Method and apparatus for predicting changes in protein complex affinity due to mutations

Abstract

The purpose of the present application is to provide a method and device for predicting the change of protein complex affinity caused by mutation, which comprises: using a first sub-network in a twin network, determining wild-type multi-modal encoding information including wild-type sequence encoding information and wild-type geometric feature encoding information discretely encoded based on wild-type protein complex information and mutation information; then determining the affinity information of the wild type through a decoder; determining the affinity information of the mutant through a corresponding second sub-network; and then determining the affinity change information. Through the use of discrete encoding and multi-modal information, the model can cope with the situation that the local structure optimization is not good, so that the model can obtain more accurate affinity change prediction results using complex information with low precision. Moreover, the introduction of protein surface information helps the model to better learn the influence of the cavity formed by the protein interaction interface on the affinity change, further improving the accuracy of affinity change prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and in particular to a technology for predicting changes in the affinity of a protein complex caused by mutations. BACKGROUND

[0002] Proteins are the most important macromolecules in living organisms, and their functions mainly depend on protein-protein interactions, such as signal transduction, metabolic processes, immune responses, etc. The strength of protein-protein interactions (i.e., affinity) affects the strength of protein functions. Therefore, predicting the effects of mutations on the strength of protein-protein interactions (i.e., changes in affinity) plays a crucial role in understanding the mechanisms of life activities, promoting drug development, and achieving precise disease treatment.

[0003] With the development of artificial intelligence models, researchers can use models such as AlphaFold2, 3, RoseTTAFold-All-Atom (RFAA) to accurately predict protein complex structures. However, the current models have low prediction accuracy for changes in the structure of the mutated complex caused by mutations, which also affects the prediction accuracy of the affinity changes caused by mutations. SUMMARY

[0004] An object of the present application is to provide a method and device for predicting changes in the affinity of a protein complex caused by mutations.

[0005] According to one aspect of the present application, a method for predicting changes in the affinity of a protein complex caused by mutations is provided, which comprises:

[0006] In the first sub-network under the twin network architecture, based on the wild-type protein complex information and the corresponding mutation information, the wild-type sequence encoding information and the wild-type geometric feature encoding information are determined, wherein the wild-type geometric feature encoding information includes wild-type surface encoding information and wild-type structure encoding information; the wild-type geometric feature encoding information is discretely encoded using a corresponding encoder model, and then combined with the wild-type sequence encoding information to determine the wild-type multi-modal encoding information; based on the wild-type multi-modal encoding information, the affinity information of the wild-type protein complex is determined using a corresponding decoder model;

[0007] The mutant protein complex information and the corresponding mutation information are input into the second sub-network under the twin network architecture to determine the affinity information of the mutant protein complex;

[0008] Based on the affinity information of the wild-type protein complex and the affinity information of the mutant protein complex, the protein complex affinity change information is determined.

[0009] According to an aspect of the present application, there is provided a computer device for predicting the change of protein complex affinity caused by mutation, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0010] According to an aspect of the present application, there is provided a computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0011] According to an aspect of the present application, there is provided a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0012] According to an aspect of the present application, there is provided a device for predicting the change of protein complex affinity caused by mutation, comprising:

[0013] A first module is configured to determine wild-type sequence encoding information and wild-type geometric feature encoding information based on wild-type protein complex information and corresponding mutation information in a first sub-network under a twin network architecture, wherein the wild-type geometric feature encoding information comprises wild-type surface encoding information and wild-type structure encoding information; perform discrete coding on the wild-type geometric feature encoding information by using a corresponding encoder model, and then determine wild-type multi-modal encoding information by combining the wild-type sequence encoding information; and determine affinity information of the wild-type protein complex based on the wild-type multi-modal encoding information by using a corresponding decoder model.

[0014] A second module is configured to input mutant protein complex information and corresponding mutation information into a second sub-network under the twin network architecture, and determine affinity information of the mutant protein complex.

[0015] A third module is configured to determine protein complex affinity change information based on the affinity information of the wild-type protein complex and the affinity information of the mutant protein complex.

[0016] Compared with the prior art, in the first sub-network under the twin network architecture, based on wild type protein complex information and corresponding mutation information, wild type sequence encoding information and wild type geometric feature encoding information are determined, wherein the wild type geometric feature encoding information includes wild type surface encoding information and wild type structure encoding information; the wild type geometric feature encoding information is discretely encoded by using a corresponding encoder model, and then combined with the wild type sequence encoding information to determine wild type multi-modal encoding information; based on the wild type multi-modal encoding information, an affinity information of the wild type protein complex is determined by using a corresponding decoder model; the mutant protein complex information and corresponding mutation information are input into the second sub-network under the twin network architecture to determine the affinity information of the mutant protein complex; and based on the affinity information of the wild type protein complex and the affinity information of the mutant protein complex, the protein complex affinity change information is determined. By discretely encoding the geometric feature encoding information, the model can cope with the situation that the local structure optimization is not good, so that the model can obtain a more accurate affinity change prediction result by using the protein complex information after mutation with low precision. Moreover, the surface information of the protein complex is introduced in the scheme to help the model better learn the influence of the cavity formed by the protein interaction interface on the affinity change, thereby further improving the affinity change prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0017] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments thereof as read with the accompanying drawings:

[0018] Figure 1 A flow chart of a method for predicting protein complex affinity change caused by mutation is shown according to an embodiment of the application;

[0019] Figure 2 A twin network architecture diagram for predicting protein complex affinity change caused by mutation is shown according to an embodiment of the application;

[0020] Figure 3 A device structure diagram for predicting protein complex affinity change caused by mutation is shown according to an embodiment of the application;

[0021] Figure 4 An exemplary system that can be used to implement various embodiments described herein is shown.

[0022] The same or similar reference numerals in the drawings represent the same or similar components. DETAILED DESCRIPTION

[0023] The application will be further described in detail below with reference to the accompanying drawings.

[0024] In one illustrative configuration, the terminal, the device of the service network, and the trusted party each include one or more processors (e.g., central processing units (CPUs)), input / output interfaces, network interfaces, and memories.

[0025] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) or flash memory. The memory is an example of computer-readable media.

