Methods, apparatus, devices, media, and products for determining protein interactions

Abstract

The application relates to a protein interaction determination method, device, equipment, medium and product, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing feature extraction processing on a first amino acid sequence and a second amino acid sequence to obtain sequence features; the first amino acid sequence is an amino acid sequence of a first protein, and the second amino acid sequence is an amino acid sequence of a second protein; performing feature extraction processing on molecular structure information of the first protein and molecular structure information of the second protein to obtain first structure features; performing feature extraction processing on molecular structure information of a complex to obtain second structure features, the complex being composed of the first protein and the second protein; and acquiring protein interaction information based on the sequence features, the first structure features and the second structure features, the protein interaction information being used for indicating the interaction between the first protein and the second protein.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium, and product for determining protein-protein interactions. Background Technology

[0002] In the medical field, determining the interactions between proteins is an important tool for drug development and auxiliary diagnosis.

[0003] Taking the determination of protein-protein interactions between antigens and antibodies as an example, in related technologies, sequence features are extracted from the amino acid sequences of antigens and antibodies respectively, and the interaction between antigens and antibodies is predicted by using the sequence features of antigens and antibodies respectively.

[0004] However, the above-mentioned scheme for determining the interaction between antigens and antibodies by analyzing the sequence characteristics of their amino acid sequences provides limited information, resulting in low accuracy in identifying protein interactions. Summary of the Invention

[0005] This application provides a method, apparatus, device, medium, and product for determining protein-protein interactions, which can improve the accuracy of determining protein-protein interactions. The technical solution is as follows:

[0006] On the one hand, a method for determining protein-protein interactions is provided, the method comprising:

[0007] Feature extraction processing is performed on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein; the sequence features are features used to characterize the first amino acid sequence and the second amino acid sequence.

[0008] Feature extraction processing is performed on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain a first structural feature; the first structural feature is a feature used to characterize the molecular structure of the first amino acid sequence and the molecular structure of the second amino acid sequence.

[0009] The molecular structure information of the complex is subjected to feature extraction processing to obtain a second structural feature, wherein the complex is composed of the first protein and the second protein;

[0010] Based on the sequence features, the first structural features, and the second structural features, protein interaction information is obtained, which is used to indicate the interaction between the first protein and the second protein.

[0011] On the other hand, an apparatus for determining protein interactions is provided, the apparatus comprising:

[0012] The sequence feature acquisition module is used to perform feature extraction processing on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein; the sequence features are features used to characterize the first amino acid sequence and the second amino acid sequence.

[0013] The first structural feature acquisition module is used to perform feature extraction processing on the molecular structural information of the first protein and the molecular structural information of the second protein to obtain a first structural feature; the first structural feature is a feature used to characterize the molecular structure of the first amino acid sequence and the molecular structure of the second amino acid sequence.

[0014] The second structural feature acquisition module is used to perform feature extraction processing on the molecular structure information of the complex to obtain the second structural feature, wherein the complex is composed of the first protein and the second protein;

[0015] An interaction information acquisition module is used to acquire protein interaction information based on the sequence features, the first structural features, and the second structural features, wherein the protein interaction information is used to indicate the interaction between the first protein and the second protein.

[0016] In one possible implementation, the second structural feature acquisition module is used for,

[0017] Based on the molecular structure information of the complex, protein interaction site information is obtained, which is used to indicate the structure of the binding site between the first protein and the second protein in the complex;

[0018] The protein-protein interaction site information is subjected to feature extraction processing to obtain the second structural feature.

[0019] In one possible implementation, the molecular structure information of the complex includes the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex.

[0020] The second structural feature acquisition module is used for,

[0021] Based on the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex, the distances between each atom in the first protein and each atom in the second protein are obtained.

[0022] Based on the distances between each atom in the first protein and each atom in the second protein, the atoms contained in the binding portion are determined;

[0023] Based on the atoms contained in the binding region, the protein interaction site information is obtained.

[0024] In one possible implementation, the second structural feature acquisition module is used to input the protein interaction site information into the all-atom modeling network in the protein interaction prediction model to obtain the second structural feature output by the all-atom modeling network.

[0025] In one possible implementation, the sequence feature acquisition module is used to,

[0026] The first amino acid sequence and the second amino acid sequence are input into the sequence feature extraction network in the protein function prediction model to obtain the sequence feature vector output by the sequence feature extraction network.

[0027] The sequence features are obtained based on the sequence feature vector.

[0028] In one possible implementation, the first amino acid sequence comprises a heavy chain amino acid sequence and a light chain amino acid sequence; the sequence feature acquisition module is used for,

[0029] The heavy chain amino acid sequence is input into the first sequence feature extraction subnetwork of the sequence feature extraction network to obtain the first sequence feature vector output by the first sequence feature extraction subnetwork.

[0030] The light chain amino acid sequence is input into the second sequence feature extraction subnetwork of the sequence feature extraction network to obtain the second sequence feature vector output by the second sequence feature extraction subnetwork.

[0031] The second amino acid sequence is input into the third sequence feature extraction subnetwork in the sequence feature extraction network to obtain the third sequence feature vector output by the third sequence feature extraction subnetwork.

[0032] In one possible implementation, the sequence feature acquisition module is used to input the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector into the coding network of the protein action prediction model to obtain the sequence features output by the coding network.

[0033] In one possible implementation, the sequence features include features encoded by the encoding network from the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector, respectively; or, the sequence features include features fused by the encoding network from the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector.

[0034] In one possible implementation, the device further includes:

[0035] The molecular structure information prediction module is used to perform feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein by the first structure feature acquisition module. Before obtaining the first structural feature, the first amino acid sequence and the second amino acid sequence are input into the structure prediction network in the protein function prediction model to obtain the molecular structure information of the first protein and the molecular structure information of the second protein output by the structure prediction network.

[0036] The first structural feature acquisition module is used for,

[0037] The molecular structure information of the first protein and the molecular structure information of the second protein are input into the geometric graph learning network in the protein function prediction model to obtain the molecular structure features of the first protein and the second protein output by the geometric graph learning network.

[0038] The first structural feature is obtained based on the molecular structural features of the first protein and the second protein respectively.

[0039] In one possible implementation, the molecular structure information prediction module is used for,

[0040] The first amino acid sequence is input into the first structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the first protein output by the first structure prediction subnetwork.

[0041] The second amino acid sequence is input into the second structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the second protein output by the second structure prediction subnetwork.

[0042] In one possible implementation, the first structural feature acquisition module is used to input the molecular structural features of the first protein and the second protein into the attention processing network in the protein function prediction model to obtain the first structural feature output by the attention processing network.

[0043] In one possible implementation, the first structural feature includes features obtained by the attention processing network processing the molecular structural features of the first protein and the second protein respectively; or, the first structural feature includes features obtained by the attention processing network fusing the molecular structural features of the first protein and the second protein respectively.

[0044] In one possible implementation, the interaction information acquisition module is used to input the sequence features, the first structural features, and the second structural features into the interaction prediction network in the protein interaction prediction model to obtain the protein interaction information output by the interaction prediction network.

[0045] In one possible implementation, the protein interaction information is used to indicate the specificity between the first protein and the second protein.

[0046] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the above-described method for determining protein interactions.

[0047] On the other hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer program, which is loaded and executed by a processor to implement the above-described method for determining protein interactions.

[0048] On the other hand, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the protein interaction determination method provided in the various alternative implementations described above.

[0049] The technical solution provided in this application may include the following beneficial effects:

[0050] This application provides a scheme for predicting protein-protein interactions based on the sequence characteristics of protein amino acid sequences, the structural characteristics of proteins, and the structural characteristics of protein compositions. Specifically, the scheme extracts the sequence characteristics of proteins from their amino acid sequences, extracts the structural characteristics of proteins from their molecular structure information, and extracts the structural characteristics of protein compositions from their molecular structure. These three characteristics are then combined to predict protein-protein interactions. In predicting protein-protein interactions, this scheme references not only the characteristics of the protein's amino acid sequence but also the molecular structure of the protein and the molecular structure of the protein composition, expanding the available information and thus improving the accuracy of predicting protein-protein interactions. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0052] Figure 1 This is a schematic diagram of the system used in the method for determining protein interactions provided in this application;

[0053] Figure 2 This is a flowchart illustrating a method for determining protein interactions according to an exemplary embodiment of this application;

[0054] Figure 3 This is a flowchart of a method for determining protein interactions provided in an exemplary embodiment of this application;

[0055] Figure 4 This is a schematic diagram of the structure of a protein action prediction model according to an embodiment of this application;

[0056] Figure 5 This is a schematic diagram of the structure of a sequence processing network according to an embodiment of this application;

[0057] Figure 6 This is a flowchart of a method for determining protein interactions provided in an exemplary embodiment of this application;

[0058] Figure 7 This is a schematic diagram of the structure of a first structure processing network according to an embodiment of this application;

[0059] Figure 8 This is a model framework diagram for antibody-antigen specificity prediction based on an embodiment of this application;

[0060] Figure 9 This is a flowchart illustrating a method for predicting the binding properties of an enzyme to a substrate, as shown in one embodiment of this application;

[0061] Figure 10 This is a flowchart illustrating a method for predicting the binding properties of enzymes and inhibitors according to an embodiment of this application;

[0062] Figure 11 This is a block diagram of a protein interaction determination apparatus provided in one embodiment of this application;

[0063] Figure 12 A structural block diagram of a computer device illustrated in an exemplary embodiment of this application is shown. Detailed Implementation

[0064] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent 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.

