Unified protein modification method, device and computer equipment

Abstract

The present application relates to a unified protein modification method, apparatus and computer equipment, which fine-tune the initial protein representation model based on a text-protein pair data set, align the human language large model and the initial protein representation model in the semantic space, and obtain a target protein representation model; input the obtained optimization target into the human language large model to obtain a first representation; based on the obtained key mutation sites, generate a set of protein mutants; input the set of protein mutants into the target protein representation model to obtain a second representation corresponding to each of the protein mutants; based on the second representation matching the first representation, obtain a protein mutant that meets the optimization target, solve the problem of poor generalization of protein function prediction, improve the function prediction performance of unknown proteins and mutants, and achieve efficient protein modification effect.

Description

Technical Field

[0001] The present application relates to the field of protein engineering and biomedical technology, and in particular to a unified protein modification method, device and computer equipment. Background Art

[0002] Protein is one of the most important biological macromolecules in living organisms. With the rapid development of modern biology and biotechnology, protein engineering has become a key field, aiming to obtain proteins with improved or new functions by designing and modifying natural proteins. Protein modification is not only of great significance in basic scientific research, but also shows great application potential in the fields of medicine, agriculture, and industry. For example, by modifying enzymes, their catalytic efficiency and stability can be improved, so that they can be used to produce more efficient biocatalysts; by modifying antibodies, their specificity and affinity can be enhanced, so that more effective therapeutic drugs can be developed.

[0003] In recent years, with the rapid development of computational biology, machine learning and artificial intelligence, the field of protein modification has also ushered in new opportunities and challenges. Common protein modification methods mainly use protein language models (PLMs) to predict protein functions, thereby ensuring that the predicted new proteins meet the target functional requirements.

[0004] However, existing protein language model-based function prediction usually adopts the "pre-training-fine-tuning" paradigm, which requires fine-tuning the model separately for each downstream task. The process is cumbersome and resource-consuming, and lacks sufficient generalization ability when facing new tasks, resulting in inefficient protein function prediction and restricting the development of protein modification. Summary of the invention

[0005] In this embodiment, a unified protein modification method, apparatus and computer device are provided to solve the problems of poor generalization and low efficiency of protein function prediction in related technologies.

[0006] In a first aspect, a unified protein modification method is provided in this embodiment, the method comprising:

[0007] Based on the text-protein pair dataset, fine-tune the initial protein representation model so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model;

[0008] Inputting the obtained optimization target into the human language macromodel to obtain a first representation;

[0009] Based on the acquired key mutation sites, a set of protein mutants is generated; the set of protein mutants is input into the target protein characterization model to obtain a second characterization corresponding to each of the protein mutants;

[0010] Based on the second characterization matching the first characterization, a protein mutant meeting the optimization goal is obtained.

[0011] In some embodiments, the set of protein mutants is input into the target protein characterization model to obtain the second characterization corresponding to each of the protein mutants, including:

[0012] Based on the set of protein mutants, obtaining sequence information and structural information of each of the protein mutants;

[0013] Inputting the sequence information and the structure information into the target protein representation model, wherein the target protein representation model includes a protein language model and a protein structure encoder;

[0014] Based on the protein language model, obtaining a sequence representation from the sequence information;

[0015] Based on the protein structure encoder, obtain a structural representation from the structural information;

[0016] The sequence representation and the structure representation are fused to obtain a second representation.

[0017] In some of the embodiments, obtaining a structural representation from the structural information based on the protein structure encoder comprises:

[0018] The structural information is represented as an undirected graph, the undirected graph includes edge features and node features, the node features include amino acid type, spatial position and the sequence representation;

[0019] The undirected graph is input into the protein structure encoder to obtain a structural representation.

[0020] In some of the embodiments, the sequence representation and the structure representation are fused to obtain a second representation, including:

[0021] Encoding the sequence representation into sequence global features and sequence local features;

[0022] encoding the structural representation into structural global features and structural local features;

[0023] The sequence global feature, the sequence local feature, the structure global feature and the structure local feature are weightedly concatenated to obtain a second representation.

