Protein modification method based on amino acid knowledge graph and active learning

Abstract

The present invention discloses a protein modification method based on amino acid knowledge graph and active learning, comprising: constructing an amino acid knowledge graph based on the biochemical properties of amino acids; performing data enhancement on protein data in combination with the amino acid knowledge graph to obtain protein enhanced data and perform representation learning to obtain a first protein enhanced representation; performing representation learning on protein data, or protein data and amino acid knowledge graph using a pre-trained protein model to obtain a second protein enhanced representation; combining the first protein enhanced representation and the second protein enhanced representation to obtain the protein enhanced representation; using the protein enhanced representation as a sample, using active learning to screen representative samples from the samples and manually annotating protein properties, and using the manually annotated representative samples to train a protein property prediction model; using the protein property prediction model to perform protein modification, which can achieve rapid and accurate modification of proteins.

Description

Technical Field

[0001] The present invention belongs to the field of protein characterization learning, and specifically relates to a protein modification method based on amino acid knowledge graph and active learning. Background Art

[0002] The knowledge graph aims to objectively describe entities in the objective world and the relationships between them in the form of a graph. The amino acid knowledge graph describes the various properties and attributes of various amino acids. Based on the amino acid knowledge graph combined with deep learning methods, an embedded representation of proteins can be obtained to accelerate the prediction of various downstream protein properties, accelerate scientific research and bioindustrialization. Traditional supervised learning methods require a large amount of labeled data to learn the representation of proteins in low-dimensional space. However, obtaining protein labels requires expensive equipment and trained experts, and is time-consuming and resource-consuming. The introduction of prior knowledge of knowledge graphs can make up for the disadvantage of insufficient labeled data and the easy underfitting of the model, and at the same time assist active learning to achieve the purpose of training a model with higher prediction accuracy with less labeled data and lower cost.

[0003] Active learning is usually applied to scenarios with abundant unlabeled data and high labeling costs. Through active learning, the most representative sample set can be selected from a large number of unlabeled samples and labeled, and then the labeled samples can be applied to model training, which can achieve higher training efficiency and save manpower, material resources and time costs. Fisher Kernel, as an important kernel method that uses probabilistic generative models, has been widely used in scenarios such as protein homology detection and speech recognition. The learnable Fisher Kernel learns by making the gradients of samples of the same class as close as possible and the gradients of samples of different classes as different as possible.

[0004] Protein property prediction driven by machine learning methods avoids the traditional idea of ​​proposing hypotheses and conducting experimental verification. This method requires the experimenter to have a lot of knowledge in the field and needs years of training. At the same time, experimental verification requires a lot of manual participation, and many experiments also require expensive instruments, which is time-consuming, labor-intensive and costly. Through machine learning methods, using the experimental data accumulated from past experiments to train a machine learning model, in the subsequent exploration, a target protein sequence with good performance can be found with less or no manual experiments, which meets the experimenter's expected protein function.

[0005] The embedded representation of proteins is used as the basic data for training protein property models. Usually, the embedded representation of proteins can be based on open source pre-trained models, such as MSA-transformer, ESM, etc., or it can be mainly based on expert knowledge, such as using Geogiev coding to represent protein sequences. However, these methods have obvious disadvantages: First, directly using pre-trained models can only focus on the potential semantic knowledge of global protein sequences, and the representation effect of each specific protein sample is not necessarily good, and it does not utilize the regular knowledge that human experts have obtained for a long time; second, using human expert knowledge may cause the embedded representation of proteins to fall into the local optimum of human understanding, and it is impossible to further improve the representation ability of proteins, which limits the representation ability of proteins. Summary of the invention

[0006] In view of the above, the purpose of the present invention is to provide a protein modification method based on amino acid knowledge graph and active learning, which combines the knowledge in the amino acid knowledge graph and adopts active learning to select representative proteins, and uses representative proteins and their manual annotations to assist in the training of protein property prediction models.