[0065] It should be understood that "multiple" as used in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0066] This application provides a method for determining protein-protein interactions, which can improve the accuracy of protein-protein interactions. For ease of understanding, several terms used in this application are explained below.

[0067] 1) Protein

[0068] Proteins are substances with a specific spatial structure formed by polypeptide chains composed of amino acids through dehydration condensation and folding. Proteins are the framework and main components of human or animal tissues and organs, playing a vital role in human life activities.

[0069] 2) Antibodies

[0070] Antibodies are large, Y-shaped proteins secreted by plasma cells (effector B cells) and used by the immune system to identify and neutralize foreign substances such as bacteria and viruses. They are found in the blood and other bodily fluids of vertebrates, as well as on the cell membrane surface of their B cells. The level of antibodies in the blood of vaccine recipients is commonly used to measure the effectiveness of the vaccine and whether the recipient has developed protective immunity against the virus.

[0071] 3) Antigen

[0072] An antigen (abbreviated as Ag) is a substance that can stimulate the body to produce an immune response. It can be foreign substances such as viruses and bacteria, as well as proteins from dead cells or cells composed of proteins. Foreign molecules can be recognized by immunoglobulins on B cells or processed by antigen-presenting cells and combined with the major histocompatibility complex to form a complex that reactivates T cells, triggering a continuous immune response.

[0073] Antibodies, as immunoglobulins, are a key component of the immune system. Generally, a complex structure containing both heavy and light chains is called an antibody, while a single-chain structure containing only one peptide chain is called a nanobody (also known as a single-domain antibody). Specific antibodies can recognize specific antigens and undergo specific binding reactions with them in vivo or in vitro. Predicting the structure of antigen-antibody complexes helps to better understand the mechanisms of the immune system and can be applied to antibody screening and evaluation, thus providing strong support for auxiliary diagnosis and adjuvant therapy, and promoting the development of antibody drugs.

[0074] Figure 1 A schematic diagram of the system used in the method for determining protein interactions provided in an exemplary embodiment of this application is shown, as follows: Figure 1 As shown, the system includes: server 110 and terminal 120.

[0075] Among them, the aforementioned server 110 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0076] The aforementioned server 110 may include a server that deploys a protein interaction prediction system and provides users with prediction services for protein-protein interactions through the protein interaction prediction system; alternatively, the aforementioned server 110 may also include a protein interaction prediction system and train or update the protein interaction prediction system.

[0077] The aforementioned terminal 120 can be a terminal device with network connectivity and data processing capabilities. For example, terminal 120 can be a smartphone, tablet computer, e-book reader, smart glasses, smartwatch, smart TV, laptop computer, desktop computer, etc.

[0078] The terminal 120 may include a user terminal that receives a prediction service for protein-protein interactions, or the terminal 120 may include a development terminal used by the developers of the protein-protein interaction prediction system.

[0079] Optionally, a protein interaction prediction system may also be deployed in the aforementioned terminal 120.

[0080] Optionally, the system described above includes one or more servers 110 and multiple terminals 120. This application embodiment does not limit the number of servers 110 and terminals 120.

[0081] The terminal and server are connected via a communication network. Optionally, the communication network can be a wired network or a wireless network.

[0082] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to Local Area Networks (LANs), Metropolitan Area Networks (MANs), Wide Area Networks (WANs), mobile, wired or wireless networks, private networks, or any combination of virtual private networks. In some embodiments, technologies and / or formats such as Hypertext Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Networks (VPNs), and Internet Protocol Security (IPsec) can also be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies. This application does not impose any limitations.

[0083] Figure 2 This application illustrates a flowchart of a method for determining protein interactions according to an exemplary embodiment. The method is performed by a computer device, which can be implemented as a terminal or a server. Figure 1 The terminal or server shown, such as Figure 2 As shown, the method for determining this protein-protein interaction includes the following steps:

[0084] Step 210: Perform feature extraction processing on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein.

[0085] The sequence features mentioned above are used to characterize the first amino acid sequence and the second amino acid sequence.

[0086] In this embodiment of the application, the computer device may provide an input interface for receiving the first amino acid sequence of the first protein and the second amino acid sequence of the second protein.

[0087] After the computer device obtains the first amino acid sequence and the second amino acid sequence, it can perform feature extraction processing on the first amino acid sequence and the second amino acid sequence to obtain the above sequence features; in one possible implementation, the feature extraction process can be implemented by a machine learning model.

[0088] The second protein mentioned above can be an antigen, and the first protein mentioned above can be an antibody corresponding to the antigen.

[0089] Antibodies have a symmetrical structure with four polypeptide chains, including two longer, relatively large, identical heavy chains (H chains) and two shorter, relatively small, identical light chains (L chains). Each heavy chain and light chain forms a pair. The chains are linked by disulfide bonds and non-covalent bonds to form a monomer molecule consisting of four polypeptide chains. The entire antibody molecule can be divided into two parts: a constant region and a variable region.

[0090] When the first protein mentioned above is an antibody, the first amino acid sequence can be an amino acid sequence pair. This amino acid sequence pair can be a sequence pair composed of the amino acid sequences of one heavy chain and one light chain that make up a pair in an antibody. Since the antibody itself has symmetry, an amino acid sequence pair can represent the amino acid sequence of the entire antibody.

[0091] Step 220: Perform feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain the first structural feature.

[0092] The aforementioned first structural feature is a feature used to characterize the molecular structure of the first protein and the molecular structure of the second protein.

[0093] The molecular structure information of the protein mentioned above can be information used to indicate the molecular structure of the protein; for example, the molecular structure information of the protein may include, but is not limited to, at least one of the positional information of each atom in the protein molecule (such as three-dimensional position coordinates) and the information of the chemical bonds between each atom (such as the type of chemical bond, bond length, etc.).

[0094] In this embodiment of the application, the computer device can obtain the molecular structure information of the first protein and the molecular structure information of the second protein.

[0095] For example, a computer device can provide an input interface to receive at least one of the molecular structure information of the first protein and the molecular structure information of the second protein.

[0096] Alternatively, computer equipment can predict the molecular structure of a protein based on its amino acid sequence; for example, it can predict the molecular structure of a first protein based on its first amino acid sequence, and vice versa. The latter method of predicting the molecular structure of a protein based on its amino acid sequence can be achieved using machine learning models.

[0097] After the computer device obtains the molecular structure information of the first protein and the molecular structure information of the second protein, it can perform feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain the aforementioned first structural features; in one possible implementation, the feature extraction process can be implemented through a machine learning model.

[0098] Step 230: Perform feature extraction processing on the molecular structure information of the complex to obtain the second structural features; the complex is composed of a first protein and a second protein.

[0099] The second structural feature mentioned above is a feature used to characterize the molecular structure of the complex.

[0100] The molecular structure information of the aforementioned complex can be information used to indicate the molecular structure of the complex; for example, the molecular structure information of the aforementioned complex may include, but is not limited to, at least one of the following: position information of each atom in the molecule of the complex (such as three-dimensional position coordinates), information of the chemical bonds between each atom (such as the type of chemical bond, bond length, etc.), and information of the protein to which each atom in the complex belongs.

[0101] In this embodiment of the application, a computer device can obtain the molecular structure information of the above-mentioned complex.

[0102] For example, computer equipment can provide an input interface to receive molecular structure information of the aforementioned complex.

[0103] Alternatively, a computer device can predict the molecular structure information of the complex using all or part of the information such as the amino acid sequence of the first protein, the amino acid sequence of the second protein, the molecular structure information of the first protein, and the molecular structure information of the second protein. The process of predicting the molecular structure information of the complex can be implemented using a machine learning model. For example, the machine learning model used to predict the molecular structure information of the complex can be an open-source molecular prediction tool.

[0104] After obtaining the molecular structure information of the complex, the computer device can perform feature extraction processing on the molecular structure information of the complex to obtain the aforementioned second structural features; in one possible implementation, this feature extraction process can be implemented through a machine learning model.

[0105] Step 240: Based on sequence features, first structural features, and second structural features, obtain protein interaction information, which is used to indicate the interaction between the first protein and the second protein.

[0106] In this embodiment of the application, the computer device can combine the execution results of steps 210 to 230 to comprehensively predict the interaction between the first protein and the second protein.

[0107] In one possible implementation, the process of acquiring protein interaction information can be achieved through a machine learning model.

[0108] In one possible implementation, protein interaction information is used to indicate the specificity between the first and second proteins.

[0109] In the embodiments of this application, the specificity between the first protein and the second protein may refer to the specificity of the binding between the first protein and the second protein.

[0110] Taking the first protein as the antibody and the second protein as the antigen as an example, for the first protein, this specificity can be manifested as the ability of the hypervariable region of the first protein to form a complementary binding with the epitope of the second protein. For the second protein, this specificity can be manifested as the characteristic that the second protein can only trigger a matching immune response and bind to the first protein.

[0111] In addition to the specificity mentioned above, the protein-protein interaction information can also be used to indicate other properties between the first and second proteins, such as reversibility, proportionality, etc.

[0112] Taking the first protein as the antibody and the second protein as the antigen as an example, the above reversibility refers to the property that the antigen and antibody are bound by non-covalent bonds (such as hydrogen bonds and van der Waals forces), and the resulting complex can dissociate under low pH or high salt conditions, and both still retain their original activities after dissociation; the above proportionality refers to the specific concentration ratio required for the antigen and antibody to form a visible precipitation or agglutination reaction.