[0024] In some embodiments, based on the text-protein pair dataset, fine-tuning the initial protein representation model so that the human language model and the initial protein representation model are aligned in the semantic space to obtain the target protein representation model includes:

[0025] Constructing a text-protein pair dataset based on existing protein data; the text-protein pair dataset includes protein sequence and structure data and corresponding performance description text;

[0026] Extracting text representation from the performance description text based on a large model of human language;

[0027] extracting a protein representation from the protein sequence and structure data based on an initial protein representation model;

[0028] By using the matching text representation and the protein representation as well as the unmatched text representation and the protein representation, based on contrastive learning training, the initial protein representation model is fine-tuned to align the human language large model and the initial protein representation model in the semantic space, thereby obtaining a target protein representation model.

[0029] In some of the embodiments, the method for generating the performance description text includes:

[0030] Obtain label data of existing protein data;

[0031] Based on a predefined text template, the label data is converted into performance description text.

[0032] In some embodiments, the obtained optimization target is input into the human language macro model to obtain a first representation, including:

[0033] In response to the optimization target submitted by the user, convert the optimization target into a target text description;

[0034] Based on the human language large model, a first representation is extracted from the target text description.

[0035] In some embodiments, obtaining a protein mutant that meets the optimization goal based on the second characterization that matches the first characterization includes:

[0036] calculating a similarity score between the first representation and the second representation;

[0037] Based on the similarity scores, the plurality of protein mutants are ranked and the modified proteins meeting the optimization goal are screened.

[0038] In a second aspect, a unified protein modification device is provided in this embodiment, the device comprising:

[0039] A model alignment module is used to fine-tune the initial protein representation model based on the text-protein pair dataset, so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model;

[0040] A first representation extraction module, used for inputting the obtained optimization target into the human language macro model to obtain a first representation;

[0041] A second representation extraction module is used to generate a set of protein mutants based on the acquired key mutation sites; input the set of protein mutants into the target protein representation model to obtain a second representation corresponding to each of the protein mutants;

[0042] A matching screening module is used to obtain a protein mutant that meets the optimization goal based on the second representation that matches the first representation.

[0043] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the unified protein modification method described in the first aspect is implemented.

[0044] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the unified protein modification method described in the first aspect is implemented.

[0045] Compared with the related art, the unified protein modification method, apparatus and computer equipment provided in this embodiment fine-tune the initial protein representation model based on the text-protein pair data set, so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model; the obtained optimization target is input into the human language large model to obtain a first representation; based on the obtained key mutation sites, a set of protein mutants is generated; the set of protein mutants is input into the target protein representation model to obtain a second representation corresponding to each of the protein mutants; based on the second representation matching the first representation, a protein mutant that meets the optimization target is obtained, which solves the problem of poor generalization of protein function prediction, improves the function prediction performance of unknown proteins and mutants, and achieves efficient protein modification effect.

[0046] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 The hardware structure block diagram of the terminal of the unified protein modification method in the embodiment of the present application;

[0049] Figure 2 This is a schematic diagram of the unified protein modification method in the examples of this application;

[0050] Figure 3 This is a schematic diagram of the process of modifying proteins based on text matching in one embodiment of the present application;

[0051] Figure 4 A schematic diagram of a process for model training for protein performance prediction in one embodiment of the present application;

[0052] Figure 5 This is a schematic diagram of a unified protein modification method in a preferred embodiment of the present application;

[0053] Figure 6 This is a structural block diagram of a unified protein modification device in an embodiment of the present application.

[0054] Figure numerals: 102, processor; 104, memory; 106, transmission device; 108, input and output device; 61, model alignment module; 62, first representation extraction module; 63, second representation extraction module; 64, matching screening module. DETAILED DESCRIPTION

[0055] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0056] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with general skills in the technical field to which this application belongs. The words "one", "a", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" may mean: A exists alone, A and B exist at the same time, and B exists alone. Generally, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0057] The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. For example, running on a terminal, Figure 1 : is a hardware structure block diagram of a terminal of the unified protein modification method of this embodiment. Figure 1 As shown, the terminal may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 and a memory 104 for storing data, wherein the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA. The above terminal may also include a transmission device 106 and an input and output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0058] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the unified protein modification method in the present embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0059] The transmission device 106 is used to receive or send data via a network. The above network includes a wireless network provided by the communication provider of the terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

[0060] In this embodiment, a unified protein engineering method is provided. Figure 2 is a flow chart of the unified protein modification method of this embodiment, such as Figure 2 As shown, the process includes the following steps:

[0061] Step S210, based on the text-protein pair dataset, fine-tune the initial protein representation model so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model.

