Protein supplement method, device, computer device, and storage medium

Abstract

The application relates to a protein completion method and device, a computer device and a storage medium. The method comprises the following steps: acquiring a training protein set; the training protein set comprises a plurality of training protein data; each training protein data comprises a protein structure with a missing functional domain, a complete protein structure, a protein sequence with a missing functional domain and a complete protein sequence; based on a preset initial model, the initial model is trained according to the training protein set to determine a target completion model; acquiring to-be-predicted protein data; the to-be-predicted protein data comprises a to-be-predicted missing protein structure, a to-be-predicted complete protein structure and a to-be-predicted missing protein sequence; based on the target completion model, protein sequence completion is performed on the to-be-predicted protein data to determine a complete protein sequence, so that the protein sequence is completely and accurately completed.

Description

Technical Field

[0001] This application relates to the field of molecular information technology, and in particular to a protein completion method, apparatus, computer device, and storage medium. Background Technology

[0002] Proteins are fundamental to life, and protein sequence completion is a core aspect of protein research in biology, crucial for understanding protein structure and function. With advancements in the biological field, artificial intelligence models have emerged, leveraging their powerful capabilities in processing large-scale data to complete protein sequences lacking functional domains.

[0003] However, current technologies cannot fully and accurately complete protein sequences, and there is an urgent need for a method that can fully and accurately complete protein sequences. Summary of the Invention

[0004] Therefore, it is necessary to provide a protein completion method, apparatus, computer equipment, and storage medium that can comprehensively and accurately complete protein sequences, addressing the aforementioned technical problems.

[0005] In a first aspect, this application provides a protein completion method. The method includes: acquiring a training protein set; the training protein set includes multiple training protein data sets; each training protein data set includes: a protein structure with a missing functional domain, a complete protein structure, a protein sequence with the missing functional domain, and a complete protein sequence; training the initial model based on a preset initial model using the training protein set to determine a target completion model; acquiring protein data to be predicted; the protein data to be predicted includes: a missing protein structure to be predicted, a complete protein structure to be predicted, and a missing protein sequence to be predicted; and performing protein sequence completion on the protein data to be predicted based on the target completion model to determine a complete protein sequence.

[0006] In one embodiment, obtaining the training protein set includes: obtaining initial protein data; the initial protein data includes: multiple complete protein structures and multiple complete protein sequences; removing the protein structure corresponding to the target functional domain from the multiple complete protein structures to obtain protein structures with missing functional domains corresponding to the multiple complete protein structures; removing the protein sequence corresponding to the target functional domain from the multiple complete protein sequences to obtain protein sequences with missing functional domains corresponding to the multiple complete protein sequences; determining the protein structures with missing functional domains, complete protein structures, protein sequences with missing functional domains, and complete protein sequences as training protein data; and constructing a training protein set based on the multiple training protein data.

[0007] In one embodiment, the step of training the initial model based on the training protein set to determine the target completion model includes: obtaining the preset initial model; converting the protein structure of the missing functional domain, the complete protein structure, and the protein sequence of the missing functional domain in the target training protein data into feature vectors; the target training protein data being any training protein data in the training protein set; determining the completed protein sequence based on the initial model and the feature vectors; and adjusting the parameters of the initial model based on the completed protein sequence and the complete protein sequence in the target training protein data to determine the target completion model.

[0008] In one embodiment, converting the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains in the target training protein data into feature vectors includes: sequentially inputting the protein structure with missing functional domains into the structural encoder and multilayer sensor adapter of the initial model to obtain a first feature vector; sequentially inputting the complete protein structure into the structural encoder and multilayer sensor adapter of the initial model to obtain a second feature vector; and inputting the protein sequence with missing functional domains into the sequence encoder of the initial model to obtain a third feature vector.

[0009] In one embodiment, determining the complete protein sequence based on the initial model and feature vectors includes: determining a target feature vector based on multiple feature vectors; determining a graph position code based on the complete protein structure using Laplacian feature vector encoding; determining a sequence position code based on the position of each amino acid in the complete protein structure; and determining the complete protein sequence by prediction based on the target feature vector, graph position code, and sequence position code using an attention mechanism and a feedforward neural network.

[0010] In one embodiment, determining the target feature vector based on multiple feature vectors includes: concatenating the first feature vector, the second feature vector, and the third feature vector to determine the target feature vector.

[0011] In one embodiment, the step of determining the graph position encoding based on the Laplacian eigenvector encoding and the complete protein structure includes: constructing an adjacency matrix based on the relative positions of each amino acid in the complete protein structure; determining the number of neighboring amino acids of each amino acid based on a preset distance threshold and the relative positions of each amino acid in the complete protein structure, and constructing a degree matrix based on the number of neighboring amino acids of each amino acid; solving the adjacency matrix and the degree matrix to determine the Laplacian eigenvector matrix and the eigenvalue diagonal matrix; a preset correspondence exists between the Laplacian eigenvector matrix and the eigenvalue diagonal matrix; and determining the graph position encoding based on the Laplacian eigenvector matrix.

