Compound library processing method, device, equipment, storage medium and program product

Abstract

The present application provides a compound processing method, device, electronic device, computer-readable storage medium and computer program product based on artificial intelligence; it relates to artificial intelligence technology; the method includes: obtaining multiple protein prediction structures of a target protein sequence; performing anti-clustering processing on multiple protein prediction structures to obtain the group to which each protein prediction structure belongs, wherein each group includes multiple different protein prediction structures; performing graph structure construction processing based on the spatial structure of the protein prediction structure in each group and the spatial structure of the target protein sequence to obtain the graph structure of the protein prediction structure in each group; performing structure quality prediction processing based on the graph structure of each protein prediction structure to obtain the structure quality index of each protein prediction structure. Through the present application, multiple spatial structures can be combined to improve the accuracy of structure quality prediction.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an artificial intelligence-based compound processing method, device, electronic device and computer-readable storage medium. Background Art

[0002] Artificial Intelligence (AI) is a comprehensive technology in computer science. By studying the design principles and implementation methods of various intelligent machines, machines are given the functions of perception, reasoning and decision-making. Artificial intelligence technology is an interdisciplinary subject that covers a wide range of fields, such as natural language processing technology and machine learning / deep learning. With the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0003] Protein structure quality assessment is to predict the quality of predicted protein structure through artificial intelligence technology, so as to screen compounds and obtain targeted lead compounds. Compared with biological experiments, protein structure quality assessment has the advantages of low cost and high efficiency.

[0004] However, protein structure quality assessment in related technologies relies on all protein predicted structures in the structure prediction model pool, and the effective information captured is relatively limited, which leads to deviations in protein structure quality assessment. Summary of the invention

[0005] The embodiments of the present application provide an artificial intelligence-based compound processing method, device, electronic device and computer-readable storage medium, which can combine multiple spatial structures to improve the accuracy of structural quality prediction.

[0006] The technical solution of the embodiment of the present application is implemented as follows:

[0007] The present application embodiment provides a compound processing method based on artificial intelligence, comprising:

[0008] Obtain multiple protein predicted structures of the target protein sequence;

[0009] Performing anti-clustering processing on the multiple predicted protein structures to obtain a group to which each predicted protein structure belongs, wherein each group includes multiple different predicted protein structures;

[0010] Performing graph structure construction processing based on the spatial structure of the predicted protein structure in each of the groups and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each of the groups;

[0011] A structural quality prediction process is performed based on the graph structure of each of the predicted protein structures to obtain a structural quality index of each of the predicted protein structures.

[0012] The present application embodiment provides a compound processing device based on artificial intelligence, comprising:

[0013] An acquisition module, used to obtain multiple protein predicted structures of a target protein sequence;

[0014] an anti-clustering module, used for performing anti-clustering processing on the plurality of predicted protein structures to obtain a group to which each predicted protein structure belongs, wherein each group includes a plurality of different predicted protein structures;

[0015] A construction module, used for performing graph structure construction processing based on the spatial structure of the predicted protein structure in each of the groups and the spatial structure of the target protein sequence, to obtain the graph structure of the predicted protein structure in each of the groups;

[0016] The prediction module is used to perform structure quality prediction processing based on the graph structure of each of the predicted protein structures to obtain a structure quality index for each of the predicted protein structures.

[0017] In the above technical solution, the anti-clustering module is further used to perform clustering processing on the multiple predicted protein structures to obtain the category to which each predicted protein structure belongs, wherein the category includes multiple different predicted protein structures;

[0018] A predicted protein structure is sampled from each of the categories, and a set of the sampled predicted protein structures is used as a group.

