Protein design methods, apparatuses, devices, media, and program products

Abstract

The application discloses a protein design method, device, equipment, medium and program product, and is applied to the field of artificial intelligence. The method comprises the following steps: obtaining a protein framework, and extracting protein features of the protein framework; obtaining a target protein sequence, and extracting target protein features of the target protein sequence; fusing the protein features and the target protein features to obtain fused features; the fused features are used to represent sequence information and structure information of the protein indicated by the protein framework and the target protein indicated by the target protein sequence under interaction; performing protein sequence recovery on the protein framework based on the fused features to recover a predicted protein sequence corresponding to the protein framework; and performing protein structure prediction on the predicted protein sequence to obtain a predicted protein structure corresponding to the predicted protein sequence.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a protein design method, apparatus, device, medium, and program product. Background Technology

[0002] Proteins are important macromolecular organic compounds in living organisms. Protein design, including sequence design and structure design, has wide applications in medicine and drug development, biotechnology and industrial applications, materials science, diagnostics and therapy, and other fields.

[0003] In related technologies, protein design models are trained based on a large amount of sequence information corresponding to proteins, and these models are fine-tuned based on specific target proteins to generate proteins that bind to those target proteins. For example, a target protein is an antigen (Ag), which is a substance that can induce antibody production, such as bacteria, viruses, and pollen. An antibody (Ab) is a protective protein produced in response to antigen stimulation.

[0004] However, the methods used in this field are insufficient to capture the interaction between proteins and target proteins, resulting in inaccurate protein design. Summary of the Invention

[0005] This application provides a protein design method, apparatus, device, medium, and program product. The technical solution is as follows:

[0006] On the one hand, a protein design method is provided, the method comprising:

[0007] Obtain a protein framework and extract protein features from the protein framework;

[0008] Obtain the target protein sequence and extract the target protein features from the target protein sequence;

[0009] The protein features and the target protein features are fused to obtain fused features; the fused features are used to characterize the sequence and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence.

[0010] Based on the fusion features, protein sequence recovery is performed on the protein framework to recover the predicted protein sequence corresponding to the protein framework, and protein structure prediction is performed on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence.

[0011] In some embodiments, the sample protein features include: sample protein sequence features, sample protein pairing features, and sample protein structure; the sample target protein features include: sample target protein MSA features, sample target protein pairing features, and sample target protein structure; and the sample fusion features include: sample fusion sequence features and sample fusion pairing features.

[0012] The fusion of the sample protein features and the sample target protein features to obtain sample fusion features includes:

[0013] The sample fusion sequence features are obtained by fusing the protein sequence features of the sample and the MSA features of the target protein of the sample.

[0014] Based on the sample protein structure and the sample target protein structure, the sample protein pairing feature, the sample target protein pairing feature, the sample protein sequence feature, and the sample target protein MSA feature are fused to obtain the sample fusion pairing feature.

[0015] In some embodiments, fusing the sample protein sequence features and the sample target protein MSA features to obtain the sample fused sequence features includes:

[0016] The sample protein sequence features are subjected to a linear transformation to obtain the transformed sample protein sequence features; and the first row of sample target protein MSA features in the sample target protein MSA features are subjected to a linear transformation to obtain the transformed first row of sample target protein MSA features.

[0017] The transformed sample protein sequence features and the transformed first-row sample target protein MSA features are fused based on cross-attention to obtain the sample fused sequence features.

[0018] In some embodiments, the step of fusing the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features based on the sample protein structure and the sample target protein structure to obtain the sample fused pairing features includes:

[0019] Based on the carbon atom coordinates of each amino acid residue in the sample protein structure, determine the protein coordinates corresponding to each amino acid residue in the sample protein structure; and based on the carbon atom coordinates of each amino acid residue in the sample target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the sample target protein structure.

[0020] Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample target protein structure.

[0021] Based on the protein pairing distance and the target protein pairing distance, the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features are fused to obtain the sample fused pairing features.

[0022] In some embodiments, the step of fusing the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features based on the protein pairing distance and the target protein pairing distance to obtain the sample fused pairing features includes:

[0023] Based on the sample protein pairing features and the protein pairing distance, a mapping is performed to calculate a first part of the fusion pairing features; and based on the sample target protein pairing features and the target protein pairing distance, a mapping is performed to calculate a second part of the fusion pairing features.

[0024] The outer product average is calculated based on the sample protein sequence characteristics and the sample target protein MSA characteristics to obtain the third part of the fusion pairing characteristics; and the transpose of the third part of the fusion pairing characteristics is determined as the fourth part of the fusion pairing characteristics.

[0025] The sample fusion pairing features are obtained by splicing the first part of the fusion pairing features, the second part of the fusion pairing features, the third part of the fusion pairing features, and the fourth part of the fusion pairing features.

[0026] In some embodiments, the step of performing mapping based on the sample protein pairing features and the protein pairing distance to calculate a first portion of fusion pairing features includes:

[0027] A linear transformation is performed on the protein pairing features of the samples to obtain the transformed protein pairing features; and a linear transformation is performed on the protein pairing distances to obtain the transformed protein pairing distances.

[0028] The first part of the fusion pairing feature is obtained by adding the transformed sample protein pairing feature and the transformed protein pairing distance.

[0029] In some embodiments, the step of performing mapping based on the sample target protein pairing features and the target protein pairing distance to calculate the second part of the fusion pairing features includes:

[0030] A linear transformation is performed on the sample target protein pairing features to obtain the transformed sample target protein pairing features; and a linear transformation is performed on the target protein pairing distance to obtain the transformed target protein pairing distance.

[0031] The transformed sample target protein pairing features and the transformed target protein pairing distance are added together to obtain the second part of the fusion pairing features.

[0032] In some embodiments, the step of calculating the outer product average based on the sample protein sequence features and the sample target protein MSA features to obtain the third part of the fusion pairing features includes:

[0033] The sample protein sequence features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized protein sequence features; and the sample target protein MSA features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized target protein MSA features.

[0034] Calculate the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and perform a linear transformation on the average value of the outer products to obtain the third part of the fusion pairing features.

[0035] In some embodiments, the method further includes:

[0036] In the presence of inter-sample chain features or sample interaction features between the sample protein and the sample target protein, sample update sequence features and sample update pairing features are obtained based on the inter-sample chain features or sample interaction features, as well as the sample protein structure and the sample target protein structure.

[0037] The sample fusion sequence features are updated based on the sample update sequence features to obtain the updated sample fusion sequence features; and the sample fusion pairing features are updated based on the sample update pairing features to obtain the updated sample fusion pairing features.

[0038] In some embodiments, when inter-chain or inter-interaction features exist between the sample protein and the sample target protein, obtaining sample update sequence features and sample update pairing features based on the inter-chain or inter-interaction features, as well as the sample protein structure and the sample target protein structure, includes:

[0039] If the sample chain features exist, the sample interaction features are determined based on the sample chain features and the chain connection relationships; or, if the sample interaction features exist, the sample chain features are determined based on the sample interaction features and the chain pseudo-connection relationships.

[0040] Perform a linear transformation on the sample interaction features to obtain the sample update sequence features; and perform a linear transformation on the sample chain features to update the sample pairing features.

[0041] On the other hand, a protein design apparatus is provided, the apparatus comprising:

[0042] An acquisition module is used to acquire a protein framework and extract protein features from the protein framework.

[0043] The acquisition module is used to acquire the target protein sequence and extract the target protein features from the target protein sequence.

[0044] A fusion module is used to fuse the protein features and the target protein features to obtain fusion features; the fusion features are used to characterize the sequence and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence.

[0045] The processing module is configured to perform protein sequence recovery on the protein framework based on the fusion features, recover the predicted protein sequence corresponding to the protein framework, and perform protein structure prediction on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence.

[0046] On the other hand, a computer device is provided, the computer device comprising: a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the protein design method as described above.

[0047] On the other hand, a computer-readable storage medium is provided that stores a computer program, which is loaded and executed by a processor to implement the protein design method described above.

[0048] On the other hand, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium, wherein a processor retrieves the computer instructions from the computer-readable storage medium, causing the processor to load and execute them to implement the protein design method as described above.

[0049] The beneficial effects of the technical solutions provided in this application include at least the following:

[0050] This invention provides a co-design method for protein sequences and structures. It can utilize pre-constructed protein frameworks and, based on specified target protein sequences, fuse the protein features of the protein framework and the target protein features of the target protein sequence to fully express the sequence and structural information of the interaction between the target protein indicated by the target protein sequence and the protein corresponding to the protein framework. This allows for the design of proteins with the ability to bind to and interact with the target protein, improving the accuracy of protein design and ensuring the rationality of the protein structure while designing the protein sequence. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a structural block diagram of a computer system provided in an exemplary embodiment of this application;

[0053] Figure 2 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application;

[0054] Figure 3 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application;

[0055] Figure 4 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application;

[0056] Figure 5 This is a flowchart of a protein design method provided in an exemplary embodiment of this application;

[0057] Figure 6 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application;

[0058] Figure 7 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application;

[0059] Figure 8 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application;

[0060] Figure 9 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application;

[0061] Figure 10 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application;

[0062] Figure 11 This is a distribution diagram of the three-round cyclic ipTM score provided in an exemplary embodiment of this application;

[0063] Figure 12 This is a block diagram of a protein design apparatus provided in an exemplary embodiment of this application;

[0064] Figure 13 This is a structural block diagram of a computer device provided in an exemplary embodiment of this application. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0066] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0067] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0068] It should be understood that although the terms first, second, etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, a first parameter may also be referred to as a second parameter, and similarly, a second parameter may also be referred to as a first parameter. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0069] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user-related data (e.g., proteins, target proteins, antigens, antibodies, genes, etc.). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their relevant data is being collected. This ensures that the application only begins the steps for collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0070] First, a brief introduction to the terms used in the embodiments of this application:

[0071] Proteins, also known as protein molecules, are important macromolecular organic compounds in living organisms, mainly composed of elements such as carbon (C), hydrogen (H), oxygen (O), and nitrogen (N). Proteins are composed of m amino acids and n peptide chains, each protein having an independent and defined amino acid sequence. Amino acids are the basic building blocks of proteins, and the sequence of amino acids and the resulting three-dimensional structure constitute the diversity of protein structure. Proteins possess primary, secondary, tertiary, and quaternary structures. The sequence of amino acid residues in the peptide chain of a protein is called the primary structure of the protein. A peptide chain, formed by multiple amino acids linked together, contains multiple peptide bonds. Because some groups of amino acids participate in the formation of peptide bonds and lose a water molecule, the amino acid units in a polypeptide are called amino acid residues, or simply residues.

[0072] Target proteins are proteins located inside or on the surface of cells that can be effectively acted upon by drugs or small molecule compounds, and are used in fields such as protein research and drug research. For example, a target protein can be at least one of antigens, viral target proteins, or receptor proteins.

[0073] In the embodiments of this application, the target protein is an antigen (Ag), which refers to a substance that can induce antibody production, such as bacteria, viruses, pollen, etc. The protein is an antibody (Ab), which is a protective protein produced in response to antigen stimulation.

[0074] Complementarity Determining Region (CDR): This is a part of the antibody, exhibiting high variability and specificity, capable of recognizing and binding to various different antigens. The CDR is located in the V region of the antibody, with three CDRs each on the light and heavy chains: CDR1, CDR2, and CDR3. CDR3 exhibits the highest degree of variability.

[0075] In the embodiments of this application, the protein framework is obtained by masking the amino acid sequences in the complementarity-determining regions of the protein sequence, and protein sequence recovery is performed on the protein framework, that is, the amino acid sequences in these complementarity-determining regions are recovered.

[0076] Protein language models are a class of large language models based on deep learning, specifically designed for processing and analyzing protein sequences. The working principle of protein language models is similar to that of natural language models; their core idea is to treat the protein sequence as a language, each amino acid as a character, and to learn the rules of this language through autoregression.

[0077] In this embodiment, the protein language model can be a large language model based on the Transformer architecture, such as an ESM (Evolutionary Scale Modeling) model. The input of the ESM model is a protein sequence, and the output is a feature representation, which includes at least one piece of information such as protein structure and function. This information can be further mapped or used in downstream models to achieve various predictions and analyses of the protein.

[0078] Invariant point attention mechanism: This is a way to process invariant features in input data that have invariant properties. It can take advantage of the rotation and translation invariance property, which means that the protein structure may be located at any spatial position in three-dimensional space. No matter how the protein structure is rotated or translated in three-dimensional space, the attention weights calculated by the invariant point attention mechanism remain consistent.

[0079] Protein-protein interaction (PPI) refers to the process by which two or more protein molecules form a protein complex through non-covalent bonds. It can reveal information such as the synergistic effects, signal transduction, and biological responses between proteins.

[0080] Protein structure prediction unit: Specifically, this includes the Evoformer unit, which is used to process and optimize protein sequence and protein structure information. The Evoformer unit takes multiple sequence alignment (MSA) data as input, which includes the protein sequence and its associated homologous sequences. Through these homologous sequences, the Evoformer unit can capture the structural and functional features that proteins have retained during evolution.

[0081] Confidence score: A metric used to evaluate the accuracy and reliability of predicting protein sequences, protein structures, and complex structures. It includes at least one of the following: prediction template modeling score (pTM), interface prediction template modeling score (ipTM), local geometry quality prediction score (pLDDT), interface local geometry quality prediction score (ipLDDT), design score, and edit distance.

[0082] Figure 1 This is a structural block diagram of a computer system provided in an exemplary embodiment of this application. The computer system 100 can be implemented as a system architecture for a protein design method. The computer system 100 includes a terminal 120 and a server 140.

[0083] Terminal 120 can be an electronic device such as a mobile phone, tablet computer, vehicle terminal (vehicle system), wearable device, PC (Personal Computer), or unmanned reservation terminal. Terminal 120 stores protein design models, or a client application for a target application can be installed and run on terminal 120. This target application can provide protein design functions or training functions for protein design models; this embodiment does not limit the specific form of the target application. Furthermore, this embodiment does not limit the form of the target application, including but not limited to apps, mini-programs, etc., installed on terminal 120, and can also be in web page form.

[0084] Server 140 stores protein design models, or it can provide protein design functions or protein model training functions; this embodiment is not limited in this respect. Server 140 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Server 140 can be the backend server of the aforementioned target application, used to provide backend services to the client of the target application. In some optional embodiments, server 140 can also be implemented as a node in a blockchain system.

[0085] Terminal 120 and server 140 can communicate via a network, such as a wired or wireless network.

[0086] The protein design method provided in this application can be executed by a computer device, which refers to an electronic device capable of data computation, processing, and storage. Figure 1 Taking the implementation environment of the scheme shown as an example, the protein design method can be executed by the terminal 120. For example, the protein design method can be executed by the client of the target application installed and running on the terminal 120, or by the server 140, or by the interaction and cooperation between the terminal 120 and the server 140. This application embodiment does not limit this.

[0087] Those skilled in the art will understand that the number of terminals 120 described above can be more or less. For example, there may be only one terminal 120, or there may be dozens or hundreds, or more. This application does not limit the number or type of terminals 120 in its embodiments.

[0088] There is no clear solution for de novo end-to-end protein design for target proteins without any constraints or reference sequence information.

[0089] In related technologies, protein design includes two implementation methods:

[0090] 1. A protein design model is trained based on a large amount of sequence information corresponding to proteins, and then fine-tuned based on specific target proteins to generate proteins that bind to those target proteins. For example, the target protein might be an antigen (Ag), and the protein might be an antibody (Ab). The protein design model can design antibodies targeting their complementarity-determining regions, particularly the CDRH3 region. It undergoes pre-training based on a large number of antibodies and fine-tuning based on specific antigens to generate antibodies that bind to those antigens and determine the antibody sequence.

[0091] 2. Protein design models that coordinate protein sequence and structure design are trained based on the structural information of the protein. For example, an end-to-end protein design model based on isovariant graph neural networks can perform co-design of protein sequences and structures, with protein structures serving as training data. For instance, the target protein might be an antigen, and the protein an antibody. The protein design model can perform co-design of antibody structures and complementarity-determining regions in the antibody sequence, particularly the CDRH3 region.

[0092] However, the limitations of the two implementation methods of the relevant technology include at least the following:

[0093] 1. First, protein design models do not explicitly consider structural information. The binding of target proteins to proteins is often accompanied by conformational changes and dynamic interactions, which are difficult to capture using only protein sequence information. Second, without structural information constraints, protein design models are prone to generating a large number of infeasible protein sequence combinations. Furthermore, protein design models lack structure-based screening mechanisms, making it difficult to effectively evaluate whether the designed proteins possess the intended function.

[0094] 2. First, the training data for protein design models relies solely on structural information, without considering pre-learning / transfer learning of specific binding modes from a broad range of neutralizing protein sequences. Second, it is limited to designing complementarity-determining regions, neglecting the compatibility of frame region sequences. Without prior experimentation with given frame regions, it is difficult to select suitable frame region sequences.