[0012] Secondly, this application also provides a protein completion device. The device includes:

[0013] An acquisition module is used to acquire a training protein set; the training protein set includes multiple training protein data; each training protein data includes: a protein structure with missing functional domains, a complete protein structure, a protein sequence with missing functional domains, and a complete protein sequence;

[0014] The training module is used to train the initial model based on a preset initial model and the training protein set to determine the target completion model;

[0015] The acquisition module is also used to acquire protein data to be predicted; the protein data to be predicted includes: the missing protein structure to be predicted, the complete protein structure to be predicted, and the missing protein sequence to be predicted.

[0016] The completion module is used to complete the protein sequence of the protein to be predicted data based on the target completion model, and determine the complete protein sequence.

[0017] Thirdly, this application also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any of the methods in the first aspect above.

[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods in the first aspect described above.

[0019] The aforementioned protein completion method, apparatus, computer equipment, and storage medium acquire a training protein set, which includes multiple training protein datasets. Each training protein dataset includes: a protein structure with a missing functional domain, a complete protein structure, a protein sequence with the missing functional domain, and a complete protein sequence. Then, based on a preset initial model, the initial model is trained using the training protein set to determine a target completion model. Next, protein data to be predicted is acquired, including: a missing protein structure to be predicted, a complete protein structure to be predicted, and a missing protein sequence to be predicted. Finally, based on the target completion model, the protein sequence of the protein data to be predicted is completed to determine the complete protein sequence, thereby achieving comprehensive and accurate protein sequence completion. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the application environment of a protein completion method in one embodiment.

[0021] Figure 2 This is a flowchart illustrating a protein completion method in one embodiment;

[0022] Figure 3 This is a flowchart illustrating the protein completion method in another embodiment;

[0023] Figure 4 This is a flowchart illustrating the protein completion method in other embodiments;

[0024] Figure 5 This is a schematic diagram illustrating the process of training the initial model in other embodiments;

[0025] Figure 6 This is a flowchart illustrating the verification of the initial model in other embodiments;

[0026] Figure 7 This is a structural block diagram of a protein completion device in one embodiment;

[0027] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] The protein completion method provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Server 104 is used to execute the protein completion method. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0030] To address the aforementioned problems, in one embodiment of this application, such as Figure 2 As shown, a method for protein completion is provided, including the following steps:

[0031] Step 201: Obtain the training protein set.

[0032] The training protein set includes multiple training protein datasets; each training protein dataset includes: protein structure with missing functional domains, complete protein structure, protein sequence with missing functional domains, and complete protein sequence.

[0033] Of the four main organizational levels of proteins, primary structure is the most basic structural level, consisting of polypeptide chains formed by amino acids linked linearly in a specific sequence via peptide bonds. For example, VDSPOERASLDEN... represents such an amino acid sequence. Secondary structure is a localized, regular, and repeating folding pattern formed by hydrogen bonds between the backbone atoms (C=O and NH groups) of neighboring amino acid residues in the primary structure. Tertiary structure is the overall three-dimensional spatial structure formed by further folding and coiling of a single polypeptide chain based on the secondary structure. Quaternary structure is a functional protein complex formed by multiple polypeptide chains (called subunits or monomers), each with its own independent tertiary structure, linked together through non-covalent interactions.

[0034] It's important to note that a protein sequence is the sequential arrangement of amino acids within a protein. A protein can be characterized by its sequence; for example, a protein might be "VDSPQERASLDEN," where each letter represents a type of amino acid. A protein's 3D structure refers to the 3D structure it folds into in space, such as the quaternary structure mentioned above, where each amino acid (or even each atom within an amino acid) has x, y, and z coordinates. A protein's function refers to the biological function it performs, such as being an enzyme, performing gene editing, or reacting with certain molecules. A domain is a relatively stable substructure within a protein; for example, if a protein consists of several relatively independent globular clusters, each globular cluster is considered a domain. A functional domain is one or more domains within a protein that collectively perform a specific function.

[0035] It should be noted that in other embodiments of this application, the data in the training protein set corresponds to the target protein data. The target protein data refers to a series of proteins collected for a specific protein function to be studied, after filtering out irrelevant proteins. The protein data to be collected varies depending on the research objective. For example, to study a certain gene editing task, a Cas9 protein dataset might be collected.

[0036] It should be noted that a protein structure with a missing functional domain is the 3D structure of a protein that lacks one or more domains corresponding to a specific function. A complete protein structure is the 3D structure of the complete protein corresponding to the protein structure with a missing functional domain. A protein sequence with a missing functional domain is the protein sequence lacking the amino acid corresponding to a specific function. A complete protein sequence is the complete protein sequence corresponding to the protein sequence with a missing functional domain. Furthermore, the protein structure with a missing functional domain, the complete protein structure, the protein sequence with a missing functional domain, and the complete protein sequence belong to the same protein.