[0007] In order to achieve the above-mentioned invention object, the present invention provides the following technical solutions:

[0008] A protein modification method based on amino acid knowledge graph and active learning, comprising the following steps:

[0009] Step 1: construct an amino acid knowledge graph based on the biochemical properties of amino acids;

[0010] Step 2, combining the amino acid knowledge graph to perform data enhancement on the protein data, obtain protein enhanced data and perform representation learning to obtain the first protein enhanced representation;

[0011] Step 3, using the pre-trained protein model to perform representation learning on the protein data, or the protein data and the amino acid knowledge graph, to obtain a second protein enhanced representation;

[0012] Step 4, combining the first protein enhancement representation and the second protein enhancement representation to obtain a protein enhancement representation;

[0013] Step 5, using the protein enhanced representation as a sample, using active learning to select representative samples from the sample and manually annotating the protein properties, and using the manually annotated representative samples to train a protein property prediction model;

[0014] Step 6: Use the protein property prediction model to perform protein modification.

[0015] Preferably, in the amino acid knowledge graph constructed in step 1, each triple is (amino acid, relationship, biochemical attribute value), wherein the relationship is the relationship between the amino acid and the biochemical attribute value.

[0016] Preferably, in step 2, data enhancement is performed on the protein data in combination with the amino acid knowledge graph, including: for each amino acid in each piece of protein data, a triple containing the amino acid is found from the amino acid knowledge graph, and the biochemical attributes corresponding to the amino acid in the triple are connected to the protein structure as new nodes, and the biochemical attribute values ​​are used as attribute values ​​of the new nodes. The protein data connected with the biochemical attribute values ​​are protein enhanced data.

[0017] Preferably, in step 2, a pluggable representation model is used to perform representation learning on the protein enhancement data to obtain a first protein enhancement representation, wherein the pluggable representation model includes a graph neural network model and a Transformer model.

[0018] Preferably, in step 3, when using the pre-trained protein model to perform representation learning on the protein data and the amino acid knowledge graph, the triples (amino acids, relationships, biochemical attribute values) in the amino acid knowledge graph are used as token-level additional information of the pre-trained protein model, and are combined with the protein data as input for representation learning to obtain a second protein enhanced representation.

[0019] Preferably, in step 4, the first protein enhanced expression and the second protein enhanced expression are combined in a splicing manner to obtain a protein enhanced expression.

[0020] Preferably, in step 5, during the active learning process, multiple rounds of screening of representative samples, manual annotation of protein properties of representative samples, and training of protein property prediction models are performed in an iterative cycle, wherein each iterative cycle includes:

[0021] (a) For each unlabeled sample in the sample space, calculate its Fisher Kernel distance with all labeled samples, and select the unlabeled sample farthest from all labeled samples as the representative sample according to the Fisher Kernel distance metric; repeat step (a) until k representative samples are obtained for manual labeling of protein properties, and labeled samples are obtained; in the initial round, the sample closest to the median point in the sample space is used as the initial labeled sample;

[0022] (b) using the k representative samples manually annotated in the current round of screening to train a protein property prediction model, and using the protein property prediction model trained in the current round to predict the labels of unlabeled samples to obtain the predicted labels of the unlabeled samples;

[0023] (c) Based on the Fisher Kernel distance metric between samples, k1 samples with the largest Fisher Kernel distance metric are screened from the sample space, and the Fisher Kernel is updated with the goal of making the labeled samples in the k1 samples as dissimilar as possible and the unlabeled samples as similar as possible, where k1 is the number of currently labeled samples in the sample space.