[0062] Specifically, the human language large model (LLM) is a model pre-trained on large-scale natural language data. It has good language understanding capabilities and can handle complex text semantic tasks. The text-protein pair dataset is used as the training dataset to train the initial protein representation model to achieve semantic alignment, so that the model can understand and capture the intrinsic connection between the protein and its text description. During the training process, the human language large model with more parameters can be locked, and only the initial protein representation model with a relatively small parameter scale can be fine-tuned to improve training efficiency. The training method can adopt contrastive learning, joint training, and knowledge distillation.

[0063] Step S220: input the obtained optimization target into the human language model to obtain a first representation.

[0064] Specifically, the optimization target can be submitted by the user in real time, or can be generated uniformly by the program based on the research task. The specific optimization target acquisition method is not limited in this embodiment. Among them, the first representation is the embedded representation, that is, the description is encoded through the target corpus representation, and the text embedding is extracted to ensure that the model understands the functional information contained in the text.

[0065] Step S230, generating a set of protein mutants based on the acquired key mutation sites; inputting the set of protein mutants into the target protein characterization model to obtain a second characterization corresponding to each protein mutant.

[0066] Specifically, based on expert experience or artificial intelligence models, the mutation position is determined in the existing protein, and the original amino acid at the mutation position is replaced by the candidate amino acid to complete the mutation transformation. For the batch-modified protein sequences, a protein mutation database can be built in advance. Once the optimization target is obtained, the protein mutation database is screened based on the optimization target to obtain matching high-quality modified proteins. Among them, the second representation is the embedded representation, that is, the protein is encoded through the target protein representation model and the opal embedding is extracted.

[0067] Step S240, obtaining a protein mutant that meets the optimization goal based on the second representation that matches the first representation.

[0068] Specifically, the protein mutants that meet the optimization goals can be applied in basic scientific research, medicine, agriculture, industry, etc. For example, by modifying enzymes, their catalytic efficiency and stability can be improved, so as to produce more efficient biocatalysts; by modifying antibodies, their specificity and affinity can be enhanced, so as to develop more effective therapeutic drugs.

[0069] By fine-tuning the initial protein representation model based on the text-protein pair dataset, the human language large model and the initial protein representation model are aligned in the semantic space to obtain the target protein representation model; the obtained optimization target is input into the human language large model to obtain the first representation; based on the obtained key mutation sites, a set of protein mutants is generated; the set of protein mutants is input into the target protein representation model to obtain the second representation corresponding to each protein mutant; based on the second representation that matches the first representation, protein mutants that meet the optimization target are obtained, which solves the problem of poor generalization of protein function prediction. There is no need to fine-tune or train each task separately. The model after training can realize function prediction of unknown proteins and mutants, which solves the shortcomings of the existing technology in generalization and achieves efficient protein modification effect.

[0070] In some of the embodiments, see Figure 3, input the set of protein mutants into the target protein characterization model to obtain the second characterization corresponding to each protein mutant, including:

[0071] Step S231, based on the set of protein mutants, obtaining sequence information and structural information of each protein mutant.

[0072] Step S232, inputting the sequence information and the structure information into a target protein representation model, wherein the target protein representation model includes a protein language model and a protein structure encoder.

[0073] Step S233, obtaining a sequence representation from the sequence information based on the protein language model.

[0074] Step S234, obtaining a structural representation from the structural information based on the protein structure encoder.

[0075] Step S235: fuse the sequence representation and the structure representation to obtain a second representation.

[0076] In some of the embodiments, the structural representation is obtained from the structural information based on the protein structure encoder, including:

[0077] Step S234-a, the structural information is represented as an undirected graph, the undirected graph includes edge features and node features, the node features include amino acid type, spatial position and sequence representation.

[0078] Step S234-b, input the undirected graph into the protein structure encoder to obtain a structural representation.