[0095] The existing techniques make it difficult for protein design models to capture the interactions between proteins and target proteins, resulting in inaccurate protein design. This application provides a method for the co-design of protein sequences and structures. By pre-constructing a protein framework library, using the protein framework and target protein sequences, and combining the prediction of the amino acid sequences corresponding to the complementarity-determining regions in the protein framework, the co-design of the entire protein sequence and structure is completed. Figure 2This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application. The protein design method is executed by a computer device, which may be... Figure 1 Terminal 120 and / or server 140, wherein terminal 120 and / or server 140 store protein design models. The protein design method can be specifically implemented based on the protein design model.

[0096] Step 1: Obtain the protein framework 210 and extract protein features 230 from it. The protein sequence in the protein framework 210 is represented as “H: QIQLVO…CARxxxxxxxWGQ...VSS; L: DVLMTQ...YKDDDK”, where “H” represents the heavy chain, “L” represents the light chain, and each letter in “QILVOCARWGSDMTYK” represents a type of amino acid. “xxx” represents the masked amino acid sequence within the protein framework 210, i.e., the amino acid sequence to be recovered. Protein features 230 include: protein sequence features, protein pairing features, and protein structure. The term “features” can also be referred to as “representations”.

[0097] Step 2: Obtain the target protein sequence 220 and extract the target protein features 240 from the target protein sequence. The target protein includes an antigen. The target protein sequence 220 is represented as "A:RVQPTE…", where "A" represents the antigen and one letter in "RVQPTE" represents a type of amino acid. The target protein features 240 include: target protein MSA features, target protein pairing features, and target protein structure.

[0098] Step 3: Fuse protein feature 230 with target protein feature 240 to obtain fusion feature 250; fusion feature 250 is used to characterize the sequence and structural information of the interaction between the protein indicated by protein frame 210 and the target protein indicated by target protein sequence 220. Fusion feature 250 includes: fusion sequence feature and fusion pairing feature.

[0099] Step 4: Based on the fusion feature 250, perform protein sequence recovery on the protein framework 210 to recover the predicted protein sequence 260 corresponding to the protein framework 210. Then, perform protein structure prediction on the predicted protein sequence 260 to obtain the predicted protein structure 270 corresponding to the predicted protein sequence 260. The protein sequence in the predicted protein sequence 260 is “H:QIQLVO…CARGYYGYYAFDYWGQ...VSS;L:DVLMTQ...YKDDDK”. GYYGYYAFD"This refers to the masked amino acid sequence in the recovered protein framework 210. Optionally, complex structure prediction can also be performed based on fusion feature 250 to obtain the predicted complex structure, and the predicted protein structure 270 can also be isolated from the predicted complex structure. Confidence scores 280 corresponding to the predicted protein sequence 260 and the predicted protein structure 270 can also be generated: ipTM = 0.892, iplDDT = 0.706, used to screen the predicted protein sequence 260 and the predicted protein structure 270.

[0100] In summary, the protein design method provided in this application utilizes a protein framework and a specified target protein sequence. Since proteins and target proteins interact and influence each other, fusion features fully represent this interaction, thereby achieving co-design of protein sequences and structures and improving the accuracy of protein design. Detailed examples are provided below.

[0101] The protein design method provided in this application can be implemented based on a protein design model. The following embodiments illustrate the model structure, application stage, and training stage of the protein design model.

[0102] Part 1: Model Structure of Protein Design Models

[0103] Figure 3 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application.

[0104] Optionally, the model structure of the protein design model 200 used to perform the protein design method includes: a target protein feature generation unit 20 and a protein feature generation unit 10 connected in parallel; a feature fusion unit 30 connected to both the target protein feature generation unit 20 and the protein feature generation unit 10; a complex structure prediction unit 40 and a protein sequence recovery unit 50 connected to the feature fusion unit 30; and a protein structure prediction unit 60 connected to the protein sequence recovery unit 50.

[0105] The target protein feature generation unit 20 is used to extract target protein features from the target protein sequence and to extract sample target protein features from the sample target protein sequence; the protein feature generation unit 10 is used to extract protein features from the protein framework and to extract sample protein features from the sample protein; the feature fusion unit 30 is used to fuse protein features and target protein features, and to fuse sample protein features and sample target protein features; the complex structure prediction unit 40 is used to perform complex structure prediction; the protein sequence recovery unit 50 is used to perform protein sequence recovery; and the protein structure prediction unit 60 is used to perform protein structure prediction.

[0106] Figure 4 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application.

[0107] Optionally, the model structure of the protein design model 200 further includes at least one of a feature recycling unit 70, a structure recycling unit 80, and a sequence recycling unit 90. The input of the feature recycling unit 70 is connected to the output of the target protein feature generation unit 20, and the output of the feature recycling unit 70 is connected to the input of the target protein feature generation unit 20. The feature recycling unit 70 is used to iteratively use previously generated target protein features or sample target protein features to improve the accuracy of the target protein features or sample target protein features. The input of the structure recycling unit 80 is connected to the output of the complex structure prediction unit 40, and the output of the structure recycling unit 80 is connected to the input of the feature fusion unit 30. The structure recycling unit 80 is used to enable the feature fusion unit 30 to iteratively use previously generated predicted complex structures to improve the accuracy of the fusion features or sample fusion features. The input of the sequence looping unit 90 is connected to the output of the protein sequence recovery unit 50, and the output of the sequence looping unit 90 is connected to the input of the protein feature generation unit 10. The sequence looping unit 90 is used to iteratively perform masking on candidate protein sequences to generate protein frames, and the protein feature generation unit 10 uses these iteratively generated protein frames to achieve iterative prediction of protein sequences and screening of protein sequences and protein frames.

[0108] In one possible implementation, the protein feature generation unit 10 in the protein design model 200 can be implemented as: a tandem pretrained ESM-PPI module and a pretrained tFold-Ab module. The target protein feature generation unit 20 can be implemented as a pretrained protein structure prediction module. The feature fusion unit 30 can also be called a feature fusion module. The complex structure prediction unit 40 can also be called a complex structure prediction module. The complex structure prediction unit 40 and the protein structure prediction unit 60 can each be implemented as Evoformer modules. The feature fusion unit 30, the complex structure prediction unit 40, the protein sequence recovery unit 50, and the protein structure prediction unit 60 can be combined into an AI-driven flexible docking module.

[0109] Part Two: The Application Stage of Protein Design Models

[0110] Figure 5 This is a flowchart of a protein design method provided in an exemplary embodiment of this application.

[0111] This method is executed by a computer device, which can be... Figure 1 The method includes at least a portion of steps 320, 340, 360, and 380: terminal 120 and / or server 140, wherein terminal 120 and / or server 140 store protein design models.

[0112] Step 320: Obtain the protein framework and extract the protein features of the protein framework.

[0113] A protein sequence is a sequence composed of peptide chains and the amino acids within those chains. Protein sequences include the amino acid sequences corresponding to the heavy chain and the light chain. For example, a protein sequence might be represented as "H: QIQLVO…;L: DVLMTQ...YKDDDK", where "H" represents the heavy chain, "L" represents the light chain, and one letter in "QILVOCARWGSDMTYK" represents a type of amino acid.

[0114] A protein framework, also known as a protein framework template or protein framework region, is obtained by masking at least a portion of the amino acid sequence within the entire protein sequence. Protein frameworks are crucial for protein stability and solubility.

[0115] In some embodiments, the protein sequence includes a complementarity-determining region (CDR), and the protein frame is obtained by masking the amino acid sequences corresponding to the CDRs in the protein sequence. For example, the protein sequence in the protein frame is represented as "H: QIQLVO…CARxxxxxxxWGQ...VSS; L: DVLMTQ...YKDDDK", where "xxx" represents the masked amino acid sequence contained in the protein frame, i.e., the amino acid sequence that needs to be recovered.

[0116] In some embodiments, a protein framework library is pre-constructed, comprising several protein frameworks. Exemplarily, protein frameworks are obtained from the pre-constructed protein framework library, and protein features of the protein frameworks are extracted. Optionally, a protein framework is obtained from the protein framework library, which is the protein framework with the highest matching degree to one of the target protein's type, activity, or origin. Alternatively, at least a portion of the protein frameworks are obtained from the protein framework library, and at least a portion of the protein frameworks correspond to at least one of the target protein's type, activity, or origin. Here, the corresponding protein frameworks are pre-defined based on the target protein's type, activity, and origin. Alternatively, all protein frameworks are obtained from the protein framework library. When obtaining at least a portion or all of the protein frameworks, screening is also performed based on the confidence scores of the predicted protein sequences and predicted protein structures corresponding to each protein framework.

[0117] Protein features are the features corresponding to the protein sequence indicated by the protein framework. In this embodiment, protein features of the protein framework are extracted from at least two dimensions: sequence and structure. Optionally, protein features include: protein sequence representation (s), protein pairwise representation (z), and protein structure (x). Protein sequence features are used to indicate the sequence information of the protein framework. Protein pairwise representation, also known as protein pairwise features, are features corresponding to any two amino acid residues located on the same peptide chain, used to indicate the interaction between any two amino acid residues located on the same peptide chain. Protein structure is represented by the atomic coordinates of each amino acid residue in the protein, used to indicate the structural information of the protein. Atomic coordinates can be three-dimensional coordinates.

[0118] Step 340: Obtain the target protein sequence and extract the target protein features from the target protein sequence.

[0119] A target protein sequence is a sequence consisting of the peptide chain and the amino acids on the peptide chain of a target protein. For example, target proteins include antigens, and the target protein sequence is represented as "A:RVQPTE…", where "A" represents the antigen and one letter in "RVQPTE" represents a type of amino acid.

[0120] In some embodiments, the target protein sequence is obtained; a sequence database is searched based on the target protein sequence to determine at least one homologous sequence related to the target protein sequence; the target protein sequence and at least one homologous sequence related to the target protein sequence are combined to obtain target protein multiple sequence alignment (MSA) data; and target protein features are extracted from the target protein MSA data. Optionally, when combining the target protein sequence and at least one homologous sequence related to the target protein sequence to obtain target protein MSA data, the target protein sequence can be placed in the first row of the target protein MSA data.

[0121] The target protein MSA data includes the target protein sequence and at least one homologous sequence associated with it. These homologous sequences allow us to capture the structural and functional characteristics of the target protein that have been preserved during evolution. For example, if a target protein is an antigen, and its sequence is represented as "A:RVQPTE…", after searching a sequence database, three homologous sequences are identified, represented as "RVVPSG…", "RVQPTV…", and "RVSPST…". Combining the target protein sequence with these three homologous sequences yields the target protein MSA data.

[0122] Target protein features are the features corresponding to the target protein sequence. In this embodiment, target protein features are extracted from the target protein sequence from at least two dimensions: sequence and structure. Optionally, target protein features include: target protein MSA features (MSA Representation, m), target protein pairwise representation (z), and target protein structure (x). The target protein MSA feature is used to indicate the sequence information of the target protein sequence and its homologous sequences. The target protein pairwise representation, also known as the target protein pairwise feature, is the feature corresponding to any two amino acid residues located on the same peptide chain, used to indicate the interaction between any two amino acid residues located on the same peptide chain. The target protein structure is represented by the atomic coordinates of each amino acid residue in the target protein, used to indicate the structural information of the target protein. The atomic coordinates can be three-dimensional coordinates.

[0123] Step 360: Fuse protein features and target protein features to obtain fusion features; fusion features are used to characterize the sequence and structural information of the interaction between the protein framework-indicated protein and the target protein sequence-indicated target protein.

[0124] Complex representation is obtained by performing feature fusion between protein features and target protein features. Because there are interactions between proteins and target proteins, and they influence each other, complex representation is used to characterize the sequence and structural information resulting from the interaction between the protein framework-indicated protein and the target protein sequence-indicated target protein.

[0125] In this embodiment, the fusion features include: fusion sequence features and fusion pairing features. The fusion sequence features indicate sequence information related to the interaction between the protein and the target protein. The fusion pairing features indicate information about the amino acid residues involved in the intrachain or interchain interaction between any two amino acid residues on the same peptide chain in the protein and any two amino acid residues on the same peptide chain in the target protein. The fusion pairing features also include structural and sequence information corresponding to the protein framework and the fusion protein sequence, which will be explained in subsequent embodiments.

[0126] Step 380: Perform protein sequence recovery on the protein framework based on fusion features to recover the predicted protein sequence corresponding to the protein framework, and perform protein structure prediction on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence.

[0127] Protein sequence recovery refers to predicting the amino acid sequences corresponding to masked complementarity-determining regions (CDRs) in a protein framework to restore the protein framework to a complete protein sequence; this is called protein sequence prediction. Protein structure prediction refers to generating protein structures based on predicted protein sequences; this is called protein structure prediction.

[0128] Fusion features include fusion sequence features and fusion pairing features. In some embodiments, protein sequence recovery is performed on the protein framework based on the fusion sequence features and fusion pairing features to predict the amino acid sequences corresponding to the masked complementarity determination regions in the protein framework, thereby recovering the predicted protein sequence corresponding to the protein framework. Furthermore, the predicted sequence features and predicted pairing features of the predicted protein sequence are extracted, and a mapping is performed on the predicted sequence features and predicted pairing features using an invariant point attention mechanism to obtain the predicted protein structure.

[0129] In one possible implementation, an invariant point attention mechanism is used to map the predicted sequence features and predicted pairing features, predicting the relative rotation angles and relative distances of each amino acid residue. Based on the relative rotation angles and translation information of each amino acid residue, the coordinates and angles of each atom are predicted. Based on the coordinates and angles of each atom, a predicted protein structure is generated. In this embodiment, the invariant point attention mechanism can map the predicted sequence features and predicted pairing features to the predicted atomic-level three-dimensional protein structure. It should also be noted that the method for extracting the predicted sequence features and predicted pairing features of the predicted protein sequence can be consistent with the method for extracting the protein sequence features and protein pairing features of the protein framework.

[0130] In some embodiments, the method further includes performing complex structure prediction based on fusion features to obtain a predicted complex structure. The predicted complex structure is a combination of the predicted protein structure and the updated target protein structure. Specifically, the fusion features include fusion sequence features and fusion pairing features. The fusion sequence features and fusion pairing features are mapped using an invariant point attention mechanism to obtain the predicted complex structure. Optionally, the predicted protein structure can also be obtained by separating it from the predicted complex structure. Therefore, the protein design model provided in this application embodiment can calculate the complex structure of the protein indicated by the protein framework and the target protein, update the original target protein structure, recover the predicted protein sequence corresponding to the protein framework, and generate the predicted protein structure corresponding to the predicted protein sequence.

[0131] In summary, the protein design method provided in this application involves: obtaining a protein framework and extracting protein features from the protein framework; obtaining a target protein sequence and extracting target protein features from the target protein sequence; fusing the protein features and target protein features to obtain fused features; using the fused features to characterize the sequence and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence; performing protein sequence recovery on the protein framework based on the fused features to recover the predicted protein sequence corresponding to the protein framework; and performing protein structure prediction on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence. Accordingly, a collaborative design method for protein sequences and protein structures is provided. This method utilizes a pre-constructed protein framework and, based on a specified target protein sequence, by fusing the protein features of the protein framework and the target protein features of the target protein sequence, it fully expresses the sequence and structural information of the interaction between the target protein indicated by the target protein sequence and the protein corresponding to the protein framework. It fully utilizes features of different dimensions and types to design proteins with binding and interaction capabilities with the target protein, improving the accuracy of protein design and ensuring the rationality of the protein structure while designing the protein sequence.

[0132] In some embodiments, protein features include: protein sequence features, protein pairing features, and protein structure; target protein features include: target protein MSA features, target protein pairing features, and target protein structure; and fusion features include: fusion sequence features and fusion pairing features. For example, step 360 is specifically implemented as steps 362 and 364:

[0133] Step 362: Fuse the protein sequence features and the target protein MSA features to obtain the fused sequence features.

[0134] Step 364: Based on the protein structure and target protein structure, fuse protein pairing features, target protein pairing features, protein sequence features, and target protein MSA features to obtain fused pairing features.

[0135] Feature fusion methods include at least one of the following: addition, multiplication, splicing, linear combination, weighted summation, and attention mechanism fusion. Fusion sequence features are obtained by fusing protein sequence features and target protein MSA features. Fusion paired features are obtained by fusing protein paired features, target protein paired features, protein sequence features, and target protein MSA features, based on protein structure and target protein structure. Through feature fusion, information from different dimensions such as sequence and structure can be combined, improving the accuracy of fused features.

[0136] This embodiment provides a method for fusing various features from protein sequence features and target protein features. The fused sequence features indicate the sequence information of both the protein framework and the target protein sequence. The fused pairing features indicate the intrachain or interchain interactions between any two amino acid residues on the same peptide chain in the protein and any two amino acid residues on the same peptide chain in the target protein, and also include the structural and sequence information corresponding to both the protein framework and the fused protein sequence. This allows for a full expression of the interactions between the protein and the target protein, improving the accuracy of protein design.