[0113] In summary, the embodiments of this application provide a scheme for predicting protein-protein interactions based on the sequence features of protein amino acid sequences, the structural features of proteins, and the structural features of protein compositions. Specifically, sequence features of proteins are extracted from amino acid sequences, structural features of proteins are extracted from molecular structure information, and structural features of protein compositions are extracted from the molecular structure of protein compositions. These three features are then combined to predict protein-protein interactions. This scheme, in predicting protein-protein interactions, references not only the amino acid sequence features of proteins but also the molecular structure of proteins and the molecular structure of protein compositions, expanding the available information and thus improving the accuracy of protein-protein interaction prediction.

[0114] Based on the above Figure 2 The proposed scheme Figure 3 A flowchart illustrating a method for determining protein interactions provided in an exemplary embodiment of this application is shown. Figure 3 In the illustrated embodiment, step 230 can be replaced by steps 230a and 230b:

[0115] Step 230a: Based on the molecular structure information of the complex, obtain the protein interaction site information, which is used to indicate the structure of the binding site of the first protein and the second protein in the complex.

[0116] In this application, protein-protein interaction sites (PPI sites) refer to key regions where protein molecules interact, and these regions are crucial for understanding protein function, intracellular signal transduction, and biological processes. Identifying PPI sites is of great significance for drug design and treatment of related diseases.

[0117] Computer equipment can determine the regional structure of the region in which the first and second proteins bind in the complex by using the molecular structure information indicated by the molecular structure information of the complex, and obtain the above-mentioned protein interaction site information.

[0118] In some embodiments, the protein-protein interaction site information may include, but is not limited to, at least one of the following: the positional information of the atoms contained in the binding portion, the protein to which the atoms contained in the binding portion belong, and the information of the chemical bonds between the atoms contained in the binding portion.

[0119] Step 230b: Perform feature extraction processing on the protein interaction site information to obtain the second structural features.

[0120] In some embodiments, after a computer device acquires protein interaction site information, it can perform feature extraction processing on the protein interaction site information. This feature extraction process can be implemented through a machine learning model.

[0121] In the above embodiments of this application, the computer device can obtain information about the structure of the binding region of the first protein and the second protein in the complex from the molecular structure information of the complex, and extract the second structural features through the information of the binding region. In the subsequent process of determining the interaction between proteins, the computer device can accurately refer to the binding region between proteins to predict the interaction, thereby improving the accuracy of the prediction of the interaction between proteins.

[0122] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, the molecular structure information of the complex includes the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex; step 230a can be implemented as the following steps:

[0123] Step 230a1: Based on the coordinates of each atom in the first protein in the complex and the coordinates of each atom in the second protein in the complex, obtain the distances between each atom in the first protein and each atom in the second protein.

[0124] In some embodiments, for each atom in the complex belonging to the first protein (taking the first atom as an example), the computer device can calculate the three-dimensional spatial distance between the first atom and each atom in the complex belonging to the second protein using the three-dimensional coordinates of the first atom in the complex and the three-dimensional coordinates of each atom in the complex belonging to the second protein. In this way, the three-dimensional spatial distance between each atom in the complex belonging to the first protein and each atom in the complex belonging to the second protein can be obtained.

[0125] For example, suppose the complex contains atoms a1, a2, and a3 belonging to a first protein, and atoms b1, b2, and b3 belonging to a second protein; the computer device calculates the three-dimensional spatial distances between atom a1 and atoms b1, b2, and b3 respectively using the three-dimensional coordinates of atom a1 and the three-dimensional coordinates of atoms b1, b2, and b3. Similarly, the computer device can also calculate the three-dimensional spatial distances between atom a2 and atoms b1, b2, and b3 respectively, and the three-dimensional spatial distances between atom a3 and atoms b1, b2, and b3 respectively.

[0126] Step 230a2: Based on the distances between each atom in the first protein and each atom in the second protein, determine the atoms contained in the binding region.

[0127] In some embodiments, the computer device can compare the distance between each atom in the first protein and each atom in the second protein with a distance threshold to determine atom pairs whose distance to each other is not greater than the distance threshold. Atom pairs contain two atoms belonging to the first protein and the second protein, respectively. The atoms contained in the determined atom pairs whose distance to each other is not greater than the distance threshold are determined as the atoms contained in the aforementioned binding portion.

[0128] In some embodiments, the aforementioned distance threshold is a threshold preset by the developer.

[0129] In some embodiments, the distance threshold is determined by the structure of the complex; for example, the distance threshold is positively correlated with the total number of atoms contained in the complex; or, for example, the distance threshold is positively correlated with the distance between the two farthest atoms in the complex.

[0130] Step 230a3: Obtain protein interaction site information based on the atoms contained in the binding region.

[0131] In some embodiments, a computer device can generate the protein interaction site information by identifying the atoms contained in the binding region.

[0132] In the solution shown in the above embodiments of this application, the computer device can locate the binding site of the first protein and the second protein by the distance between atoms belonging to different proteins in the complex, and obtain the protein interaction site information through the atoms of the binding site. This process can be achieved with only simple distance calculation and does not require a complicated measurement process, which can ensure the efficiency of obtaining protein interaction site information, and thus ensure the efficiency of determining protein interactions.

[0133] Based on any one or more of the embodiments / implementations of this application above, in an exemplary embodiment, the method for determining protein interactions can be implemented by a computer device through a protein interaction prediction model. This protein interaction prediction model can be a machine learning model trained on protein samples. For example, the protein sample may include a first amino acid sequence sample corresponding to a first protein sample, a second amino acid sequence sample corresponding to a second protein sample, molecular structure information of the complex sample composed of the first and second protein samples, and labeled protein interaction information. This labeled protein interaction information is used to indicate the interaction between the first and second proteins. Optionally, the protein sample may also include molecular structure information of the first protein sample and molecular structure information of the second protein sample.

[0134] For example, please refer to Figure 4 This illustrates a structural schematic diagram of a protein function prediction model according to an embodiment of this application. Figure 4 As shown, the protein interaction prediction model includes a sequence processing network 410, a first structure processing network 420, a second structure processing network 430, and an interaction prediction network 440.

[0135] Combination Figure 4 As shown in the model structure, step 210 above can be: the computer device performs feature extraction processing on the first amino acid sequence and the second amino acid sequence through the sequence processing network 410 to obtain sequence features.

[0136] Step 220 above can be implemented as follows: performing feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein through the first structure processing network 420 to obtain the first structural features.

[0137] Step 230 above can be implemented as follows: performing feature extraction processing on the molecular structure information of the complex through the second structure processing network 430 to obtain the second structure features.

[0138] Step 240 above can be implemented as follows: the interaction prediction network 440 processes the sequence features, the first structural features, and the second structural features to obtain protein interaction information.

[0139] In some embodiments, the training process of the protein action prediction model described above can be as follows:

[0140] The sequence processing network 410 performs feature extraction processing on the first amino acid sequence sample and the second amino acid sequence sample to obtain sequence feature samples.

[0141] The first structure processing network 420 performs feature extraction processing on the molecular structure information of the first protein sample and the molecular structure information of the second protein sample to obtain the first structure feature sample.

[0142] The second structure processing network 430 performs feature extraction processing on the molecular structure information of the complex sample to obtain the second structure feature sample;

[0143] The protein-protein interaction information samples are obtained by processing the sequence feature samples, the first structural feature samples, and the second structural feature samples through the interaction prediction network 440.

[0144] The computer device updates the parameters of the sequence processing network 410, the first structure processing network 420, the second structure processing network 430, and the interaction prediction network 440 based on the differences between the protein interaction information samples and the labeled protein interaction information. For example, the computer device calculates a loss function value based on the differences between the protein interaction information samples and the labeled protein interaction information, and updates the parameters of the sequence processing network 410, the first structure processing network 420, the second structure processing network 430, and the interaction prediction network 440 based on the loss function value.

[0145] The above process can be executed iteratively until the parameters of the sequence processing network 410, the first structure processing network 420, the second structure processing network 430, and the interaction prediction network 440 are updated to convergence, or the number of updates reaches the threshold.

[0146] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 230b can be implemented as follows:

[0147] By inputting the protein interaction site information into the all-atom modeling network in the protein interaction prediction model, the second structural feature output by the all-atom modeling network is obtained.

[0148] In this embodiment of the application, the above-mentioned all-atom modeling network is used to perform all-atom modeling of the binding site of the first protein and the second protein in the complex through protein interaction site information, and to obtain the structural features of the binding site in the complex, namely the above-mentioned second structural features.

[0149] The aforementioned all-atom modeling network can be one of the above. Figure 4 The second structure in the process network 430 may be part or all of it.

[0150] In some embodiments, the above-described all-atom modeling network can be a machine learning network based on the AlphaFold 3 model.

[0151] In this embodiment, the computer device takes the protein-protein interaction site information of the complex as input, performs full-atom modeling of the binding sites in the complex through a full-atom modeling network, and outputs the structural features of the complex, ensuring the accuracy of the structural features of the binding sites of the complex, thereby improving the accuracy of subsequent protein-protein interaction prediction.

[0152] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 210 can be implemented as steps 210a and 210b.

[0153] Step 210a: Input the first amino acid sequence and the second amino acid sequence into the sequence feature extraction network in the protein function prediction model to obtain the sequence feature vector output by the sequence feature extraction network.

[0154] In this embodiment of the application, the sequence feature extraction network described above may be a part of the sequence processing network 410.