[0019] In the above technical solution, the anti-clustering module is also used to take one of the predicted protein structures as one of the categories;

[0020] The following processing is performed iteratively:

[0021] Performing similarity processing based on predicted protein structures included in the plurality of categories to obtain similarities between the categories;

[0022] According to the order of the similarities from high to low, the categories corresponding to the obtained similarities are merged;

[0023] When the number of predicted protein structures included in the merged category is less than or equal to the quantity threshold, taking the merged category as the new category;

[0024] When the number of predicted protein structures included in the merged categories is greater than the quantity threshold, the categories corresponding to the next similarity of the similarity are merged in descending order of the similarity until the number of predicted protein structures included in all merged categories is greater than the quantity threshold and the iteration is stopped.

[0025] In the above technical solution, the anti-clustering module is further used to perform the following processing on any two of the multiple categories:

[0026] Obtain any first protein predicted structure in one of any two of the categories, and obtain a plurality of target atom pairs in the first protein predicted structure, wherein two target atoms in the target atom pairs belong to different residues, and a first Euclidean distance between the two target atoms is less than a distance threshold;

[0027] Taking the set of the plurality of target atom pairs as an initial set;

[0028] Obtain any second protein predicted structure in another category of any two of the categories, and obtain a candidate atom pair in the second protein predicted structure that is identical to the target atom pair;

[0029] When the error between the second Euclidean distance between two candidate atoms in the candidate atom pair and the first Euclidean distance is less than an error threshold, deleting the target atom pair corresponding to the candidate atom pair from the initial set to obtain an updated set;

[0030] Using the ratio of the number of target atom pairs included in the updated set to the number of target atom pairs included in the initial set as the local distance difference between the first protein predicted structure and the second protein predicted structure;

[0031] An average value of the plurality of local distance differences is taken as the similarity between any two of the categories.

[0032] In the above technical solution, the construction module is also used to construct the vertices of the graph structure of each predicted protein structure based on the amino acid residues of the predicted protein structure in each of the groups;

[0033] Determining a first distance matrix for each of the predicted protein structures based on the spatial structure of each of the predicted protein structures;

[0034] Determining a second distance matrix of the target protein sequence based on the spatial structure of the target protein sequence;

[0035] Based on the first distance matrix and the second distance matrix, edges of a graph structure of predicted protein structures in each of the groups are constructed.

[0036] In the above technical solution, the building module is also used to obtain the Euclidean distance between the amino acid residues in each of the predicted protein structures;

[0037] Based on the Euclidean distances between the amino acid residues, a first distance matrix is ​​constructed for each of the predicted protein structures.

[0038] In the above technical solution, the construction module is also used to call the protein structure prediction model based on the target protein sequence to perform the following processing:

[0039] Performing multi-data source fusion processing on the target protein sequence to obtain co-evolution information of the target protein sequence;

[0040] Deep cross-attention processing is performed on the co-evolution information of the target protein sequence to obtain a second distance matrix of the target protein sequence.

[0041] In the above technical solution, the construction module is also used to perform template search processing based on the spatial structure of the target protein sequence to obtain the template sequence of the target protein sequence;

[0042] Aligning the target protein sequence and the template sequence to obtain the aligned template sequence;

[0043] Based on the aligned Euclidean distances between the amino acid residues in the template sequence and the amino acid residues in the target protein sequence, a second distance matrix of the target protein sequence is constructed.

[0044] In the above technical solution, the construction module is also used to perform the following processing on any two vertices in the predicted protein structure in each of the groups:

[0045] Determining a first Euclidean distance between the two vertices from the first distance matrix;

[0046] Determining a second Euclidean distance between the two vertices from the second distance matrix;

[0047] Perform intersection logic processing on the first Euclidean distance and the second Euclidean distance to obtain an edge between the two vertices.

[0048] In the above technical solution, the construction module is also used to determine the similarity between each of the predicted protein structures and other predicted protein structures in the same group;

[0049] Determining vertex features of vertices in the graph structure based on an average value of the similarities and energies of the vertices in the graph structure;

[0050] Performing distance discretization processing on the edges of the graph structure based on the first distance matrix and the second distance matrix to obtain discretized edges;

[0051] Gaussian distribution processing is performed on the discretized edges to obtain edge features of the edges.