[0024] Preferably, in step 6, protein modification is performed using a protein property prediction model, comprising:

[0025] The amino acid sequence of the original protein data is changed to obtain a plurality of new protein data, and the enhanced representation of the new protein corresponding to the new protein data is obtained according to step 2 to step 4;

[0026] Using the protein property prediction model to predict the properties of the new protein enhanced representation, and obtain the predicted protein properties;

[0027] The new protein data whose predicted protein properties differ from the original protein properties corresponding to the original protein data within a threshold range are screened as modified proteins.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] An amino acid knowledge graph is constructed based on various important biochemical properties of amino acids. The amino acid knowledge graph represents the connection between the microscopic biochemical properties of amino acids, and the biochemical properties of amino acids represented by the amino acid knowledge graph are connected to protein data to enhance protein data. This allows the protein enhanced representation corresponding to the protein enhanced data to have both the data-driven protein semantic representation capability of the pre-trained protein model and the knowledge-driven protein property representation capability through the amino acid knowledge graph.

[0030] Taking protein enhanced representation as samples, active learning is used to select representative samples from the samples and manually annotate protein properties. The protein property prediction model is trained using the manually annotated representative samples, which can achieve a protein property prediction model with high prediction accuracy with as few manual sample annotations as possible.

[0031] Furthermore, since the samples of active learning are protein enhanced representations that contain both semantic representation and attribute representation, active learning can not only utilize the semantic information contained in the huge unsupervised corpus, but also capture the connection between the microscopic biochemical properties of each amino acid, thereby enhancing the effect of active learning. That is, the selected representative samples are more representative and of higher quality, which is better than training protein property prediction models;

[0032] Using protein property prediction models to modify proteins can achieve rapid and accurate protein modification. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 is a flow chart of a protein modification method based on amino acid knowledge graph and active learning provided in an embodiment;

[0035] Figure 2 is a flow chart of combining active learning to train a protein property prediction model provided by an embodiment;

[0036] Figure 3 It is a schematic diagram of data enhancement of protein data in combination with an amino acid knowledge graph provided in an embodiment. DETAILED DESCRIPTION

[0037] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0038] Figure 1 is a flow chart of a protein modification method based on amino acid knowledge graph and active learning provided in an embodiment. Figure 1 As shown, the protein modification method based on amino acid knowledge graph and active learning provided in the embodiment comprises the following steps:

[0039] Step 1: Construct an amino acid knowledge graph based on the biochemical properties of amino acids.

[0040] In the embodiments, the biochemical properties of amino acids refer to important biochemical properties selected by experts, which are generally collected from published papers, and specifically include: polarity, hydropathic index, aromatic or aliphatic, volume, isoelectric point, pKr value, molecular weight, dissociation constant (carboxyl), dissociation constant (amino), flexibility, etc., and an amino acid knowledge graph is constructed based on these biochemical property information to obtain the microscopic biochemical property connection between each amino acid for subsequent protein encoding embedding representation.

[0041] In the amino acid knowledge graph, the microscopic biochemical property relationship between amino acids is represented in the form of a triple, specifically expressed as (amino acid, relationship, biochemical property value), where the relationship is the relationship between the amino acid and the biochemical property value. For example, the triples (Aromatic, isFamilyOf, Histidine) and (Arginine, hasPKrValue, 12.48) respectively indicate that histidine belongs to the aromatic family and the pKr value of arginine is 12.48.

[0042] Step 2: Perform data enhancement on the protein data in combination with the amino acid knowledge graph to obtain protein enhanced data and perform representation learning to obtain the first protein enhanced representation.

[0043] Protein data refers to an amino acid sequence composed of several amino acids, and each piece of protein data exhibits different protein properties due to the presence of several special amino acids. Protein modification means changing the special amino acids and / or positions to make the resulting protein properties different.

[0044] like Figure 3 As shown, in the embodiment, the protein data is enhanced in combination with the amino acid knowledge graph, including: for each amino acid in each protein data, a triple containing the amino acid is found from the amino acid knowledge graph, and the biochemical attribute corresponding to the amino acid in the triple is connected to the protein structure as a new node, and the biochemical attribute value is used as the attribute value of the new node. The protein data connected with the biochemical attribute value is the protein enhanced data.

