features of the entire protein sequence, so that the protein language model can understand the influence of the mutation on the local structure of the protein and the influence of the mutation on the global structure, thereby improving the accuracy of predicting the stability of the protein.

[0223] Please refer to Figure 5 The embodiments of the application also provide a protein stability prediction device, which can implement the protein stability prediction method described above, and the protein stability prediction device comprises:

[0224] The first obtaining module 51 is configured to obtain a first to-be-sequenced sequence of an original protein and a second to-be-sequenced sequence of a mutant protein corresponding to the original protein;

[0225] The first extraction module 52 is configured to determine a plurality of first target residue features from the first to-be-sequenced sequence according to an attention mechanism of a pre-trained protein language model, and perform extraction of global features on the plurality of first target residue features to obtain first virtual structure microenvironment features.

[0226] The second extraction module 53 is configured to determine a plurality of second target residue features from the second to-be-sequenced sequence according to an attention mechanism of a protein language model, and perform extraction of global features on the plurality of second target residue features to obtain second virtual structure microenvironment features.

[0227] The second acquisition module 54 is configured to acquire, by a protein language model, first global sequence features of the first to-be-sequenced sequence and second global sequence features of the second to-be-sequenced sequence.

[0228] The determination module 55 is configured to determine a first difference value based on the first virtual structure microenvironment features and the second virtual structure microenvironment features, and determine a second difference value based on the first global sequence features and the second global sequence features.

[0229] The prediction module 56 is configured to perform protein stability prediction on the mutant protein according to the first difference value and the second difference value to obtain a protein stability prediction result.

[0230] The specific implementation of the protein stability prediction device is basically the same as that of the above-mentioned protein stability prediction method, and will not be repeated here. The protein stability prediction device can also be provided with other functional modules to realize the protein stability prediction method in the above-mentioned embodiments, as long as the requirements of the embodiments of the present application are met.

[0231] The embodiments of the present application also provide a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to realize the above-mentioned protein stability prediction method. The computer device can be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.

[0232] Please refer to Figure 6 , Figure 6 The hardware structure of the computer device of another embodiment is illustrated, which includes:

[0233] The processor 61 can be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, etc., and is used to execute related programs to realize the technical solutions provided by the embodiments of the present application.

[0234] The memory 62 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), etc. The memory 62 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present application are implemented by software or firmware, the related program codes are stored in the memory 62 and are called and executed by the processor 61 to implement the protein stability prediction method of the embodiments of the present application.

[0235] The input / output interface 63 is configured to realize information input and output.

[0236] The communication interface 64 is configured to realize the communication interaction between the device and other devices. The communication can be realized by a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0237] The bus 65 is configured to transmit information between various components (for example, the processor 61, the memory 62, the input / output interface 63, and the communication interface 64) of the device.

[0238] The processor 61, the memory 62, the input / output interface 63, and the communication interface 64 are connected to each other through the bus 65 to realize the communication connection between the device.

[0239] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the above-mentioned protein stability prediction method.

[0240] The memory is a non-transitory computer readable storage medium, which can be used to store a non-transitory software program and a non-transitory computer executable program. In addition, the memory can include a high-speed random access memory and can also include a non-transitory memory, for example, at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor. These remote memories can be connected to the processor through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0241] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0242] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation to the embodiments of the present application, and can include more or fewer steps than the figures, or combine certain steps, or different steps.

[0243] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, that is, can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments.

[0244] Those skilled in the art can understand that all or some steps in the above disclosed method, functional modules / units in the system and device can be implemented as software, firmware, hardware and their appropriate combinations.

[0245] The terms "first", "second", "third", "fourth" and the like (if any) in the description and the above drawings of the present application are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0246] It should be understood that in the present application, "at least one" and "several" refer to one or more, and "multiple" refers to two or more. "And / or" is used to describe the association between the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases: only A, only B, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0247] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other manners. For example, the system embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0248] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0249] In addition, each functional unit in the various embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0250] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes multiple instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0251] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the rights of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the rights of the embodiments of the present application.

Claims

1. A method for predicting protein stability, characterized in that, The method includes: Obtain the first test sequence corresponding to the original protein, and the second test sequence of the mutant protein corresponding to the original protein; Based on the attention mechanism of the pre-trained protein language model, multiple first target residue features are determined from the first test sequence, and global features are extracted from the multiple first target residue features to obtain the first virtual structure microenvironment features. Based on the attention mechanism of the protein language model, multiple second target residue features are determined from the second test sequence, and global features are extracted from the multiple second target residue features to obtain the second virtual structure microenvironment features; The first global sequence feature of the first test sequence and the second global sequence feature of the second test sequence are obtained through the protein language model. A first difference is determined based on the first virtual structure microenvironment features and the second virtual structure microenvironment features, and a second difference is determined based on the first global sequence features and the second global sequence features; Based on the first difference and the second difference, the protein stability of the mutant protein is predicted to obtain the protein stability prediction result.