[0155] In some embodiments, the sequence processing network described above can be a machine learning network based on a pre-trained protein language model.

[0156] Step 210b: Obtain sequence features based on sequence feature vectors.

[0157] In this embodiment of the application, the computer device can obtain the sequence feature vector as the sequence feature, or the computer device can perform further feature extraction processing on the sequence feature vector to obtain the sequence feature. This feature extraction processing can be implemented through a machine learning model.

[0158] The solution described in the above embodiments of this application involves a computer device extracting sequence feature vectors and obtaining sequence features from the first and second amino acid sequences using a pre-trained machine learning model network (sequence feature extraction network). This ensures the accuracy of sequence feature extraction and, consequently, the accuracy of subsequent protein-protein interaction prediction.

[0159] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, the first amino acid sequence comprises a heavy chain amino acid sequence and a light chain amino acid sequence; step 210a can be implemented as follows:

[0160] The heavy chain amino acid sequence is input into the first sequence feature extraction subnetwork in the sequence feature extraction network to obtain the first sequence feature vector output by the first sequence feature extraction subnetwork.

[0161] The light chain amino acid sequence is input into the second sequence feature extraction subnetwork of the sequence feature extraction network to obtain the second sequence feature vector output by the second sequence feature extraction subnetwork.

[0162] The second amino acid sequence is input into the third sequence feature extraction subnetwork in the sequence feature extraction network to obtain the third sequence feature vector output by the third sequence feature extraction subnetwork.

[0163] In some embodiments, the first protein may be an antibody, and the first amino acid sequence may be an amino acid sequence pair consisting of the amino acid sequences corresponding to the heavy chain and light chain of the antibody, respectively.

[0164] Since the structures of the heavy chain, light chain, and amino acid sequences of the antigen may differ significantly, using a network with identical structure and parameters for feature extraction may result in poor accuracy of the extracted features. Therefore, in this embodiment, the sequence feature extraction network may include sub-networks corresponding to the heavy chain, light chain, and antigen of the antibody (i.e., the first, second, and third sequence feature extraction sub-networks). When performing feature extraction on an amino acid sequence, the sequence is input into the corresponding sub-network for processing. The heavy chain amino acid sequence, after processing by the first sub-network, outputs the first sequence feature vector; the light chain amino acid sequence, after processing by the second sub-network, outputs the second sequence feature vector; and the second amino acid sequence, after processing by the third sub-network, outputs the third sequence feature vector.

[0165] For example, the first and second sequence feature extraction subnetworks mentioned above can be implemented using a pre-trained language model based on RoFormer, and the third sequence feature extraction subnetwork mentioned above can be implemented using a pre-trained Evolutionary Scale Modeling (ESM) model (such as ESM1 model, ESM2 model, ESM-C model, etc.).

[0166] RoFormer employs an encoder-based architecture inspired by Bidirectional Encoder Representations from Transformers (BERT), enhancing the traditional BERT framework by introducing Rotary Position Embedding (RoPE). Compared to the absolute position encoding implemented in traditional BERT, RoPE enhances the model's ability to capture long-range dependencies within sequences by computing position embeddings based on rotation matrices of relative positions. The structure of RoFormer is explained in detail below.

[0167] set up Let w represent a sequence consisting of N input labels, where w i It is identified as the i-th element of the sequence. (And S) N The relevant embedding features are represented as Where x i ∈R d Indicates token w i The d-dimensional word embedding vectors. The self-attention mechanism initially integrates positional information into the word embeddings and transforms them into representations of q, k, and v.

[0168] q m =f q (x m ,m) (1)

[0169] k n =f k (x n ,n) (2)

[0170] v n =f v (x n ,n) (3)

[0171] The m-th and n-th positions are connected by q. m k n and v n with f q f k and f v Integrate the results. Then calculate the attention weights using q and k, and define the output as a weighted sum.

[0172]

[0173] To integrate relative location information, RoPE will use q m and k n The inner product is defined as a function g, which accepts only word embeddings x. m ,x nThe relative positions m and n are used as input.

[0174] <f q (x m ,m),f k (x n ,n)>=g(x m ,x n (5)

[0175] Assuming location information is encoded only in a relative format, RoPE is represented as follows:

[0176] f q (x m ,m)=(W q x m )e imθ (6)

[0177] f k (x n ,n)=(W k x n )e inθ (7)

[0178] g(x m ,x n ,mn)=Re[(W q x m (W) k x n )*e i(m-n)θ (8)

[0179] Where Re[·] and θ∈R represent the real part of the complex number and a pre-defined nonzero constant, (W k x n ) * W k x n The conjugate of complex numbers.

[0180] The aforementioned RoFormer-based pre-trained language model (first sequence feature extraction sub-network and second sequence feature extraction sub-network) includes several improved Transformers, which replace the encoding part with RoPE on the basis of the traditional Transformer.

[0181] Specifically, taking the first protein as an antibody as an example, in this embodiment of the application, the input of the first sequence feature extraction subnetwork is the heavy chain amino acid sequence of the antibody (corresponding to the S mentioned above). N The first sequence feature extraction subnetwork extracts the embedding vector of the heavy chain amino acid sequence (corresponding to the above E). NThe PoRE part performs rotational position encoding on the embedding vector of the heavy chain amino acid sequence (corresponding to Formulas 6-8 above) to introduce the information of the absolute position of each element in the heavy chain amino acid sequence and the information of the relative position between each element into the encoding result. Then, the encoding result is processed by a self-attention mechanism, and finally the first sequence feature vector is output.

[0182] Correspondingly, the input of the second sequence feature extraction subnetwork is the light chain amino acid sequence. The second sequence feature extraction subnetwork extracts the embedding vector of the light chain amino acid sequence. The embedding vector of the light chain amino acid sequence is rotated and encoded by the PoRE part to introduce the information of the absolute position of each element in the light chain amino acid sequence and the information of the relative position between each element into the encoding result. Then, the encoding result is processed by the self-attention mechanism, and finally the second sequence feature vector is output.

[0183] Because RoPE utilizes absolute position encoding to establish relative position encoding, it avoids the use of an attention matrix. Given the folded conformation of antibody proteins, relative position encoding can be more effective in constructing protein-protein interactions, thereby improving the accuracy of sequence feature extraction from antibodies. Therefore, embodiments of this application can utilize RoFormer to pre-train large-scale language models (i.e., the aforementioned first and second sequence feature extraction subnetworks) for antibody data.

[0184] For antigen sequences, embodiments of this application fine-tuned the ESM2 model (e.g., ESM2-150M) on paired antibody-antigen datasets. The ESM2 model is pre-trained using UniRef50 protein sequences, enabling it to identify evolutionary patterns from millions of sequences. Because the ESM2 model has strong expression capabilities for general proteins, embodiments of this application utilize it to extract embedding features from antigen sequences and use them to predict antibody-antigen binding specificity.

[0185] Among them, the ESM2 model (such as the third sequence feature extraction sub-network mentioned above) contains multiple task processing sub-networks with shared embedding weights and a sequence combination network.

[0186] Taking the second protein as an antigen as an example, in this embodiment of the application, the amino acid sequence of the antigen is input into multiple task processing subnetworks in the third sequence feature extraction subnetwork for processing and outputting the corresponding task processing results (specifically, feature vectors). The sequence combination network combines the task processing results of the multiple task processing subnetworks and outputs the aforementioned third sequence feature vector.

[0187] The above approach allows for the use of different networks to perform feature extraction on different types of amino acid sequences, thereby ensuring the accuracy of the extracted sequence feature vectors and consequently the accuracy of subsequent protein-protein interaction predictions.

[0188] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 210b is implemented as follows:

[0189] The first sequence feature vector, the second sequence feature vector, and the third sequence feature vector are input into the coding network of the protein action prediction model to obtain the sequence features output by the coding network.

[0190] The aforementioned encoding network may be a part of the sequence processing network 410.

[0191] In this embodiment of the application, after the computer device obtains the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector, it further performs encoding processing on the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector through an encoding network, thereby obtaining more accurate sequence features and ensuring the accuracy of subsequent protein-protein interaction prediction.

[0192] Based on any one or more of the embodiments / implementations of this application above, in an exemplary embodiment, the sequence features include features obtained by encoding the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector by the encoding network respectively.

[0193] In this embodiment of the application, when the computer device performs encoding processing on the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector through the encoding network, the encoding network can be used to perform encoding processing on the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector respectively to obtain partial features of the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector respectively in the above sequence features. This scheme can maintain the independence of the sequence features corresponding to different amino acid sequences, better preserve the features of different amino acid sequences, thereby obtaining more accurate sequence features and ensuring the accuracy of subsequent protein-protein interaction prediction.

[0194] Based on any one or more of the embodiments / implementations described above in this application, in an exemplary embodiment, the sequence features include features obtained by the encoding network fusing the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector.

[0195] In this embodiment of the application, when the computer device performs encoding processing on the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector through an encoding network, it can simultaneously perform encoding processing on the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector through the encoding network to obtain the sequence feature after the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector are fused. This scheme can extract the connection between different amino acid sequences, strengthen the interrelationship between amino acid sequences, thereby obtaining more accurate sequence features and ensuring the accuracy of subsequent protein-protein interaction prediction.

[0196] For example, please refer to Figure 5 This illustrates a schematic diagram of the structure of a sequence processing network 410 according to an embodiment of this application. Figure 5 As shown, the sequence processing network 410 includes a sequence feature extraction network 410a and an encoding network 410b.