[0052] In the above technical solution, the prediction module is also used to fuse the vertex features and edge features of the vertices in the graph structure of each protein prediction structure to obtain the fused features of the vertex;

[0053] Performing vertex update processing on the fused features to obtain updated features of the vertices;

[0054] The updated features of the vertices are subjected to nonlinear quality mapping processing to obtain a structural quality index of each vertex in the predicted protein structure.

[0055] In the above technical solution, the acquisition module is also used to screen the protein sequences in the first protein structure database based on the similarity between the protein sequences to obtain candidate protein sequences;

[0056] Based on the template sequence in the second protein structure database, a template search is performed on the candidate protein sequence to obtain the target protein sequence;

[0057] The target protein sequence is processed for protein structure generation to obtain a plurality of predicted protein structures of the target protein sequence.

[0058] An embodiment of the present application provides an electronic device for compound processing, the electronic device comprising:

[0059] A memory for storing executable instructions;

[0060] The processor is used to implement the artificial intelligence-based compound processing method provided in the embodiment of the present application when executing the executable instructions stored in the memory.

[0061] An embodiment of the present application provides a computer-readable storage medium storing executable instructions for causing a processor to execute and implement the artificial intelligence-based compound processing method provided in the embodiment of the present application.

[0062] An embodiment of the present application provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the artificial intelligence-based compound processing method provided in the embodiment of the present application.

[0063] The embodiments of the present application have the following beneficial effects:

[0064] By combining the spatial structure of the predicted protein structure and the spatial structure of the target protein sequence to construct a graph structure, the accuracy of the structural quality prediction is improved, and the predicted protein structure is reversed clustered so that the predicted structures of proteins in the same group are not similar, so that the structural quality prediction is performed according to the group. Compared with the related technology that relies on the predicted structures of all proteins, the efficiency of the structural quality prediction can be improved while ensuring the accuracy of the structural quality prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of an application scenario of a compound analysis system provided in an embodiment of the present application;

[0066] Figure 2 is a schematic diagram of the structure of an electronic device for compound processing provided in an embodiment of the present application;

[0067] Figure 3-Figure 6 It is a schematic flow chart of the compound processing method based on artificial intelligence provided in the embodiment of the present application;

[0068] Figure 7 It is a schematic flow chart of the artificial intelligence-based compound processing method provided in the embodiments of the present application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0070] In the following description, the terms "first\second" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0072] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0073] 1) Three-dimensional (3D) conformation: the countless specific images of the atoms or groups of a compound arranged in three-dimensional space.

[0074] 2) Multiple Sequence Alignment (MSA): Multiple (3 or more) amino acid sequences (also called protein sequences) or nucleic acid sequences of protein molecules with systematic evolutionary relationships are aligned, and the same bases or amino acid residues are arranged in the same column as much as possible. The site homology of amino acids or nucleotides in different sequences is detected. When sequence alignment is used for evolutionary analysis, amino acids or nucleotide sites in the same position are considered to be evolutionarily homologous and have a common ancestor.

[0075] 3) Local distance difference (LDDT): an indicator for evaluating the difference between the predicted protein structure and the experimentally resolved structure. The value range is between 0 and 1, where 1 indicates a perfect match between the predicted protein structure and the experimentally resolved structure.

[0076] 4) Protein prediction structure model: also known as protein prediction structure. The protein prediction structure model in the embodiments of the present application refers to a number of entities containing protein structure information generated using protein structure prediction technology, rather than an algorithmic model for predicting protein structure.

[0077] 5) Anti-clustering: the reverse operation of clustering. Clustering is used to group similar objects into the same class, that is, to make objects in the same class as similar as possible, while anti-clustering has the opposite effect of clustering. Anti-clustering is used to group dissimilar objects into the same class, that is, to make objects in the same class as dissimilar as possible. Anti-clustering in the embodiment of the present application is used to group dissimilar protein prediction structures into the same group, that is, to make protein prediction structures in the same group as dissimilar as possible.