[0045] After obtaining the protein enhancement data, representation learning is performed on the protein enhancement data to obtain a first protein enhancement representation corresponding to the protein enhancement data. In an embodiment, a pluggable representation model is used to perform representation learning on the protein enhancement data to obtain a first protein enhancement representation after knowledge enhancement, and the first protein enhancement representation contains both protein topological structure and biological protein domain knowledge. Among them, the pluggable representation model includes a graph neural network model, a Transformer model, and the like. The first protein enhancement representation extracted using the graph neural network model (GNN model) includes not only the topological structure knowledge in the protein data, but also can capture the microscopic connections between amino acids that are not connected by peptide bonds.

[0046] Step 3: Use the pre-trained protein model to perform representation learning on the protein data, or the protein data and the amino acid knowledge graph, to obtain a second protein enhanced representation.

[0047] A pre-trained protein model refers to a model specifically used to extract protein embedded representations. A pre-trained MSA-transformer model can be used to perform representation learning on protein data using pre-trained protein models such as a pre-trained MSA-transformer model to obtain a second protein enhanced representation, which contains topological structure knowledge in the protein data.

[0048] Of course, the pre-trained protein model can also be used to perform representation learning on protein data and amino acid knowledge graphs, that is, the triples (amino acids, relationships, biochemical attribute values) in the amino acid knowledge graph are used as token-level additional information of the pre-trained protein model, and combined with the protein data as input for representation learning to obtain a second protein enhanced representation, which contains both protein topological structure and biological protein domain knowledge.

[0049] Step 4, combining the first protein enhancement representation and the second protein enhancement representation to obtain a protein enhancement representation.

[0050] In the embodiment, the first protein enhancement expression and the second protein enhancement expression are combined in a splicing manner to obtain a protein enhancement expression. The formula is:

[0051] x=concat(T(s),f(s,kg))

[0052] or

[0053] x=concat(T(s,kg),f(s,kg))

[0054] Among them, s represents each protein data, that is, the amino acid sequence, kg represents the amino acid knowledge graph information, f() represents the pluggable representation model, f(s,kg) represents the first protein enhanced representation obtained by learning the pluggable representation model by adding kg to s, T() represents the pre-trained protein model, T(s) represents the second protein enhanced representation of s learned by the pre-trained protein model, T(s,kg) represents the second protein enhanced representation of s learned by the pre-trained protein model, concat() represents the concatenation operation, and x represents the protein enhanced representation.

[0055] In the embodiment, the enhanced protein representation obtained by the splicing operation contains both the latent semantic information of a large amount of unsupervised protein data and the biological domain knowledge (biochemical property prior knowledge) contained in the amino acid expert knowledge graph to better characterize the protein.

[0056] Step 5: Using the protein enhanced representation as a sample, active learning is used to select representative samples from the sample and manually annotate the protein properties, and the protein property prediction model is trained using the manually annotated representative samples.

[0057] In the embodiment, each protein enhancement representation is taken as one sample, and all samples together constitute a sample space. Active learning is performed in the sample space to screen representative samples for manual annotation, and the protein property prediction model is trained using the manually annotated representative samples to efficiently improve the robustness of the protein property prediction model.

[0058] like Figure 2 As shown, in the active learning process, multiple rounds of representative sample screening, manual annotation of protein properties of representative samples, and training of protein property prediction models are performed in an iterative cycle, wherein each round of iterative cycle includes:

[0059] (a) For each unlabeled sample in the sample space, calculate its Fisher Kernel distance with all labeled samples, and select the unlabeled sample farthest from all labeled samples as the next representative sample based on the Fisher Kernel distance metric; repeat step (a) until k representative samples are obtained for manual annotation of protein properties, thus obtaining labeled samples.

[0060] It should be noted that, in the initial round, the samples closest to the median point in the sample space are selected as the initial labeled samples, and then the Fisher Kernel distances between all remaining samples in the sample space and the initial labeled samples are calculated, where k is a natural number and is set according to the application selection.