[0137] In some embodiments, step 362 may be implemented as steps 3621 and 3622:

[0138] Step 3621: Perform a linear transformation on the protein sequence features to obtain the transformed protein sequence features; and perform a linear transformation on the first row of target protein MSA features in the target protein MSA features to obtain the transformed first row of target protein MSA features.

[0139] Linear transformations can be used for at least one of feature extraction, dimensionality reduction, and data transformation. The transformed protein sequence features are obtained by performing a linear transformation on the protein sequence. The transformed first-row target protein MSA features are obtained by performing a linear transformation on the first row of target protein MSA features. In the target protein MSA data, the target protein sequence can be located in the first row.

[0140] Step 3622: Based on cross-attention, the transformed protein sequence features and the transformed first-row target protein MSA features are fused to obtain the fused sequence features.

[0141] The cross-attention mechanism focuses on the interaction between the transformed protein sequence features and the transformed first-row target protein MSA features. For example, cross-attention weights are calculated for each of the transformed protein sequence features and the transformed first-row target protein MSA features. Using these weights, a weighted sum is performed on the transformed protein sequence features and the transformed first-row target protein MSA features to obtain the fused sequence features.

[0142] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0143]

[0144] Where Linear is a linear transformation. It refers to protein sequence characteristics, specifically antibody sequence characteristics. This refers to the MSA features of the first-row target protein, specifically the MSA features of the first-row antigen. CrossAttention refers to cross-attention. It is a feature of the fusion sequence.

[0145] In one example, the target protein is the antigen, and the protein is the antibody. For cross-attention, the following code representation exists:

[0146]

[0147] 1.

[0148] 2.

[0149] 3.

[0150] 4.

[0151] 5.

[0152] Where Linear is a linear transformation. It refers to protein sequence characteristics, specifically antibody sequence characteristics. "Concat" means to assemble. It refers to the MSA features of the first-row target protein, specifically the MSA features of the first-row antigen. MultiHeadAttention is multi-head attention.

[0153] This embodiment provides a method for obtaining fusion sequence features, which can improve the accuracy of fusion sequence feature expression and facilitate subsequent protein sequence recovery.

[0154] In some embodiments, step 364 may be implemented as steps 420, 440, and 460:

[0155] Step 420: Based on the carbon atom coordinates of each amino acid residue in the protein structure, determine the protein coordinates corresponding to each amino acid residue in the protein structure; and based on the carbon atom coordinates of each amino acid residue in the target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the target protein structure.

[0156] The carbon (C) atoms of an amino acid residue include α-carbon atoms and β-carbon atoms. In this embodiment, the coordinates of the β-carbon atom of each amino acid residue in the protein are determined as the protein coordinates corresponding to that amino acid residue, and the coordinates of the β-carbon atom of each amino acid residue in the target protein are determined as the target protein coordinates corresponding to that amino acid residue. Both protein coordinates and target protein coordinates can be three-dimensional coordinates.

[0157] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0158]

[0159] Here, GetCbAtom is used to obtain the coordinates of the β carbon atom, which is on the left side of the equation. This refers to the protein coordinates corresponding to the i-th amino acid residue in the protein, specifically the antibody coordinates corresponding to the i-th amino acid residue in the antibody. It is the target protein coordinate corresponding to the i-th amino acid residue in the target protein, specifically the antigen coordinate corresponding to the i-th amino acid residue in the antigen.

[0160] Step 440: Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the target protein structure.

[0161] In some embodiments, the distance includes Euclidean distance, which can be calculated based on the L2 norm. Protein-protein pairing distance can be obtained by first calculating the Euclidean distance between protein pairs, and then sequentially performing discretization and one-hot encoding on the Euclidean distance between protein pairs. Target-protein pairing distance can also be obtained by first calculating the Euclidean distance between target-protein pairs, and then sequentially performing discretization and one-hot encoding on the target-protein pairing distance.

[0162] Optionally, the Euclidean distance between protein coordinates corresponding to any two amino acid residues on the same peptide chain in the protein structure is calculated, and the protein pairing Euclidean distance is discretized to obtain a discretized protein pairing Euclidean distance. This discretized protein pairing Euclidean distance is then subjected to one-hot encoding to obtain the protein pairing distance. Similarly, the Euclidean distance between target protein coordinates corresponding to any two amino acid residues on the same peptide chain in the target protein structure is calculated, and the target protein pairing Euclidean distance is discretized to obtain a discretized target protein pairing Euclidean distance. This discretized target protein pairing Euclidean distance is then subjected to one-hot encoding to obtain the target protein pairing distance.

[0163] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0164]

[0165] Where ||·||² is the Euclidean distance, one-hot encoding is the one-hot encoding, and v represents discretization. This refers to the protein coordinates corresponding to the i-th amino acid residue in the protein, specifically the antibody coordinates corresponding to the i-th amino acid residue in the antibody. This refers to the protein coordinates corresponding to the j-th amino acid residue in the protein, specifically the antibody coordinates corresponding to the j-th amino acid residue in the antibody. These are the coordinates of the target protein corresponding to the i-th amino acid residue, or specifically, the coordinates of the antigen corresponding to the i-th amino acid residue. It is the target protein coordinate corresponding to the j-th amino acid residue in the target protein, specifically the antigen coordinate corresponding to the j-th amino acid residue in the antigen. It refers to the Euclidean distance between protein pairs, specifically the Euclidean distance between antibody pairs. It is the protein pairing distance. It refers to the Euclidean distance between target proteins, specifically the Euclidean distance between antigens. It is the target protein pairing distance.

[0166] Step 460: Based on the protein pairing distance and the target protein pairing distance, fuse the protein pairing features, target protein pairing features, protein sequence features, and target protein MSA features to obtain the fused pairing features.

[0167] This embodiment provides a method for obtaining fusion pairing features, which can improve the accuracy of fusion pairing features and fully express the interaction between different amino acid residues. Discretization and one-hot encoding can also reduce time complexity and improve data processing efficiency.

[0168] In some embodiments, the fused pairing feature can be a matrix. In this embodiment, the fused pairing feature matrix is ​​divided into four matrix blocks. After calculating the four matrix blocks separately, the four matrix blocks are concatenated to obtain the fused pairing feature. Step 460 is specifically implemented as steps 462, 464, and 466:

[0169] Step 462: Perform mapping based on protein pairing features and protein pairing distance to calculate the first part of fusion pairing features; and perform mapping based on target protein pairing features and target protein pairing distance to calculate the second part of fusion pairing features.

[0170] Feature mapping is used to map features from a high-dimensional feature space to a low-dimensional feature space. The first part of the fused pairing features is obtained by mapping based on protein pairing features and protein pairing distances, and is the top left matrix of the matrix corresponding to the fused pairing features. Specifically, step 462 performs mapping based on protein pairing features and protein pairing distances to calculate the first part of the fused pairing features, which is implemented as steps 462A and 462B: Step 462A, performs a linear transformation on the protein pairing features to obtain the transformed protein pairing features; and performs a linear transformation on the protein pairing distances to obtain the transformed protein pairing distances; Step 462B, adds the transformed protein pairing features and the transformed protein pairing distances to obtain the first part of the fused pairing features.

[0171] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0172]

[0173] in, The first part is the fusion of paired features; Linear is a linear transformation. These are protein-protein pairing characteristics, specifically antibody pairing characteristics. This refers to the protein pairing distance, specifically the antibody pairing distance. The size of the matrix block corresponding to the first part of the fusion pairing features is L. ab ×L abL ab It is the protein structure x ab The sequence length is specifically the sequence length of the antigen structure.

[0174] The second part, the fusion pairing feature, is obtained by mapping the target protein pairing feature and the target protein pairing distance. It is the bottom right matrix block of the matrix corresponding to the fusion pairing feature. Specifically, step 462 performs mapping based on the target protein pairing feature and the target protein pairing distance to calculate the second part of the fusion pairing feature, which is implemented as steps 462C and 462D: Step 462C, performs a linear transformation on the target protein pairing feature to obtain the transformed target protein pairing feature; and performs a linear transformation on the target protein pairing distance to obtain the transformed target protein pairing distance; Step 462D, adds the transformed target protein pairing feature and the transformed target protein pairing distance to obtain the second part of the fusion pairing feature.

[0175] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0176]

[0177] in, The second part is the fusion of paired features. It refers to the target protein pairing distance, specifically the antigen pairing distance. This refers to the target protein pairing distance, specifically the antigen pairing distance. The size of the matrix block corresponding to the fusion pairing features in the second part is L. ab ×L ab L ab It is the protein structure x ab The sequence length is specifically the sequence length of the antigen structure.

[0178] Step 464: Calculate the average value of the outer product based on the protein sequence features and the target protein MSA features to obtain the third part of the fusion pairing features; and determine the transpose of the third part of the fusion pairing features as the fourth part of the fusion pairing features.

[0179] The OuterProductMean is used to update the paired features with information from protein sequence features and target protein MSA features, enabling protein design models to better understand and predict protein structures. The calculation of the OuterProductMean typically includes: outer product calculation, mean calculation, and linear transformation. The third part, the fused paired features, is obtained by calculating the outer product mean based on protein sequence features and target protein MSA features; it is the top-right matrix block of the matrix corresponding to the fused paired features.

[0180] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0181]

[0182] in, The third part is the fusion of paired features. It refers to protein sequence characteristics, specifically antibody sequence characteristics. It refers to the MSA characteristics of the target protein, specifically the MSA characteristics of the antigen. OuterProductMean is the average value of the outer product.

[0183] Specifically, step 464 is implemented as steps 464A and 464B: Step 464A involves sequentially performing layer normalization and linear transformation on the protein sequence features to obtain the transformed normalized protein sequence features; and sequentially performing layer normalization and linear transformation on the target protein MSA features to obtain the transformed normalized target protein MSA features; Step 464B involves calculating the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and performing a linear transformation on the average value of the outer product to obtain the third part of the fusion pairing features.

[0184] In one example, the target protein is the antigen, and the protein is the antibody. The following code represents the calculation of the outer product mean:

[0185]

[0186] Input:number of dimensions in hidden embeddings c

[0187] 1.

[0188] 2.

[0189] 3.

[0190] 4.

[0191] 5.

[0192] 6.z ij =Linear(o ij )

[0193] 7. return{z ij}

[0194] in, It refers to protein sequence characteristics, specifically antigen sequence characteristics. It refers to the MSA characteristics of the target protein, specifically the MSA characteristics of the antigen. LayerNorm is layer normalization, and Linear is linear transformation. si b si These are elements in the transformed normalized protein sequence features and the transformed normalized target protein MSA features. ij It is the average of the outer product, z ij This is the third part, which integrates paired features.

[0195] The fourth part, the fusion pairing feature, is determined based on the symmetric property and is the bottom left matrix of the matrix corresponding to the fusion pairing feature. Specifically, the transpose of the third part, the fusion pairing feature, is used to determine the fourth part, the fusion pairing feature.

[0196] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0197]

[0198] in, The third part is the fusion of paired features. It is the fourth part and the paired features, where T represents transpose.

[0199] Step 466: Combine the first part of the fused pairing features, the second part of the fused pairing features, the third part of the fused pairing features, and the fourth part of the fused pairing features to obtain the fused pairing features.

[0200] This embodiment provides specific implementation steps for obtaining fusion pairing features, which improves the accuracy of fusion pairing features and achieves full expression of the interaction between different amino acid residues.

[0201] In some embodiments, in addition to protein features and target protein features, additional features—inter-chain features or PPI features—can be input into the protein design model to provide inter-chain connectivity and interaction information between different peptide chains of the target protein and the protein, thereby further updating the fusion pairing features and fusion sequence features. Optionally, after step 360, the method further includes steps 372 and 374:

[0202] Step 372: In the presence of interchain features or interaction features between the protein and the target protein, updated sequence features and updated pairing features are obtained based on the interchain features or interaction features, as well as the protein structure and the target protein structure.

[0203] Step 374: Update the fusion sequence features based on the updated sequence features to obtain the updated fusion sequence features; and update the fusion pairing features based on the updated pairing features to obtain the updated fusion pairing features.

[0204] Feature updates can be achieved through summation and weighted summation. Specifically, a linear transformation is performed on the updated sequence features to obtain the transformed updated sequence features. The transformed updated sequence features are then added to the fused sequence features to obtain the updated fused sequence features. Similarly, a linear transformation is performed on the updated paired features to obtain the transformed updated paired features. The transformed updated paired features are then added to the fused paired features to obtain the updated fused paired features.

[0205] In one example, let's consider a target protein as an antigen and a protein as an antibody. The following code represents this:

[0206]

[0207] in, It updates sequence features. It updates the paired features. It refers to protein structure, specifically antibody structure. It refers to the target protein structure, specifically the antigen structure. It is an interaction characteristic. These are inter-chain features; Linear is a linear transformation. It is a fusion sequence feature. It is a fusion of paired features.

[0208] In some embodiments, inter-chain features and interaction features can be converted, and one feature can be used to determine the other. The method of using inter-chain features or interaction features, protein structure, and target protein structure to update fusion sequence features and fusion pairing features is also called inter-chain feature embedding (IcfEmbedding). Optionally, step 372 is specifically implemented as steps 3721 and 3722:

[0209] Step 3721: If inter-chain features exist, determine the interaction features based on the inter-chain features and the inter-chain connectivity between the protein structure and the target protein structure; or, if interaction features exist, determine the inter-chain features based on the interaction features and the pseudo-inter-chain connectivity between the protein structure and the target protein structure.

[0210] In some embodiments, where inter-chain features exist, the protein structure and the target protein structure have an inter-chain contact. In this case, the inter-chain features and this inter-chain contact are used to generate an interaction feature (PPI). Alternatively, where interaction features exist, the protein structure and the target protein structure may have an inter-chain contact, called a pseudo-contact. In this case, the interaction features and this pseudo-contact are used to generate an inter-chain feature (PPI).

[0211] Step 3722: Perform a linear transformation on the interaction features to obtain updated sequence features; and perform a linear transformation on the inter-chain features to update the pairing features.

[0212] In one example, the target protein is the antigen, and the protein is the antibody. For inter-strand coding, the following code representation exists:

[0213]

[0214] 1.

[0215] 2.

[0216] 3.else

[0217] 4.

[0218] 5.end if

[0219] 6.

[0220] 7.

[0221] 8.else

[0222] 9.

[0223] 10.end if

[0224] 11.

[0225] 12.

[0226] 13.

[0227] 14.

[0228] 15.

[0229] 16.

[0230] in, It updates sequence features. It updates the paired features. It refers to protein structure, specifically antibody structure. It refers to the target protein structure, specifically the antigen structure. It is an interaction characteristic. These are inter-chain features; Linear is a linear transformation. It is a fusion sequence feature. It is a fusion of paired features.

[0231] In this embodiment, if additional features exist, these additional features can be converted into updated sequence features and updated pairing features in a format similar to the fused sequence features and fused pairing features, so as to update the fused sequence features and fused pairing features and improve their accuracy.

[0232] In some embodiments, in conjunction with the foregoing examples, the target protein is an antigen (Ag) and the protein is an antibody (Ab). The entire process of feature fusion is represented by the following code:

[0233]

[0234] Input:number of dimensions in hidden embeddings c

[0235] 1.

[0236] 2.

[0237] 3.

[0238] 4.

[0239] 5. Calculate the top left matrix block by adding the projections of the one-hot embedding of distance and the pairwise representation.

[0240] 6. The block size is L ab ×L ab ,where L ab is the sequence length of x ab .

[0241] 7.

[0242] 8. Calculate the bottom right matrix block by adding the projections of the one-hot embedding of distance and the pairwise representation.

[0243] 9.

[0244] 10. Calculate the top right matrix block by using OuterProductMeanbetween the single representation and / / The matrix block in the upper right corner is calculated by using the outer product mean between the individual feature representations of antibody sequence features and antigen MSA sequence features.

[0245] 11.

[0246] 12. Assign the bottom left matrix block by using symmetric propertyof / / Use the symmetry attribute to specify the bottom left matrix block.

[0247] 13.

[0248] 14.

[0249] 15.

[0250] 16.

[0251] 17. If additional inter-chain feature provided,update via residual connection. / / Updated via the linking relationships between amino acid residues when additional inter-chain features are provided. as well as

[0252] 18.

[0253] 19.

[0254] 20.

[0255] twenty one.

[0256] 22.end if

[0257] twenty three.

[0258] In other embodiments, in addition to interchain and interaction features, additional features may include epitope features and contact features. Epitope and contact features are input into the protein design model to provide information on key locations for target protein binding and spatial contact relationships. An epitope is a region on a target protein that can bind to another protein; it is typically located on the surface of the target protein and is a key location for protein recognition and binding. Epitope features can be used to indicate the region where the target protein epitope is located and the corresponding amino acid sequence. Contact features can be used to indicate the spatial contact relationships between different amino acid residues in a protein; spatial contact relationships can be determined by calculating the distance between two amino acid residues.