[0079] Specifically, the protein structure of the input model is represented as an undirected graph G=(V,E)G=(V,E): Node VV: Each amino acid is a node, and the node features include amino acid type, position (spatial coordinates), and ESM sequence embedding. Edge EE: The connection relationship of the edge is defined according to the spatial distance (such as atom pairs within 8 Å) or chemical bonds. The edge features can include distance information and bond type. At the input stage of the model, the initial features of the node are designed as the concatenation of the ESM sequence representation Eseq\mathbf{E}_{\text{seq}} and geometric features:

[0080] Xnode=Concat(Eseq,Estruct)\mathbf{X}_{\text{node}}=\text{Concat}(\mathbf{E}_{\text{seq}},\mathbf{E}_{\text{struct}}).

[0081] The formula is: X node =Concat(E seq ,Estruct ), where X node represents the node feature, E seq represents the sequence representation, E struct Indicates structural characteristics (amino acid type, spatial position).

[0082] In some of the embodiments, the sequence representation and the structure representation are fused to obtain a second representation, including:

[0083] Step S235 - a , encoding the sequence representation into sequence global features and sequence local features.

[0084] Step S235 - b , encoding the structural representation into structural global features and structural local features.

[0085] Step S235 - c , weighted concatenation of the sequence global features, the sequence local features, the structure global features, and the structure local features to obtain a second representation.

[0086] Specifically, the formula for weighted splicing fusion is expressed as:

[0087] E fused =α[E seq,global ,E struct,global ]+β[E seq,local ,E struct,local ], where E fused For the second representation, E seq,global represents the global characteristics of the sequence, E struct,global Represents the global characteristics of the structure, E seq,local represents the local features of the sequence, E struct,local represents the local structural features, and α and β are weights.

[0088] This embodiment is a hierarchical fusion. First, the sequence feature Eseq\mathbf{E}_{\text{seq}} and the structure feature Estruct\mathbf{E}_{\text{struct}} are encoded into global and local features respectively.

[0089] Eseq,global\mathbf{E}_{\text{seq,global}} and Estruct,global\mathbf{E}_{\text{struct,global}} represent global context features.

[0090] Eseq,local\mathbf{E}_{\text{seq,local}} and Estruct,local\mathbf{E}_{\text{struct,local}} represent local detail features.

[0091] Use the weighted splicing fusion method: Efused=α[Eseq,global,Estruct,global]+β[Eseq,local,Estruct,local]\mathbf{E}_{\text{fused}}=\alpha\cdot[\mathbf{E}_{\text{seq,global}},\mathbf{E}_{\text{struct,global}}]+\beta\cdot[\mathbf{E}_{\text{seq,local}},\mathbf{E}_{\text{struct,local}}], where α\alpha and β\beta are weights dynamically adjusted during task optimization.

[0092] In some of the embodiments, see Figure 4 Based on the text-protein pair dataset, fine-tune the initial protein representation model to align the human language model and the initial protein representation model in the semantic space to obtain the target protein representation model, including:

[0093] Step S211, construct a text-protein pair data set based on existing protein data; the text-protein pair data set includes protein sequence and structure data and corresponding performance description text. The structure data comes from database query, and if it does not exist in the database, the Alphafold2 software is used to predict the protein structure.

[0094] Specifically, classification tasks and regression tasks are performed on existing protein data to obtain protein function labels; function labels include classification labels or numerical labels. Use text templates to convert numerical or category labels into natural language descriptions to further generate text-protein pair data sets. For example, a protein sequence: XXXX...XXXX; a text description of the function of this protein: This protein has medium stability, and the text description and protein sequence constitute a text-protein pair. In one embodiment, for the same function label, text templates with different description methods are expanded to improve the diversity of description texts.

[0095] Step S212: extracting text representation from the performance description text based on the human language macro model.

[0096] Step S213: extracting protein representation from protein sequence and structure data based on the initial protein representation model.

[0097] Step S214, using the matched text representations and protein representations and the unmatched text representations and protein representations, based on contrastive learning training, fine-tune the initial protein representation model, align the human language large model and the initial protein representation model in the semantic space, and obtain the target protein representation model.

[0098] The protein representation and text representation in this embodiment, as well as the first representation and the second representation in the above embodiment are embeddings. Specifically, the extracted text and protein embeddings are used to train the model through a comparative learning strategy, that is, to minimize the embedding distance between matching text-protein pairs and maximize the embedding distance between unmatched pairs, so as to train the alignment ability between the protein representation model and the human language large model.