[0026] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology for storing information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PCM), programmable random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0027] The device referred to in the present application includes but is not limited to a user device, a network device, or a device formed by integrating a user device and a network device through a network. The user device includes but is not limited to any kind of mobile electronic product capable of human-computer interaction (for example, human-computer interaction through a touch panel), such as a smart phone, a tablet computer, etc. The mobile electronic product can adopt any operating system, such as an Android operating system, an iOS operating system, etc. The network device includes an electronic device capable of automatically performing numerical calculation and information processing according to a pre-set or stored instruction. The hardware of the network device includes but is not limited to a microprocessor, an application specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc. The network device includes but is not limited to a computer, a network host, a single network server, a plurality of network servers, or a cloud formed by a plurality of servers. The cloud is formed by a large number of computers or network servers based on cloud computing. The cloud computing is a kind of distributed computing, which is a virtual supercomputer formed by a group of loosely coupled computer clusters. The network includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, a wireless Ad Hoc network, etc. Preferably, the device can also be a program running on the user device, the network device, or a device formed by integrating a user device and a network device, a network device, a touch terminal, or a device formed by integrating a network device and a touch terminal through a network.

[0028] Of course, those skilled in the art should understand that the above device is only an example, and other existing or future devices, such as devices that can be applicable to the present application, should also be included in the protection scope of the present application, and are hereby included by reference.

[0029] In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0030] Reference Figure 2 The illustrated architecture diagram, Figure 1A method flowchart for predicting the change in affinity of a protein complex caused by a mutation is shown according to one embodiment of the present application, which includes steps S11, S12 and S13. In step S11, the device 1 determines wild-type sequence encoding information and wild-type geometric feature encoding information in the first subnetwork under the twin network architecture based on wild-type protein complex information and corresponding mutation information, wherein the wild-type geometric feature encoding information includes wild-type surface encoding information and wild-type structure encoding information; uses a corresponding encoder model to discretely encode the wild-type geometric feature encoding information, and then combines the wild-type sequence encoding information to determine wild-type multi-modal encoding information; and uses a corresponding decoder model to determine the affinity information of the wild-type protein complex based on the wild-type multi-modal encoding information. In step S12, the device 1 inputs the mutant protein complex information and corresponding mutation information into the second subnetwork under the twin network architecture to determine the affinity information of the mutant protein complex. In step S13, the device 1 determines the protein complex affinity change information based on the affinity information of the wild-type protein complex and the affinity information of the mutant protein complex.

[0031] In some embodiments, the device 1 includes but is not limited to a user device, a network device with information processing or computing capabilities, such as a tablet computer, a computer, a server. The device 1 is deployed with a twin network model that can be used for affinity change information prediction. The twin network used in the present application includes two identical subnetworks, i.e., a first subnetwork and a second subnetwork. The affinity information of the wild-type protein complex and the affinity information of the mutant protein complex are predicted by the two subnetworks respectively, and the difference between the two prediction results is calculated to obtain the protein complex affinity change information caused by the mutation.

[0032] In step S11, the device 1 determines wild-type sequence encoding information and wild-type geometric feature encoding information in the first subnetwork under the twin network architecture based on wild-type protein complex information and corresponding mutation information, wherein the wild-type geometric feature encoding information includes wild-type surface encoding information and wild-type structure encoding information; uses a corresponding encoder model to discretely encode the wild-type geometric feature encoding information, and then combines the wild-type sequence encoding information to determine wild-type multi-modal encoding information; and uses a corresponding decoder model to determine the affinity information of the wild-type protein complex based on the wild-type multi-modal encoding information.

[0033] In some embodiments, the wild-type protein complex information includes, but is not limited to, the wild-type protein complex structure information and the wild-type protein complex sequence information. In some embodiments, the wild-type protein complex structure information includes, but is not limited to, the three-dimensional coordinate information of each atom in the wild-type protein complex. In some embodiments, the wild-type protein complex sequence information includes the amino acid sequence information corresponding to the wild-type protein complex and the label information of the protein to which the amino acid belongs. Through the label information, the model focuses on the interaction between proteins rather than the interaction between each chain within the protein, thereby better predicting the strength of the interaction between proteins, i.e., the affinity information of the protein complex. For example, if a first protein (including chain A and chain B) and a second protein (including chain C and chain D) combine to form the wild-type protein complex, the label information distinguishes different proteins, thereby determining which chain interaction is used for calculation. In this example, the focus is on the interaction between chains AB-CD rather than the interaction between chains A-B or chains C-D, thereby more accurately predicting the affinity information of the protein complex. In some embodiments, the mutation information includes mutation site information. By inputting the mutation site information, the model pays more attention to the information related to the mutation site when performing feature encoding, thereby more accurately predicting the affinity change caused by the mutation. In some embodiments, the wild-type protein complex information and the mutation information are input into the first sub-network to determine the wild-type sequence encoding information and the wild-type geometric feature encoding information corresponding to the wild-type protein complex. And the wild-type geometric feature encoding information is discretely encoded to obtain a discrete representation of the wild-type geometric feature encoding information to help the model deal with the situation that the local structure is not optimized well. Further determine the wild-type multi-modal encoding information, so as to determine the affinity information of the wild-type protein complex by using the wild-type multi-modal encoding information.

[0034] In some embodiments, based on the wild-type protein complex information and the corresponding mutation information, the wild-type sequence encoding information and the wild-type geometric feature encoding information are determined, wherein the wild-type geometric feature encoding information includes wild-type surface encoding information and wild-type structure encoding information, including: based on the wild-type protein complex information, determining the wild-type sequence encoding information by using a protein language model; based on the wild-type protein complex information and the corresponding mutation information, determining the wild-type geometric feature encoding information, wherein the wild-type geometric feature encoding information matches the mutation information.

[0035] In some embodiments, the protein language model includes, but is not limited to, ESM2, ESM3, ProtT5-XL-UniRef50, ProGen. For example, for a wild-type protein complex sequence of length L, an embedding matrix of L*d is generated using the aforementioned protein language model as the wild-type sequence encoding information, where the vector of length d at the i-th position in the embedding matrix is the sequence encoding information corresponding to the i-th amino acid in the wild-type protein complex sequence. In some embodiments, the protein language model can also be used to distinguish different proteins in the wild-type protein complex in combination with the label information, obtaining the embedding matrix corresponding to each protein. The wild-type sequence encoding information includes the embedding matrix corresponding to each protein.