[0197] The sequence feature extraction network 410a includes a first sequence feature extraction subnetwork 410a1, a second sequence feature extraction subnetwork 410a2, and a third sequence feature extraction subnetwork 410a3. The first sequence feature extraction subnetwork 410a1 and the second sequence feature extraction subnetwork 410a2 are pre-trained language models based on RoFormer, and the third sequence feature extraction subnetwork 410a3 is a pre-trained ESM2 model.

[0198] The aforementioned encoding network 410b includes an encoder 410b1 and a multilayer perceptron (MLP) network 410b2.

[0199] In the process of predicting protein-protein interactions, the heavy chain amino acid sequence of the first amino acid sequence is input into the first sequence feature extraction subnetwork 410a1, and the first sequence feature vector 501 is output; the light chain amino acid sequence of the first amino acid sequence is input into the second sequence feature extraction subnetwork 410a2, and the second sequence feature vector 502 is output; the second amino acid sequence is input into the third sequence feature extraction subnetwork 410a3, and the third sequence feature vector 503 is output; the first sequence feature vector 501, the second sequence feature vector 502, and the third sequence feature vector 503 are respectively input into the encoder 41. 0b1, the encoder 410b1 outputs the first encoded feature vector 504 corresponding to the first sequence feature vector 501, the second encoded feature vector 505 corresponding to the second sequence feature vector 502, and the third encoded feature vector 506 corresponding to the third sequence feature vector 503; the first encoded feature vector 504, the second encoded feature vector 505, and the third encoded feature vector 506 are input into the MLP network 410b2, and the MLP network 410b2 fuses the first encoded feature vector 504, the second encoded feature vector 505, and the third encoded feature vector 506 to output the sequence features.

[0200] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, please refer to... Figure 6 This illustrates a flowchart of a method for determining protein interactions provided in an exemplary embodiment of this application. Figure 6 As shown, before step 220, there is also step 212. Step 220 may include steps 220a and 220b.

[0201] Step 212: Input the first amino acid sequence and the second amino acid sequence into the structure prediction network in the protein function prediction model to obtain the molecular structure information of the first protein and the molecular structure information of the second protein output by the structure prediction network.

[0202] The aforementioned structure prediction network may be a part of the aforementioned first structure processing network 420.

[0203] In this embodiment of the application, when the molecular structure information of the first protein and the second protein is not obtained, the computer device can predict the molecular structure of the protein by the amino acid sequence of the protein. Specifically, the computer device can input the first amino acid sequence and the second amino acid sequence into the above-mentioned structure prediction network to obtain the molecular structure information of the first protein and the second protein output by the structure prediction network.

[0204] In some embodiments, the above-described structure prediction network is a machine learning network based on the ESMFold model.

[0205] The ESMFold model adds a folded subnetwork and a structural subnetwork to the ESM2 model. The folded subnetwork is used to generate descriptive features from the feature vectors output by the ESM2 model. These descriptive features are one-dimensional features. The structural subnetwork can be an isovariant Transformer, whose input is the aforementioned descriptive features and whose output is the three-dimensional spatial coordinates of atoms.

[0206] For example, after inputting the first amino acid sequence into the ESMFold model, the first amino acid sequence is first processed by the ESM2 model to obtain the sequence feature vector corresponding to the first amino acid sequence. Then, the sequence feature vector corresponding to the first amino acid sequence is processed by the folding sub-network to generate the descriptive features of the first amino acid sequence. Finally, the descriptive features of the first amino acid sequence are processed by the structural sub-network to output the information of the three-dimensional spatial coordinates of the atoms in the first protein (i.e., the molecular structure information of the first protein).

[0207] For example, after inputting the second amino acid sequence into the ESMFold model, the second amino acid sequence is first processed by the ESM2 model to obtain the sequence feature vector corresponding to the second amino acid sequence. Then, the sequence feature vector corresponding to the second amino acid sequence is processed by the folding sub-network to generate the descriptive features of the second amino acid sequence. Finally, the descriptive features of the second amino acid sequence are processed by the structural sub-network to output the information of the three-dimensional spatial coordinates of the atoms in the second protein (i.e., the molecular structure information of the second protein).

[0208] The structure prediction network used to predict the molecular structure information of the first protein and the molecular structure information of the second protein can be the same structure prediction network or a structure prediction network with the same parameters, or the structure prediction networks used to predict the molecular structure information of the first protein and the molecular structure information of the second protein can be different structure prediction networks.

[0209] Step 220a: Input the molecular structure information of the first protein and the molecular structure information of the second protein into the geometric learning network in the protein function prediction model to obtain the molecular structure features of the first protein and the second protein output by the geometric learning network.

[0210] The aforementioned geometric graph learning network may be a part of the aforementioned first structure processing network 420.

[0211] In this embodiment of the application, after obtaining the molecular structure information of the first protein and the second protein, the computer device can extract the features of the molecular structure of the first protein and the second protein using the molecular structure information of the first protein and the second protein. Specifically, the computer device can input the molecular structure information of the first protein and the second protein into the above-mentioned geometric graph learning network to obtain the molecular structure features of the first protein and the second protein output by the geometric graph learning network.

[0212] In some embodiments, a computer device can extract molecular structural features from the molecular structural information of a protein through geometric learning.

[0213] Geometric Graph Learning: This application uses geometric graph learning to learn features with multiple individual structures. To learn multi-scale structural information, node updates, edge updates, and global context attention mechanisms can be utilized during training.

[0214] Node update: Define the feature vectors of node i and edge j→i in layer l as follows and Then convert it to a d-dimensional space. The message passing operation for updating node i in layer l is as follows:

[0215]

[0216] Among them, attention weight The calculation is as follows:

[0217]

[0218] and These are three weight matrices used to transform the query, key, and value representations. Edge vectors are passed through the weight matrices. Further enhance the representation of key values, while NB i This represents the neighbors of node i.

[0219] Edge update: Update edge features based on neighboring nodes to enhance the representation of interactions between residues.

[0220]

[0221] Where || represents a join operation, and EdgeMLP represents an MLP operation performed on an edge.

[0222] Specifically, in the embodiments of this application, the geometric graph learning network may include multiple information update network layers. After the molecular structure information of the protein is input into the geometric graph learning network, each information update network layer sequentially performs node update, edge update and global context attention mechanism processing, and then inputs the processing result into the next information update network layer until the molecular structure features of the protein are output.

[0223] For example, in the process of predicting the molecular structural features of the first protein, the geometric graph learning network first generates the initial node information of the first protein and the edge information between each pair of nodes based on the molecular structural information of the first protein. Each of the above nodes is a residue in the first protein (each residue contains one or more atoms), and the edge between the above two nodes refers to the connection relationship (or distance relationship) between the two residues. Each information update network layer in the geometric graph learning network updates the nodes of this layer with the input node information and edge information (corresponding to the above formula 9). Then, it updates the edge information of this layer with the input edge information, the node information of this layer, and the input edge information (corresponding to the above formula 11). After performing attention processing on the node information and edge information of this layer (corresponding to the above formula 10), it inputs the next information update network layer. The output layer can process the processing result of the last information update network layer and input the molecular structural features of the first protein (which can be a feature vector).

[0224] For example, in the process of predicting the molecular structural features of a second protein, the geometric graph learning network first generates the initial node information of the second protein and the edge information between each pair of nodes based on the molecular structural information of the second protein. Each information update network layer in the geometric graph learning network updates the nodes of this layer using the input node information and edge information, and then updates the edge information of this layer using the input edge information, the node information of this layer, and the input edge information. After performing attention processing on the node information and edge information of this layer, it is input into the next information update network layer. The output layer can process the processing result of the last information update network layer and input the molecular structural features of the second protein.

[0225] Taking the antibody-antigen interaction determination scheme as an example, through geometric graph learning, the scheme shown in the embodiments of this application can obtain monomer structure embedding features from the monomer structures of antibodies and antigens. These embedding features can capture key information within the molecular structure that affects the interaction between proteins.

[0226] Step 220b: Obtain the first structural feature based on the molecular structural features of the first protein and the second protein respectively.

[0227] In this embodiment of the application, after the computer device obtains the molecular structural features of the first protein and the second protein respectively, it can generate the first structural feature based on the molecular structural features of the first protein and the second protein respectively.

[0228] In some embodiments, the process of generating the first structural feature from the molecular structural features of the first protein and the second protein can be implemented using a machine learning model.

[0229] The solution shown in the embodiments of this application allows a computer device to predict the molecular structure information of the first and second proteins based on their respective amino acid sequences, and to extract the molecular structure features of the first and second proteins from their respective molecular structure information, thereby obtaining the first structural feature. This solution does not require the user to prepare the molecular structure information of the first and second proteins in advance, and can be applied to scenarios where the molecular structure information of proteins is lacking, thus ensuring the efficiency and applicability of obtaining the structural features of proteins, and ensuring the efficiency and applicability of subsequent protein-protein interaction prediction.

[0230] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 212 can be implemented as follows:

[0231] The first amino acid sequence is input into the first structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the first protein output by the first structure prediction subnetwork.

[0232] The second amino acid sequence is input into the second structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the second protein output by the second structure prediction subnetwork.