[0099] In this embodiment, by contrastive learning to optimize the alignment of text and representation, the model successfully overcomes the problem of difficulty in balancing the weights of different tasks in multi-task learning and improves the unified prediction capability for various protein function tasks.

[0100] In some of the embodiments, see Figure 4 ,The initial protein representation model includes the initial protein language model (Pre-trained Language Mode,PLM) and the initial protein structure encoder (ProteinMPNN).

[0101] The protein language model (PLM) is a machine learning model used to learn representations from large amounts of protein sequence data and can effectively capture the semantic information of proteins.

[0102] The protein structure encoder uses graph neural networks (GNNs) to model the three-dimensional structure of proteins and use this structural information to predict the possible amino acid sequence at each position.

[0103] In some of the embodiments, see Figure 4 , based on the above step S213, extracting protein representation from protein sequence and structure data based on the initial protein representation model, including:

[0104] Step S213a, based on the initial protein language model, the sequence data of the protein sequence is processed to obtain an initial sequence representation.

[0105] Step S213b, processing the structural data of the protein sequence based on the initial protein structure encoder to obtain an initial structural representation.

[0106] Specifically, we use the open source protein structure encoding model (ProteinMPNN), and the protein structure of the input model is represented as an undirected graph G=(V,E)G=(V,E): Node VV: Each amino acid is a node, and the node features include amino acid type, position (spatial coordinates), and ESM sequence embedding. Edge EE: The connection relationship of the edge is defined according to the spatial distance (such as atomic pairs within 8 Å) or chemical bonds. The edge features can include distance information and bond type. At the input stage of the model, the initial features of the node are designed as a splicing of sequence representation and geometric features: X node =Concat(E seq ,E struct ), where X node represents the node feature, E seq represents the sequence representation, E struct Indicates structural characteristics (amino acid type, spatial position).

[0107] Step S213c, fusing the initial sequence representation and the initial structure representation to obtain a protein representation.

[0108] Specifically, the initial sequence representation and the initial structure representation are encoded as global and local features respectively, and the global and local features are weightedly spliced ​​and fused to obtain the protein representation.

[0109] In this embodiment, the understanding of protein structure is improved by combining a protein language model and a protein structure encoder.

[0110] In some of the embodiments, the method for generating a performance description text includes:

[0111] Step S310, obtaining label data of existing protein data.

[0112] Step S320: converting the label data into performance description text based on a predefined text template.

[0113] Specifically, existing protein data usually contain functional labels of proteins (classification labels or numerical labels). According to the characteristics of proteins, multiple templates are pre-designed to ensure that the generated performance text description has rich semantic information and can cover different types of protein functions.

[0114] In the classification task of protein data, the function of the protein is usually given in the form of a category label, such as "cell membrane localization", "nuclear localization", etc. In order to convert these classification labels into text descriptions, different text templates are designed for each classification task to increase the diversity of text descriptions and the richness of expression. For example, for the label "cell membrane localization", the following different text descriptions can be generated: "This protein is localized in the cell membrane"; "This protein is mainly present on the cell membrane"; "This protein is involved in the formation and maintenance of the cell membrane".

[0115] In the regression task of protein data, the function of the protein is usually given in numerical form, such as the fitness value represents fluorescence intensity, stability, etc. Since the numerical label itself does not have semantic information, it may be difficult for the model to understand its meaning when used directly for training. Therefore, the uniform quantization method can be used to convert the numerical label into discrete categories and generate corresponding text descriptions for each category.

[0116] Uniform quantization: The normalized numerical range (usually between 0 and 1) is divided into several equal segments (such as 5 segments), each of which corresponds to a category. For example, suppose the numerical range is divided into 5 segments: [0, 0.2) corresponds to "very low"; [0.2, 0.4) corresponds to "low"; [0.4, 0.6) corresponds to "normal"; [0.6, 0.8) corresponds to "high"; [0.8, 1.0] corresponds to "very high". Then, a corresponding text description is generated for each category, for example: "The fluorescence intensity of this protein is very low"; "The fluorescence intensity of this protein is low"; "The fluorescence intensity of this protein is normal"; "The fluorescence intensity of this protein is high"; "The fluorescence intensity of this protein is very high". In this way, the numerical labels are converted into text descriptions with clear semantics, so that the model can better understand the functional information in the regression task.