[0259] In some embodiments, a protein framework comprises several protein frameworks. In this embodiment, a multi-stage iterative prediction and screening method is adopted. Protein sequence recovery and protein structure prediction are performed on several protein frameworks respectively. Predicted protein sequences are screened based on confidence scores, and the screened predicted protein sequences are reused to continue generating new protein frameworks and performing iterative prediction. Finally, after the iteration ends, the predicted protein sequence with the highest confidence score and its predicted protein structure are determined as the final predicted protein sequence and its predicted protein structure. For example, after step 380, the method further includes steps 392 and 394:

[0260] Step 392: Perform protein sequence recovery and protein structure prediction sequentially on the protein framework of stage i. Based on the confidence scores of the predicted protein sequence and predicted protein structure corresponding to the protein framework of stage i, candidate protein sequences are selected from the predicted protein sequences of stage i; i is greater than 0.

[0261] Step 394: Mask the amino acid sequence corresponding to the complementarity determination region in the candidate protein sequence of stage i to obtain the protein framework of stage i+1; update i to i+1, and repeat the steps of protein sequence recovery and protein structure prediction for the protein framework of stage i in sequence until the set number of iterations is reached to obtain the predicted protein sequence and predicted protein structure with the highest confidence score.

[0262] Since both the heavy and light chains of a protein have three complementarity-determining regions (CDRs): CDR1, CDR2, and CDR3, the number of iterations in this embodiment can also be set to 3, i.e., i is greater than 0 and less than or equal to 3. The protein frames in the protein frame library are obtained by masking the amino acid sequences corresponding to the three CDRs of both the heavy and light chains. In subsequent iterations, the protein frames are obtained by masking the amino acid sequences corresponding to at least one of the three CDRs in the candidate protein sequence.

[0263] When i=1, the protein frames in the first stage are: at least some or all of the protein frames in the protein frame library, obtained by masking the amino acid sequences corresponding to CDR1, CDR2, and CDR3 of the protein sequence. Protein sequence recovery and protein structure prediction are performed sequentially on the protein frames in the first stage to obtain the predicted protein sequence, predicted protein structure, and confidence score corresponding to the protein frames in the first stage. Next, based on the confidence scores of the predicted protein sequences and predicted protein structures corresponding to the protein frames in the first stage, candidate protein sequences are selected from the predicted protein sequences in the first stage. Optionally, the candidate protein sequences can be predicted protein sequences with confidence scores greater than a threshold, or the top N predicted protein sequences with confidence scores, where N is greater than 0.

[0264] Next, when i=2, the amino acid sequences corresponding to CDR2 and CDR3 of the predicted protein sequence from the first stage are masked to obtain the protein framework for the second stage. Protein sequence recovery and protein structure prediction are then performed sequentially on the protein framework for the second stage, yielding the predicted protein sequence, predicted protein structure, and confidence score corresponding to the second-stage protein framework. Next, based on the confidence scores of the predicted protein sequence and predicted protein structure corresponding to the second-stage protein framework, candidate protein sequences are selected from the predicted protein sequences of the second stage. Optionally, the candidate protein sequences can be predicted protein sequences with confidence scores greater than a threshold, or the top N predicted protein sequences with confidence scores, where N is greater than 0.

[0265] Finally, when i=3, the amino acid sequence corresponding to CDR3 of the predicted protein sequence in the second stage is masked to obtain the protein framework in the third stage. Protein sequence recovery and protein structure prediction are then performed sequentially on the protein framework in the third stage, yielding the predicted protein sequence, predicted protein structure, and confidence score corresponding to the third-stage protein framework. If the set number of iterations has been reached, the predicted protein sequence and predicted protein structure with the highest confidence scores are selected.

[0266] In some embodiments, after obtaining the predicted protein sequence and predicted protein structure with the highest confidence scores, the method may optionally include: evaluating the efficiency and stability of the predicted protein sequence and predicted protein structure using molecular dynamics simulation (MD).

[0267] This embodiment provides a method for designing protein sequences and structures through multiple iterations and loops, as well as for screening these protein structures and sequences. Multiple iterations can improve the accuracy of the final predicted protein structure and sequence, thereby increasing the precision of protein design.

[0268] In some embodiments, a protein framework library is pre-constructed, the protein framework library including several protein frameworks. Before step 320, the method further includes steps 312, 314, and 316:

[0269] Step 312: Encode the protein sequences in the protein sequence database to obtain the encoded protein sequences.

[0270] Protein sequence databases refer to publicly available databases, such as SWISS-PROT, UniPot, and SCOP. Taking antibodies as an example, antibody sequence databases could include the Observed Antibody Space database (OAS). Encoding can be achieved through numbering, for example, numbering each amino acid in the protein sequence. The encoded protein sequence can represent the position, structure, and type of each amino acid.

[0271] Step 314: Align the encoded protein sequence with a standard database to obtain the variant region-connector region genes and frequencies corresponding to the encoded protein sequence.

[0272] Standard databases are dedicated to immunological and genetic research, providing extensive gene and protein sequence information, along with related annotations and functional predictions. Standard databases include the IMGT database. Taking antibodies as an example, antibody structures include a variant region (V region) and a linker region (J region). The variant region is used to recognize and bind to specific types of antigens, while the linker region ensures the stability of the antibody structure. The variant-linker region genes (VJ genes) and their frequencies can indicate the complementarity-determining regions in antibodies.

[0273] Step 316: Determine the protein framework based on the encoded protein sequences with frequencies greater than a threshold.

[0274] In some embodiments, regions where the frequency of variant-connector genes is greater than a threshold are identified as complementation-determining regions (CDGs). For the original protein sequence corresponding to the encoded protein sequence, the CDGs in the protein sequence are masked to obtain the protein framework.

[0275] This embodiment provides a method for pre-constructing protein frameworks, which are crucial for protein stability and solubility and can be used for protein sequence recovery, thereby enabling protein design to be implemented as a high-throughput protein design method and improving the efficiency of protein design.

[0276] Part Three: Training Phase of Protein Design Model

[0277] In some embodiments, the training phase of the protein design model is performed by a computer device, which may be... Figure 1 The terminal 120 and / or server 140 store the protein design model. Prior to step 320, the method further includes at least a portion of steps 520, 540, 560, and 580:

[0278] Step 520: Obtain the sample protein sequence and extract the sample protein features from the sample protein sequence.

[0279] The sample protein sequence is the complete protein sequence used in the training phase of the protein design model, providing complete sequence information for the model's learning. The sample protein sequence consists of the peptide chain and the amino acids within that chain. It includes the amino acid sequences corresponding to the heavy chain and the light chain. For example, a sample protein sequence might be represented as "H: QIQLVO…;L: DVLMTQ...YKDDDK", where "H" represents the heavy chain, "L" represents the light chain, and one letter in "QILVOCARWGSDMTYK" represents a type of amino acid.

[0280] Sample protein features are the features corresponding to the sample protein sequence. In this embodiment, sample protein features are extracted from at least two dimensions: sequence and structure. Optionally, sample protein features include: sample protein sequence features (S), sample protein pairwise features (Z), and sample protein structure (X). The sample protein sequence features are used to indicate the sequence information of the sample protein. Sample protein pairwise features, also known as sample protein pairwise features, are features corresponding to any two amino acid residues located on the same peptide chain, used to indicate the interaction between any two amino acid residues located on the same peptide chain. The sample protein structure is represented by the atomic coordinates of each amino acid residue in the sample protein, used to indicate the structural information of the sample protein. Atomic coordinates can be three-dimensional coordinates.

[0281] Step 540: Obtain the target protein sequence of the sample and extract the target protein features of the target protein sequence.

[0282] The sample target protein sequence is the target protein sequence used in the training phase of the protein design model. The sample target protein sequence is the sequence composed of the peptide chain and the amino acids within that peptide chain. For example, if the sample target protein includes an antigen, the sample target protein sequence might be represented as "A:RVQPTE…", where "A" represents the antigen and each letter in "RVQPTE" represents a type of amino acid.

[0283] In some embodiments, the sample target protein sequence is obtained; a sequence database is searched based on the sample target protein sequence to determine at least one homologous sequence related to the sample target protein sequence; the sample target protein sequence and at least one homologous sequence related to the sample target protein sequence are combined to obtain sample target protein multiple sequence alignment (MSA) data; and sample target protein features are extracted from the sample target protein MSA data. Optionally, when combining the sample target protein sequence and at least one homologous sequence related to the sample target protein sequence to obtain sample target protein MSA data, the sample target protein sequence can be placed in the first row of the sample target protein MSA data.

[0284] The MSA data for the target protein includes: the target protein sequence and at least one homologous sequence related to the target protein sequence. These homologous sequences allow us to capture the structural and functional characteristics of the target protein that have been preserved during evolution. For example, if the target protein is an antigen, and its sequence is represented as "A:RVQPTE…", after searching a sequence database, three homologous sequences are identified, represented as "RVVPSG…", "RVQPTV…", and "RVSPST…". Combining the target protein sequence and the three homologous sequences yields the target protein MSA data.

[0285] The target protein features are the features corresponding to the target protein sequence. In this embodiment, the target protein features are extracted from the target protein sequence from at least two dimensions: sequence and structure. Optionally, the target protein features include: target protein MSA features (MSA Representation, m), target protein pairwise representation (z), and target protein structure (x). The target protein MSA features are used to indicate the sequence information of the target protein sequence and its homologous sequences. The target protein pairwise representation, also known as the target protein pairwise feature, is the feature corresponding to any two amino acid residues located on the same peptide chain, used to indicate the interaction between any two amino acid residues located on the same peptide chain. The target protein structure is represented by the atomic coordinates of each amino acid residue in the target protein, used to indicate the structural information of the target protein. The atomic coordinates can be three-dimensional coordinates.

[0286] Step 560: Fuse the sample protein features and the sample target protein features to obtain the sample fusion features; the sample fusion features are used to characterize the sequence information and structural information of the interaction between the sample protein indicated by the sample protein sequence and the sample target protein indicated by the sample target protein sequence.

[0287] Complex representation is obtained by fusing sample protein features with sample target protein features. Since there are interactions between sample proteins and sample target proteins, and they influence each other, complex representation is used to characterize the sequence and structural information resulting from the interaction between the sample protein sequence (indicated by the sample protein sequence) and the sample target protein sequence (indicated by the sample target protein sequence).

[0288] In this embodiment, the sample fusion features include: sample fusion sequence features and sample fusion pairing features. The sample fusion sequence features indicate the sequence information of the interaction between the sample protein and the target protein. The sample fusion pairing features indicate the information of the amino acid residues of any two amino acid residues in the same peptide chain of the sample protein and any two amino acid residues in the same peptide chain of the target protein, under intra-chain or inter-chain interactions. The sample fusion pairing features also include the structural and sequence information corresponding to the sample protein sequence and the sample fusion protein sequence, which will be explained in subsequent embodiments.

[0289] Step 580: Perform prediction on the sample fusion features to obtain the prediction results. With the goal of reducing the difference between the prediction results and the sample results, adjust the model parameters of the protein design model.

[0290] This embodiment performs multi-task prediction, with each task including a corresponding prediction result. The model parameters of the protein design model are adjusted to reduce the difference between the prediction results and the sample results during training.

[0291] In summary, the protein design method provided in this application involves: acquiring a sample protein sequence and extracting sample protein features from the sample protein sequence; acquiring a sample target protein sequence and extracting sample target protein features from the sample target protein sequence; fusing the sample protein features and the sample target protein features to obtain sample fusion features; using the sample fusion features to characterize the sequence information and structural information under the interaction between the sample protein indicated by the sample protein sequence and the sample target protein indicated by the sample target protein sequence; performing prediction on the sample fusion features to obtain prediction results; and adjusting the model parameters of the protein design model with the training objective of reducing the difference between the prediction results and the sample results. Accordingly, a training method for a protein design model is provided. This method utilizes sample protein sequences and sample target protein sequences, fusing the sample protein features of the sample protein sequences and the sample target protein features of the sample target protein sequences. This allows the protein design model to fully learn the interaction between the sample target protein indicated by the sample target protein sequence and the sample protein corresponding to the sample protein sequence. By fully learning various features, the accuracy of the protein design model is improved. This enables the protein design model to design proteins with the ability to bind to and interact with target proteins during the application stage, thus improving the accuracy of protein design. Furthermore, it ensures the rationality of the protein structure while designing the protein sequence.

[0292] In some embodiments, sample protein features include: sample protein sequence features, sample protein pairing features, and sample protein structure; sample target protein features include: sample target protein MSA features, sample target protein pairing features, and sample target protein structure; and sample fusion features include: sample fusion sequence features and sample fusion pairing features. For example, step 560 is specifically implemented as steps 562 and 564:

[0293] Step 562: Fuse the sample protein sequence features and the sample target protein MSA features to obtain the sample fused sequence features.

[0294] Step 564: Based on the sample protein structure and the sample target protein structure, fuse the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features to obtain the sample fused pairing features.

[0295] Feature fusion methods include at least one of the following: addition, multiplication, concatenation, linear combination, weighted summation, and attention mechanism fusion. Sample fusion sequence features are obtained by fusing sample protein sequence features and sample target protein MSA features. Sample fusion paired features are obtained by fusing sample protein paired features, sample target protein paired features, sample protein sequence features, and sample target protein MSA features, based on the sample protein structure and sample target protein structure. Through feature fusion, information from different dimensions such as sequence and structure can be combined, improving the accuracy of sample fusion features.

[0296] This embodiment provides a method for fusing various features from the sample protein sequence features and the sample target protein features. The sample fusion sequence features are used to indicate the sequence information of the sample protein sequence and the sample target protein sequence. The sample fusion pairing features are used to indicate the intrachain or interchain interaction information between any two amino acid residues in the same peptide chain of the sample protein and any two amino acid residues in the same peptide chain of the sample target protein. They also include the structural and sequence information corresponding to the sample protein sequence and the sample fusion protein sequence, respectively. This allows for a full expression of the interaction between the sample protein and the sample target protein, which is beneficial for the protein design model to achieve sufficient learning.

[0297] In some embodiments, step 562 may be implemented as steps 620 and 640:

[0298] Step 620: Perform a linear transformation on the sample protein sequence features to obtain the transformed sample protein sequence features; and perform a linear transformation on the first row of sample target protein MSA features in the sample target protein MSA features to obtain the transformed first row of sample target protein MSA features.

[0299] Linear transformations can be used for at least one of feature extraction, dimensionality reduction, and data transformation. The transformed sample protein sequence features are obtained by performing a linear transformation on the sample protein sequence. The transformed first row of sample target protein MSA features is obtained by performing a linear transformation on the first row of sample target protein MSA features. In the sample target protein MSA data, the sample target protein sequence can be located in the first row.

[0300] Step 640: Based on cross-attention, the transformed sample protein sequence features and the transformed first-row sample target protein MSA features are fused to obtain the sample fused sequence features.

[0301] The cross-attention mechanism focuses on the interaction between the transformed sample protein sequence features and the transformed first-row sample target protein MSA features. For example, cross-attention weights are calculated for each of the transformed sample protein sequence features and the transformed first-row sample target protein MSA features. Using these weights, a weighted sum is performed on the transformed sample protein sequence features and the transformed first-row sample target protein MSA features to obtain the fused sample sequence features.

[0302] This embodiment provides a method for obtaining sample fusion sequence features, which can improve the accuracy of sample fusion sequence feature expression and facilitate protein sequence recovery in the application stage of protein design models.

[0303] In some embodiments, step 564 may be implemented as steps 620, 640, and 660:

[0304] Step 620: Based on the carbon atom coordinates of each amino acid residue in the sample protein structure, determine the protein coordinates corresponding to each amino acid residue in the sample protein structure; and based on the carbon atom coordinates of each amino acid residue in the sample target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the sample target protein structure.

[0305] The carbon (C) atoms of an amino acid residue include α-carbon atoms and β-carbon atoms. In this embodiment, the coordinates of the β-carbon atom of each amino acid residue in the sample protein are determined as the protein coordinates corresponding to that amino acid residue, and the coordinates of the β-carbon atom of each amino acid residue in the target protein are determined as the target protein coordinates corresponding to that amino acid residue. Both protein coordinates and target protein coordinates can be three-dimensional coordinates.

[0306] Step 640: Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues on the same peptide chain in the sample protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues on the same peptide chain in the sample target protein structure.

[0307] In some embodiments, the distance includes Euclidean distance, which can be calculated based on the L2 norm. Protein-protein pairing distance can be obtained by first calculating the Euclidean distance between protein pairs, and then sequentially performing discretization and one-hot encoding on the Euclidean distance between protein pairs. Target-protein pairing distance can also be obtained by first calculating the Euclidean distance between target-protein pairs, and then sequentially performing discretization and one-hot encoding on the target-protein pairing distance.