[0233] Since the structures of different proteins (such as antibodies and antigens) may vary significantly, using a network with identical structure and parameters to predict the molecular structure information of proteins may result in poor accuracy of the predicted molecular structure information. Therefore, in this embodiment, the structure prediction network may include separate structure prediction sub-networks (i.e., the first and second structure prediction sub-networks) corresponding to different types of proteins. When predicting the molecular structure information of a protein, the amino acid sequence is input into a corresponding structure prediction sub-network for processing. The first amino acid sequence, after processing by the first structure prediction sub-network, outputs the molecular structure information of the first protein, and the second amino acid sequence, after processing by the second structure prediction sub-network, outputs the molecular structure information of the second protein.

[0234] The above approach allows for the use of different networks to predict molecular structures for different types of amino acid sequences, thereby ensuring the accuracy of the predicted molecular structure information and consequently the accuracy of subsequent protein-protein interaction predictions.

[0235] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 220a can be implemented as follows:

[0236] The molecular structure information of the first protein is input into the first geometric graph learning subnetwork in the geometric graph learning network to obtain the molecular structure features of the first protein output by the first geometric graph learning subnetwork.

[0237] The molecular structure information of the second protein is input into the second geometric graph learning subnetwork in the geometric graph learning network to obtain the molecular structure features of the second protein output by the second geometric graph learning subnetwork.

[0238] Since the structures of different proteins (such as antibodies and antigens) may vary significantly, using a network with identical structure and parameters to extract molecular structural features may result in poor accuracy of the extracted features. Therefore, in this embodiment, the geometric learning network can include sub-networks corresponding to different types of proteins (i.e., the first and second geometric learning sub-networks). When extracting the molecular structural features of a protein, the protein's molecular structural information is input into a corresponding geometric learning sub-network for processing. Specifically, the molecular structural information of the first protein is processed by the first geometric learning sub-network, and the molecular structural features of the first protein are output. Similarly, the molecular structural information of the second protein is processed by the second geometric learning sub-network, and the molecular structural features of the second protein are output.

[0239] The above approach allows for the use of different networks to extract molecular structural features from different types of proteins, thereby ensuring the accuracy of the extracted molecular structural features and consequently the accuracy of subsequent protein-protein interaction predictions.

[0240] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 220b can be implemented as follows:

[0241] The molecular structural features of the first and second proteins are input into the attention processing network in the protein function prediction model to obtain the first structural feature output by the attention processing network.

[0242] The attention processing network described above may be a part of the first structure processing network 420 described above.

[0243] In this embodiment of the application, after the computer device obtains the molecular structural features of the first protein and the second protein respectively, it further performs attention mechanism-based processing on the molecular structural features of the first protein and the second protein respectively through an attention processing network, thereby obtaining a more accurate first structure and ensuring the accuracy of subsequent protein-protein interaction prediction.

[0244] Based on any one or more of the embodiments / implementations of this application above, in an exemplary embodiment, the first structural feature includes features obtained by the attention processing network processing the molecular structural features of the first protein and the second protein respectively.

[0245] In this embodiment of the application, when the computer device processes the molecular structural features of the first protein and the second protein through the attention processing network, it can perform attention-based processing on the molecular structural features of the first protein and the second protein respectively, thereby obtaining partial features of the first protein and the second protein respectively in the first structural features. This scheme can maintain the independence of the structural features corresponding to different proteins, better preserve the structural features of different proteins, thereby obtaining more accurate first structural features and ensuring the accuracy of subsequent protein-protein interaction prediction.

[0246] Based on any one or more of the embodiments / implementations of this application above, in an exemplary embodiment, the first structural feature includes the feature obtained by the attention processing network fusing the molecular structural features of the first protein and the second protein respectively.

[0247] In this embodiment of the application, when the computer device processes the molecular structural features of the first protein and the second protein using an attention processing network, it can simultaneously perform attention-based processing on the molecular structural features of the first protein and the second protein, thereby obtaining a first structural feature after the fusion of the molecular structural features corresponding to the first protein and the second protein. This approach can extract the connections between the structures of different proteins, strengthen the interrelationships between the structures of proteins, thereby obtaining a more accurate first structural feature and ensuring the accuracy of subsequent protein-protein interaction predictions.

[0248] For example, please refer to Figure 7 This illustrates a schematic diagram of the structure of a first structure processing network 420 according to an embodiment of this application. Figure 7As shown, the first structure processing network 420 includes a structure prediction network 420a, a geometry learning network 420b, and an attention processing network 420c; wherein, the attention processing network 420c includes an attention layer 420c1 and an MLP network 420c2.

[0249] like Figure 7 As shown, in the process of predicting protein-protein interactions, the first amino acid sequence and the second amino acid sequence are respectively input into the structure prediction network 420a, and the molecular structure information 701 of the first protein and the molecular structure information 702 of the second protein are output. The molecular structure information 701 of the first protein and the molecular structure information 702 of the second protein are respectively input into the geometric graph learning network 420b, and the molecular structure features 703 and 704 of the first protein are output. The molecular structure features 703 and 704 of the first protein are respectively input into the attention layer 420c1, and the attention features 705 and 706 of the first protein are output. The attention features 705 and 706 of the first protein are input into the MLP network 420c2, and the MLP network 420c2 fuses the attention features 705 and 706 of the first protein to output the first structural feature.

[0250] Based on any one or more embodiments / implementations of the above-described embodiments / implementations of this application, in an exemplary embodiment, step 240 can be implemented as follows:

[0251] The sequence features, first structural features, and second structural features are input into the interaction prediction network in the protein interaction prediction model to obtain the protein interaction information output by the interaction prediction network.

[0252] The aforementioned interaction prediction network can be the above-mentioned Figure 4 Interaction prediction network 440 in the middle.

[0253] In this embodiment of the application, the computer device takes sequence features, first structural features and second structural features as input, and predicts the interaction between the first protein and the second protein through an interaction prediction network, which can ensure the accuracy of protein interaction prediction.

[0254] Taking the scenario of predicting the specificity between antibodies and antigens as an example, the solution shown in the embodiments of this application extracts the features of antibody and antigen sequences based on a pre-trained language model and uses full-atom modeling to analyze the structure, thereby achieving accurate prediction of antibody-antigen. For a given antibody and antigen, it can be determined whether a certain antibody can bind to the antigen.

[0255] For example, please refer to Figure 8 This diagram illustrates a model framework for antibody-antigen specificity prediction based on embodiments of this application. Figure 8 As shown, the protein interaction prediction model includes a sequence feature extraction network 801, an encoder 802, a first MLP network 803, a structure prediction network 804, a geometric graph learning network 805, an attention layer 806, a second MLP network 807, a full-atom modeling network 808, and an interaction prediction network 809.

[0256] refer to Figure 8 The amino acid sequences of the antibody (including the amino acid sequences of the heavy chain Ab_H and the light chain Ab_L) and the amino acid sequences of the antigen Ag are input into the sequence feature extraction network 801 and the structure prediction network 804, respectively. For the sequence feature extraction network 801, the amino acid sequences of the antibody and antigen are processed sequentially through the sequence feature extraction network 801, the encoder 802, and the first MLP network 803 to obtain sequence features. For the structure prediction network 804, the amino acid sequences of the antibody and antigen are processed by the structure prediction network 804 to obtain the molecular structure information of the antibody and antigen respectively. The molecular structure information of the antibody and antigen is then processed sequentially through the geometric graph learning network 805, the attention layer 806, and the second MLP network 807 to obtain the first structural features. In addition, the molecular structure information 810 of the antibody-antigen complex is processed by atomic distance calculation and comparison to determine the PPI sites information 811 of the antibody-antigen complex. The site information 811 can be processed by the all-atom modeling network 808 to obtain the second structure information; the above sequence information, first structure information and second structure information can be processed by the interaction prediction network 809 to obtain the specificity prediction results between the antibody and the antigen.

[0257] The solution presented in this application proposes an antibody-antigen interaction prediction system (AAIP), which runs via terminal code and can have the following functions:

[0258] (1) Antibody-antigen specificity prediction: Supports input of the amino acid sequence and structure of antibody and antigen to predict the binding specificity between antibody and antigen, breaking through the limitation of only being able to predict for a certain type of antigen. AAIP can identify any antigen and achieve a wider range of applications.

[0259] (2) Screening and discovery of new antibodies: Using pre-trained language models and full-atom modeling of complex structures, the binding strength of antibodies to antigens can be quantitatively evaluated, and new antibodies for a certain target can be analyzed and discovered. This can be used in antibody screening, antibody design and other fields, and provide technical support for further industrial production.

[0260] For example, the solution presented in this application first uses a large language model pre-trained for antibodies to extract key information from the amino acid sequences of antigens and antibodies, solving the problem that previous sequence-based methods failed to fully utilize the sequence. In addition to amino acid sequences, this application also introduces a molecular structure feature extraction module to fully extract the feature information of the molecular structures of antibodies and antigens. For the monomeric structures of antibodies and antigens, geometric graph networks are used for modeling to mine important information within the monomeric structures. For the antibody-antigen complex structure, AlphaFold 3 is referenced for full-atom modeling to further extract interaction information within the complex structure, solving the problem of insufficient precision in using molecular structures. Furthermore, to predict antibodies lacking known three-dimensional structures, this application innovatively proposes using a structure prediction tool for complex structure prediction, enabling accurate prediction of antigen specificity even for newly discovered antibodies.

[0261] The solution shown in this application can solve the problems of performance, efficiency, and scalability in antibody-antigen interaction prediction methods, achieving high-accuracy antibody-antigen interaction prediction, and also has the potential to discover novel antibodies; specifically, it can be as follows:

[0262] 1) Accurate prediction of antibody-antigen interactions: In terms of model performance, the scheme presented in this application combines the advantages of sequence-based and structure-based methods. It uses a pre-trained large language model to extract features from protein sequences and uses geometric graph networks and whole-atom modeling methods to capture key interaction information from antibody and antigen structures, which greatly improves the performance of the scheme presented in this application.