2. The protein stability prediction method according to claim 1, characterized in that, The step of predicting the protein stability of the mutant protein based on the first difference and the second difference, and obtaining the protein stability prediction result, includes: The first difference between the first target residue feature and the second target residue feature is input into a pre-trained prediction model to predict the protein folding free energy of the residue feature change dimension, and a first prediction result is obtained. The second difference between the first test sequence and the second test sequence is input into the prediction model to predict the protein folding free energy of the global change dimension, and a second prediction result is obtained. Based on the average of the first prediction result and the second prediction result, the protein stability prediction result is obtained.

3. The protein stability prediction method according to claim 1, characterized in that, The determination of multiple first target residue features from the first test sequence based on the attention mechanism of a pre-trained protein language model includes: The first overall attention map corresponding to the first test sequence is generated based on the attention mechanism of the pre-trained protein language model. Global pooling is performed on the first overall attention map to obtain the first target overall attention map; From the first test sequence, determine multiple first positions of multiple mutant residue features; Based on the matching of the multiple first positions in the overall attention map of the first target, multiple first attention score vectors corresponding to the multiple mutated residue features are obtained; Obtain the number of first residue features selected, and determine the same number of first target residue features from the plurality of mutant residue features based on the plurality of first attention score vectors.

4. The protein stability prediction method according to claim 3, characterized in that, The extraction of global features from the features of the plurality of first target residues to obtain the first virtual structural microenvironment features includes: A preset first residue feature initial vector is obtained, and the first residue feature initial vector is concatenated with the plurality of first target residue features to obtain a first target feature sequence; wherein the first residue feature initial vector is located at the first position of the first target feature sequence; Based on the first residue feature initial vector, feature aggregation is performed on each first target residue feature corresponding to the first target feature sequence, as well as the interrelationships between the multiple first target residue features, to obtain a first virtual structural microenvironment feature used to characterize the global information of the first target feature sequence.

5. The protein stability prediction method according to claim 1, characterized in that, The step of determining multiple second target residue features from the second test sequence based on the attention mechanism of the protein language model includes: A second overall attention map corresponding to the second test sequence is generated based on the attention mechanism of the protein language model. The second overall attention map is subjected to global pooling to obtain the second target overall attention map; From the second test sequence, determine multiple second positions of multiple mutant residue features; Based on the matching of the multiple second positions in the overall attention map of the second target, multiple second attention score vectors corresponding to the multiple mutated residue features are obtained; The number of second residue features selected is obtained, and based on the plurality of second attention score vectors, the same number of second target residue features as the number of second residue features selected are determined from the plurality of mutant residue features.

6. The protein stability prediction method according to claim 5, characterized in that, The extraction of global features from the features of the plurality of second target residues to obtain the second virtual structure microenvironment features includes: A preset initial vector of second residue features is obtained, and the initial vector of second residue features is concatenated with the plurality of second target residue features to obtain a second target feature sequence; wherein, the initial vector of second residue features is located at the first position of the second target feature sequence; Based on the second residue feature initial vector, feature aggregation is performed on each second target residue feature corresponding to the second target feature sequence, as well as the interrelationships between the multiple second target residue features, to obtain a second virtual structural microenvironment feature used to characterize the global information of the second target feature sequence.

7. The protein stability prediction method according to claim 1, characterized in that, The step of obtaining the first global sequence features of the first test sequence and the second global sequence features of the second test sequence through the protein language model includes: A preset first global initial vector is obtained, and based on the first global initial vector, the protein language model performs feature aggregation on multiple first features in the first test sequence to obtain the first global sequence features; A preset second global initial vector is obtained, and based on the second global initial vector, the protein language model performs feature aggregation on multiple second features in the second test sequence to obtain the second global sequence features.

8. A protein stability prediction device, characterized in that, The device includes: The first acquisition module is used to acquire the first test sequence corresponding to the original protein and the second test sequence of the mutant protein corresponding to the original protein. The first extraction module is used to determine multiple first target residue features from the first test sequence based on the attention mechanism of the pre-trained protein language model, and to extract global features from the multiple first target residue features to obtain first virtual structure microenvironment features. The second extraction module is used to determine multiple second target residue features from the second test sequence according to the attention mechanism of the protein language model, and to extract global features from the multiple second target residue features to obtain the second virtual structure microenvironment features. The second acquisition module is used to acquire the first global sequence features of the first test sequence and the second global sequence features of the second test sequence through the protein language model. The determining module is used to determine a first difference based on the first virtual structure microenvironment features and the second virtual structure microenvironment features, and to determine a second difference based on the first global sequence features and the second global sequence features; The prediction module is used to predict the protein stability of the mutant protein based on the first difference and the second difference, and obtain the protein stability prediction result.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the protein stability prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the protein stability prediction method according to any one of claims 1 to 7.

Images