[0078] 6) Graph structure construction: The physical structure (i.e., 3D conformation) of the protein structure is constructed into a graph structure, which includes vertices and edges between vertices. In the embodiment of the present application, the amino acid residues of the protein structure are used as the vertices of the graph structure of the protein structure, and the edges of the graph structure of the protein structure are constructed based on the distances between the amino acid residues.

[0079] 7) Structural quality prediction: The quality of the protein predicted structure is evaluated through a neural network model to determine the structural quality of the generated protein predicted structure. The higher the structural quality of the generated protein predicted structure, the more the generated protein predicted structure matches the experimental analyzed structure, that is, the more authentic and reliable the generated protein predicted structure is.

[0080] 8) Structural quality index: an index used to evaluate the structural quality of the generated predicted protein structure. Subsequent R&D personnel can conduct molecular research and analysis based on the structural quality index of the predicted protein structure. For example, the structural quality index can be the difference between the predicted protein structure and the experimentally resolved structure. The smaller the difference, the more authentic and reliable the predicted protein structure is. The structural quality index can also be the lDDT value of each amino acid residue in the predicted protein structure. The closer the lDDT value of the amino acid residue is to the lDDT value of each amino acid residue in the real structure, the more authentic and reliable the predicted protein structure is.

[0081] 9) Template search: Based on the target protein sequence, a template sequence matching the protein sequence is searched in the protein structure database, where the template sequence in the protein structure database is an existing protein sequence. For example, a protein sequence search module (such as hhblits) is used to perform template search.

[0082] 10) Alignment: Multiple sequence alignment (MSA) technology is used to align protein sequences so that the residues at each position are identical or similar and there is a one-to-many relationship at each position.

[0083] 11) Protein structure generation: Use protein structure prediction technology (such as tFold) to generate multiple entities containing protein structure information for the protein sequence, namely, protein predicted structures, where the protein predicted structure has the same amino acid sequence as the protein sequence but a different 3D conformation.

[0084] 12) Co-evolution information: used to identify protein factors that affect biological evolution. If an amino acid in a protein sequence mutates during biological evolution and destroys the interaction between amino acid residues, the protein sequence will become unstable, and the organism with this mutation will not be able to survive easily. However, when an amino acid that interacts with the mutated amino acid also mutates, and the two mutated amino acids can interact well with each other, so that the structure of the protein sequence continues to stabilize, this situation is called protein co-evolution. Co-evolution information includes information about two mutated, interacting amino acids, and is used to determine protein factors that affect biological evolution.

[0085] Protein prediction structure model evaluation methods in related technologies can be mainly divided into three categories: single model evaluation method, consensus model evaluation method and quasi-single model evaluation method.

[0086] Among them, the single model evaluation method only considers the current single protein prediction structure model (also called prediction structure model) in each evaluation, calculates the corresponding molecular potential energy and water solubility and other characteristics according to the geometric distribution inside the protein prediction structure model, and linearly weights or inputs the various indicators that can be extracted from the single model to obtain a total energy, and statistically analyzes the geometric probability distribution of experimental analytical structures at various scales, so as to predict the rationality of the protein prediction structure model to be evaluated. Since the characteristic accuracy of the single model evaluation method itself has certain errors, the error in the final score calculation may be amplified.

[0087] For the consensus model evaluation method, each evaluation will consider all other structural models in the entire pool to be evaluated. The score of a protein prediction structure model to be evaluated depends on its similarity with all other protein prediction structure models. The overall time complexity of the consensus model evaluation method is O(n 2 ), which will incur huge time overhead when there are a large number of predicted structures in the evaluation pool.

[0088] For the quasi-single model evaluation method, for a protein prediction structure model to be evaluated, it will not refer to other models to be evaluated in the pool, but will generate several structural models with the same residue sequence through the internal structure prediction module. The score of the protein prediction structure model to be evaluated will refer to the relationship between these generated structural models. The quasi-single model evaluation method relies on its own internal structure prediction module. If the accuracy of the structure prediction result is low, the accuracy of the structure model evaluation will also be affected.