[0036] In some embodiments, based on the mutation site information in the mutation information, the amino acids in the wild-type protein complex that match the mutation site information can be selected. For example, the amino acid in the wild-type protein complex corresponding to the mutation site information and other amino acids with a spatial distance less than the corresponding spatial distance threshold (usually set to ) from the amino acid. Based on the information of these amino acids, the wild-type geometric feature encoding information matching the mutation information is determined.

[0037] In some embodiments, the wild-type geometric feature encoding information is determined based on the wild-type protein complex information and the corresponding mutation information, wherein the matching between the wild-type geometric feature encoding information and the mutation information includes: determining a plurality of protein surface point cloud information based on the wild-type protein complex information, wherein each protein surface point cloud information corresponds to one protein that combines to form the wild-type protein complex; determining the wild-type surface encoding information based on the plurality of protein surface point cloud information and the mutation information, wherein the wild-type surface encoding information includes surface encoding information corresponding to each protein surface point cloud information, and the surface encoding information corresponding to each protein surface point cloud information matches the mutation information; determining the wild-type structure encoding information using an attention mechanism-based graph model based on the wild-type protein complex information and the mutation information, wherein the wild-type structure encoding information includes node information and edge information, the node information includes amino acid information matching the mutation information, and the edge information includes amino acid pair information matching the mutation information.

[0038] In some embodiments, the wild-type protein complex is formed by a plurality of proteins. Based on the wild-type protein complex structure information, the structure information corresponding to different proteins in the wild-type protein complex can be distinguished in this application in combination with the label information, so as to determine the protein surface point cloud information corresponding to each protein respectively.

[0039] In some embodiments, the determining, based on the wild-type protein complex information, a plurality of protein surface point cloud information, wherein each protein surface point cloud information corresponds to one protein that binds to form the wild-type protein complex, comprises: extracting a plurality of points respectively centered on heavy atoms of each protein that binds to form the wild-type protein complex to form point cloud information corresponding to each protein; and screening and determining protein surface point cloud information corresponding to each protein based on the point cloud information corresponding to each protein and using a smooth distance function.

[0040] In some embodiments, the heavy atoms include non-hydrogen atoms such as carbon, oxygen, nitrogen, sulfur, etc. The heavy atoms belonging to the same protein can be screened based on the aforementioned labeling information. Based on the wild-type protein complex structure information, a plurality of points are extracted using corresponding rules centered on these heavy atoms to combine to form point cloud information. For example, a plurality of points can be uniformly or randomly extracted within a preset extraction range; or a plurality of points can be extracted according to a normal distribution based on a preset standard deviation with the coordinates of the heavy atoms as the mean value. In some embodiments, a corresponding smooth distance function is defined as follows:

[0041]

[0042] wherein, represents the coordinates of the jth heavy atom in the protein, and a is used to represent that x represents atomic coordinates, represents the coordinates of the ith extracted point, and S is used to represent that x represents point cloud coordinates, represents the atomic radius of the jth heavy atom in the protein, is the calculated corresponding atomic radius. Based on the above formula, the smooth distance function value of each point in the point cloud information is calculated, and points whose smooth distance function values satisfy the smooth distance function value range are screened to form protein surface point cloud information. Generally, points whose values are between 1 and 1.1 are screened to form protein surface point cloud information. Compared with the algorithm that uses a virtual probe to scan the surface of the protein to generate surface point cloud, the scheme adopted in the present application has higher calculation efficiency and saves more computing resources.

[0043] In the foregoing embodiments, a plurality of points are extracted centered on each heavy atom in the protein, which only considers the position of a single heavy atom. In order to make the obtained point cloud information consider the positions of multiple heavy atoms at the same time, in some embodiments, the point cloud information can also be optimized in combination with a corresponding loss function before screening the protein surface point cloud information. For example, the point cloud information is optimized by minimizing the following square loss function ​

[0044]

[0045] Where η is the number of points sampled centered on each heavy atom, N is the total number of heavy atoms in the protein, and r is the standard radius. In current practice, r is taken as... This usually yields good optimization results.

[0046] In some embodiments, determining multiple protein surface point cloud information based on the wild-type protein complex information, wherein each protein surface point cloud information corresponds to a protein that binds to form the wild-type protein complex, further includes: smoothing the protein surface point cloud information. In some embodiments, the smoothing process includes smoothing the normal vector corresponding to each point in the protein surface point cloud information. Smoothing is performed. The midpoint of the protein surface point cloud information can be calculated based on the gradient of the aforementioned loss function. corresponding normal vector The normal vector is smoothed using Gaussian kernel smoothing, as shown in the following formula:

[0047]

[0048] Where M is the number of points in the protein surface point cloud information, σ n The standard deviation parameter, for The corresponding weights and The closer the distance, the higher the weight. Those skilled in the art should understand that the above Gaussian kernel smoothing method is merely an example; other existing or future smoothing methods that are applicable to this application should also be included within the scope of protection of this application, and are hereby incorporated by reference.

[0049] After determining the point cloud information of each protein surface, these point cloud information can be encoded to obtain the surface coding information corresponding to the protein surface point cloud information. To better learn the effect of mutations on the affinity of protein complexes, based on the mutation site information in the mutation information, points corresponding to amino acids in the protein surface point cloud information that match the mutation site information (e.g., the amino acid corresponding to the mutation site information in the wild-type protein complex and other amino acids whose spatial distance from the amino acid is less than the corresponding spatial distance threshold) can be determined. Then, based on these points, the surface coding information corresponding to the protein surface point cloud information can be determined.

[0050] In some embodiments, determining the wild-type surface coding information based on the plurality of protein surface point cloud information and the mutation information, wherein the wild-type surface coding information includes surface coding information corresponding to each protein surface point cloud information, and matching the surface coding information corresponding to each protein surface point cloud information with the mutation information includes: determining a plurality of block information corresponding to each protein surface point cloud information by clustering based on the surface distance between each point in each protein surface point cloud information; determining a plurality of target block information matching the mutation information from the plurality of block information based on the mutation information; determining the block coding information corresponding to each target block information, and using the block coding information as the surface coding information corresponding to the corresponding protein surface point cloud information, wherein the block coding information includes physicochemical feature integrated coding information and relative position coding information corresponding to the target block information.