[0308] Optionally, the Euclidean distance between protein coordinates corresponding to any two amino acid residues on the same peptide chain in the sample protein structure is calculated. This protein-protein pairing Euclidean distance is then discretized to obtain a discretized protein-protein pairing Euclidean distance. This discretized protein-protein pairing Euclidean distance is then subjected to one-hot encoding to obtain the protein-protein pairing distance. Similarly, the Euclidean distance between target protein coordinates corresponding to any two amino acid residues on the same peptide chain in the sample target protein structure is calculated. This target protein-protein pairing Euclidean distance is then discretized to obtain a discretized target protein-protein pairing Euclidean distance. This discretized target protein-protein pairing Euclidean distance is then subjected to one-hot encoding to obtain the target protein-protein pairing distance.

[0309] Step 660: Based on the protein pairing distance and the target protein pairing distance, fuse the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features to obtain the sample fused pairing features.

[0310] This embodiment provides a method for obtaining sample fusion pairing features, which can improve the accuracy of sample fusion pairing features and fully express the interaction between different amino acid residues. Discretization and one-hot encoding can also reduce time complexity and improve data processing efficiency.

[0311] In some embodiments, the sample fusion pairing feature can be a matrix. In this embodiment, the sample fusion pairing feature matrix is ​​divided into four matrix blocks. After calculating the four matrix blocks separately, the four matrix blocks are concatenated to obtain the sample fusion pairing feature. Step 660 is specifically implemented as steps 662, 664, and 666:

[0312] Step 662: Perform mapping based on sample protein pairing features and protein pairing distance to calculate the first part of fusion pairing features; and perform mapping based on sample target protein pairing features and target protein pairing distance to calculate the second part of fusion pairing features.

[0313] Feature mapping is used to map features from a high-dimensional feature space to a low-dimensional feature space. The first part of the fused pairing features is obtained by mapping based on the sample protein pairing features and protein pairing distances, and is the top left matrix block of the matrix corresponding to the sample fused pairing features. Specifically, step 662 performs mapping based on the sample protein pairing features and protein pairing distances to calculate the first part of the fused pairing features, which is implemented as steps 662A and 662B: Step 662A, performs a linear transformation on the sample protein pairing features to obtain the transformed sample protein pairing features; and performs a linear transformation on the protein pairing distances to obtain the transformed protein pairing distances; Step 662B, adds the transformed sample protein pairing features and the transformed protein pairing distances to obtain the first part of the fused pairing features.

[0314] The second part, the fusion pairing feature, is obtained by mapping the sample target protein pairing feature and the target protein pairing distance. It is the bottom right matrix block of the matrix corresponding to the sample fusion pairing feature. Specifically, step 662 performs mapping based on the sample target protein pairing feature and the target protein pairing distance to calculate the second part of the fusion pairing feature, which is implemented as steps 662C and 662D: Step 662C, performs a linear transformation on the sample target protein pairing feature to obtain the transformed sample target protein pairing feature; and performs a linear transformation on the target protein pairing distance to obtain the transformed target protein pairing distance; Step 662D, adds the transformed sample target protein pairing feature and the transformed target protein pairing distance to obtain the second part of the fusion pairing feature.

[0315] Step 664: Calculate the average of the outer product based on the sample protein sequence characteristics and the sample target protein MSA characteristics to obtain the third part of the fusion pairing characteristics; and determine the transpose of the third part of the fusion pairing characteristics as the fourth part of the fusion pairing characteristics.

[0316] The OuterProductMean is used to update the paired features with information from the sample protein sequence features and the MSA features of the sample target protein, enabling protein design models to better understand and predict protein structures. The calculation of the OuterProductMean typically includes: outer product calculation, mean calculation, and linear transformation. The third part, the fused paired features, is obtained by calculating the outer product mean based on the sample protein sequence features and the sample target protein MSA features; it is the top-right matrix block of the matrix corresponding to the sample fused paired features.

[0317] Specifically, step 664 is implemented as steps 664A and 664B: Step 664A involves sequentially performing layer normalization and linear transformation on the sample protein sequence features to obtain the transformed normalized protein sequence features; and sequentially performing layer normalization and linear transformation on the sample target protein MSA features to obtain the transformed normalized target protein MSA features; Step 664B involves calculating the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and performing a linear transformation on the average value of the outer product to obtain the third part of the fusion pairing features.

[0318] The fourth part, the fusion pairing feature, is determined based on the symmetric property and is the bottom left matrix of the matrix corresponding to the sample fusion pairing feature. Specifically, the transpose of the third part of the fusion pairing feature is used to determine the fourth part of the fusion pairing feature.

[0319] Step 666: Concatenate the first part of the fused pairing features, the second part of the fused pairing features, the third part of the fused pairing features, and the fourth part of the fused pairing features to obtain the sample fused pairing features.

[0320] This embodiment provides specific implementation steps for obtaining sample fusion pairing features, which improves the accuracy of sample fusion pairing features and achieves full expression of the interaction between different amino acid residues.

[0321] In some embodiments, in addition to sample protein features and sample target protein features, additional features—inter-chain features or sample interaction features (PPI features)—can be input into the protein design model to provide inter-chain connectivity and interaction information between different peptide chains of the sample target protein and the sample protein, thereby further updating the sample fusion pairing features and sample fusion sequence features. Optionally, after step 560, the method further includes steps 572 and 574:

[0322] Step 572: In the presence of inter-sample chain features or sample interaction features between sample proteins and sample target proteins, sample update sequence features and sample update pairing features are obtained based on the inter-sample chain features or sample interaction features, as well as the sample protein structure and sample target protein structure.

[0323] Step 574: Update the sample fusion sequence features based on the sample update sequence features to obtain the updated sample fusion sequence features; and update the sample fusion pairing features based on the sample update pairing features to obtain the updated sample fusion pairing features.

[0324] Feature updates can be achieved through summation and weighted summation. Specifically, a linear transformation is performed on the sample update sequence features to obtain transformed sample update sequence features. These transformed features are then added to the sample fusion sequence features to obtain updated sample fusion sequence features. Similarly, a linear transformation is performed on the sample update paired features to obtain transformed sample update paired features. These transformed features are then added to the sample fusion paired features to obtain updated sample fusion paired features.

[0325] In some embodiments, inter-strand features and sample interaction features can be converted, and one feature can be used to determine the other. The method of using inter-strand features or sample interaction features, sample protein structure, and sample target protein structure to update sample fusion sequence features and sample fusion pairing features is also called inter-strand coding. Optionally, step 572 is implemented as steps 5721 and 5722:

[0326] Step 5721: If there are inter-chain features, determine the sample interaction features based on the inter-chain features and inter-chain connection relationships; or, if there are inter-chain features, determine the inter-chain features based on the sample interaction features and inter-chain pseudo-connection relationships.

[0327] In some embodiments, where inter-chain features exist, indicating an inter-chain contact between the sample protein structure and the target protein structure, a sample interaction feature (PI) is generated using the inter-chain features and this inter-chain contact relationship. Alternatively, where sample interaction features exist, an inter-chain contact, termed a pseudo-contact, may exist between the sample protein structure and the target protein structure; in this case, a pseudo-contact is generated using the sample interaction features and this pseudo-contact relationship.

[0328] Step 5722: Perform a linear transformation on the sample interaction features to obtain the sample update sequence features; and perform a linear transformation on the sample chain features to update the sample pairing features.

[0329] In this embodiment, if additional features exist, these additional features can be converted into sample update sequence features and sample update pairing features with formats similar to sample fusion sequence features and sample fusion pairing features, so as to update the sample fusion sequence features and sample fusion pairing features and improve their accuracy.

[0330] In other embodiments, in addition to inter-strand and interaction features, additional features may include epitope and contact features. These features are input into the protein design model to provide information on key binding sites and spatial contact relationships between the target protein and the sample protein. An epitope is a region on a target protein that can bind to another protein; it is typically located on the surface of the target protein and is a key site for protein recognition and binding. Epitope features can indicate the region where the epitope is located and the corresponding amino acid sequence. Contact features can indicate the spatial contact relationships between different amino acid residues in the sample protein; these relationships can be determined by calculating the distance between two amino acid residues.

[0331] It should also be noted that the code representation of the feature fusion part in the training phase of the protein design model can refer to the code representation of the feature fusion part in the application phase of the protein design model in the aforementioned embodiment. The code representation in the training phase of the protein design model will not be described again.

[0332] In some embodiments, the protein design model is a multi-task model trained through multi-task learning. These tasks include: a complex structure prediction task, a protein sequence recovery task, and a protein structure prediction task. The protein sequence recovery task and the protein structure prediction task can be performed as co-tasks executed simultaneously. During the training phase of the protein design model, prediction is performed as a multi-task prediction. Prediction results include: predicted complex structure, predicted protein sequence, and predicted protein structure; sample results include: sample complex structure, sample protein sequence, and sample protein structure. For example, step 580 is specifically implemented as steps 582, 584, 586, and 588:

[0333] Step 582: Perform complex structure prediction on the sample fusion features to obtain the predicted complex structure.

[0334] Step 584: Mask the amino acid sequence corresponding to the complementarity-determining region in the sample protein sequence to obtain the masked sample protein sequence.

[0335] Step 586: Perform protein sequence recovery on the sample fusion features to recover the predicted protein sequence corresponding to the masked sample protein sequence, and perform protein structure prediction on the predicted protein sequence to obtain the predicted protein structure.

[0336] Step 588: Adjust the model parameters of the protein design model with the training objectives of reducing the differences between the predicted complex structure and the sample complex structure, the differences between the predicted protein sequence and the sample protein sequence, and the differences between the predicted protein structure and the sample protein structure.

[0337] Specifically, for the complex structure prediction task: complex structure prediction is performed on the sample fusion features to obtain the predicted complex structure. For the protein sequence recovery and protein structure prediction tasks: the amino acid sequences corresponding to the complementarity-determining regions (CDMs) in the sample protein sequence are masked to obtain the masked sample protein sequence, wherein the amino acid sequences corresponding to at least one CDM in the sample protein sequence are masked. Protein sequence recovery is performed on the sample fusion features to recover the predicted protein sequence corresponding to the masked sample protein sequence, and protein structure prediction is performed on the predicted protein sequence to obtain the predicted protein structure. Finally, the model parameters of the protein design model are adjusted with the training objectives of reducing the differences between the predicted complex structure and the sample complex structure, the differences between the predicted protein sequence and the sample protein sequence, and the differences between the predicted protein structure and the sample protein structure.

[0338] In some embodiments, for multi-task learning of the protein design model, each task can be executed separately or simultaneously. When each task is executed separately, the model parameters of the protein design model are adjusted with the training objective of reducing the loss function corresponding to that task. When each task is executed simultaneously, the model parameters of the protein design model are adjusted with the training objective of reducing the total loss function corresponding to multiple tasks.

[0339] This embodiment provides a multi-task learning method for protein design models. By utilizing the correlation between different tasks, the protein design model can learn multiple related tasks simultaneously, thereby improving the performance of the protein design model on each task and enhancing its learning and generalization abilities.

[0340] In some embodiments, step 588 is specifically implemented as steps 5881, 5882, and 5883:

[0341] Step 5881: Based on the predicted complex structure and the sample complex structure, determine the complex loss function. The complex loss function is obtained by weighted summation of at least one of the following losses: inter-amino acid residue geometric loss for reducing the distance and angle between amino acid residues, frame alignment error loss for preventing local structural damage, interface frame alignment error loss for constraining inter-chain amino acid residue pairs, root mean square error loss for aligning coordinates, confidence loss for estimating confidence scores, and structural violation loss for handling structural violations.

[0342] Inter-Residue Geometric Loss is used to reduce the distance and angle between amino acid residues. It is calculated based on the distance and angle between the sample α-carbon atom and sample β-carbon atom of each amino acid residue in the sample complex structure and the predicted α-carbon atom and predicted β-carbon atom of each amino acid residue in the predicted complex structure. Frame Aligned Point Error is used to prevent local structural damage. It is obtained by projecting each atom in the predicted complex structure onto both the actual sample complex structure and the predicted complex structure. Interface Frame Aligned Point Error is used to constrain inter-chain amino acid residue pairs. Coordinate RMSD Loss is used to align coordinates and is determined based on the root mean square difference (RMSD) between the aligned actual sample complex structure and the predicted complex structure. Confidence Loss is a loss function used to estimate the confidence score, which can be either ipLDDT or ipTM score. Structure Violation Loss is used to address structural violations, which refer to incorrect peptide bond lengths and angles, as well as spatial conflicts between non-bonded atoms.

[0343] In one example, the complex loss function has the following representation:

[0344] L1 = L geo +L fape +L ifape +L rmsd +λ conf ·L conf +λ viol ·L viol

[0345] Where L1 is the complex loss function, L geo It is the geometric loss between amino acid residues, L fape It is the frame alignment point error loss, L ifape It is the interface frame alignment point error loss, L rmsd It is the root mean square error loss of coordinates, L conf It is confidence loss, L viol This is a structural violation loss. λ conf This is the confidence loss weight, for example, set to 0.1, λ viol This is the weight of the structural violation loss, for example, set to 0.01.

[0346] Step 5882: Based on the predicted protein structure and the sample protein structure, determine the structure loss function; and based on the predicted protein sequence and the sample protein sequence, determine the sequence loss function.

[0347] The structure loss function is used to reduce the discrepancy between the predicted protein structure and the sample protein structure. The sequence loss function is used to reduce the discrepancy between the predicted protein sequence and the sample protein sequence. During the training phase, the protein design model simultaneously optimizes protein structure prediction and protein sequence recovery.

[0348] In one example, the structural loss function and the sequence loss function coexist as follows:

[0349] L2=λ struture ·L struture +λ sequence ·L sequence

[0350] Where L2 is the structural sequence collaborative loss function, L struture It is the structural loss function, λ struture It is the weight of the structural loss function, L sequence It is the sequence loss function, λ sequence These are the weights of the sequence loss function.

[0351] In some embodiments, the structural loss function may also be obtained by weighted summation of at least one of the following losses: inter-amino acid residue geometric loss for reducing the distance and angle between amino acid residues, frame alignment error loss for preventing local structural damage, interface frame alignment error loss for constraining inter-chain amino acid residue pairs, root mean square error loss for aligning coordinates, confidence loss for estimating confidence scores, and structural violation loss for handling structural violations.

[0352] The inter-amino acid residue geometry loss is used to reduce the distance and angle between amino acid residues. It is calculated based on the distance and angle between the sample α-carbon atom and sample β-carbon atom of each amino acid residue in the sample protein structure and the predicted α-carbon atom and predicted β-carbon atom of each amino acid residue in the predicted protein structure. The frame alignment point error loss is used to prevent local structural damage. It is obtained by projecting each atom in the predicted protein structure onto the actual sample protein structure and the predicted protein structure. The interface frame alignment point error loss is used to constrain inter-chain amino acid residue pairs. The root mean square error loss is used to align coordinates and is determined based on the root mean square error of coordinates between the aligned actual sample protein structure and the predicted protein structure. The confidence loss is a loss function used to estimate the confidence score, which can be an ipLDDT score or an ipTM score. The structural violation loss is used to handle structural violations, which refer to incorrect peptide bond lengths and angles, as well as spatial conflicts between non-bonded atoms.

[0353] Step 5883: Perform a weighted summation of the complex loss function, structure loss function, and sequence loss function to obtain the total loss function; adjust the model parameters of the protein design model with the goal of reducing the total loss function.

[0354] The total loss function is obtained by weighted summation of the complex loss function, structure loss function, and sequence loss function, and is used to guide the multi-task learning of the protein design model. The model parameters of the protein design model are adjusted with the goal of reducing the total loss function. Specifically, the model parameters corresponding to the protein feature generation unit, target protein feature generation unit, feature fusion unit, complex structure prediction unit, protein sequence recovery unit, and protein structure prediction unit of the protein design model are adjusted.

[0355] In one example, the total loss function can be represented as follows:

[0356] L total =λ1·L1+λ2·L2

[0357] Where, L total λ1 is the total loss function, L1 is the complex loss function weight, L2 is the structure sequence collaborative loss function, and λ2 is the structure sequence collaborative loss function weight.

[0358] In this embodiment, the loss function and total loss function for each task of the multi-task learning of the protein design model are provided. With the goal of reducing the total loss function, the multi-task learning of the protein design model can be realized. This can ensure the rationality of the predicted protein structure and the feasibility of the biological function while generating predicted protein sequences.

[0359] As an example, taking the target protein as an antigen (Ag) and the protein as an antibody (Ab) in the aforementioned embodiments as examples, the protein design method will be explained in detail with reference to the various schematic diagrams.