[0089] In order to solve the above problems, the embodiments of the present application provide an artificial intelligence-based compound processing method, device, electronic device and computer-readable storage medium, which can combine multiple spatial structures to improve the accuracy of structural quality prediction.

[0090] The artificial intelligence-based compound processing method provided in the embodiments of the present application can be implemented by the terminal / server alone; it can also be implemented by the terminal and the server in collaboration. For example, the terminal alone undertakes the artificial intelligence-based compound processing method described below, or the terminal sends a quality assessment request for the predicted protein structure (including the target protein sequence) to the server, and the server executes the artificial intelligence-based compound processing method according to the received quality assessment request for the predicted protein structure, obtains multiple predicted protein structures of the target protein sequence, performs anti-clustering processing on the multiple predicted protein structures to obtain the group to which each predicted protein structure belongs, performs graph structure construction processing based on the spatial structure of the predicted protein structure in each group and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each group, performs structure quality prediction processing based on the graph structure of each predicted protein structure to obtain the structural quality index of each predicted protein structure, so that R&D personnel can quickly conduct subsequent molecular research, analysis, etc. according to the structural quality index of each predicted protein structure.

[0091] The electronic device for compound processing provided in the embodiments of the present application may be various types of terminal devices or servers, wherein the server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services; the terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the present application.

[0092] Taking the server as an example, it can be a server cluster deployed in the cloud, opening artificial intelligence cloud services (AI as a Service, AIaaS) to users. The AIaaS platform will split several common AI services and provide independent or packaged services in the cloud. This service model is similar to an AI theme mall. All users can access one or more artificial intelligence services provided by the AIaaS platform through an application programming interface.

[0093] For example, one of the artificial intelligence cloud services can be a compound processing service, that is, the server in the cloud is encapsulated with a compound processing program provided by the embodiment of the present application. The user calls the compound processing service in the cloud service through a terminal (running a client, such as a compound screening client, etc.), so that the server deployed in the cloud calls the encapsulated compound processing program, performs anti-clustering processing on multiple protein prediction structures, obtains the group to which each protein prediction structure belongs, and performs graph structure construction processing based on the spatial structure of the protein prediction structure in each group and the spatial structure of the target protein sequence, obtains the graph structure of the protein prediction structure in each group, performs structural quality prediction processing based on the graph structure of each protein prediction structure, and obtains the structural quality index of each protein prediction structure, so that the research and development personnel can quickly carry out subsequent molecular research and analysis based on the structural quality index of each protein prediction structure.

[0094] See also Figure 1 , Figure 1 It is a schematic diagram of an application scenario of the compound analysis system 10 provided in an embodiment of the present application. The terminal 200 is connected to the server 100 via a network 300. The network 300 may be a wide area network or a local area network, or a combination of the two.

[0095] Terminal 200 (running a client, such as a compound screening client, etc.) can be used to obtain a quality assessment request for a predicted protein structure (including a target protein sequence). For example, a research and development personnel inputs a target protein sequence through the input interface of terminal 200, and a quality assessment request for a predicted protein structure is automatically generated.

[0096] In some embodiments, a compound processing plug-in may be implanted in the client running in the terminal to implement a compound processing method based on artificial intelligence locally on the client. For example, after the terminal 200 obtains a quality assessment request for a protein predicted structure, the compound processing plug-in is called to implement a compound processing method based on artificial intelligence, and a plurality of protein predicted structures are reversely clustered to obtain the group to which each protein predicted structure belongs, and a graph structure construction process is performed based on the spatial structure of the protein predicted structure in each group and the spatial structure of the target protein sequence to obtain the graph structure of the protein predicted structure in each group, and a structural quality prediction process is performed based on the graph structure of each protein predicted structure to obtain a structural quality index of each protein predicted structure, so that R&D personnel can quickly conduct subsequent molecular research, analysis, etc. according to the structural quality index of each protein predicted structure.