[0037] It should be noted that the data in the training protein set was obtained by searching and organizing publicly available protein databases.

[0038] Step 202: Based on the preset initial model, train the initial model according to the training protein set to determine the target completion model.

[0039] In other embodiments of this application, the preset initial model is the Evolutionary Scale Model (ESM3).

[0040] That is, the protein structures of missing functional domains, complete protein structures, and protein sequences of missing functional domains in the training protein set are predicted along with a preset initial model to obtain the completed protein sequence. Then, the initial model is adjusted based on the completed protein sequence and the complete protein sequence to finally obtain the target completed model.

[0041] Step 203: Obtain the protein data to be predicted.

[0042] The protein data to be predicted includes: the missing protein structure to be predicted, the complete protein structure to be predicted, and the missing protein sequence to be predicted.

[0043] It should be noted that the protein data to be predicted refers to protein data that needs to be supplemented in actual application scenarios.

[0044] The predicted deleted protein structure is the 3D structure of a protein that is missing one or more domains corresponding to a specific function. The predicted complete protein structure is the 3D structure of the complete protein corresponding to the predicted deleted protein structure. The predicted deleted protein sequence is the protein sequence that is missing an amino acid corresponding to a specific function.

[0045] Among them, the predicted missing protein structure, the predicted complete protein structure, and the predicted missing protein sequence belong to the same protein.

[0046] Step 204: Based on the target completion model, perform protein sequence completion on the protein data to be predicted to determine the complete protein sequence.

[0047] Among them, the complete protein sequence is the protein sequence predicted by the target completion model based on the protein data to be predicted.

[0048] That is, the protein data to be predicted is input into the target completion model, and the target completion model completes the protein sequence based on the predicted protein data, and finally obtains the complete protein sequence.

[0049] In the above protein completion method, a training protein set is acquired, comprising multiple training protein datasets. Each training protein dataset includes: a protein structure with a missing functional domain, a complete protein structure, a protein sequence with the missing functional domain, and a complete protein sequence. Then, based on a preset initial model, the initial model is trained using the training protein set to determine a target completion model. Next, protein data to be predicted is acquired, comprising: a missing protein structure to be predicted, a complete protein structure to be predicted, and a missing protein sequence to be predicted. Finally, based on the target completion model, protein sequence completion is performed on the protein data to be predicted to determine the complete protein sequence, thereby achieving comprehensive and accurate protein sequence completion.

[0050] In other embodiments of this application, obtaining the training protein set includes:

[0051] Step 1: Obtain initial protein data.

[0052] The initial protein data includes: multiple complete protein structures and multiple complete protein sequences.

[0053] The initial protein data was obtained by searching and organizing publicly available protein databases.

[0054] It should be noted that each complete protein structure in a plurality of complete protein structures corresponds to a different protein type, and each complete protein sequence in a plurality of complete protein sequences corresponds to a different protein type. Furthermore, each complete protein structure corresponds to a specific complete protein sequence. That is, for any one complete protein structure in a plurality of complete protein structures, there exists a complete protein in each of the plurality of complete protein sequences that belongs to the same protein type.

[0055] Step 2: In multiple intact protein structures, remove the protein structure corresponding to the target functional domain to obtain the protein structure with the missing functional domain corresponding to the multiple intact protein structures.

[0056] That is, the protein structures corresponding to the target functional domains in each complete protein structure are removed, thereby obtaining protein structures with missing functional domains corresponding to multiple complete protein structures.

[0057] The target functional domain is one or more structural domains that work together to achieve a specific function.

[0058] Step 3: Remove the protein sequence corresponding to the target functional domain from multiple complete protein sequences to obtain protein sequences with missing functional domains corresponding to multiple complete protein sequences.

[0059] That is, the protein sequences corresponding to the target functional domains in each complete protein sequence are removed to obtain protein sequences with missing functional domains corresponding to multiple complete protein structures.

[0060] For example, taking the protein “VDSPQERASLDEN” as an example, each letter represents a type of amino acid. The protein sequence corresponding to the target functional domain is SPQ. The protein sequence corresponding to the missing functional domain of the protein “VDSPQERASLDEN” is “VD---ERASLDEN”, where “-” indicates the amino acid deletion at this point.

[0061] Step 4: Identify the protein structures with missing functional domains, complete protein structures, protein sequences with missing functional domains, and complete protein sequences as training protein data.

[0062] It should be noted that each training protein data point corresponds to the same protein. That is, the protein structure with missing functional domains, the complete protein structure, the protein sequence with missing functional domains, and the complete protein sequence contained in a training protein data point all belong to the same protein.

[0063] Step 5: Construct a training protein set based on multiple training protein data.

[0064] It should be noted that in other embodiments of this application, multiple training protein data will be divided according to a preset ratio to obtain a training set and a validation set. The training set is used to train the initial model, and the validation set is used to verify the performance of the trained model.