[0360] (I) Application Scenarios

[0361] Antibodies are a major type of protein therapy. Antibody design, including antibody sequence design and antibody structure design, has wide applications in medical and drug development, biotechnology and industrial applications, materials science, diagnostics and therapy. In related technologies, the development of therapeutic antibodies still relies on animal immunization or screening antibody libraries to identify candidate molecules that bind to the target antigen. These methods are time-consuming and labor-intensive, and may not accurately produce antibodies that interact with the target antigen. Using the methods described in this application, antibody sequences and structures can be accurately designed.

[0362] (II) Protein Design Model

[0363] (1) Brief description of model structure and function

[0364] Figure 6 This is a schematic block diagram of a protein design model provided in an exemplary embodiment of this application.

[0365] When the target protein is an antigen and the protein is an antibody, the protein design model is called: Antibody Design Model (tFold-Ag) 300. The antibody design model (tFold-Ag) 300 includes: a tandem pre-trained language module (ESM-PPI) 11, a pre-trained sequence-based protein structure prediction module (tFold-Ab) 12; a pre-trained protein structure prediction module 21, wherein the pre-trained protein structure prediction module 21 is connected in parallel with the pre-trained ESM-PPI module 11 and the pre-trained tFold-Ab module 12; a feature fusion module 31 connected to both the pre-trained tFold-Ab module 12 and the pre-trained protein structure prediction module 21; a complex structure prediction module 41 and a sequence recovery module 51 connected to the feature fusion module 31, wherein the complex structure prediction module 41 and the sequence recovery module 51 are connected in parallel; and a structure prediction module 61 connected to the sequence recovery module 51. Optional features include: a recycling module 71 connected to both ends of the pre-trained protein structure prediction module 21; a recycling module 81 connected to both ends of the complex structure prediction module 41; and a recycling module 91 connected between the sequence recovery module 51 and the pre-trained ESM-PPI module 11. In some examples, a recycling module connected to the pre-trained ESM-PPI module 11 and the pre-trained tFold-Ab module 12 may also be optionally included. Figure 6 Not shown in the diagram. The feature fusion module 31, complex structure prediction module 41, sequence recovery unit 51, and structure prediction unit 61 can be combined into an AI-driven flexible docking module.

[0366] Among them, the pre-trained ESM-PPI module 11 and the pre-trained tFold-Ab module 12 are used to extract antibody features; the pre-trained protein structure prediction module 21 is used to extract antigen features; the feature fusion module 31 is used to fuse antigen features and antibody features; the complex structure prediction module 41 is used to perform complex structure prediction; the sequence recovery module 51 is used to perform antibody sequence recovery; the structure prediction module 61 is used to perform antibody structure prediction; the loop module 71 and the loop module 81 are used to reuse features and structures from previous iterations; and the loop module 91 is used to perform masking and multi-stage iterative loops.

[0367] In one possible implementation, the pre-trained protein structure prediction module 21 is implemented as the Evoformer module and structure module in the AlphaFold2 model. The sequence recovery module 51 can be implemented as a multilayer perceptron (MLP). The complex structure prediction module 41 and structure prediction module 61 can be implemented as the Evoformer module and structure module in the AlphaFold2 model, respectively. Alternatively, the complex structure prediction module 41 can also be implemented using a diffusion model to sample more possible docking postures of the antibody and antigen. The pre-trained ESM-PPI module 11 and the pre-trained tFold-Ab module 12 are implemented as modules in the tFold-Ab model. Optionally, the pre-trained ESM-PPI module 11 is used to extract the sequence information and intrachain information of the antibody and generate features for downstream structure prediction.

[0368] Sequence feature extraction:

[0369] h = ESM - PPI(x)

[0370] Where x is the input antibody sequence and h is the extracted sequence feature.

[0371] Extraction of information within the chain:

[0372] h ij =Attention(h i ,h j )

[0373] Among them, h i and h j This represents the sequence characteristics of the i-th and j-th amino acid residues located on the same peptide chain.

[0374] Optionally, the pre-trained tFold-Ab module 12 can be specifically implemented as an Evoformer-Single module. The Evoformer-Single module is used to update and optimize the sequence features input from the pre-trained ESM-PPI module 11, generating more informative sequence features and paired features.

[0375] Sequence feature updating:

[0376] h′=Evoformer-Single(h)

[0377] Where h′ is the sequence feature updated by the Evoformer-Single module.

[0378] Update of paired features:

[0379] z ij =PairUpdate(z ij ,h i ,h j )

[0380] Among them, z ij It is the pairing feature of the i-th amino acid residue and the j-th amino acid residue located on the same peptide chain, h i and h j This represents the sequence characteristics of the i-th and j-th amino acid residues located on the same peptide chain.

[0381] Optionally, the pre-trained tFold-Ab module 12 can also be implemented as a structure module. The structure module uses an invariant point attention (IPA) mechanism to map the optimized sequence features and paired features to the predicted atomic-level three-dimensional structure, that is, to obtain the predicted antibody structure. This antibody structure serves as the initial structure of the antibody so that the feature fusion module 31 can continue to perform subsequent feature fusion.

[0382] Invariant point attention mechanism:

[0383]

[0384] Among them, R i The coordinates of the i-th atom in the antibody are w. i It is the weight corresponding to the i-th atom, which is calculated through the attention mechanism.

[0385] Structure mapping:

[0386] R′=IPA(h′,R)

[0387] R′ is the antibody structure.

[0388] Optionally, when a recurrent module is also present, connected to the pre-trained ESM-PPI module 11 and the pre-trained tFold-Ab module 12, the recurrent module allows the reuse of sequence features and antibody structures from previous iterations to improve the quality of the determined antibody structures.

[0389] Cyclic Update:

[0390] h (t+1) =Recycle(h (t) ,R (t) )

[0391] R (t+1) =StructureUpdate(h (t+1) )

[0392] Among them, h (t+1) and R (t+1) These represent the sequence characteristics and antibody structure after the (t+1)th iteration, respectively.

[0393] (2) Training phase

[0394] The antibody design model (tFold-Ag)300 is a multi-task model, which includes: complex structure prediction, antibody sequence recovery, and antibody structure prediction. Among these, the antibody sequence recovery and antibody structure prediction tasks can be performed simultaneously as synergistic tasks.

[0395] 2.1: Complex Structure Prediction Task

[0396] Figure 7 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application.

[0397] The input data for the antibody design model (tFold-Ag) 300 includes complete antibody sequences and antigen sequences. The output data includes predicted complex structures and their confidence scores. Optionally, a sequence database can be queried based on the antigen sequence to obtain homologous sequences of the antigen sequence. The antigen sequence and its homologous sequences can then be further combined to form antigen MSA data and input.

[0398] Specifically, the antibody design model (tFold-Ag) 300 extracts antibody features from the antibody sequence. These antibody features include: antibody sequence features (s H-L Antibody pairing characteristics (z) H-L ), antibody structure (x H-L ), where H represents the heavy chain and L represents the light chain; and, the antigenic characteristics extracted from the antigen sequence, including: antigen MSA characteristics (m A Antigen pairing characteristics (z) A ), antigen structure (x A ), where A represents the antigen. Fusion antibody characteristics and antigen characteristics are combined to obtain fusion characteristics, which include: fusion sequence characteristics (s H-L:A ), fusion pairing features (z) H-L:A Based on the fusion features, complex structure prediction is performed to obtain the predicted complex structure (x). H-L:A The system also generates corresponding confidence scores: plddt = 0.871, ptm = 0.880, and iptm = 0.898. Based on the predicted complex structure and the sample complex structure, the complex loss function for the complex structure prediction task is obtained, and the model parameters are updated. The formula is shown below for reference:

[0399] Extraction of antibody sequence features:

[0400] h Ab =tFold-Ab(x) Ab )

[0401] Where, x Ab It is an antibody sequence, h Ab These are the extracted antibody sequence characteristics.

[0402] Extraction of initial antibody structure:

[0403] R Ab =Structure(h Ab )

[0404] Among them, R Ab This is the initial antibody structure.

[0405] Extraction of antigen sequence features:

[0406] h Ag =Alphafold2(x Ag )

[0407] Where, x Ag It is the antigen sequence, h Ag It is an antigen sequence characteristic.

[0408] Extraction of the initial antigen structure:

[0409] R Ag =Structure(h Ag )

[0410] Among them, R Ag It is the initial antigenic structure of the antigen.

[0411] Feature fusion:

[0412] h fusion =Fusion(h Ab ,h Ag )

[0413] Among them, h fusion This is a fusion feature; for specific code representations, please refer to the aforementioned method implementation examples.

[0414] Structural prediction: After updating the fused features using the Evoformer module, mapping is performed.

[0415] h′=Evoformer(h fusion )

[0416] Where h′ is the fusion feature updated by the Evoformer module, used to predict the complex structure:

[0417] r complex =StructureModule(h′)

[0418] Among them, R complex This is the predicted structure of the antibody-antigen complex.

[0419] 2.2: Synergistic Task of Antibody Sequence Recovery and Antibody Structure Prediction

[0420] Figure 8 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application.

[0421] The input data for the antibody design model (tFold-Ag) 300 includes: masked antibody sequences and antigen sequences. Optional inputs also include: PPI (Protein Interaction) information, epitope information, and contact information. The output data includes: predicted antibody sequences, predicted complex structures, and their confidence scores. Specifically, a masking technique is used to mask the amino acid sequences corresponding to the complementarity-determining regions (CUDs) in the antibody sequences to obtain masked antibody sequences. The predicted antibody sequences are obtained by recovering the amino acid sequences corresponding to the CUDs in these masked antibody sequences.

[0422] It should also be noted that when the protein design model executes both 2.1 and 2.2, the masked antibody sequence used in 2.2 can also be obtained by masking the amino acid sequence corresponding to the complementarity-determining region in the complete antibody sequence used in 2.1. In this way, the antibody features corresponding to the complete antibody sequence in 2.1 can be fully and effectively utilized in 2.2, which is beneficial for the antibody sequence recovery task.

[0423] Specifically, the antibody design model (tFold-Ag) 300, in addition to predicting the complex structure based on fusion features, can also perform antibody sequence recovery based on fusion features to obtain the predicted antibody sequence, and further predict the antibody structure based on the predicted antibody sequence, generating corresponding confidence scores: iplddt = 0.706 and iptm = 0.892. Based on the predicted antibody structure and sample antibody structure, and the predicted antibody sequence and sample antibody sequence, the sequence loss function and structure loss function are obtained, and the model parameters are updated. The formulas are shown below for reference:

[0424] Antibody sequence mask:

[0425] X masker =mask(X Ab )

[0426] Among them, X Ab It is the input antibody sequence, X masker It is a masked antibody sequence.

[0427] Antibody sequence recovery:

[0428] X′ Ab =SequenceRecovery(X masked ,h Ag )

[0429] Where, X′ Ab It is the predicted antibody sequence for recovery, h Ag It is an antigenic characteristic.

[0430] Antibody structure prediction:

[0431] R Ab =StructureModule(X′) Ab ,h Ag )

[0432] Among them, R Ab It is based on the predicted antibody sequence X′ Ab The generated predicted antibody structure.

[0433] Loss function for collaborative tasks:

[0434] L=λ struture ·L struture +λ sequence ·L sequence

[0435] Among them, L struture It is the structural loss function, L sequence It is the sequence loss function, λ struture The weights of the structural loss function, λ sequence These are the sequence loss function weights. Through this loss function, the antibody design model (tFold-Ag)300 can ensure the rationality of its structure and the feasibility of its biological function while generating new antibody sequences.

[0436] (3) Application stage

[0437] The antibody design model (tFold-Ag) 300 is used to perform the co-design of antibody sequences and antibody structures, which can realize antibodies with the ability to bind to the antigen indicated by the antigen sequence based on the antigen sequence.

[0438] In some embodiments, antibody framework templates are crucial for antibody stability and solubility, enabling them to predict the amino acid sequences corresponding to complementarity-determining regions (CDRs) in antibody sequences. Furthermore, antibody frameworks are typically conserved, exhibiting only limited variation. In this embodiment, an antibody framework library is pre-constructed. Specifically, antibodies from the OAS database are encoded using Chothia encoding to obtain encoded antibody sequences. These encoded antibody sequences are then compared with the IMGT standard library to obtain the corresponding VJ genes and record their frequencies. The top 2500 most frequent antibody sequences are selected, and the amino acid sequences corresponding to the CDRs of these sequences are masked to determine the antibody frameworks and form the antibody framework library. Converting antibody sequences to VJ genes ignores artificially modified amino acid mutation sites and merges similar homologous antibodies, thus ensuring the universality of the antibody framework.

[0439] Figure 9 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application. Figure 10 This is a schematic diagram of a protein design method provided in an exemplary embodiment of this application.

[0440] By inputting these antibody frameworks into the antibody design model (tFold-Ag)300, a complete predicted antibody sequence and structure can be predicted for each framework. Next, candidate antibody sequences can be selected based on confidence scores for subsequent iterative predictions. This embodiment sets the number of iterations to three, meaning the antibody design includes three stages:

[0441] In the first stage, all antibody frames are obtained from the antibody frame library. These antibody frames are obtained by masking the amino acid sequences corresponding to the CDR1, CDR2, and CDR3 regions of the antibody sequence. The antibody sequence recovery is performed using the antibody design model (tFold-Ag) 300 to obtain the predicted antibody sequence, its predicted antibody structure, and its confidence score. Predicted antibody sequences with confidence scores higher than a threshold are used as candidate antibody sequences and will proceed to the second stage.

[0442] In the second stage, the amino acid sequences corresponding to the CDR2 and CDR3 regions in the masked candidate antibody sequences are used to obtain the antibody framework. The antibody design model (tFold-Ag) 300 is used to perform antibody sequence recovery to obtain the predicted antibody sequence, its predicted antibody structure, and confidence score. The predicted antibody sequences with confidence scores higher than the threshold are used as candidate antibody sequences and will enter the third stage.

[0443] In the third stage, only the amino acid sequence corresponding to the CDR3 region in the candidate antibody series is masked to obtain the antibody framework. The antibody design model (tFold-Ag) 300 is used to perform antibody sequence recovery, obtaining the predicted antibody sequence, its predicted antibody structure, and confidence score. The predicted antibody sequence and predicted antibody structure with the highest confidence can be selected as the final output of the antibody design model (tFold-Ag) 300.

[0444] Optionally, the antibody framework can be divided into multiple framework clusters, each containing multiple antibody frameworks. The aforementioned steps are performed on the antibody frameworks within each framework cluster, and the predicted antibody sequence with the highest confidence score is selected as the representative of that cluster. Finally, molecular dynamics simulations can be optionally used to evaluate the efficacy and stability of the predicted antibody sequence with the highest confidence score and the predicted antibody structure. In this embodiment, using antibody frameworks to perform antibody sequence recovery and antibody structure prediction can improve the accuracy of antibody design and enhance biocompatibility and manufacturability.

[0445] (III) Beneficial Effects

[0446] The antibody design model (tFold-Ag) is a multi-task model that can not only calculate the structure of antibody-antigen complexes, but also determine the structure of the updated antibody itself and antibody structure based on the structure of the antibody-antigen complexes. It can also realize antibody sequence recovery and antibody structure prediction.

[0447] Referring to Table 1, in the performance evaluation, the antibody design model (tFold-Ag) of this embodiment was compared with end-to-end methods in related technologies such as AlphaFold-Multimer, and docking-based methods such as ZDock and ClusPro. The results show that the antibody design model (tFold-Ag) of this embodiment performs superiorly in antibody-antigen complex structure prediction, with a DockQ score of 0.217 and a TM-score of 0.708. In comparison, AlphaFold-Multimer's DockQ score is 0.158 and its TM-score is 0.665. Furthermore, the antibody design model (tFold-Ag) also performs better in the success rate (SR) of antibody-antigen complex prediction, reaching 0.283, while AlphaFold-Multimer's is only 0.182.

[0448] Table 1. Comparison of antibody-antigen complex quality predicted by different models

[0449]

[0450]

[0451] Referring to Table 2, the antibody design model (tFold-Ag) of this embodiment still achieved good results in terms of recovery rate. Computer model evaluation was performed using amino acid recovery rate (AAR) and contact amino acid recovery rate (CAAR). In SAbDab-22-DesignAb, the antibody design model (tFold-Ag) of this embodiment achieved AARs of 0.309 and 0.267 for CDR-H3 without utilizing any prior information. This recovery rate is comparable to the performance of the dyMEAN method, but dyMEAN requires prior epitope information. When evaluating the recovery rate of contact residues, the CAAR of the antibody design model (tFold-Ag) of this embodiment was 0.212. The effect of utilizing the interchain features of the antibody design model (tFold-Ag) of this embodiment in antibody design was also evaluated. Combining PPI information can significantly improve design performance. Figure 11 This is a distribution diagram of the three-round cyclic ipTM score provided in an exemplary embodiment of this application. (Reference) Figure 11 In full-length antibody design, the antibody design model (tFold-Ag) undergoes three rounds of iterations to filter out unreasonable framework regions. With increasing iteration counts, complex structures with higher confidence scores are obtained.