[0117] In some of the embodiments, see Figure 3 , input the obtained optimization target into the human language model to obtain the first representation, including:

[0118] Step S221, responding to the optimization target submitted by the user, converting the optimization target into a target text description.

[0119] Step S222: extracting a first representation from the target text description based on a large human language model.

[0120] Specifically, the user submits the optimization goal of the protein through natural language description (such as "improving enzyme activity" or "enhancing thermal stability"), and the large language model LLM converts the optimization goal into a target text description (such as "this enzyme activity is {high}" or "this thermal stability is {high}"), and then uses the large language model LLM to extract the first representation to match it with the second representation.

[0121] In this embodiment, by describing the optimization goal in natural language, the user can easily guide the model to generate mutants that meet specific functional requirements through text prompts, which greatly reduces the technical threshold for protein optimization and design, making the technology easier to be adopted by biological researchers in non-computing fields, and overcoming the shortcomings of existing methods with high complexity and high usage thresholds.

[0122] In some of the embodiments, see Figure 3 , based on the second representation matching the first representation, a protein mutant meeting the optimization goal is obtained, including:

[0123] Step S241: Calculate a similarity score between the first representation and the second representation.

[0124] Step S241, sorting multiple protein mutants based on similarity scores, and screening out modified proteins that meet the optimization goal.

[0125] In this example, the protein function prediction task is converted into a text matching problem, which avoids personalized fine-tuning and enables prediction of different protein functions through a unified model, demonstrating strong generalization capabilities.

[0126] The present embodiment is described and illustrated below through preferred embodiments.

[0127] Figure 5 Flow chart of the unified protein modification method of the preferred embodiment.

[0128] S1. Training data preparation: Construct a text-protein pair dataset based on existing protein data; the text-protein pair dataset includes performance description text, sequence data and structure data of protein sequences.

[0129] S2. Model training: Based on the human language big model, text representation is extracted from the performance description text; based on the initial protein language model, the sequence data of the protein sequence is processed to obtain the initial sequence representation; based on the initial protein structure encoder, the structure data of the protein sequence is processed to obtain the initial structure representation; the initial sequence representation and the initial structure representation are fused to obtain the protein representation. Through the comparative learning strategy, the alignment ability of the human language big model, the initial protein language model and the initial protein structure encoder is trained by using the matching text representation and protein representation as well as the mismatching text representation and protein representation, and the target protein language model and the target protein structure encoder are obtained.

[0130] S3, protein modification: respond to the optimization text prompts submitted by users, convert the optimization text prompts into target text descriptions; extract the first representation from the target text description based on the human language macromodel. Perform mutation modification on the existing protein sequence to obtain protein mutants; process the sequence information and structural information of the protein mutants based on the target protein language model and the target protein structure encoder, and extract the second representation from the protein mutants; calculate the similarity score between the first representation and the second representation; sort multiple protein mutants based on the similarity score, and screen for modified proteins that meet the optimization target.

[0131] In this preferred embodiment, a unified framework for protein multifunctional prediction and mutant optimization is realized by combining text matching and text prompt guidance. After training, the model no longer needs additional fine-tuning for each new protein task, and can directly achieve zero-shot prediction and optimization through simple human language descriptions. This efficient zero-shot adaptability solves the problem of cumbersome personalized fine-tuning and lack of generalization ability in the prior art.

[0132] It should be noted that the steps shown in the above process or the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0133] In this embodiment, a unified protein modification device is also provided, which is used to implement the above embodiments and preferred embodiments, and the descriptions that have been made will not be repeated. The terms "module", "unit", "subunit" and the like used below can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware is also possible and conceivable.

[0134] Figure 6 is a structural block diagram of the unified protein modification device of this embodiment, such as Figure 6 As shown, the device includes: a model alignment module 61, a first representation extraction module 62, a second representation extraction module 63 and a matching screening module 64.

[0135] The model alignment module 61 is used to fine-tune the initial protein representation model based on the text-protein pair dataset, so as to align the human language large model and the initial protein representation model in the semantic space to obtain the target protein representation model.

[0136] The first representation extraction module 62 is used to input the obtained optimization target into the human language model to obtain the first representation.