[0065] In other embodiments of this application, the step of training the initial model based on the training protein set to determine the target completion model includes:

[0066] Step 1: Obtain the preset initial model.

[0067] In other embodiments of this application, the preset initial model can be a preset neural network model, a deep learning model, or an evolutionary scale model (ESM3).

[0068] Step 2: Convert the protein structures with missing functional domains, complete protein structures, and protein sequences with missing functional domains in the target training protein data into feature vectors.

[0069] The target training protein data is any training protein data in the training protein set.

[0070] It should be noted that in the process of this embodiment, the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains in the target training protein data are converted into feature vectors and executed only on the target training protein data. However, in actual application scenarios, the relevant steps of this embodiment will be executed on all training protein data in the training protein set.

[0071] It should be noted that the feature vector is a vector used to characterize the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains. An encoder can convert the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains into feature vectors.

[0072] It should be noted that, in a specific embodiment of this application, the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains are converted into three feature vectors, and then the three feature vectors are summarized to obtain the target feature vector.

[0073] Step 3: Based on the initial model and feature vectors, determine the complete protein sequence.

[0074] The feature vector is used as input to the initial model, which then makes predictions based on the feature vector to determine the complete protein sequence.

[0075] Step 4: Adjust the parameters of the initial model based on the completed protein sequence and the complete protein sequence in the target training protein data to determine the target completion model.

[0076] That is, based on the difference between the completed protein sequence and the complete protein sequence in the target training protein data, the relevant parameters of the initial model are adjusted to obtain the target completion model.

[0077] For example, in other embodiments of this application, the relevant parameters of the initial model are adjusted using LoRA fine-tuning, an efficient and lightweight method for model training. Specifically, a delta matrix is ​​added to the existing model parameter matrix. Taking the parameter matrix as an example:

[0078] ,in This is the original parameter matrix. It is a low-rank matrix, meaning it represents the change in the original parameter matrix, and its parameter dimension is low. This is the new parameter matrix. During fine-tuning, only the low-rank matrix is ​​used. The original matrix will be adjusted. It will not be directly adjusted. This method is an improvement on parameter matrix adjustment.

[0079] It is worth noting that in this embodiment, by converting the protein structure of missing functional domains, the complete protein structure, and the protein sequence of missing functional domains in the target training protein data into feature vectors, and determining the complete protein sequence based on the initial model and the feature vectors, the parameters of the initial model are finally adjusted according to the completed protein sequence and the complete protein sequence in the target training protein data to determine the target completion model, thereby improving the accuracy of predicting the complete protein sequence and improving the overall efficiency.

[0080] In other embodiments of this application, converting the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains in the target training protein data into a feature vector includes:

[0081] Step 1: Input the protein structure with missing functional domains into the structural encoder and multilayer sensor adapter of the initial model in sequence to obtain the first feature vector.

[0082] The first feature vector is a feature vector representing the protein structure with missing functional domains. The structure encoder generates a corresponding structure code based on the protein structure. The multilayer perceptron adapter generates a corresponding feature vector based on the code.

[0083] That is, the protein structure with missing functional domains is transformed into a structural code corresponding to the protein structure with missing functional domains by the structural encoder of the initial model, and then the first feature vector is generated by the multilayer perceptron adapter.

[0084] Step 2: Input the complete protein structure into the structure encoder and multilayer sensor adapter of the initial model in sequence to obtain the second feature vector.

[0085] The second feature vector is a feature vector that characterizes the complete protein structure.

[0086] That is, the complete protein structure is transformed into a structural code corresponding to the complete protein structure by the structural encoder of the initial model, and then a second feature vector is generated by the multilayer perceptron adapter.

[0087] For example, in one embodiment of this application, the multilayer perceptron adapter is an MLP adapter composed of multiple neural networks, capable of performing linear transformation operations on the input vector. For each amino acid, the formula is as follows: Where h is the input vector and W is the parameter matrix of the MLP adapter. This is the output vector.

[0088] Step 3: Input the protein sequence with missing functional domains into the sequence encoder of the initial model to obtain the third feature vector.

[0089] The third feature vector is a feature vector representing the protein sequence with missing functional domains. The sequence encoder is used to convert the protein sequence into the corresponding feature vector.

[0090] That is, the protein sequence lacking functional domains is processed by the sequence encoder of the initial model to obtain the third feature vector.

[0091] It is worth noting that by converting the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains into the first feature vector, the second feature vector, and the third feature vector, respectively, we can better characterize the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains, thus laying the foundation for determining the target completion model.

[0092] In other embodiments of this application, determining the complete protein sequence based on the initial model and feature vectors includes:

[0093] Step 1: Determine the target feature vector based on multiple feature vectors.

[0094] That is, the first feature vector, the second feature vector, and the third feature vector are concatenated to determine the target feature vector.