[0061] In the embodiment, calculating the Fisher Kernel distance between each unlabeled sample and all labeled samples includes:

[0062] According to the protein enhancement representation corresponding to the sample, all the first Fisher Kernel distances of each unlabeled sample and all the labeled samples are calculated, and the minimum value is selected from all the first Fisher Kernel distances as the first Fisher Kernel distance of each unlabeled sample. The larger the first Fisher Kernel distance of the sample, the greater the amount of information, which can be expressed as:

[0063]

[0064]

[0065] Where N is the total number of samples in the sample space, n is the index of the unlabeled sample number, m is the signal index of the labeled sample, k is the number of labeled samples, ||x n -x m || fk Represents the distance metric ||x n -x m ||After being processed by the Fisher Kernel (fk) method, It represents the first Fisher Kernel distance of the nth unlabeled sample from the mth labeled sample under the Fisher Kernel condition. Min() means taking the minimum value. represents the first Fisher Kernel distance of the nth unlabeled sample, x n and x m They represent the protein enhancement representation of the nth unlabeled sample and the mth labeled sample respectively;

[0066] According to the manually labeled labels of the labeled samples and the predicted labels of the unlabeled samples, all the second Fisher Kernel distances of each unlabeled sample and all the labeled samples are calculated, and the minimum value is selected from all the second Fisher Kernel distances as the second Fisher Kernel distance of each unlabeled sample. The larger the second Fisher Kernel distance of the sample, the greater the amount of information, which can be expressed as:

[0067]

[0068]

[0069] Among them, y n and m They represent the predicted label of the nth unlabeled sample and the manually labeled label of the mth labeled sample respectively; It represents the second Fisher Kernel distance of the nth unlabeled sample from the mth labeled sample under the Fisher Kernel condition. Represents the second Fisher Kernel distance of the nth unlabeled sample.

[0070] The Fisher Kernel distance of each unlabeled sample includes the first Fisher Kernel distance Second Fisher Kernel distance

[0071] In the embodiment, two methods can be used to select the unlabeled sample farthest from all labeled samples as the representative sample according to the Fisher Kernel distance metric;

[0072] Method 1: Fuse the first Fisher Kernel distance and the second Fisher Kernel distance of each unlabeled sample to obtain the first fused Fisher Kernel distance of each unlabeled sample, and select the unlabeled sample corresponding to the maximum first fused Fisher Kernel distance as the representative sample, which is expressed as follows:

[0073]

[0074] Among them, F() represents the fusion operation, Represents the first fused Fisher Kernel distance.

[0075] Method 2: Fuse the first Fisher Kernel distance and the second Fisher Kernel distance of each unlabeled sample relative to each labeled sample to obtain the second fused Fisher Kernel distance of each unlabeled sample relative to each labeled sample, and select the unlabeled sample corresponding to the largest second fused Fisher Kernel distance as the representative sample, which is expressed as follows:

[0076]

[0077] in, Represents the second fused Fisher Kernel distance.

[0078] (b) Use the k manually annotated representative samples screened in the current round to train a protein property prediction model, and use the protein property prediction model trained in the current round to predict the labels of unlabeled samples to obtain the predicted labels of the unlabeled samples.

[0079] In the embodiment, k representative samples screened in each round are used for training the protein property prediction model after manual annotation. After each round of training, the trained protein property prediction model is used to predict the labels of unlabeled samples to obtain the predicted labels of the unlabeled samples. The predicted labels are also used to construct the loss function of the current round of active learning to update the Fisher Kernel parameters of the active learning.

[0080] In the embodiment, the protein property prediction model is used to predict the properties of the protein, and a variety of pluggable regression models can be used, or a custom model framework can be used, or a regression model encapsulated and directly called in keras, sklearn, or xgboost to predict the properties of the protein.