[0051] In some embodiments, to improve computational efficiency, points in the protein surface point cloud information can be clustered into blocks, allowing subsequent calculations to be performed on a block-by-block basis. In some embodiments, the surface distance between points is calculated by combining the normal vectors corresponding to each point in the protein surface point cloud information.

[0052]

[0053] Where, d ij For point and Inter-surface distance, n i For point Corresponding normal vector, n j For point Corresponding normal vector. <n i ,n j >To perform n i With n j The dot product operation is performed. In some embodiments, based on the surface distance, multiple protein surface point clusters are determined using a corresponding clustering algorithm, with each protein surface point cluster corresponding to a block of information. For example, in the protein surface point cloud information, multiple representative points are selected using farthest point sampling; based on the representative points, multiple protein surface point clusters are determined, with one representative point corresponding to one protein surface point cluster, and the surface distance between each point in a protein surface point cluster and the corresponding representative point of that protein surface point cluster is less than the surface distance between each point and other representative points.

[0054] In some embodiments, based on the mutation site information in the mutation information, amino acids in the protein surface point cloud information that match the mutation site information (e.g., amino acids in the wild-type protein complex corresponding to the mutation site information and other amino acids that have a spatial distance less than a corresponding spatial distance threshold from the amino acids) are determined. The block information corresponding to these amino acids that match the mutation site information is determined as the target block information. For example, in combination with the wild-type protein complex structure information, the amino acids corresponding to each block information can be determined according to the principle of the closest distance, and then the target block information corresponding to the amino acids that match the mutation site information is selected from them.

[0055] In some embodiments, for each target block information in the protein surface point cloud information, the corresponding physicochemical feature integrated encoding information can be obtained based on the physicochemical features of the points in the target block information, wherein the physicochemical features of the atoms closest to the points can be used as the physicochemical features corresponding to the points. The physicochemical features include but are not limited to the amino acid type information, polarity, electrical property, side chain size, atomic size, etc. to which the atom belongs. In some embodiments, the amino acid type information used for integrated encoding is encoded by a protein language model, rather than simple one-hot encoding. In addition, in order to preserve the positional relationship of each target block in the protein surface space, the positional relationship of each target block is also encoded to obtain the corresponding relative position encoding information.

[0056] In some embodiments, the determination of the block encoding information corresponding to each target block information, and the use of the block encoding information as the surface encoding information corresponding to the corresponding protein surface point cloud information, wherein the block encoding information includes the physicochemical feature integrated encoding information corresponding to the target block information and the relative position encoding information, includes: determining the integrated encoding information corresponding to each point in the protein surface point cloud information based on the points in the protein surface point cloud information and other points corresponding to the points that satisfy a first surface distance condition; determining the physicochemical feature encoding information corresponding to each block information based on the integrated encoding information corresponding to each point in the protein surface point cloud information; determining the physicochemical feature integrated encoding information corresponding to the target block information based on the target block information and other block information corresponding to the target block information that satisfies a second surface distance condition, in combination with the physicochemical feature encoding information of each block information; and determining the relative position encoding information of the target block information in a one-dimensional space based on the relative position information of the target block information on the protein surface using Morton coding.

[0057] In some embodiments, for each point in the protein surface point cloud information, the first surface distance condition can be to filter a preset number of other points in the protein surface point cloud information that are closest to the point in surface distance; or to filter other points in the protein surface point cloud information that have a surface distance less than a point-to-point surface distance threshold. Based on each point in the protein surface point cloud information and the other points corresponding to the point that satisfy the first surface distance condition, the physicochemical characteristics of each point and the corresponding other points are calculated by weighting to obtain the integrated encoding information corresponding to each point in the protein surface point cloud information. For example, a neural network model such as a multi-layer perception (MLP) can be used to perform the calculation to obtain the integrated encoding information corresponding to each point in the protein surface point cloud information. Here, those skilled in the art should understand that the above-mentioned multi-layer perception is only an example, and other existing or future models that can be applicable to the present application should also be included within the protection scope of the present application, and are hereby incorporated by reference.

[0058] In some embodiments, based on the integrated encoding information corresponding to the points included in each block information in the protein surface point cloud information, the physicochemical feature encoding information corresponding to each block information in the protein surface point cloud information can be determined by a pooling operation. The pooling operation includes but is not limited to max pooling, average pooling, mix pooling, and soft pooling.

[0059] In some embodiments, for each target block information in the protein surface point cloud information, the second surface distance condition can be to filter a preset number of other block information in the protein surface point cloud information that are closest to the target block information in surface distance; or to filter other block information in the protein surface point cloud information that have a surface distance less than a block-to-block surface distance threshold. The surface distance between each block information can be determined based on the surface distance between the cluster centers corresponding to each block information. In some embodiments, based on each target block information in the protein surface point cloud information and the other block information corresponding to the target block information that satisfy the second surface distance condition, the physicochemical feature integrated encoding information corresponding to each target block information is determined by weighting calculation combined with the corresponding physicochemical feature encoding information. Here, a multi-head self-attention mechanism can be introduced to determine the weight. Specifically, the weight information can be obtained by performing a row-level softmax calculation on the attention score in the model. Based on the weight information corresponding to each target block information and the other block information that satisfy the second surface distance condition, the corresponding physicochemical feature integrated encoding information is determined by weighting calculation combined with the corresponding physicochemical feature encoding information.

[0060] Here, by layer-by-layer information integration, the final obtained wild-type surface encoding information contains richer feature information, which helps to improve the accuracy of the prediction results. In some embodiments, in order to preserve the positional relationship of each target block in the corresponding protein surface space, Morton coding can be used to sort the block information located on the protein surface to obtain one-dimensional relative position encoding information. The wild-type surface encoding information includes the physicochemical feature integrated encoding information and the relative position encoding information of the target block information corresponding to each protein that forms the wild-type protein complex.