[0095] For example, the formula for concatenating the first feature vector, the second feature vector, and the third feature vector to determine the target feature vector is as follows: ,in, The first eigenvector, This is the second feature vector. The third eigenvector, This indicates vector concatenation. The target feature vector.

[0096] Step 2: Based on Laplacian feature vector encoding, determine the graph position encoding according to the complete protein structure.

[0097] Among them, the graph position encoding is used to characterize the relative positions of each amino acid in a protein.

[0098] That is, the Laplacian matrix is ​​constructed based on the complete protein structure and solved to finally determine the graph position encoding.

[0099] Step 3: Determine the sequence position code based on the position of each amino acid in the complete protein structure.

[0100] Sequence position encoding is used to characterize the absolute position of each amino acid in a protein, that is, the coordinates of each amino acid in space. For example, the coordinates of a certain amino acid are (X1, Y1, Z1).

[0101] Step 4: Based on the attention mechanism and feedforward neural network, predict and determine the complete protein sequence according to the target feature vector, graph position encoding and sequence position encoding.

[0102] For example, in one embodiment of this application, the attention mechanism includes a geometric attention mechanism and a multi-head attention mechanism. In the geometric attention mechanism, the target feature vector is concatenated with the graph position code, and then the concatenated result is adjusted according to the sequence position code and normalized before being used as the input of the multi-head attention mechanism.

[0103] In the geometric attention mechanism, the state corresponds to the target feature vector and the graph position encoding. This involves the fusion of structural and sequence information, where sequence information and 3D structural information are given equal importance and both are used as feature vectors throughout the model. Backbone frames correspond to the sequence position encoding r, used to calculate rotation and translation. Finally, the output is the Attention output. As input to multi-head attention.

[0104] Output of multi-head attention mechanism and geometric attention mechanism , which serves as the input to the multi-head attention mechanism, where n is the number of amino acids and d is the vector dimension. For each head... There are H headers in total, each generating an independent query matrix (Q), key matrix (K), and value matrix (V). The calculation formulas are as follows:

[0105] ,

[0106] in, , , This is a preset parameter matrix.

[0107] For each head The calculation formula is as follows: Where T represents the transpose, , where is the column dimension of the key matrix (K), i.e., the number of columns in the key matrix (K).

[0108] Finally, merge all the heads: Here, Concat() means concatenation function. The parameter matrix represents the final output of the multi-head attention mechanism. As input to a feedforward neural network.

[0109] Feedforward neural network: consists of multiple feedforward propagation layers (the number of layers can be customized), where the calculation formula for the Lth layer is: In this diagram, X is the input of the layer (i.e., the output of the upper layer), W and B are parameter matrices, and Y is the output of the layer (i.e., the input of the lower layer). The function of a feedforward neural network is to map input data to the target output through multiple layers of nonlinear transformations; it is a fundamental building block of deep learning networks.

[0110] In other embodiments of this application, determining the target feature vector based on multiple feature vectors includes: concatenating the first feature vector, the second feature vector, and the third feature vector to determine the target feature vector.

[0111] In other embodiments of this application, the step of determining the graph position encoding based on the Laplacian feature vector encoding and the complete protein structure includes:

[0112] Step 1: Construct an adjacency matrix based on the relative positions of the amino acids in the complete protein structure.

[0113] That is, the elements in the adjacency matrix represent the relative distances between each amino acid. The relative distance is the distance between each amino acid in the complete protein structure. For example, the element in the first row and first column of the adjacency matrix represents the relative distance between the first amino acid and the first amino acid, and the element in the first row and second column represents the relative distance between the first amino acid and the second amino acid.

[0114] Step 2: Based on a preset distance threshold, determine the number of neighboring amino acids of each amino acid according to the relative positions of each amino acid in the complete protein structure, and construct a degree matrix based on the number of neighboring amino acids of each amino acid.

[0115] The degree matrix is ​​a diagonal matrix, and the non-zero elements in the degree matrix represent the number of neighboring amino acids of the corresponding amino acid. For example, the elements in the first row and first column of the degree matrix represent the number of neighboring amino acids of the first amino acid.

[0116] For example, the distance between the first amino acid and each other amino acid is obtained, and each distance is compared with a preset proximity threshold. The number of amino acids that are less than the preset proximity threshold is counted and used as the number of neighboring amino acids of the first amino acid.

[0117] Step 3: Solve based on the adjacency matrix and degree matrix to determine the Laplacian eigenvector matrix and eigenvalue diagonal matrix; there is a pre-defined correspondence between the Laplacian eigenvector matrix and the eigenvalue diagonal matrix.

[0118] Step 4: Determine the graph position encoding based on the Laplacian eigenvector matrix.