[0081] (c) Based on the Fisher Kernel distance metric between samples, k1 samples with the largest Fisher Kernel distance metric are screened from the sample space, and the Fisher Kernel is updated with the goal of making the labeled samples in the k1 samples as dissimilar as possible and the unlabeled samples as similar as possible, where k1 is the number of currently labeled samples in the sample space.

[0082] In step (c), the calculation method of the Fisher Kernel distance between samples is basically the same as the calculation method of the Fisher Kernel distance between each unlabeled sample and all labeled samples in step (a). The difference is that the calculation object of the Fisher Kernel distance is different. In step (c), it is not confirmed whether the sample is labeled. Specifically, the Fisher Kernel distance measurement between samples includes:

[0083] According to the protein enhancement representation corresponding to the sample, all the first Fisher Kernel distances between each sample and all other samples are calculated, and the minimum value is selected from all the first Fisher Kernel distances as the first Fisher Kernel distance of each sample, where the sample includes labeled samples and unlabeled samples, which is expressed as:

[0084]

[0085]

[0086] Among them, N is the total sample size of the sample space, i and j are sample serial number indexes, ‖x i -x j ‖ fk Denotes the distance metric ‖x i -x j ‖After being processed by the Fisher Kernel (fk) method, It represents the first Fisher Kernel distance of the Fisher Kernel condition of the i-th sample to the j-th sample. Min() represents the minimum value. represents the first Fisher Kernel distance of the i-th sample, x i and x jThey represent the protein enhancement representation of the i-th sample and the j-th sample respectively;

[0087] According to the manually labeled labels of the labeled samples and / or the predicted labels of the unlabeled samples, all the second Fisher Kernel distances of each sample with all other samples are calculated, and the minimum value is selected from all the second Fisher Kernel distances as the second Fisher Kernel distance of each sample, which is expressed as follows:

[0088]

[0089]

[0090] Among them, y i and j They represent the label of the i-th sample (manually labeled label or predicted label) and the label of the j-th sample (manually labeled label or predicted label). It represents the second Fisher Kernel distance of the ith sample from the jth sample under the Fisher Kernel condition. Represents the second Fisher Kernel distance of the i-th sample;

[0091] The Fisher Kernel distance of each sample includes the first Fisher Kernel distance Second FisherKernel distance

[0092] Based on this, in the embodiment, two methods can be used to screen out k1 samples with the largest Fisher Kernel distance metric from the sample space, including:

[0093] Method 1: The first Fisher Kernel distance and the second Fisher Kernel distance of each sample are fused to obtain the first fused Fisher Kernel distance of each sample. The samples corresponding to the first fused Fisher Kernel distance with the largest k1 before screening are the k1 samples obtained by screening.

[0094] Method 2: The first Fisher Kernel distance and the second Fisher Kernel distance of each sample relative to each other sample are fused to obtain the second fused Fisher Kernel distance of each sample relative to each other sample. The unlabeled samples corresponding to the second fused Fisher Kernel distance with the largest k1 before screening are the k1 samples obtained by screening.

[0095] In the embodiment, the metric learning method combined with Fisher Kernel can be used to better distinguish the vectorized data and continuously update the Fisher Kernel in the model. The use of different learnable Fisher Kernel for calculation makes the differentiation of Fisher Kernel distance metrics of different samples more sensitive. The expert knowledge contained in the protein knowledge graph and the latent semantic knowledge represented by the pre-trained protein model are combined to assist active learning. The samples selected in this way are more representative. These representative samples can be manually labeled and then used to train the protein property prediction model to improve training efficiency and reduce the time and cost of labeling.

[0096] Step 6: Use the protein property prediction model to perform protein modification.

[0097] In the embodiment, protein modification is performed using a protein property prediction model, including:

[0098] The amino acid sequence of the original protein data is changed to obtain a plurality of new protein data, and the enhanced representation of the new protein corresponding to the new protein data is obtained according to step 2 to step 4;

[0099] Using the protein property prediction model to predict the properties of the new protein enhanced representation, and obtain the predicted protein properties;

[0100] The new protein data whose predicted protein properties differ from the original protein properties corresponding to the original protein data within a threshold range are screened as modified proteins.