[0061] In some embodiments, based on the mutation information, the amino acids matching the mutation site information, i.e., the amino acids corresponding to the mutation site in the wild-type protein complex, and other amino acids with a spatial distance less than the corresponding spatial distance threshold are determined. Each amino acid matching the mutation site information is taken as a node. Further, the wild-type structure encoding information matching the mutation information is obtained. The node information includes the species information corresponding to the amino acid, the three-dimensional coordinate information of the heavy atom contained in the amino acid, the dihedral angle information formed by the amino acid, the physicochemical characteristics of the amino acid (such as side chain size, electrical property, polarity, etc.), the protein information to which the amino acid belongs, etc. The edge information includes the relative distance of the amino acid pair corresponding to the amino acid matching the mutation site information in the wild-type protein complex sequence, the relative direction and distance information between the alpha-carbon atoms (Ca).

[0062] Here, the present application does not limit the execution order of determining the wild-type sequence encoding information, the wild-type surface encoding information, and the wild-type structure encoding information, which can be executed in sequence, simultaneously, or in other orders.

[0063] In order to have better prediction performance on protein complex structures with low accuracy, the wild-type surface encoding information and the wild-type structure encoding information are respectively discretely encoded (Vector Quantised) in the present application. The encoder model selects the embedding vector that best approximates the wild-type surface encoding information from the code book corresponding to the surface encoding information, thereby obtaining the discretely encoded wild-type surface encoding information, and the code book is composed of discrete embedding vectors. The discrete encoding of the wild-type structure encoding information is similar to the foregoing discrete encoding process. The encoder model selects the embedding vector that best approximates the wild-type structure encoding information from the code book corresponding to the structure encoding information, thereby obtaining the discretely encoded wild-type structure encoding information.

[0064] The wild type sequence encoding information and the discrete coded wild type surface encoding information and the discrete coded wild type structure encoding information are spliced together to obtain wild type multi-modal encoding information. The wild type multi-modal encoding information is input into the trained decoder model, and the affinity information of the corresponding wild type protein complex can be determined. The wild type multi-modal encoding information combines the wild type protein complex sequence, the protein surface information matched with the mutation site, and the three-dimensional structure information, and can more accurately predict the affinity information. Moreover, the protein surface information is introduced through the wild type surface encoding information. Since the protein surface cavity is formed by wrapping the protein surface, the protein surface information can help the model better learn / predict the influence of the cavity on the affinity, thereby improving the prediction accuracy of the affinity change.

[0065] In step S12, the device 1 inputs the mutant protein complex information and the corresponding mutation information into the second sub-network under the twin network architecture to determine the affinity information of the mutant protein complex. The second sub-network uses the same model architecture and parameters as the first sub-network. In the second sub-network, the mutant sequence encoding information and the mutant geometric feature encoding information are also determined based on the mutant protein complex information and the mutation information, wherein the mutant geometric feature encoding information includes mutant surface encoding information and mutant structure encoding information; the mutant geometric feature encoding information is discretely coded using a corresponding encoder model, and then combined with the mutant sequence encoding information to determine the mutant multi-modal encoding information; and the affinity information of the mutant protein complex is determined based on the mutant multi-modal encoding information using a corresponding decoder model. Here, the mutant sequence encoding information and the mutant geometric feature encoding information are determined, the mutant geometric feature encoding information is discretely coded, and then the mutant multi-modal encoding information is obtained to determine the affinity information of the mutant protein complex, which is the same as the method in the foregoing step S11, and thus is incorporated herein by reference. The execution order of step S11 and step S12 is not limited in the present application, and they can be executed in sequence, simultaneously, or in other order.

[0066] In step S13, the device 1 determines the protein complex affinity change information based on the affinity information of the wild type protein complex and the affinity information of the mutant protein complex. For example, the affinity information of the wild type protein complex and the affinity information of the mutant protein complex can be subtracted to obtain the final predicted protein complex affinity change information affected by the mutation.

[0067] In some embodiments, the method further comprises: step S14 (not shown), determining, by the device 1, mutant protein complex information based on the wild-type protein complex information and corresponding mutation information. For example, the mutation information comprises mutation site information and mutated amino acid type information. The mutant protein complex information comprises, but is not limited to, mutant protein complex structure information and mutant protein complex sequence information. Based on the mutation information and the wild-type protein complex sequence information, the mutant protein complex sequence information is determined. Based on the wild-type protein complex structure information and the mutation information, the mutant protein complex structure information is determined by using a corresponding protein complex structure prediction model (for example, FoldX, AlphaFold2, 3, etc.). In this application, the complex surface and three-dimensional structure are discretely encoded, and the affinity is predicted by using multi-modal encoding information, so that even in the case of poor local structure optimization, the method of this application can still obtain more accurate affinity change prediction results.

[0068] Reference Figure 2 In some embodiments, as shown in the architecture diagram, the method further comprises: step S15 (not shown), determining, by the device 1, wild-type sample sequence encoding information and wild-type sample geometric feature encoding information in the first sub-network under the twin network architecture based on wild-type protein complex sample information and mutation sample information, wherein the wild-type sample geometric feature encoding information comprises wild-type sample surface encoding information and wild-type sample structure encoding information; discretely encoding the wild-type sample geometric feature encoding information by using the encoder model, and then determining wild-type sample multi-modal encoding information in combination with wild-type sample sequence encoding information; determining the affinity information of the wild-type protein complex sample by using the decoder model; determining the affinity information of the mutant protein complex sample by using the second sub-network under the twin network architecture based on the mutant protein complex sample information and the mutation sample information; determining the protein complex sample affinity change information based on the affinity information of the wild-type protein complex sample and the affinity information of the mutant protein complex sample; reconstructing the discretely encoded wild-type sample geometric feature encoding information to determine wild-type sample reconstructed encoding information; and optimizing the encoder model and the decoder model in combination with the corresponding loss function based on the wild-type sample reconstructed encoding information, the protein complex sample affinity change information, and the corresponding wild-type sample geometric feature encoding information, the real affinity change information of the mutant protein complex sample, wherein the loss function comprises reconstruction error and affinity prediction error.

[0069] The above embodiments demonstrate the training process of the twin network described in the present application. Here, the method for predicting the affinity change information of the protein complex sample based on the wild-type protein complex sample information, the mutant protein complex sample information, and the mutation sample information is the same as or similar to the method described in the foregoing steps S11 to S13, and thus will not be described again, and is included herein by reference.