[0452] Table 2. Comparison of CDR region recovery rates of antibodies designed with different models

[0453]

[0454] Referring to Table 3, among the 50 antibodies designed in the laboratory, all 50 antibodies were successfully generated and expressed using the antibody design model (tFold-Ag) of this embodiment. In comparison, the expression success rate of the related EBM model was only 42%, and the expression success rate of the dWJS model was 97%.

[0455] Table 3. Comparison of expression success rates of antibodies designed with different models in wet experiments

[0456] method <![CDATA[Expression Success Rate Total expressed > tFold-Ag 1 dWJS 0.97 EBM 0.42 Random 0

[0457] Figure 12 This is a block diagram of a protein design apparatus provided in an exemplary embodiment of this application. Exemplarily, the protein design apparatus 800 includes at least some of the following modules: an acquisition module 820, a fusion module 840, and a processing module 860.

[0458] The acquisition module 820 is used to acquire a protein framework and extract protein features from the protein framework.

[0459] The acquisition module 820 is used to acquire the target protein sequence and extract the target protein features of the target protein sequence.

[0460] The fusion module 840 is used to fuse the protein features and the target protein features to obtain fusion features; the fusion features are used to characterize the sequence information and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence.

[0461] The processing module 860 is configured to perform protein sequence recovery on the protein framework based on the fusion features, recover the predicted protein sequence corresponding to the protein framework, and perform protein structure prediction on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence.

[0462] In some embodiments, the protein features include: protein sequence features, protein pairing features, and protein structure; the target protein features include: target protein MSA features, target protein pairing features, and target protein structure; and the fusion features include: fusion sequence features and fusion pairing features.

[0463] Fusion module 840 is used for:

[0464] The fused sequence feature is obtained by fusing the protein sequence feature and the target protein MSA feature;

[0465] Based on the protein structure and target protein structure, the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature are fused to obtain the fused pairing feature.

[0466] In some embodiments, the fusion module 840 is configured to:

[0467] The protein sequence features are subjected to a linear transformation to obtain the transformed protein sequence features; and the first row of target protein MSA features in the target protein MSA features are subjected to a linear transformation to obtain the transformed first row of target protein MSA features.

[0468] The transformed protein sequence features and the transformed first-row target protein MSA features are fused based on cross-attention to obtain the fused sequence features.

[0469] In some embodiments, the fusion module 840 is configured to:

[0470] Based on the carbon atom coordinates of each amino acid residue in the protein structure, determine the protein coordinates corresponding to each amino acid residue in the protein structure; and based on the carbon atom coordinates of each amino acid residue in the target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the target protein structure.

[0471] Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the target protein structure.

[0472] Based on the protein pairing distance and the target protein pairing distance, the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature are fused to obtain the fused pairing feature.

[0473] In some embodiments, the fusion module 840 is configured to:

[0474] Based on the protein pairing features and the protein pairing distance, a mapping is performed to calculate a first part of the fusion pairing features; and based on the target protein pairing features and the target protein pairing distance, a mapping is performed to calculate a second part of the fusion pairing features.

[0475] The outer product average value is calculated based on the protein sequence features and the target protein MSA features to obtain the third part of the fusion pairing features; and the transpose of the third part of the fusion pairing features is determined as the fourth part of the fusion pairing features.

[0476] The first part of the fusion pairing feature, the second part of the fusion pairing feature, the third part of the fusion pairing feature, and the fourth part of the fusion pairing feature are spliced ​​together to obtain the fusion pairing feature.

[0477] In some embodiments, the fusion module 840 is configured to:

[0478] A linear transformation is performed on the protein pairing features to obtain the transformed protein pairing features; and a linear transformation is performed on the protein pairing distance to obtain the transformed protein pairing distance.

[0479] The first part of the fusion pairing feature is obtained by adding the transformed protein pairing feature and the transformed protein pairing distance.

[0480] In some embodiments, the fusion module 840 is configured to:

[0481] A linear transformation is performed on the target protein pairing features to obtain the transformed target protein pairing features; and a linear transformation is performed on the target protein pairing distance to obtain the transformed target protein pairing distance.

[0482] The transformed target protein pairing feature and the transformed target protein pairing distance are added together to obtain the second part of the fusion pairing feature.

[0483] In some embodiments, the fusion module 840 is configured to:

[0484] The protein sequence features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized protein sequence features; and the target protein MSA features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized target protein MSA features.

[0485] Calculate the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and perform a linear transformation on the average value of the outer products to obtain the third part of the fusion pairing features.

[0486] In some embodiments, the fusion module 840 is further configured to:

[0487] In the presence of interchain features or interaction features between the protein and the target protein, updated sequence features and updated pairing features are obtained based on the interchain features or interaction features, as well as the protein structure and the target protein structure.

[0488] The fusion sequence features are updated based on the updated sequence features to obtain the updated fusion sequence features; and the fusion pairing features are updated based on the updated pairing features to obtain the updated fusion pairing features.

[0489] In some embodiments, the fusion module 840 is configured to:

[0490] In the presence of the inter-chain features, interaction features are determined based on the inter-chain features and the inter-chain connectivity between the protein structure and the target protein structure;

[0491] Alternatively, in the presence of the aforementioned interaction features, inter-chain features are determined based on the interaction features and the inter-chain pseudo-connection relationship between the protein structure and the target protein structure;

[0492] Perform a linear transformation on the interaction features to obtain the updated sequence features; and perform a linear transformation on the inter-chain features to update the pairing features.

[0493] In some embodiments, the processing module 860 is further configured to:

[0494] For the protein framework in stage i, the protein sequence recovery and protein structure prediction are performed sequentially. Based on the confidence scores of the predicted protein sequence and predicted protein structure corresponding to the protein framework in stage i, candidate protein sequences are selected from the predicted protein sequences in stage i; where i is greater than 0.

[0495] Masking is performed on the amino acid sequences corresponding to the complementarity-determining regions in the candidate protein sequences of the i-th stage to obtain the protein framework of the (i+1)-th stage; i is updated to i+1, and the steps of sequentially performing protein sequence recovery and protein structure prediction on the protein framework of the i-th stage are repeated until the set number of iterations is reached to obtain the predicted protein sequence and predicted protein structure with the highest confidence score.

[0496] In some embodiments, the processing module 860 is configured to:

[0497] Encode the protein sequences in the protein sequence database to obtain the encoded protein sequences;

[0498] The encoded protein sequence is compared with a standard database to obtain the variant region-connector region genes and frequencies corresponding to the encoded protein sequence.

[0499] The protein framework is determined based on the encoded protein sequences whose frequencies are greater than the threshold.

[0500] In some embodiments, the model structure of a protein design model for performing a protein design method includes:

[0501] Parallel target protein feature generation unit and protein feature generation unit; feature fusion unit connected simultaneously with the target protein feature generation unit and the protein feature generation unit; complex structure prediction unit and protein sequence recovery unit connected with the feature fusion unit; protein structure prediction unit connected with the protein sequence recovery unit;

[0502] The target protein feature generation unit is used to extract target protein features from the target protein sequence and to extract sample target protein features from the sample target protein sequence; the protein feature generation unit is used to extract protein features from the protein framework and to extract sample protein features from the sample protein; the feature fusion unit is used to fuse the protein features and the target protein features, and to fuse the sample protein features and the sample target protein features; the complex structure prediction unit is used to perform complex structure prediction; the protein sequence recovery unit is used to perform protein sequence recovery; and the protein structure prediction unit is used to perform protein structure prediction.

[0503] In some embodiments, the acquisition module 820 is further configured to:

[0504] Obtain the sample protein sequence and extract the sample protein features from the sample protein sequence;

[0505] Obtain the target protein sequence of the sample, and extract the target protein features of the target protein sequence of the sample;

[0506] The fusion module 840 is also used for:

[0507] The sample protein features and the sample target protein features are fused to obtain sample fusion features; the sample fusion features are used to characterize the sequence information and structural information of the sample protein indicated by the sample protein sequence and the sample target protein indicated by the sample target protein sequence under the interaction.

[0508] It also includes a training module; the training module is used for:

[0509] Prediction is performed on the sample fusion features to obtain prediction results. The model parameters of the protein design model are adjusted with the training objective of reducing the difference between the prediction results and the sample results.

[0510] In some embodiments, the sample protein features include: sample protein sequence features, sample protein pairing features, and sample protein structure; the sample target protein features include: sample target protein MSA features, sample target protein pairing features, and sample target protein structure; and the sample fusion features include: sample fusion sequence features and sample fusion pairing features.

[0511] Fusion module 840 is used for:

[0512] The sample fusion sequence features are obtained by fusing the protein sequence features of the sample and the MSA features of the target protein of the sample.

[0513] Based on the sample protein structure and the sample target protein structure, the sample protein pairing feature, the sample target protein pairing feature, the sample protein sequence feature, and the sample target protein MSA feature are fused to obtain the sample fusion pairing feature.

[0514] In some embodiments, the fusion module 840 is configured to:

[0515] The sample protein sequence features are subjected to a linear transformation to obtain the transformed sample protein sequence features; and the first row of sample target protein MSA features in the sample target protein MSA features are subjected to a linear transformation to obtain the transformed first row of sample target protein MSA features.

[0516] The transformed sample protein sequence features and the transformed first-row sample target protein MSA features are fused based on cross-attention to obtain the sample fused sequence features.

[0517] In some embodiments, the fusion module 840 is configured to:

[0518] Based on the carbon atom coordinates of each amino acid residue in the sample protein structure, determine the protein coordinates corresponding to each amino acid residue in the sample protein structure; and based on the carbon atom coordinates of each amino acid residue in the sample target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the sample target protein structure.

[0519] Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample target protein structure.

[0520] Based on the protein pairing distance and the target protein pairing distance, the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features are fused to obtain the sample fused pairing features.

[0521] In some embodiments, the fusion module 840 is configured to:

[0522] Based on the sample protein pairing features and the protein pairing distance, a mapping is performed to calculate a first part of the fusion pairing features; and based on the sample target protein pairing features and the target protein pairing distance, a mapping is performed to calculate a second part of the fusion pairing features.

[0523] The outer product average is calculated based on the sample protein sequence characteristics and the sample target protein MSA characteristics to obtain the third part of the fusion pairing characteristics; and the transpose of the third part of the fusion pairing characteristics is determined as the fourth part of the fusion pairing characteristics.

[0524] The sample fusion pairing features are obtained by splicing the first part of the fusion pairing features, the second part of the fusion pairing features, the third part of the fusion pairing features, and the fourth part of the fusion pairing features.

[0525] In some embodiments, the fusion module 840 is configured to:

[0526] A linear transformation is performed on the protein pairing features of the samples to obtain the transformed protein pairing features; and a linear transformation is performed on the protein pairing distances to obtain the transformed protein pairing distances.

[0527] The first part of the fusion pairing feature is obtained by adding the transformed sample protein pairing feature and the transformed protein pairing distance.

[0528] In some embodiments, the fusion module 840 is configured to:

[0529] A linear transformation is performed on the sample target protein pairing features to obtain the transformed sample target protein pairing features; and a linear transformation is performed on the target protein pairing distance to obtain the transformed target protein pairing distance.

[0530] The transformed sample target protein pairing features and the transformed target protein pairing distance are added together to obtain the second part of the fusion pairing features.

[0531] In some embodiments, the fusion module 840 is configured to:

[0532] The sample protein sequence features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized protein sequence features; and the sample target protein MSA features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized target protein MSA features.

[0533] Calculate the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and perform a linear transformation on the average value of the outer products to obtain the third part of the fusion pairing features.

[0534] In some embodiments, the fusion module 840 is further configured to:

[0535] In the presence of inter-sample chain features or sample interaction features between the sample protein and the sample target protein, sample update sequence features and sample update pairing features are obtained based on the inter-sample chain features or sample interaction features, as well as the sample protein structure and the sample target protein structure.

[0536] The sample fusion sequence features are updated based on the sample update sequence features to obtain the updated sample fusion sequence features; and the sample fusion pairing features are updated based on the sample update pairing features to obtain the updated sample fusion pairing features.

[0537] In some embodiments, the fusion module 840 is configured to:

[0538] If the sample chain features exist, the sample interaction features are determined based on the sample chain features and the chain connection relationships; or, if the sample interaction features exist, the sample chain features are determined based on the sample interaction features and the chain pseudo-connection relationships.

[0539] Perform a linear transformation on the sample interaction features to obtain the sample update sequence features; and perform a linear transformation on the sample chain features to update the sample pairing features.

[0540] In some embodiments, the prediction is a multi-task prediction, and the prediction results include: predicted complex structure, predicted protein sequence, and predicted protein structure; the sample results include: sample complex structure, sample protein sequence, and sample protein structure.

[0541] The training module is used for:

[0542] Perform complex structure prediction on the sample fusion features to obtain the predicted complex structure;

[0543] Masking is performed on the amino acid sequences corresponding to the complementarity-determining regions in the sample protein sequence to obtain the masked sample protein sequence.

[0544] Protein sequence recovery is performed on the sample fusion features to recover the predicted protein sequence corresponding to the masked sample protein sequence; and protein structure prediction is performed on the predicted protein sequence to obtain the predicted protein structure.

[0545] The training objective is to reduce the differences between the predicted complex structure and the sample complex structure, the differences between the predicted protein sequence and the sample protein sequence, and the differences between the predicted protein structure and the sample protein structure, thereby adjusting the model parameters of the protein design model.

[0546] In some embodiments, the training module is used for:

[0547] Based on the predicted complex structure and the sample complex structure, a complex loss function is determined. The complex loss function is obtained by weighted summation of at least one of the following losses: inter-amino acid residue geometric loss for reducing the distance and angle between amino acid residues, frame alignment error loss for preventing local structural damage, interface frame alignment error loss for constraining inter-chain amino acid residue pairs, root mean square error loss for aligning coordinates, confidence loss for estimating confidence scores, and structural violation loss for handling structural violations.

[0548] Based on the predicted protein structure and the sample protein structure, a structure loss function is determined, and based on the predicted protein sequence and the sample protein sequence, a sequence loss function is determined.

[0549] The complex loss function, the structure loss function, and the sequence loss function are weighted and summed to obtain the total loss function; the model parameters of the protein design model are adjusted with the goal of reducing the total loss function.

[0550] It should be noted that the specific limitations of the one or more protein design devices 800 provided above can be found in the limitations of the protein design method above, and will not be repeated here. Each module of the above device can be implemented entirely or partially by software, hardware, or a combination thereof. Each module can be embedded in the processor of the computer device in hardware form or independent of the processor, or it can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0551] This application also provides a computer device, which includes: a processor and a memory, wherein the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the protein design methods provided in the above-described method embodiments.

[0552] For example, Figure 13 This is a structural block diagram of a computer device provided in an exemplary embodiment of this application. Optionally, the computer device is a server 1000.

[0553] Typically, server 1000 includes a processor 1001 and memory 1002.

[0554] Processor 1001 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). Processor 1001 may also include a main processor and a coprocessor. The main processor, also known as a central processing unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1001 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1001 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0555] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 are used to store at least one instruction, which is executed by the processor 1001 to implement the protein design methods provided in the various method embodiments of this application.

[0556] In some embodiments, the server 1000 may optionally include an input interface 1003 and an output interface 1004. The processor 1001, memory 1002, and input interfaces 1003 and 1004 can be connected via a bus or signal lines. Various peripheral devices can be connected to the input interfaces 1003 and 1004 via a bus, signal lines, or a circuit board. The input interfaces 1003 and 1004 can be used to connect at least one input / output (I / O) related peripheral device to the processor 1001 and memory 1002. In some embodiments, the processor 1001, memory 1002, and input interfaces 1003 and 1004 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1001, memory 1002, and input interfaces 1003 and 1004 can be implemented on separate chips or circuit boards, and this application does not limit this.

[0557] Those skilled in the art will understand that Figure 13 The structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0558] In an exemplary embodiment, this application also provides a chip, which includes programmable logic circuits and / or program instructions, and is used to implement the protein design method provided in the above method embodiments when the chip is run on a computer device.

[0559] This application also provides a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the protein design method provided in the above-described method embodiments.

[0560] This application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the processor of the computer device to load and execute the protein design method provided in the above-described method embodiments.