[0101] In the embodiment, the threshold range is customized according to the application requirements. When the amino acid sequence of the original protein data is changed, the special amino acids that play a role in the protein properties are generally selected to replace the original position or adjust the amino acid position. In view of obtaining multiple new protein data, due to the powerful computing power of the protein property prediction model, the properties of all new protein data are predicted at the same time, and the protein structure that can be modified is screened according to the prediction results.

[0102] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A protein modification method based on amino acid knowledge graph and active learning, characterized in that: The following steps are involved: Step 1: construct an amino acid knowledge graph based on the biochemical properties of amino acids, where each triple is (amino acid, relationship, biochemical property value), and the relationship is the relationship between the amino acid and the biochemical property value; Step 2, combining the amino acid knowledge graph to perform data enhancement on the protein data, including: for each amino acid in each piece of protein data, finding a triple containing the amino acid from the amino acid knowledge graph, and connecting the biochemical attribute corresponding to the amino acid in the triple as a new node to the protein structure, and using the biochemical attribute value as the attribute value of the new node, the protein data connected with the biochemical attribute value is the protein enhanced data, and representation learning is performed on the obtained protein enhanced data to obtain a first protein enhanced representation; Step 3, using the pre-trained protein model to perform representation learning on the protein data, or the protein data and the amino acid knowledge graph, to obtain a second protein enhanced representation, wherein when using the pre-trained protein model to perform representation learning on the protein data and the amino acid knowledge graph, triples (amino acids, relationships, biochemical attribute values) in the amino acid knowledge graph are used as token-level additional information of the pre-trained protein model, and are combined with the protein data as input to perform representation learning to obtain the second protein enhanced representation; Step 4, combining the first protein enhancement representation and the second protein enhancement representation to obtain a protein enhancement representation; Step 5, using the protein enhanced representation as a sample, using active learning to select representative samples from the samples and manually annotating the protein properties, and using the manually annotated representative samples to train a protein property prediction model; Step 6: Use the protein property prediction model to perform protein modification.

2. The protein modification method based on amino acid knowledge graph and active learning according to claim 1, characterized in that: In step 2, a pluggable representation model is used to perform representation learning on the protein enhancement data to obtain a first protein enhancement representation, wherein the pluggable representation model includes a graph neural network model and a Transformer model.

3. The protein modification method based on amino acid knowledge graph and active learning according to claim 1, characterized in that: In step 4, the first protein enhancement expression and the second protein enhancement expression are combined in a splicing manner to obtain a protein enhancement expression.

4. The protein modification method based on amino acid knowledge graph and active learning according to claim 1, characterized in that: In step 5, in the active learning process, multiple rounds of representative sample screening, manual labeling of protein properties of representative samples, and training of protein property prediction models are performed in an iterative cycle, wherein each iterative cycle includes: (a) For each unlabeled sample in the sample space, calculate its Fisher Kernel distance with all labeled samples, and select the unlabeled sample farthest from all labeled samples as the representative sample according to the Fisher Kernel distance metric; repeat step (a) until k representative samples are obtained for manual labeling of protein properties, and labeled samples are obtained; in the initial round, the sample closest to the median point in the sample space is used as the initial labeled sample; (b) Using the k representative samples manually annotated in the current round of screening to train a protein property prediction model, and using the protein property prediction model trained in the current round to predict the labels of unlabeled samples to obtain the predicted labels of the unlabeled samples; (c) Based on the Fisher Kernel distance metric between samples, k1 samples with the largest Fisher Kernel distance metric are screened from the sample space, and the Fisher Kernel is updated with the goal of making the labeled samples as dissimilar as possible and the unlabeled samples as similar as possible among the k1 samples, where k1 is the number of currently labeled samples in the sample space.