[0070] In some embodiments, the decoder model is used to reconstruct the discrete encoded surface coding information and structure coding information of the wild-type sample based on the codebook corresponding to the surface coding information and the codebook corresponding to the structure coding information, respectively, to obtain wild-type sample reconstructed coding information. The wild-type sample reconstructed coding information includes wild-type sample surface coding reconstructed information and wild-type sample structure coding reconstructed information. The reconstruction aims to convert the discrete encoded information back to continuous encoded information. The model is optimized in combination with the aforementioned prediction results (i.e., wild-type sample reconstructed coding information and protein complex sample affinity change information) and corresponding true values (i.e., wild-type sample geometric feature coding information and true affinity change information of the mutant protein complex sample) in combination with a corresponding loss function. The loss function includes reconstruction error and affinity prediction error. The reconstruction error includes surface information reconstruction error and structure information reconstruction error. Each part includes a reconstruction loss (Reconstruction Loss) term, a codebook loss (Codebook Loss) term, and a commitment loss (Commitment Loss) term. The reconstruction loss term is used to make the reconstruction result close to the true value; the codebook loss term is used to optimize the codebook corresponding to the surface coding information or the codebook corresponding to the structure coding information, and the commitment loss term is used to optimize the model parameters. The affinity prediction error includes an ordered classification loss and a mean square error loss. For example, in the ordered classification loss, the affinity change information can be divided into three categories: an increase of more than 1 Kcal / mol, a change within 1 Kcal / mol, and a decrease of more than 1 Kcal / mol. Only when the prediction result and the true value do not belong to the same category, a loss is generated, which is then used to optimize the model parameters. The greater the difference between the prediction result and the true value belongs to the category, the greater the loss generated. The difference between an increase of more than 1 Kcal / mol and a change within 1 Kcal / mol is less than the difference between an increase of more than 1 Kcal / mol and a decrease of more than 1 Kcal / mol. For example, if the true value is a decrease of 2 Kcal / mol, the loss corresponding to the prediction result of an increase of 2 Kcal / mol is higher than that of an increase of 0.5 Kcal / mol. The use of the ordered classification loss can reduce the error of the model in the affinity change category. Here, those skilled in the art should understand that the above classification and numerical setting are only examples. Those skilled in the art can design corresponding parameters based on actual application requirements. In the above embodiments, considering that the two-dimensional and three-dimensional structures of the mutant protein complex are not necessarily accurate, only the wild-type corresponding reconstruction error is used for model optimization.

[0071] In some embodiments, in the above training, the use of the gradient stopping technique makes the update of the codebook corresponding to the surface coding information and the codebook corresponding to the structure coding information only affected by the reconstruction error.

[0072] Figure 3 Fig. 1 shows a device structure diagram for predicting the change of protein complex affinity caused by mutation according to one embodiment of the present application, which comprises a one-one module 11, a one-two module 12 and a one-three module 13. The one-one module 11 inputs wild-type protein complex information and corresponding mutation information into a first sub-network under a twin network architecture, to determine wild-type sequence encoding information and wild-type geometric feature encoding information based on the wild-type protein complex information and the corresponding mutation information, wherein the wild-type geometric feature encoding information comprises wild-type surface encoding information and wild-type structure encoding information; the wild-type geometric feature encoding information is discretely encoded by using a corresponding encoder model, and then combined with the wild-type sequence encoding information to determine wild-type multi-modal encoding information; and the affinity information of the wild-type protein complex is determined by using a corresponding decoder model based on the wild-type multi-modal encoding information. The one-two module 12 inputs mutant protein complex information and corresponding mutation information into a second sub-network under the twin network architecture, to determine the affinity information of the mutant protein complex. The one-three module 13 determines protein complex affinity change information based on the affinity information of the wild-type protein complex and the affinity information of the mutant protein complex. Here, the specific embodiments of the one-one module 11, the one-two module 12 and the one-three module 13 shown in the diagram correspond to the specific embodiments of the aforementioned steps S11, S12 and S13 respectively, and are not described again herein by way of reference.

[0073] In some embodiments, the device 1 further comprises a one-four module 14 (not shown). The one-four module 14 determines mutant protein complex information based on the wild-type protein complex information and corresponding mutation information. Here, the specific embodiments of the one-four module 14 are the same as or similar to the specific embodiments of the aforementioned step S14, and are not described again herein by way of reference.

[0074] In some embodiments, the device 1 further comprises a five-module 15 (not shown). The five-module 15 determines wild-type sample sequence encoding information and wild-type sample geometric feature encoding information based on wild-type protein complex sample information and mutant sample information in the first sub-network under the twin network architecture, wherein the wild-type sample geometric feature encoding information comprises wild-type sample surface encoding information and wild-type sample structure encoding information; utilizes the encoder model to discretely encode the wild-type sample geometric feature encoding information, and determines wild-type sample multi-modal encoding information in combination with wild-type sample sequence encoding information; utilizes the decoder model to determine affinity information of the wild-type protein complex sample; utilizes the second sub-network under the twin network architecture to determine affinity information of the mutant protein complex sample based on mutant protein complex sample information and mutant sample information; determines protein complex sample affinity change information based on the affinity information of the wild-type protein complex sample and the affinity information of the mutant protein complex sample; reconstructs discretely encoded wild-type sample geometric feature encoding information to determine wild-type sample reconstructed encoding information; and optimizes the encoder model and the decoder model in combination with the corresponding loss function based on the wild-type sample reconstructed encoding information, the protein complex sample affinity change information, and the corresponding wild-type sample geometric feature encoding information, the real affinity change information of the mutant protein complex sample, wherein the loss function comprises reconstruction error and affinity prediction error. Here, the corresponding specific embodiments of the five-module 15 are the same as or similar to the specific embodiments of the foregoing step S15, and thus will not be described again, and are hereby incorporated by reference.

[0075] Figure 4 An example system that can be used to implement various embodiments described herein is shown; as Figure 4 As shown, in some embodiments, system 300 can function as any of the devices in the various embodiments described. In some embodiments, system 300 can include one or more computer-readable media (e.g., system memory or NVM / storage 320) having instructions and one or more processors (e.g., processor(s) 305) coupled with the one or more computer-readable media and configured to execute the instructions to implement modules to perform the actions described herein.