[0119] For example, Laplacian eigenvectors are used to treat the entire protein as a graph. The Laplacian matrix of the graph is then calculated. Where A is the adjacency matrix and D is the degree matrix, the Laplacian matrix L is decomposed. Select the eigenvectors corresponding to the k smallest non-zero eigenvalues, where the Laplacian eigenvector matrix is... The column vectors are eigenvectors, and the eigenvalue diagonal matrix is... It is an eigenvalue diagonal matrix. and There is a pre-defined correspondence between them, where T represents the transpose of the matrix, and the graph position code is the position code corresponding to each amino acid i. .

[0120] In other embodiments of this application, D is a diagonal matrix. .

[0121] In a particular embodiment of this application, such as Figure 3 As shown, the protein completion method includes: S1: constructing a protein dataset and preprocessing the data; S2: constructing a large multimodal model of protein domains, with sub-modules including multimodal feature embedding, feature space alignment, pre-trained basic transformers, and low-rank adaptive matrices; S3: fine-tuning the large multimodal model of protein domains until convergence, saving the optimal model parameters and evaluating the accuracy; S4: model deployment and inference, generating protein missing domain completion results, saving the data and visualizing the results.

[0122] In a specific embodiment of this application, S1: Constructing a protein dataset, the data preprocessing specifically involves: searching and collecting training protein data in a publicly available protein database, separating key functional domains, and organizing them to obtain a training protein set, wherein each protein sample contains a protein structure with missing functional domains, a complete protein structure, a protein sequence with missing functional domains, and a complete protein sequence; the training protein set includes a training set and a prediction set.

[0123] In a particular embodiment of this application, such as Figure 4As shown, S2: Constructing a large multimodal model of protein domains. Sub-modules include multimodal feature embedding, feature space alignment, pre-trained basic transformers, and low-rank adaptive matrices. Specifically: S21, Feature Embedding: Using a structural encoder and a sequence encoder, 3D structures and sequences are converted into feature vectors. 3D structures with missing domains enter the structural encoder, then pass through an MLP adapter to be converted into feature vectors; 3D structures of complete proteins enter the structural encoder, then pass through an MLP adapter to be converted into feature vectors; protein sequences with missing domains enter the sequence encoder to be converted into feature vectors; all feature vectors enter S22; S22, Feature Space Alignment: Feature alignment is achieved through connection layers and positional encoding; the three sets of feature vectors from S21 are processed together... The input vector is connected to a layer, followed by graph position encoding based on protein structure and sequence position encoding based on sequence, which are then converted into aligned feature vectors and fed into S23. S23, ESM3 converter: generates a complete sequence based on geometric attention mechanism. The aligned feature vectors from S22 are fed into the geometric attention mechanism, where the weights of the geometric attention are calculated based on the complete 3D structure of the protein. After layer normalization, the QKV matrix is ​​calculated. It then passes through a multi-head attention mechanism, undergoes layer normalization, and passes through a feedforward neural network. This process is repeated N times, outputting the completed protein sequence. S24, LoRA fine-tuning module: optimizes model parameters using a low-rank matrix. For the parameters in the QKV matrix and the feedforward neural network, low-rank matrices are externalized to serve as the change in the original parameter matrix.

[0124] In other embodiments of this application, such as Figure 5 As shown, S3: Fine-tuning the protein domain multimodal large model until convergence, saving the optimal model parameters and evaluating the accuracy specifically involves: S31: Determining the training period, number of sub-expert networks, training batches, and learning rate, and initializing the entire network using pre-trained ESM3 model parameters; S32: Randomly selecting small batches of training protein data; S33: Inputting the small batches of training protein data into the protein domain multimodal large model; S34: Finally predicting and completing the protein sequence through each sub-module of the large model; S35: Backpropagating gradients, updating the adapter parameters in S2, and using LoRA to fine-tune and update the ESM converter parameters in S23; S36: Determining whether the training termination condition is met; if yes, S37: Terminating training, saving the optimal model parameters and evaluating the accuracy, and proceeding to S4; if not, re-executing S32.

[0125] In other embodiments of this application, such as Figure 6As shown, S4: Model deployment and inference, generating protein missing domain completion results, data saving and result visualization are specifically as follows: S41: Read the model and optimal parameters saved in S3; S42: Read a sample from the prediction set; S43: Input the predicted protein sample into the protein domain multimodal large model; S44: Predict and complete the protein sequence; S45: Check if there are any remaining unpredicted samples in the prediction set. If so, execute S42; otherwise, execute S46: Terminate inference, save complete data and visualize results.

[0126] It should be noted that the goal of the aforementioned protein completion method is to generate completed protein sequences that are essentially consistent with the distribution of the actual complete protein, while also identifying proteins with similar structures but significant sequence differences. Based on the protein sequences generated by this invention, new proteins with similar structures and functions to existing known proteins can be folded. This invention achieves feature space alignment where structural and sequence information interact on an equal footing, enabling protein completion and generation at the domain level. This invention uses the ESM3 model as its backbone, incorporates the idea of ​​multimodal large models with instruction fine-tuning from AI for Science, draws on the feature space alignment approach based on MLP adapters in large visual models such as MPNN, and introduces the LoRA fine-tuning method to achieve efficient protein sequence generation, effectively addressing the existing challenges mentioned above.