5. The protein modification method based on amino acid knowledge graph and active learning according to claim 4, characterized in that: In step (a), the Fisher Kernel distance between each unlabeled sample and all labeled samples is calculated, including: According to the protein enhancement representation corresponding to the sample, all the first Fisher Kernel distances of each unlabeled sample and all the labeled samples are calculated, and the minimum value is selected from all the first Fisher Kernel distances as the first Fisher Kernel distance of each unlabeled sample. The larger the first Fisher Kernel distance of the sample, the greater the amount of information; According to the manually labeled labels of the labeled samples and the predicted labels of the unlabeled samples, all the second Fisher Kernel distances of each unlabeled sample and all the labeled samples are calculated, and the minimum value is selected from all the second Fisher Kernel distances as the second Fisher Kernel distance of each unlabeled sample. The larger the second Fisher Kernel distance of the sample, the greater the amount of information; The Fisher Kernel distance of each unlabeled sample includes the first Fisher Kernel distance and the second Fisher Kernel distance; In step (a), the unlabeled samples that are farthest from all labeled samples are selected as representative samples according to the Fisher Kernel distance metric, including: The first Fisher Kernel distance and the second Fisher Kernel distance of each unlabeled sample are fused to obtain the first fused Fisher Kernel distance of each unlabeled sample, and the unlabeled sample corresponding to the maximum first fused Fisher Kernel distance is selected as the representative sample; Alternatively, the first Fisher Kernel distance and the second Fisher Kernel distance of each unlabeled sample relative to each labeled sample are fused to obtain the second fused Fisher Kernel distance of each unlabeled sample relative to each labeled sample, and the unlabeled sample corresponding to the maximum second fused Fisher Kernel distance is selected as the representative sample.

6. The protein modification method based on amino acid knowledge graph and active learning according to claim 4, characterized in that: In step (c), the Fisher Kernel distance metric between samples includes: Calculate all Fisher Kernel distances between each sample and all other samples according to the protein enhancement representation corresponding to the sample, and select the minimum value from all Fisher Kernel distances as the first Fisher Kernel distance of each sample. The larger the first Fisher Kernel distance of the sample, the greater the information content of the sample. The samples include labeled samples and unlabeled samples. Calculate all Fisher Kernel distances between each sample and all other samples based on the manually labeled labels of the labeled samples and / or the predicted labels of the unlabeled samples, and select the minimum value from all Fisher Kernel distances as the second Fisher Kernel distance of each sample. The larger the second Fisher Kernel distance of the sample, the greater the information content of the sample. The Fisher Kernel distance of each sample includes the first Fisher Kernel distance and the second Fisher Kernel distance; In step (c), k1 samples with the largest Fisher Kernel distance metric are screened in the sample space, including: The first Fisher Kernel distance and the second Fisher Kernel distance of each sample are fused to obtain the first fused Fisher Kernel distance of each sample. The samples corresponding to the first fused Fisher Kernel distance with the largest k1 before screening are the k1 samples obtained by screening. Alternatively, the first Fisher Kernel distance and the second Fisher Kernel distance of each sample relative to each other sample are fused to obtain the second fused Fisher Kernel distance of each sample relative to each other sample, and the unlabeled samples corresponding to the second fused Fisher Kernel distance with the largest k1 before screening are the k1 samples obtained by screening.

7. The protein modification method based on amino acid knowledge graph and active learning according to claim 1, characterized in that: In step 6, protein modification is performed using the protein property prediction model, including: The amino acid sequence of the original protein data is changed to obtain a plurality of new protein data, and the enhanced representation of the new protein corresponding to the new protein data is obtained according to step 2 to step 4; Using the protein property prediction model to predict the properties of the new protein enhanced representation, and obtain the predicted protein properties; The new protein data whose predicted protein properties differ from the original protein properties corresponding to the original protein data within a threshold range are screened as modified proteins.

Images