[0076] For one embodiment, system control module 310 can include any suitable interface controllers to provide any suitable interface between each of the other system components and at least one of the processor(s) 305 and / or any suitable device or component communicating with system control module 310.

[0077] The system control module 310 can include a memory controller module 330 to provide an interface to system memory 315. The memory controller module 330 can be a hardware module, a software module, and / or a firmware module.

[0078] The system memory 315 can be used, for example, to load and store data and / or instructions for the system 300. For one embodiment, the system memory 315 can include any suitable volatile memory, such as suitable DRAM. In some embodiments, the system memory 315 can include double data rate type four synchronous dynamic random access memory (DDR4 SDRAM).

[0079] For one embodiment, the system control module 310 can include one or more input / output (I / O) controllers to provide an interface to the NVM / storage device 320 and the communication interface(s) 325.

[0080] The NVM / storage device 320 can be used, for example, to store data and / or instructions. The NVM / storage device 320 can include any suitable non-volatile memory (e.g., flash memory) and / or can include any suitable non-volatile storage device(s) (e.g., one or more hard disk drives (HDDs), one or more compact disk (CD) drives, and / or one or more digital versatile disk (DVD) drives).

[0081] The NVM / storage device 320 can include a storage resource that is physically part of the device on which the system 300 is installed, or that is accessed via the device but not necessarily physically part of the device. For example, the NVM / storage device 320 can be accessed over a network via the communication interface(s) 325.

[0082] The communication interface(s) 325 can provide an interface to the system 300 to communicate over one or more networks and / or with any other suitable device. The system 300 can communicate wirelessly with one or more components of a wireless network according to any of one or more wireless network standards and / or protocols.

[0083] For one embodiment, at least one of the processor(s) 305 can be packaged together with logic for one or more controllers of the system control module 310 (e.g., a memory controller module 330). For one embodiment, at least one of the processor(s) 305 can be packaged together with logic for one or more controllers of the system control module 310 to form a system in a package (SiP). For one embodiment, at least one of the processor(s) 305 can be fabricated together with logic for one or more controllers of the system control module 310 on the same die. For one embodiment, at least one of the processor(s) 305 can be fabricated together with logic for one or more controllers of the system control module 310 on the same die to form a system on a chip (SoC).

[0084] In various embodiments, the system 300 can be, but is not limited to, a server, a workstation, a desktop computing device, or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet, a netbook, etc.). In various embodiments, the system 300 can have more or less components, and / or different architectures. For example, in some embodiments, the system 300 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including touch screen displays), non- volatile memory port, multiple antennas, a graphics chip, an application specific integrated circuit (ASIC), and speakers.

[0085] In addition to the methods and devices described in the above embodiments, the present application also provides a computer-readable storage medium, the computer-readable storage medium stores computer code, when the computer code is executed, the method of any one of the preceding is executed.

[0086] The present application also provides a computer program product, when the computer program product is executed by a computer device, the method of any one of the preceding is executed.

[0087] The present application also provides a computer device, the computer device comprises:

[0088] one or more processors;

[0089] a memory for storing one or more computer programs;

[0090] when the one or more computer programs are executed by the one or more processors, the one or more processors implement the method of any one of the preceding.

[0091] It is noted that the present application can be implemented in software and / or in a combination of software and hardware, e.g., using application specific integrated circuits (ASIC), a general purpose computer or any other similar hardware devices. In one embodiment, the software program of the present application is implemented by the processor to perform the above-described steps or functions. Also, the software program of the present application (including related data structures) can be stored in a computer-readable recording medium, e.g., RAM memory, magnetic or optical drive or diskette, and so on. Additionally, some of the steps or functions can be implemented with hardware. For example, a circuitry can be used in place of software or in combination with software to perform some of the steps or functions.

[0092] In addition, some of the steps or functions of the present application can be applied as a computer program product, e.g., a computer program instruction, which, when executed by a computer, can invoke or provide the method and / or technical solutions according to the present application through the operation of the computer. Those skilled in the art should understand that the form of the computer program instruction in the computer-readable medium includes, but is not limited to, source files, executable files, installation package files, etc., and accordingly, the way of the computer program instruction being executed by the computer includes, but is not limited to, the computer directly executing the instruction, or the computer executing the corresponding compiled program after compiling the instruction, or the computer reading and executing the instruction, or the computer executing the corresponding installed program after reading and installing the instruction. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible by the computer.

[0093] Communication media includes wired and wireless communication media including, for example, computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism. The term "modulated data signal" means a signal that has one or more of its characteristics changed or set in a manner as to encode information in the signal. Modulation can be analog, digital or a combination of the two.

[0094] By way of example, and not limitation, computer-readable storage media can include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. For example, computer-readable storage media includes, but is not limited to, RAM, such as SRAM, DRAM, or RAM; ROM, such as PROM, EPROM, EEPROM, or Flash memory; magnetic or optical storage, such as hard disks, optical disks, or magnetic tape; or other memory technology, including FeRAM, or other similar technology, as well as any other medium which can be used to store the desired information in a computer-readable fashion.

[0095] Herein, according to one embodiment of the present application comprises a device, the device comprises a memory for storing computer program instructions and a processor for executing program instructions, wherein when the computer program instructions are executed by the processor, trigger the device to run based on the method and / or technical solutions according to the foregoing embodiments of the present application.

[0096] It is apparent that the application is not limited to the details of the foregoing exemplary embodiments, and thus can be implemented in other particular forms without departing from the spirit or essential characteristics of the application. The foregoing embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No feature of the application is considered critical unless the claims expressly state otherwise. Furthermore, no element, act, or instruction is implied by the presence of a preceding or following clause. The mere fact that an element, act, or instruction is recited in the foregoing specification does not imply that it is an essential, critical component or step. In addition, the term "comprising" is used throughout the specification to mean the inclusion from the listed items with the indication that not all possibilities are included. Finally, terms of degree such as "substantially", "about", and "approximately" are used herein to convey an intended precision of a number, quantity, or degree but should be interpreted in a non-critical manner to cover some slack or variation from the stated reference quantifier.