[0127] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0128] Based on the same inventive concept, this application also provides a protein completion device for implementing the protein completion method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more protein completion device embodiments provided below can be found in the limitations of the protein completion method described above, and will not be repeated here.

[0129] In one embodiment of this application, such as Figure 7 As shown, a protein completion device is provided, comprising:

[0130] The acquisition module 100 is used to acquire a training protein set; the training protein set includes multiple training protein data; each training protein data includes: a protein structure with missing functional domains, a complete protein structure, a protein sequence with missing functional domains, and a complete protein sequence.

[0131] The training module 200 is used to train the initial model based on the preset initial model and the training protein set to determine the target completion model;

[0132] The acquisition module 100 is also used to acquire protein data to be predicted; the protein data to be predicted includes: the missing protein structure to be predicted, the complete protein structure to be predicted, and the missing protein sequence to be predicted.

[0133] The completion module 300 is used to complete the protein sequence of the protein to be predicted data based on the target completion model, and determine the complete protein sequence.

[0134] In one embodiment of this application, the acquisition module 100 is further configured to acquire initial protein data; the initial protein data includes: multiple complete protein structures and multiple complete protein sequences; among the multiple complete protein structures, the protein structure corresponding to the target functional domain is removed to obtain a protein structure with a missing functional domain corresponding to the multiple complete protein structures; among the multiple complete protein sequences, the protein sequence corresponding to the target functional domain is removed to obtain a protein sequence with a missing functional domain corresponding to the multiple complete protein sequences; the protein structure with the missing functional domain, the complete protein structure, the protein sequence with the missing functional domain, and the complete protein sequence are determined as training protein data; and a training protein set is constructed based on the multiple training protein data.

[0135] In one embodiment of this application, the training module 200 is further configured to: acquire a preset initial model; convert the protein structure of the missing functional domain, the complete protein structure, and the protein sequence of the missing functional domain in the target training protein data into a feature vector; the target training protein data is any training protein data in the training protein set; determine the completed protein sequence based on the initial model and the feature vector; and adjust the parameters of the initial model according to the completed protein sequence and the complete protein sequence in the target training protein data to determine the target completed model.

[0136] In one embodiment of this application, the training module 200 is further configured to sequentially input the protein structure with missing functional domains into the structural encoder and multilayer sensor adapter of the initial model to obtain a first feature vector; sequentially input the complete protein structure into the structural encoder and multilayer sensor adapter of the initial model to obtain a second feature vector; and input the protein sequence with missing functional domains into the sequence encoder of the initial model to obtain a third feature vector.

[0137] In one embodiment of this application, the training module 200 is further configured to: determine a target feature vector based on multiple feature vectors; determine a graph position code based on the complete protein structure using Laplacian feature vector encoding; determine a sequence position code based on the position of each amino acid in the complete protein structure; and predict and determine the complete protein sequence based on the target feature vector, graph position code, and sequence position code using an attention mechanism and a feedforward neural network.

[0138] In one embodiment of this application, the training module 200 is further configured to perform feature alignment and concatenation based on the first feature vector, the second feature vector, and the third feature vector to determine the target feature vector.

[0139] In one embodiment of this application, the training module 200 is further configured to: construct an adjacency matrix based on the relative positions of amino acids in the complete protein structure; determine the number of neighboring amino acids of each amino acid based on a preset distance threshold and the relative positions of amino acids in the complete protein structure, and construct a degree matrix based on the number of neighboring amino acids of each amino acid; solve the adjacency matrix and degree matrix to determine the Laplacian eigenvector matrix and eigenvalue diagonal matrix; a preset correspondence exists between the Laplacian eigenvector matrix and the eigenvalue diagonal matrix; and determine the graph position encoding based on the Laplacian eigenvector matrix.

[0140] Each module in the aforementioned protein completion device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0141] In one embodiment of this application, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 8As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all relevant data for executing a protein completion method. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a protein completion method.

[0142] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0143] In one embodiment of this application, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the protein completion method in the above embodiment.

[0144] In one embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored, the computer program being executed by a processor to implement the steps of the protein completion method in the above-described method embodiments.

[0145] In one embodiment of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the protein completion method in the above-described method embodiments.