[0263] 2) High-speed screening of specific antibodies: In this application, a pre-trained language model was used to extract sequence features, and a large language model could also be used to predict the structure of the complex. This avoids time-consuming multiple sequence alignment in the predicted structure, can adapt to high-throughput screening, and provides assistance for new drug development.

[0264] 3) Discovery of novel antibodies: The scheme presented in this application does not rely on known experimentally determined structures and can accurately predict newly discovered antibodies, which is of great significance for the discovery of new antibodies and vaccine development.

[0265] The solutions described in the above embodiments of this application can be applied not only to predicting interactions (such as specificity) between antibodies and antigens, but also to predicting other types of interactions between proteins (such as the properties of various other combinations / reactions), including, but not limited to, predicting the binding properties of enzymes to substrates / inhibitors, predicting the binding properties of receptors to ligands, predicting the self-assembly properties of structural proteins, and predicting the formation properties of multi-subunit complexes, etc. The following explanation uses the prediction of enzyme-substrate / inhibitor binding properties as an example.

[0266] I. Predictive Scenarios for Enzyme-Substrate Binding Properties

[0267] Please refer to Figure 9 It illustrates a flowchart of a method for predicting the binding properties of an enzyme to a substrate according to an embodiment of this application. This method is executed by a computer device, which can be implemented as a terminal or a server. Figure 1 The terminal or server shown, such as Figure 9 As shown, the method includes the following steps:

[0268] Step 910: Perform feature extraction processing on the amino acid sequences of the enzyme protein and the substrate protein to obtain enzyme-substrate sequence features.

[0269] For example, the enzyme protein mentioned above can be a kinase protein, and the substrate protein mentioned above can be a phosphorylated substrate (such as a target protein of MAP kinase phosphorylation).

[0270] Among them, the above-mentioned enzyme-substrate sequence characteristics are the characteristics of the amino acid sequence of the enzyme protein and the amino acid sequence of the substrate protein used to characterize the enzyme protein.

[0271] Step 920: Perform feature extraction processing on the molecular structure information of the enzyme protein and the substrate protein to obtain the first enzyme-substrate structural features.

[0272] Among them, the first enzyme-substrate structural feature mentioned above is a feature used to characterize the molecular structure of the enzyme protein and the molecular structure of the substrate protein.

[0273] Step 930: Perform feature extraction processing on the molecular structure information of the enzyme-substrate complex to obtain the second enzyme-substrate structural features; the enzyme-substrate complex is composed of enzyme protein and substrate protein.

[0274] Among them, the above-mentioned second enzyme-substrate structural features are features used to characterize the molecular structure of enzyme-substrate complexes.

[0275] Step 940: Based on the enzyme-substrate sequence characteristics, the first enzyme-substrate structural characteristics, and the second enzyme-substrate structural characteristics, obtain information on the enzyme-substrate binding properties.

[0276] This enzyme-substrate binding property information is used to indicate the binding affinity between the enzyme protein and the substrate protein (such as the ease of binding).

[0277] II. Predictive Scenarios Regarding the Binding Properties of Enzymes and Inhibitors

[0278] Please refer to Figure 10 It illustrates a flowchart of a method for predicting the binding properties of enzymes and inhibitors according to an embodiment of this application. This method is executed by a computer device, which can be implemented as a terminal or a server. Figure 1 The terminal or server shown, such as Figure 10 As shown, the method includes the following steps:

[0279] Step 1010: Perform feature extraction processing on the amino acid sequences of the enzyme protein and the inhibitor protein to obtain enzyme-inhibitor sequence features.

[0280] For example, the inhibitor protein mentioned above can be a serine protease inhibitor (such as α1-antitrypsin).

[0281] Among them, the above-mentioned enzyme-inhibitor sequence characteristics are the characteristics of the amino acid sequence of the enzyme protein and the amino acid sequence of the inhibitor protein used to characterize the enzyme protein.

[0282] Step 1020: Perform feature extraction processing on the molecular structure information of the enzyme protein and the inhibitor protein to obtain the first enzyme-inhibitor structural features.

[0283] Among them, the above-mentioned first enzyme-inhibitor structural features are features used to characterize the molecular structure of enzyme proteins and inhibitor proteins.

[0284] Step 1030: Perform feature extraction processing on the molecular structure information of the enzyme-inhibitor complex to obtain the second enzyme-inhibitor structural features; the enzyme-inhibitor complex is composed of an enzyme protein and an inhibitor protein.

[0285] Among them, the above-mentioned second enzyme-inhibitor structural feature is a feature used to characterize the molecular structure of the enzyme-inhibitor complex.

[0286] Step 1040: Based on the enzyme-inhibitor sequence characteristics, the first enzyme-inhibitor structural characteristics, and the second enzyme-inhibitor structural characteristics, obtain information on the enzyme-inhibitor binding properties.

[0287] The enzyme-inhibitor binding property information is used to indicate the binding affinity between the enzyme protein and the inhibitor protein (such as the ease of binding).

[0288] Figure 11This illustration shows a block diagram of a protein interaction determination apparatus according to an exemplary embodiment of this application. This apparatus can be used to perform all or part of the steps in the protein interaction determination method shown in any one or more of the above embodiments; such as Figure 11 As shown, the device includes:

[0289] The sequence feature acquisition module 1101 is used to perform feature extraction processing on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein; the sequence features are features used to characterize the first amino acid sequence and the second amino acid sequence.

[0290] The first structural feature acquisition module 1102 is used to perform feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain a first structural feature; the first structural feature is a feature used to characterize the molecular structure of the first amino acid sequence and the molecular structure of the second amino acid sequence.

[0291] The second structural feature acquisition module 1103 is used to perform feature extraction processing on the molecular structure information of the complex to obtain the second structural feature, wherein the complex is composed of the first protein and the second protein.

[0292] The interaction information acquisition module 1104 is used to acquire protein interaction information based on the sequence features, the first structural features, and the second structural features, wherein the protein interaction information is used to indicate the interaction between the first protein and the second protein.

[0293] In one possible implementation, the second structural feature acquisition module 1103 is used for,

[0294] Based on the molecular structure information of the complex, protein interaction site information is obtained, which is used to indicate the structure of the binding site between the first protein and the second protein in the complex;

[0295] The protein-protein interaction site information is subjected to feature extraction processing to obtain the second structural feature.

[0296] In one possible implementation, the molecular structure information of the complex includes the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex.

[0297] The second structural feature acquisition module 1103 is used for,

[0298] Based on the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex, the distances between each atom in the first protein and each atom in the second protein are obtained.

[0299] Based on the distances between each atom in the first protein and each atom in the second protein, the atoms contained in the binding portion are determined;

[0300] Based on the atoms contained in the binding region, the protein interaction site information is obtained.

[0301] In one possible implementation, the second structural feature acquisition module 1103 is used to input the protein interaction site information into the all-atom modeling network in the protein interaction prediction model to obtain the second structural feature output by the all-atom modeling network.

[0302] In one possible implementation, the sequence feature acquisition module 1101 is used to,

[0303] The first amino acid sequence and the second amino acid sequence are input into the sequence feature extraction network in the protein function prediction model to obtain the sequence feature vector output by the sequence feature extraction network.

[0304] The sequence features are obtained based on the sequence feature vector.

[0305] In one possible implementation, the first amino acid sequence comprises a heavy chain amino acid sequence and a light chain amino acid sequence; the sequence feature acquisition module 1101 is used for,

[0306] The heavy chain amino acid sequence is input into the first sequence feature extraction subnetwork of the sequence feature extraction network to obtain the first sequence feature vector output by the first sequence feature extraction subnetwork.

[0307] The light chain amino acid sequence is input into the second sequence feature extraction subnetwork of the sequence feature extraction network to obtain the second sequence feature vector output by the second sequence feature extraction subnetwork.

[0308] The second amino acid sequence is input into the third sequence feature extraction subnetwork in the sequence feature extraction network to obtain the third sequence feature vector output by the third sequence feature extraction subnetwork.

[0309] In one possible implementation, the sequence feature acquisition module 1101 is used to input the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector into the coding network in the protein action prediction model to obtain the sequence features output by the coding network.

[0310] In one possible implementation, the sequence features include features encoded by the encoding network from the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector, respectively; or, the sequence features include features fused by the encoding network from the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector.

[0311] In one possible implementation, the device further includes:

[0312] The molecular structure information prediction module is used to perform feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein by the first structure feature acquisition module 1102. Before obtaining the first structural feature, the first amino acid sequence and the second amino acid sequence are input into the structure prediction network in the protein function prediction model to obtain the molecular structure information of the first protein and the molecular structure information of the second protein output by the structure prediction network.

[0313] The first structural feature acquisition module 1102 is used for,

[0314] The molecular structure information of the first protein and the molecular structure information of the second protein are input into the geometric graph learning network in the protein function prediction model to obtain the molecular structure features of the first protein and the second protein output by the geometric graph learning network.

[0315] The first structural feature is obtained based on the molecular structural features of the first protein and the second protein respectively.

[0316] In one possible implementation, the molecular structure information prediction module is used for,

[0317] The first amino acid sequence is input into the first structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the first protein output by the first structure prediction subnetwork.