Claims

1. A method for predicting changes in protein complex affinity caused by mutations, wherein, The method includes: In the first sub-network of the Siamese network architecture, wild-type sequence coding information and wild-type geometric feature coding information are determined based on wild-type protein complex information and corresponding mutation information. The wild-type geometric feature coding information includes wild-type surface coding information and wild-type structural coding information. Using a corresponding encoder model, the wild-type geometric feature coding information is discretely encoded, and then combined with the wild-type sequence coding information to determine wild-type multimodal coding information. Based on the wild-type multimodal coding information, the affinity information of the wild-type protein complex is determined using a corresponding decoder model. Specifically, the determination of wild-type sequence coding information and wild-type geometric feature coding information based on wild-type protein complex information and corresponding mutation information, wherein the wild-type geometric feature coding information includes wild-type surface coding information and wild-type structure coding information, comprising: determining the wild-type sequence coding information based on the wild-type protein complex information using a protein language model; and determining the wild-type geometric feature coding information based on the wild-type protein complex information and corresponding mutation information, wherein the wild-type geometric feature coding information matches the mutation information; The step of determining the wild-type geometric feature encoding information based on the wild-type protein complex information and the corresponding mutation information, wherein matching the wild-type geometric feature encoding information with the mutation information includes: determining multiple protein surface point cloud information based on the wild-type protein complex information, wherein each protein surface point cloud information corresponds to a protein that binds to form the wild-type protein complex; determining the wild-type surface encoding information based on the multiple protein surface point cloud information and the mutation information, wherein the wild-type surface encoding information includes surface encoding information corresponding to each protein surface point cloud information, and the surface encoding information corresponding to each protein surface point cloud information matches the mutation information; and determining the wild-type structural encoding information based on the wild-type protein complex information and the mutation information using a graph model based on an attention mechanism, wherein the wild-type structural encoding information includes node information and edge information, the node information includes amino acid information matching the mutation information, and the edge information includes amino acid pairs matching the mutation information. The information on mutant protein complexes and the corresponding mutation information are input into the second sub-network under the Siamese network architecture to determine the affinity information of mutant protein complexes. Based on the affinity information of the wild-type protein complex and the affinity information of the mutant protein complex, the affinity change information of the protein complex is determined.

2. The method according to claim 1, wherein, The step of determining multiple protein surface point cloud information based on the wild-type protein complex information, wherein each protein surface point cloud information corresponds to a protein that binds to form the wild-type protein complex, includes: Several points are extracted, centered on the heavy atoms of each protein that binds to form the wild-type protein complex, to form point cloud information corresponding to each protein. Based on the point cloud information corresponding to each protein, the protein surface point cloud information corresponding to each protein is selected and determined using a smoothing distance function.

3. The method according to claim 2, wherein, The step of determining multiple protein surface point cloud information based on the wild-type protein complex information, wherein each protein surface point cloud information corresponds to a protein that binds to form the wild-type protein complex, further includes: The point cloud information on the protein surface is smoothed.

4. The method according to claim 1, wherein, The wild-type surface coding information is determined based on the plurality of protein surface point cloud information and the mutation information, wherein the wild-type surface coding information includes surface coding information corresponding to each protein surface point cloud information, and the matching of the surface coding information corresponding to each protein surface point cloud information with the mutation information includes: Based on the surface distances between points in the point cloud information of each protein surface, clustering is used to determine multiple block information corresponding to the point cloud information of each protein surface. Based on the mutation information, multiple target block information that matches the mutation information is determined from the multiple block information; Determine the block encoding information corresponding to each target block information, and use the block encoding information as the surface encoding information corresponding to the corresponding protein surface point cloud information. The block encoding information includes the physicochemical feature integrated encoding information and the relative position encoding information corresponding to the target block information.

5. The method according to claim 4, wherein, The step of determining the block encoding information corresponding to each target block information, and using the block encoding information as the surface encoding information corresponding to the protein surface point cloud information, wherein the block encoding information includes the physicochemical feature integrated encoding information and relative position encoding information corresponding to the target block information, including: Based on each point in the protein surface point cloud information, and other points corresponding to each point that satisfy the first surface distance condition, the integrated coding information corresponding to each point in the protein surface point cloud information is determined. Based on the integrated coding information corresponding to each point in the protein surface point cloud information, the physicochemical feature coding information corresponding to each block of information is determined. Based on the target block information and other block information that corresponds to the target block information and satisfies the second surface distance condition, and combined with the physicochemical feature coding information corresponding to each block information, the physicochemical feature integrated coding information corresponding to the target block information is determined; Using Morton coding, the relative position encoding information of the target block information in one-dimensional space is determined based on the relative position information of the target block information on the protein surface.

6. The method according to any one of claims 1 to 5, wherein, The method further includes: Based on the information of the wild-type protein complex and the corresponding mutation information, the information of the mutant protein complex is determined.

7. The method according to any one of claims 1 to 5, wherein, The method further includes: In the first sub-network of the twin network architecture, based on wild-type protein complex sample information and mutant sample information, wild-type sample sequence encoding information and wild-type sample geometric feature encoding information are determined. The wild-type sample geometric feature encoding information includes wild-type sample surface encoding information and wild-type sample structure encoding information. Using the encoder model, the wild-type sample geometric feature encoding information is discretely encoded, and then combined with the wild-type sample sequence encoding information to determine the wild-type sample multimodal encoding information. Finally, the decoder model is used to determine the affinity information of the wild-type protein complex sample. Based on the mutant protein complex sample information and the mutant sample information, the affinity information of the mutant protein complex sample is determined using the second sub-network under the twin network architecture. Based on the affinity information of the wild-type protein complex sample and the affinity information of the mutant protein complex sample, the affinity change information of the protein complex sample is determined. Reconstruct the geometric feature coding information of the wild-type samples with discrete coding, and determine the reconstructed coding information of the wild-type samples; Based on the reconstructed encoding information of the wild-type sample, the affinity change information of the protein complex sample and the corresponding geometric feature encoding information of the wild-type sample, and the true affinity change information of the mutant protein complex sample, the encoder model and the decoder model are optimized by combining the corresponding loss function, wherein the loss function includes reconstruction error and affinity prediction error.

8. The method according to claim 7, wherein, The reconstruction error includes surface information reconstruction error and structural information reconstruction error, and the affinity prediction error includes ordered classification loss and mean square error loss.

9. A computer device for predicting changes in the affinity of protein complexes caused by mutations, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 8.

11. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 8.

Images