[0146] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0147] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for protein completion, characterized in that, The method includes: Obtain a training protein set; the training protein set includes multiple training protein data; each training protein data includes: protein structure with missing functional domain, complete protein structure, protein sequence with missing functional domain, and complete protein sequence; Based on a preset initial model, the initial model is trained according to the training protein set to determine the target completion model; Acquire the protein data to be predicted; the protein data to be predicted includes: the missing protein structure to be predicted, the complete protein structure to be predicted, and the missing protein sequence to be predicted. Based on the target completion model, the protein sequence of the protein to be predicted is completed to determine the complete protein sequence; The process of determining the target completion model based on a preset initial model, training the initial model using the training protein set, includes: Obtain the preset initial model; The protein structures with missing functional domains, complete protein structures, and protein sequences with missing functional domains in the target training protein data are sequentially converted into a first feature vector, a second feature vector, and a third feature vector; the target training protein data is any training protein data in the training protein set. Based on the initial model, the first feature vector, the second feature vector, and the third feature vector, the complete protein sequence is determined. Based on the completed protein sequence and the complete protein sequence in the target training protein data, the parameters of the initial model are adjusted to determine the target completion model; Based on the initial model, the first feature vector, the second feature vector, and the third feature vector, the determination of the completed protein sequence includes: Based on multiple feature vectors, the determination of the target feature vector includes: Determine the target feature vector based on multiple feature vectors; Based on Laplacian feature vector encoding, the graph position encoding is determined according to the complete protein structure; The sequence position encoding is determined based on the position of each amino acid in the complete protein structure. Based on the attention mechanism and feedforward neural network, the target feature vector, graph position encoding and sequence position encoding are used to predict and determine the complete protein sequence.

2. The protein completion method according to claim 1, characterized in that, The acquisition of the training protein set includes: Acquire initial protein data; the initial protein data includes: multiple complete protein structures and multiple complete protein sequences; In the plurality of complete protein structures, the protein structure corresponding to the target functional domain is removed to obtain a protein structure with a missing functional domain corresponding to the plurality of complete protein structures. In the plurality of complete protein sequences, the protein sequence corresponding to the target functional domain is removed to obtain a protein sequence with the missing functional domain corresponding to the plurality of complete protein sequences; The protein structures with missing functional domains, complete protein structures, protein sequences with missing functional domains, and complete protein sequences are identified as training protein data. A training protein set is constructed based on multiple training protein data sets.

3. The protein completion method according to claim 1, characterized in that, The step of converting the protein structure with missing functional domains, the complete protein structure, and the protein sequence with missing functional domains in the target training protein data into a first feature vector, a second feature vector, and a third feature vector in sequence includes: The protein structure with missing functional domains is sequentially input into the structural encoder and multilayer sensor adapter of the initial model to obtain the first feature vector; The complete protein structure is sequentially input into the structural encoder and multilayer sensor adapter of the initial model to obtain the second feature vector; The protein sequence with missing functional domains is input into the sequence encoder of the initial model to obtain the third feature vector.

4. The protein completion method according to claim 1, characterized in that, The step of determining the target feature vector based on multiple feature vectors includes: The target feature vector is determined by concatenating the first feature vector, the second feature vector, and the third feature vector.

5. The protein completion method according to claim 1, characterized in that, The Laplacian feature vector-based encoding, which determines the graph position encoding based on the complete protein structure, includes: Based on the relative positions of the amino acids in the complete protein structure, an adjacency matrix is ​​constructed. Based on a preset distance threshold, the number of neighboring amino acids of each amino acid is determined according to the relative positions of each amino acid in the complete protein structure, and a degree matrix is ​​constructed based on the number of neighboring amino acids of each amino acid. The Laplacian eigenvector matrix and eigenvalue diagonal matrix are determined by solving the adjacency matrix and degree matrix; a predetermined correspondence exists between the Laplacian eigenvector matrix and the eigenvalue diagonal matrix. The graph position encoding is determined based on the Laplacian eigenvector matrix.

6. A protein completion device, characterized in that, The device includes: An acquisition module is used to acquire a training protein set; the training protein set includes multiple training protein data; each training protein data includes: a protein structure with missing functional domains, a complete protein structure, a protein sequence with missing functional domains, and a complete protein sequence; The training module is used to train the initial model based on a preset initial model and the training protein set to determine the target completion model; The acquisition module is also used to acquire protein data to be predicted; the protein data to be predicted includes: the missing protein structure to be predicted, the complete protein structure to be predicted, and the missing protein sequence to be predicted. The completion module is used to complete the protein sequence of the protein to be predicted based on the target completion model, and determine the complete protein sequence. The training module is also used to obtain a preset initial model; to convert the protein structure of missing functional domains, the complete protein structure, and the protein sequence of missing functional domains in the target training protein data into a first feature vector, a second feature vector, and a third feature vector in sequence; the target training protein data is any training protein data in the training protein set; based on the initial model, the first feature vector, the second feature vector, and the third feature vector, the completed protein sequence is determined; according to the completed protein sequence and the complete protein sequence in the target training protein data, the parameters of the initial model are adjusted to determine the target completed model; The training module is also used to determine a target feature vector based on multiple feature vectors; determine a graph position code based on the complete protein structure using Laplacian feature vector encoding; determine a sequence position code based on the position of each amino acid in the complete protein structure; and predict and determine the complete protein sequence based on the target feature vector, graph position code, and sequence position code using an attention mechanism and a feedforward neural network.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images