[0137] The second representation extraction module 63 is used to generate a set of protein mutants based on the acquired key mutation sites; input the set of protein mutants into the target protein representation model to obtain the second representation corresponding to each protein mutant.

[0138] The matching screening module 64 is used to obtain protein mutants that meet the optimization target based on the second representation that matches the first representation.

[0139] In some of the embodiments, a set of protein mutants is input into a target protein representation model to obtain a second representation corresponding to each protein mutant, including: obtaining sequence information and structural information of each protein mutant based on the set of protein mutants; inputting the sequence information and structural information into the target protein representation model, wherein the target protein representation model includes a protein language model and a protein structure encoder; obtaining a sequence representation from the sequence information based on the protein language model; obtaining a structural representation from the structural information based on the protein structure encoder; and fusing the sequence representation and the structural representation to obtain a second representation.

[0140] In some of the embodiments, a structural representation is obtained from structural information based on a protein structure encoder, including: the structural information is represented as an undirected graph, the undirected graph includes edge features and node features, the node features include amino acid type, spatial position and sequence representation; the undirected graph is input into the protein structure encoder to obtain the structural representation.

[0141] In some of the embodiments, the sequence representation and the structure representation are fused to obtain a second representation, including: encoding the sequence representation into sequence global features and sequence local features; encoding the structure representation into structure global features and structure local features; and weighted concatenating the sequence global features, sequence local features, structure global features, and structure local features to obtain the second representation.

[0142] In some of the embodiments, based on a text-protein pair dataset, an initial protein representation model is fine-tuned so that a large human language model and the initial protein representation model are aligned in a semantic space to obtain a target protein representation model, including: constructing a text-protein pair dataset based on existing protein data; the text-protein pair dataset includes protein sequence and structure data and corresponding performance description text; based on the large human language model, text representation is extracted from the performance description text; based on the initial protein representation model, protein representation is extracted from the protein sequence and structure data; using matching text representations and protein representations as well as mismatching text representations and protein representations, based on contrastive learning training, the initial protein representation model is fine-tuned so that a large human language model and the initial protein representation model are aligned in a semantic space to obtain a target protein representation model.

[0143] In some embodiments, the method for generating a performance description text includes: obtaining label data of existing protein data; converting the label data into a performance description text based on a predefined text template. In some embodiments, inputting the obtained optimization target into a human language macromodel to obtain a first representation includes: responding to the optimization target submitted by the user, converting the optimization target into a target text description; extracting the first representation from the target text description based on the human language macromodel.

[0144] In some of the embodiments, a protein mutant that meets the optimization goal is obtained based on a second representation that matches the first representation, including: calculating a similarity score between the first representation and the second representation; sorting multiple protein mutants based on the similarity score, and screening to obtain a modified protein that meets the optimization goal.

[0145] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0146] In this embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0147] Optionally, the computer device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0148] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and will not be repeated in this embodiment.

[0149] In addition, in combination with the unified protein modification method provided in the above embodiments, a storage medium can also be provided in this embodiment to implement the unified protein modification method. The storage medium stores a computer program; when the computer program is executed by a processor, any unified protein modification method in the above embodiments is implemented.

[0150] It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

[0151] Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

[0152] The term "embodiment" in this application refers to a specific feature, structure or characteristic described in conjunction with the embodiment that can be included in at least one embodiment of the present application. The appearance of this phrase in various locations in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. It is clearly or implicitly understood by those of ordinary skill in the art that the embodiments described in this application can be combined with other embodiments without conflict.

[0153] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of patent protection. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the attached claims.