[0318] The second amino acid sequence is input into the second structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the second protein output by the second structure prediction subnetwork.

[0319] In one possible implementation, the first structural feature acquisition module 1102 is used to input the molecular structural features of the first protein and the second protein into the attention processing network in the protein function prediction model to obtain the first structural feature output by the attention processing network.

[0320] In one possible implementation, the first structural feature includes features obtained by the attention processing network processing the molecular structural features of the first protein and the second protein respectively; or, the first structural feature includes features obtained by the attention processing network fusing the molecular structural features of the first protein and the second protein respectively.

[0321] In one possible implementation, the interaction information acquisition module 1104 is used to input the sequence features, the first structural features, and the second structural features into the interaction prediction network in the protein interaction prediction model to obtain the protein interaction information output by the interaction prediction network.

[0322] In one possible implementation, the protein interaction information is used to indicate the specificity between the first protein and the second protein.

[0323] Figure 12 A structural block diagram of a computer device 1200 illustrating an exemplary embodiment of this application is shown. This computer device can be implemented as a server or terminal as described above in this application. In this embodiment, the computer device is described as a server. The computer device 1200 includes a Central Processing Unit (CPU) 1201, a system memory 1204 including Random Access Memory (RAM) 1202 and Read-Only Memory (ROM) 1203, and a system bus 1205 connecting the system memory 1204 and the CPU 1201. The computer device 1200 also includes a mass storage device 1206 for storing an operating system 1209, application programs 1210, and other program modules 1211.

[0324] The mass storage device 1206 is connected to the central processing unit 1201 via a mass storage controller (not shown) connected to the system bus 1205. The mass storage device 1206 and its associated computer-readable media provide non-volatile storage for the computer device 1200. That is, the mass storage device 1206 may include computer-readable media (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0325] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 1204 and mass storage device 1206 described above can be collectively referred to as memory.

[0326] According to various embodiments of this disclosure, the computer device 1200 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1200 can be connected to a network via a network interface unit 1207 connected to the system bus 1205, or the network interface unit 1207 can be used to connect to other types of networks or remote computer systems (not shown).

[0327] The memory also includes at least one computer program stored in the memory, and the central processing unit 1201 executes the at least one computer program to implement all or part of the steps in the methods shown in the above embodiments.

[0328] In one exemplary embodiment, a computer-readable storage medium is also provided for storing at least one computer program, which is loaded and executed by a processor to implement all or part of the steps in the methods shown in the various embodiments above. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0329] In one exemplary embodiment, a computer program product is also provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform all or part of the steps in the methods shown in the various embodiments above.

[0330] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0331] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for determining protein-protein interactions, characterized in that, The method includes: Feature extraction processing is performed on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein; the sequence features are features used to characterize the first amino acid sequence and the second amino acid sequence. Feature extraction processing is performed on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain a first structural feature; the first structural feature is a feature used to characterize the molecular structure of the first amino acid sequence and the molecular structure of the second amino acid sequence. The molecular structure information of the complex is subjected to feature extraction processing to obtain a second structural feature, wherein the complex is composed of the first protein and the second protein; Based on the sequence features, the first structural features, and the second structural features, protein interaction information is obtained, which is used to indicate the interaction between the first protein and the second protein.

2. The method according to claim 1, characterized in that, The process of performing feature extraction on the molecular structure information of the complex to obtain a second structural feature includes: Based on the molecular structure information of the complex, protein interaction site information is obtained, which is used to indicate the structure of the binding site between the first protein and the second protein in the complex; The protein-protein interaction site information is subjected to feature extraction processing to obtain the second structural feature.

3. The method according to claim 2, characterized in that, The molecular structure information of the complex includes the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex. The acquisition of protein-protein interaction site information based on the molecular structure information of the complex includes: Based on the coordinates of each atom in the first protein in the complex, and the coordinates of each atom in the second protein in the complex, the distances between each atom in the first protein and each atom in the second protein are obtained. Based on the distances between each atom in the first protein and each atom in the second protein, the atoms contained in the binding portion are determined; Based on the atoms contained in the binding region, the protein interaction site information is obtained.

4. The method according to claim 2, characterized in that, The step of performing feature extraction processing on the protein-protein interaction site information to obtain the second structural feature includes: The protein interaction site information is input into the all-atom modeling network in the protein interaction prediction model to obtain the second structural feature output by the all-atom modeling network.

5. The method according to any one of claims 1 to 4, characterized in that, The feature extraction process performed on the first and second amino acid sequences to obtain sequence features includes: The first amino acid sequence and the second amino acid sequence are input into the sequence feature extraction network in the protein function prediction model to obtain the sequence feature vector output by the sequence feature extraction network. The sequence features are obtained based on the sequence feature vector.

6. The method according to claim 5, characterized in that, The first amino acid sequence comprises a heavy chain amino acid sequence and a light chain amino acid sequence; The step of inputting the first amino acid sequence and the second amino acid sequence into the sequence feature extraction network in the protein action prediction model to obtain the sequence feature vector output by the sequence feature extraction network includes: The heavy chain amino acid sequence is input into the first sequence feature extraction subnetwork of the sequence feature extraction network to obtain the first sequence feature vector output by the first sequence feature extraction subnetwork. The light chain amino acid sequence is input into the second sequence feature extraction subnetwork of the sequence feature extraction network to obtain the second sequence feature vector output by the second sequence feature extraction subnetwork. The second amino acid sequence is input into the third sequence feature extraction subnetwork in the sequence feature extraction network to obtain the third sequence feature vector output by the third sequence feature extraction subnetwork.

7. The method according to claim 6, characterized in that, The step of obtaining the sequence features based on the sequence feature vector includes: The first sequence feature vector, the second sequence feature vector, and the third sequence feature vector are input into the coding network of the protein action prediction model to obtain the sequence features output by the coding network.

8. The method according to claim 7, characterized in that, The sequence features include features encoded by the encoding network from the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector, respectively; or... The sequence features include features obtained by fusing the first sequence feature vector, the second sequence feature vector, and the third sequence feature vector by the coding network.

9. The method according to any one of claims 1 to 8, characterized in that, Before performing feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain the first structural feature, the method further includes: The first amino acid sequence and the second amino acid sequence are input into the structure prediction network in the protein function prediction model to obtain the molecular structure information of the first protein and the molecular structure information of the second protein output by the structure prediction network. The step of performing feature extraction processing on the molecular structure information of the first protein and the molecular structure information of the second protein to obtain the first structural feature includes: The molecular structure information of the first protein and the molecular structure information of the second protein are input into the geometric graph learning network in the protein function prediction model to obtain the molecular structure features of the first protein and the second protein output by the geometric graph learning network. The first structural feature is obtained based on the molecular structural features of the first protein and the second protein respectively.

10. The method according to claim 9, characterized in that, The step of inputting the first amino acid sequence and the second amino acid sequence into the structure prediction network of the protein function prediction model to obtain the molecular structure information of the first protein and the molecular structure information of the second protein output by the structure prediction network includes: The first amino acid sequence is input into the first structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the first protein output by the first structure prediction subnetwork. The second amino acid sequence is input into the second structure prediction subnetwork of the structure prediction network to obtain the molecular structure information of the second protein output by the second structure prediction subnetwork.

11. The method according to claim 9 or 10, characterized in that, The step of obtaining the first structural feature based on the molecular structural features of the first protein and the second protein includes: The molecular structural features of the first protein and the second protein are input into the attention processing network in the protein function prediction model to obtain the first structural feature output by the attention processing network.

12. The method according to claim 11, characterized in that, The first structural feature includes features obtained by the attention processing network processing the molecular structural features of the first protein and the second protein respectively; or, The first structural feature includes the feature obtained by the attention processing network through the fusion processing of the molecular structural features of the first protein and the second protein respectively.

13. The method according to any one of claims 1 to 12, characterized in that, The process of obtaining protein-protein interaction information based on the sequence features, the first structural features, and the second structural features includes: The sequence features, the first structural features, and the second structural features are input into the interaction prediction network in the protein interaction prediction model to obtain the protein interaction information output by the interaction prediction network.

14. The method according to any one of claims 1 to 13, characterized in that, The protein interaction information is used to indicate the specificity between the first protein and the second protein.

15. An apparatus for determining protein-protein interactions, characterized in that, The device includes: The sequence feature acquisition module is used to perform feature extraction processing on the first amino acid sequence and the second amino acid sequence to obtain sequence features; the first amino acid sequence is the amino acid sequence of the first protein, and the second amino acid sequence is the amino acid sequence of the second protein; the sequence features are features used to characterize the first amino acid sequence and the second amino acid sequence. The first structural feature acquisition module is used to perform feature extraction processing on the molecular structural information of the first protein and the molecular structural information of the second protein to obtain a first structural feature; the first structural feature is a feature used to characterize the molecular structure of the first amino acid sequence and the molecular structure of the second amino acid sequence. The second structural feature acquisition module is used to perform feature extraction processing on the molecular structure information of the complex to obtain the second structural feature, wherein the complex is composed of the first protein and the second protein; An interaction information acquisition module is used to acquire protein interaction information based on the sequence features, the first structural features, and the second structural features, wherein the protein interaction information is used to indicate the interaction between the first protein and the second protein.

16. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the method for determining protein interactions as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the method for determining protein interactions as described in any one of claims 1 to 14.

18. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium; the computer program is read and executed by a processor of a computer device to implement the method for determining protein interactions as described in any one of claims 1 to 14.

Images