[0561] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0562] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0563] Those skilled in the art will recognize that the functions described in the embodiments of this application in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of a computer program from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0564] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A protein design method, characterized in that, The method includes: Obtain a protein framework and extract protein features from the protein framework; Obtain the target protein sequence and extract the target protein features from the target protein sequence; The protein features and the target protein features are fused to obtain fused features; the fused features are used to characterize the sequence and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence. Based on the fusion features, protein sequence recovery is performed on the protein framework to recover the predicted protein sequence corresponding to the protein framework, and protein structure prediction is performed on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence. For the protein framework in stage i, the protein sequence recovery and protein structure prediction are performed sequentially. Based on the confidence scores of the predicted protein sequence and predicted protein structure corresponding to the protein framework in stage i, candidate protein sequences are selected from the predicted protein sequences in stage i; where i is greater than 0. Masking is performed on the amino acid sequence corresponding to the complementarity determination region (CDR) in the candidate protein sequence of the i-th stage to obtain the protein framework of the (i+1)-th stage; i is updated to i+1, and the steps of performing protein sequence recovery and protein structure prediction on the protein framework of the i-th stage are repeated until the set number of iterations is reached to obtain the predicted protein sequence and predicted protein structure with the highest confidence score. The protein frameworks are multiple; when i=1, the first-stage protein framework is obtained by masking the amino acid sequences corresponding to CDR1, CDR2, and CDR3 of the protein sequence; when i=2, the second-stage protein framework is obtained by masking the amino acid sequences corresponding to CDR2 and CDR3 of the candidate protein sequences in the first stage; when i=3, the third-stage protein framework is obtained by masking the amino acid sequence corresponding to CDR3 of the candidate protein sequences in the second stage.

2. The method according to claim 1, characterized in that, The protein features include: protein sequence features, protein pairing features, and protein structure; the target protein features include: target protein MSA features, target protein pairing features, and target protein structure; and the fusion features include: fusion sequence features and fusion pairing features. The fusion of the protein feature and the target protein feature to obtain the fused feature includes: The fused sequence feature is obtained by fusing the protein sequence feature and the target protein MSA feature; Based on the protein structure and target protein structure, the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature are fused to obtain the fused pairing feature.

3. The method according to claim 2, characterized in that, The fusion of the protein sequence features and the target protein MSA features to obtain the fused sequence features includes: The protein sequence features are subjected to a linear transformation to obtain the transformed protein sequence features; and the first row of target protein MSA features in the target protein MSA features are subjected to a linear transformation to obtain the transformed first row of target protein MSA features. The transformed protein sequence features and the transformed first-row target protein MSA features are fused based on cross-attention to obtain the fused sequence features.

4. The method according to claim 2, characterized in that, The fused pairing feature, obtained by fusing the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature based on the protein structure and the target protein structure, includes: Based on the carbon atom coordinates of each amino acid residue in the protein structure, determine the protein coordinates corresponding to each amino acid residue in the protein structure; and based on the carbon atom coordinates of each amino acid residue in the target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the target protein structure. Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the target protein structure. Based on the protein pairing distance and the target protein pairing distance, the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature are fused to obtain the fused pairing feature.

5. The method according to claim 4, characterized in that, The fused pairing feature, obtained by fusing the protein pairing feature, the target protein pairing feature, the protein sequence feature, and the target protein MSA feature based on the protein pairing distance and the target protein pairing distance, includes: Based on the protein pairing features and the protein pairing distance, a mapping is performed to calculate a first part of the fusion pairing features; and based on the target protein pairing features and the target protein pairing distance, a mapping is performed to calculate a second part of the fusion pairing features. The outer product average value is calculated based on the protein sequence features and the target protein MSA features to obtain the third part of the fusion pairing features; and the transpose of the third part of the fusion pairing features is determined as the fourth part of the fusion pairing features. The first part of the fusion pairing feature, the second part of the fusion pairing feature, the third part of the fusion pairing feature, and the fourth part of the fusion pairing feature are spliced ​​together to obtain the fusion pairing feature.

6. The method according to claim 5, characterized in that, The mapping based on the protein pairing features and the protein pairing distance, and the calculation of the first part of the fusion pairing features, include: A linear transformation is performed on the protein pairing features to obtain the transformed protein pairing features; and a linear transformation is performed on the protein pairing distance to obtain the transformed protein pairing distance. The first part of the fusion pairing feature is obtained by adding the transformed protein pairing feature and the transformed protein pairing distance.

7. The method according to claim 5, characterized in that, The mapping based on the target protein pairing features and the target protein pairing distance is used to calculate the second part of the fusion pairing features, including: A linear transformation is performed on the target protein pairing features to obtain the transformed target protein pairing features; and a linear transformation is performed on the target protein pairing distance to obtain the transformed target protein pairing distance. The transformed target protein pairing feature and the transformed target protein pairing distance are added together to obtain the second part of the fusion pairing feature.

8. The method according to claim 5, characterized in that, The third part, the fusion pairing feature, is obtained by calculating the outer product average based on the protein sequence features and the target protein MSA features, including: The protein sequence features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized protein sequence features; and the target protein MSA features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized target protein MSA features. Calculate the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and perform a linear transformation on the average value of the outer products to obtain the third part of the fusion pairing features.

9. The method according to claim 2, characterized in that, The method further includes: In the presence of interchain features or interaction features between the protein and the target protein, updated sequence features and updated pairing features are obtained based on the interchain features or interaction features, as well as the protein structure and the target protein structure. The fusion sequence features are updated based on the updated sequence features to obtain the updated fusion sequence features; and the fusion pairing features are updated based on the updated pairing features to obtain the updated fusion pairing features.

10. The method according to claim 9, characterized in that, In the presence of inter-chain features or interaction features, the method of obtaining updated sequence features and updated pairing features based on the inter-chain features or interaction features, as well as the protein structure and the target protein structure, includes: In the presence of the inter-chain features, interaction features are determined based on the inter-chain features and the inter-chain connectivity between the protein structure and the target protein structure; Alternatively, in the presence of the aforementioned interaction features, inter-chain features are determined based on the interaction features and the inter-chain pseudo-connection relationship between the protein structure and the target protein structure; Perform a linear transformation on the interaction features to obtain the updated sequence features; and perform a linear transformation on the inter-chain features to update the pairing features.

11. The method according to claim 1, characterized in that, The method further includes: Encode the protein sequences in the protein sequence database to obtain the encoded protein sequences; The encoded protein sequence is compared with a standard database to obtain the variant region-connector region genes and frequencies corresponding to the encoded protein sequence. The protein framework is determined based on the encoded protein sequences whose frequencies are greater than the threshold.

12. The method according to any one of claims 1 to 11, characterized in that, The model structure for a protein design model used to execute protein design methods includes: Parallel target protein feature generation unit and protein feature generation unit; feature fusion unit connected simultaneously with the target protein feature generation unit and the protein feature generation unit; complex structure prediction unit and protein sequence recovery unit connected with the feature fusion unit; protein structure prediction unit connected with the protein sequence recovery unit; The target protein feature generation unit is used to extract target protein features from the target protein sequence and to extract sample target protein features from the sample target protein sequence; the protein feature generation unit is used to extract protein features from the protein framework and to extract sample protein features from the sample protein; the feature fusion unit is used to fuse the protein features and the target protein features, and to fuse the sample protein features and the sample target protein features; the complex structure prediction unit is used to perform complex structure prediction; the protein sequence recovery unit is used to perform protein sequence recovery; and the protein structure prediction unit is used to perform protein structure prediction.

13. The method according to any one of claims 1 to 11, characterized in that, The method further includes: Obtain the sample protein sequence and extract the sample protein features from the sample protein sequence; Obtain the target protein sequence of the sample, and extract the target protein features of the target protein sequence of the sample; The sample protein features and the sample target protein features are fused to obtain sample fusion features; the sample fusion features are used to characterize the sequence information and structural information of the sample protein indicated by the sample protein sequence and the sample target protein indicated by the sample target protein sequence under the interaction. Prediction is performed on the sample fusion features to obtain prediction results. The model parameters of the protein design model are adjusted with the training objective of reducing the difference between the prediction results and the sample results.

14. The method according to claim 13, characterized in that, The prediction is a multi-task prediction, and the prediction results include: predicted complex structure, predicted protein sequence, and predicted protein structure. The sample results include: sample complex structure, sample protein sequence, and sample protein structure. The process of performing prediction on the sample fusion features to obtain prediction results, and adjusting the model parameters of the protein design model with the training objective of reducing the difference between the prediction results and the sample results, includes: Perform complex structure prediction on the sample fusion features to obtain the predicted complex structure; Masking is performed on the amino acid sequences corresponding to the complementarity-determining regions in the sample protein sequence to obtain the masked sample protein sequence. Protein sequence recovery is performed on the sample fusion features to recover the predicted protein sequence corresponding to the masked sample protein sequence; and protein structure prediction is performed on the predicted protein sequence to obtain the predicted protein structure. The training objective is to reduce the differences between the predicted complex structure and the sample complex structure, the differences between the predicted protein sequence and the sample protein sequence, and the differences between the predicted protein structure and the sample protein structure, thereby adjusting the model parameters of the protein design model.

15. The method according to claim 14, characterized in that, The training objective is to reduce the differences between the predicted complex structure and the sample complex structure, the differences between the predicted protein sequence and the sample protein sequence, and the differences between the predicted protein structure and the sample protein structure. The model parameters of the protein design model are adjusted, including: Based on the predicted complex structure and the sample complex structure, a complex loss function is determined. The complex loss function is obtained by weighted summation of at least one of the following losses: inter-amino acid residue geometric loss for reducing the distance and angle between amino acid residues, frame alignment error loss for preventing local structural damage, interface frame alignment error loss for constraining inter-chain amino acid residue pairs, root mean square error loss for aligning coordinates, confidence loss for estimating confidence scores, and structural violation loss for handling structural violations. Based on the predicted protein structure and the sample protein structure, a structure loss function is determined, and based on the predicted protein sequence and the sample protein sequence, a sequence loss function is determined. The complex loss function, the structure loss function, and the sequence loss function are weighted and summed to obtain the total loss function; the model parameters of the protein design model are adjusted with the goal of reducing the total loss function.

16. The method according to claim 13, characterized in that, The sample protein features include: sample protein sequence features, sample protein pairing features, and sample protein structure; the sample target protein features include: sample target protein MSA features, sample target protein pairing features, and sample target protein structure; and the sample fusion features include: sample fusion sequence features and sample fusion pairing features. The fusion of the sample protein features and the sample target protein features to obtain sample fusion features includes: The sample fusion sequence features are obtained by fusing the protein sequence features of the sample and the MSA features of the target protein of the sample. Based on the sample protein structure and the sample target protein structure, the sample protein pairing feature, the sample target protein pairing feature, the sample protein sequence feature, and the sample target protein MSA feature are fused to obtain the sample fusion pairing feature.

17. The method according to claim 16, characterized in that, The fusion of the sample protein sequence features and the sample target protein MSA features to obtain the sample fused sequence features includes: The sample protein sequence features are subjected to a linear transformation to obtain the transformed sample protein sequence features; and the first row of sample target protein MSA features in the sample target protein MSA features are subjected to a linear transformation to obtain the transformed first row of sample target protein MSA features. The transformed sample protein sequence features and the transformed first-row sample target protein MSA features are fused based on cross-attention to obtain the sample fused sequence features.

18. The method according to claim 16, characterized in that, The sample fusion pairing features are obtained by fusing the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features based on the sample protein structure and the sample target protein structure, including: Based on the carbon atom coordinates of each amino acid residue in the sample protein structure, determine the protein coordinates corresponding to each amino acid residue in the sample protein structure; and based on the carbon atom coordinates of each amino acid residue in the sample target protein structure, determine the target protein coordinates corresponding to each amino acid residue in the sample target protein structure. Calculate the protein pairing distance between the protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample protein structure; and calculate the target protein pairing distance between the target protein coordinates corresponding to any two amino acid residues located on the same peptide chain in the sample target protein structure. Based on the protein pairing distance and the target protein pairing distance, the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features are fused to obtain the sample fused pairing features.

19. The method according to claim 18, characterized in that, The sample fused pairing features are obtained by fusing the sample protein pairing features, the sample target protein pairing features, the sample protein sequence features, and the sample target protein MSA features based on the protein pairing distance and the target protein pairing distance, including: Based on the sample protein pairing features and the protein pairing distance, a mapping is performed to calculate a first part of the fusion pairing features; and based on the sample target protein pairing features and the target protein pairing distance, a mapping is performed to calculate a second part of the fusion pairing features. The outer product average is calculated based on the sample protein sequence characteristics and the sample target protein MSA characteristics to obtain the third part of the fusion pairing characteristics; and the transpose of the third part of the fusion pairing characteristics is determined as the fourth part of the fusion pairing characteristics. The sample fusion pairing features are obtained by splicing the first part of the fusion pairing features, the second part of the fusion pairing features, the third part of the fusion pairing features, and the fourth part of the fusion pairing features.

20. The method according to claim 19, characterized in that, The mapping based on the sample protein pairing features and the protein pairing distance, and the calculation of the first part of the fused pairing features, include: A linear transformation is performed on the protein pairing features of the samples to obtain the transformed protein pairing features; and a linear transformation is performed on the protein pairing distances to obtain the transformed protein pairing distances. The first part of the fusion pairing feature is obtained by adding the transformed sample protein pairing feature and the transformed protein pairing distance.

21. The method according to claim 19, characterized in that, The mapping based on the sample target protein pairing features and the target protein pairing distance is used to calculate the second part of the fusion pairing features, including: A linear transformation is performed on the sample target protein pairing features to obtain the transformed sample target protein pairing features; and a linear transformation is performed on the target protein pairing distance to obtain the transformed target protein pairing distance. The transformed sample target protein pairing features and the transformed target protein pairing distance are added together to obtain the second part of the fusion pairing features.

22. The method according to claim 19, characterized in that, The third part of the fusion pairing features is obtained by calculating the outer product average based on the sample protein sequence features and the sample target protein MSA features, including: The sample protein sequence features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized protein sequence features; and the sample target protein MSA features are subjected to layer normalization and linear transformation sequentially to obtain the transformed normalized target protein MSA features. Calculate the average value of the outer product of each element in the transformed normalized protein sequence features with each element in the transformed normalized target protein MSA features, and perform a linear transformation on the average value of the outer products to obtain the third part of the fusion pairing features.

23. The method according to claim 16, characterized in that, The method further includes: In the presence of inter-sample chain features or sample interaction features between the sample protein and the sample target protein, sample update sequence features and sample update pairing features are obtained based on the inter-sample chain features or sample interaction features, as well as the sample protein structure and the sample target protein structure. The sample fusion sequence features are updated based on the sample update sequence features to obtain the updated sample fusion sequence features; and the sample fusion pairing features are updated based on the sample update pairing features to obtain the updated sample fusion pairing features.

24. The method according to claim 23, characterized in that, In the case where there are inter-chain features or interaction features between the sample protein and the sample target protein, the sample update sequence features and sample update pairing features are obtained based on the inter-chain features or interaction features, as well as the sample protein structure and the sample target protein structure, including: If the sample chain features exist, the sample interaction features are determined based on the sample chain features and the chain connection relationships; or, if the sample interaction features exist, the sample chain features are determined based on the sample interaction features and the chain pseudo-connection relationships. Perform a linear transformation on the sample interaction features to obtain the sample update sequence features; and perform a linear transformation on the sample chain features to update the sample pairing features.

25. A protein design device, characterized in that, The device includes: An acquisition module is used to acquire a protein framework and extract protein features from the protein framework. The acquisition module is used to acquire the target protein sequence and extract the target protein features from the target protein sequence. A fusion module is used to fuse the protein features and the target protein features to obtain fusion features; the fusion features are used to characterize the sequence and structural information of the interaction between the protein indicated by the protein framework and the target protein indicated by the target protein sequence. The processing module is configured to perform protein sequence recovery on the protein framework based on the fusion features, recover the predicted protein sequence corresponding to the protein framework, and perform protein structure prediction on the predicted protein sequence to obtain the predicted protein structure corresponding to the predicted protein sequence. The processing module is further configured to sequentially perform protein sequence recovery and protein structure prediction on the protein framework of the i-th stage, and based on the confidence scores of the predicted protein sequence and predicted protein structure corresponding to the protein framework of the i-th stage, select candidate protein sequences from the predicted protein sequences of the i-th stage; where i is greater than 0; perform masking on the amino acid sequences corresponding to the complementarity determination regions (CDRs) in the candidate protein sequences of the i-th stage to obtain the protein framework of the (i+1)-th stage; update i to i+1, and repeat the steps of sequentially performing protein sequence recovery and protein structure prediction on the protein framework of the i-th stage until the set number of iterations is reached to obtain the predicted protein sequence and predicted protein structure with the highest confidence scores. The protein frameworks are multiple; when i=1, the first-stage protein framework is obtained by masking the amino acid sequences corresponding to CDR1, CDR2, and CDR3 of the protein sequence; when i=2, the second-stage protein framework is obtained by masking the amino acid sequences corresponding to CDR2 and CDR3 of the candidate protein sequences in the first stage; when i=3, the third-stage protein framework is obtained by masking the amino acid sequence corresponding to CDR3 of the candidate protein sequences in the second stage.

26. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the protein design method as described in any one of claims 1 to 24.

27. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is loaded and executed by a processor to implement the protein design method as described in any one of claims 1 to 24.

28. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium, from which a processor retrieves the computer instructions, causing the processor to load and execute them to implement the protein design method as described in any one of claims 1 to 24.

Images