Claims

1. A unified protein modification method, characterized in that: The method comprises: Based on the text-protein pair dataset, fine-tune the initial protein representation model so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model; Inputting the obtained optimization target into the human language macromodel to obtain a first representation; Based on the acquired key mutation sites, a set of protein mutants is generated; the set of protein mutants is input into the target protein characterization model to obtain a second characterization corresponding to each of the protein mutants; Based on the second characterization matching the first characterization, a protein mutant meeting the optimization goal is obtained; The method of inputting the set of protein mutants into the target protein characterization model to obtain the second characterization corresponding to each of the protein mutants comprises: Based on the set of protein mutants, sequence information and structural information of each of the protein mutants are obtained; the sequence information and the structural information are input into the target protein representation model, wherein the target protein representation model includes a protein language model and a protein structure encoder; Based on the protein language model, a sequence representation is obtained from the sequence information; based on the protein structure encoder, a structural representation is obtained from the structural information; the sequence representation and the structural representation are fused to obtain a second representation; Wherein, based on the text-protein pair dataset, the initial protein representation model is fine-tuned to align the human language model with the initial protein representation model in the semantic space to obtain the target protein representation model, including: Constructing a text-protein pair dataset based on existing protein data; the text-protein pair dataset includes protein sequence and structure data and corresponding performance description text; Extracting text representation from the performance description text based on a large model of human language; extracting a protein representation from the protein sequence and structure data based on an initial protein representation model; By using the matching text representation and the protein representation as well as the unmatched text representation and the protein representation, based on contrastive learning training, the initial protein representation model is fine-tuned to align the human language large model and the initial protein representation model in the semantic space, thereby obtaining a target protein representation model.

2. The unified protein modification method according to claim 1, characterized in that: Based on the protein structure encoder, a structural representation is obtained from the structural information, including: The structural information is represented as an undirected graph, the undirected graph includes edge features and node features, the node features include amino acid type, spatial position and the sequence representation; The undirected graph is input into the protein structure encoder to obtain a structural representation.

3. The unified protein modification method according to claim 1, characterized in that: The sequence representation and the structure representation are fused to obtain a second representation, including: Encoding the sequence representation into sequence global features and sequence local features; encoding the structural representation into structural global features and structural local features; The sequence global feature, the sequence local feature, the structure global feature and the structure local feature are weightedly concatenated to obtain a second representation.

4. The unified protein modification method according to claim 1, characterized in that: The method for generating the performance description text comprises: Obtain label data of existing protein data; Based on a predefined text template, the label data is converted into performance description text.

5. The unified protein modification method according to claim 1, characterized in that: Inputting the obtained optimization target into the human language macro model to obtain a first representation includes: In response to the optimization target submitted by the user, convert the optimization target into a target text description; Based on the human language large model, a first representation is extracted from the target text description.

6. The unified protein modification method according to claim 1, characterized in that: Based on the second characterization matching the first characterization, a protein mutant meeting the optimization goal is obtained, including: calculating a similarity score between the first representation and the second representation; Based on the similarity scores, the plurality of protein mutants are ranked and the modified proteins meeting the optimization goal are screened.

7. A unified protein modification device, characterized in that: The device comprises: A model alignment module is used to fine-tune the initial protein representation model based on the text-protein pair dataset, so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain a target protein representation model; wherein, based on the text-protein pair dataset, fine-tune the initial protein representation model so that the human language large model and the initial protein representation model are aligned in the semantic space to obtain the target protein representation model, including: Constructing a text-protein pair dataset based on existing protein data; the text-protein pair dataset includes protein sequence and structure data and corresponding performance description text; Based on the human language macromodel, text representation is extracted from the performance description text; based on the initial protein representation model, protein representation is extracted from the protein sequence and structure data; using the matching text representation and protein representation as well as the unmatched text representation and protein representation, based on contrastive learning training, the initial protein representation model is fine-tuned to align the human language macromodel and the initial protein representation model in the semantic space, thereby obtaining a target protein representation model; A first representation extraction module, used for inputting the obtained optimization target into the human language macro model to obtain a first representation; The second representation extraction module is used to generate a set of protein mutants based on the acquired key mutation sites; input the set of protein mutants into the target protein representation model to obtain the second representation corresponding to each of the protein mutants; wherein, inputting the set of protein mutants into the target protein representation model to obtain the second representation corresponding to each of the protein mutants includes: Based on the set of protein mutants, sequence information and structural information of each of the protein mutants are obtained; the sequence information and the structural information are input into the target protein representation model, wherein the target protein representation model includes a protein language model and a protein structure encoder; Based on the protein language model, a sequence representation is obtained from the sequence information; based on the protein structure encoder, a structural representation is obtained from the structural information; the sequence representation and the structural representation are fused to obtain a second representation; A matching screening module is used to obtain a protein mutant that meets the optimization goal based on the second representation that matches the first representation.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images