[0097] In some embodiments, after the terminal 200 obtains a quality assessment request for a predicted protein structure, it calls the compound processing interface of the server 100 (which can be provided in the form of a cloud service, i.e., a compound processing service), and the server 100 performs anti-clustering processing on multiple predicted protein structures to obtain the group to which each predicted protein structure belongs, and performs graph structure construction processing based on the spatial structure of the predicted protein structure in each group and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each group, and performs structural quality prediction processing based on the graph structure of each predicted protein structure to obtain the structural quality index of each predicted protein structure, so that R&D personnel can quickly conduct subsequent molecular research, analysis, etc. according to the structural quality index of each predicted protein structure.

[0098] The structure of the electronic device for compound processing provided by the embodiment of the present application is described below. Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device 500 for compound processing provided in an embodiment of the present application, taking the electronic device 500 as a server as an example, Figure 2 The electronic device 500 for compound processing shown includes: at least one processor 510, a memory 550, and at least one network interface 520. The various components in the electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 540 is not described in detail. Figure 2 Various buses are labeled as bus system 540 .

[0099] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0100] The memory 550 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 550 described in the embodiment of the present application is intended to include any suitable type of memory. The memory 550 optionally includes one or more storage devices that are physically far away from the processor 510.

[0101] In some embodiments, the memory 550 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0102] Operating system 551, including system programs for processing various basic system services and performing hardware-related tasks, such as framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0103] A network communication module 552, for reaching other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 include: Bluetooth, Wireless Compatibility Authentication (WiFi), and Universal Serial Bus (USB);

[0104] In some embodiments, the compound processing device provided in the embodiments of the present application can be implemented in software, for example, it can be a compound processing plug-in in the terminal described above, or it can be a compound processing service in the server described above. Of course, it is not limited to this, and the compound processing device provided in the embodiments of the present application can be provided as various software embodiments, including various forms including application programs, software, software modules, scripts or codes.

[0105] Figure 2 A compound processing device 555 stored in a memory 550 is shown, which may be software in the form of a program and a plug-in, such as a compound processing plug-in, and includes a series of modules, including an acquisition module 5551, an anti-clustering module 5552, a construction module 5553 and a prediction module 5554; wherein the acquisition module 5551, the anti-clustering module 5552, the construction module 5553 and the prediction module 5554 are used to implement the compound processing function provided in an embodiment of the present application.

[0106] As mentioned above, the compound processing method based on artificial intelligence provided in the embodiment of the present application can be implemented by various types of electronic devices. Figure 3 , Figure 3 is a schematic diagram of a compound processing method based on artificial intelligence provided in an embodiment of the present application, combined with Figure 3 The steps shown are explained.

[0107] In step 101, a plurality of predicted protein structures of a target protein sequence are obtained.

[0108] As an example of obtaining a predicted protein structure, a user inputs a target protein sequence through the input interface of the terminal, and a quality assessment request for the predicted protein structure (including the target protein sequence) is automatically generated and sent to the server. The server parses the quality assessment request for the predicted protein structure, obtains the target protein sequence, and performs structure prediction through a protein structure prediction tool (such as the tFold tool) to obtain multiple predicted protein structures of the target protein sequence, for example, generating 150 candidate structural models (i.e., predicted protein structures) for each sequence.

[0109] In some embodiments, obtaining multiple predicted protein structures of a target protein sequence includes: screening protein sequences in a first protein structure database based on similarities between protein sequences to obtain candidate protein sequences; performing template search on the candidate protein sequences based on template sequences in a second protein structure database to obtain the target protein sequence; performing protein structure generation processing on the target protein sequence to obtain multiple predicted protein structures of the target protein sequence.

[0110] For example, all protein sequences containing structural tags are collected in a first protein structure database (PDB), and based on the similarity between protein sequences, the protein sequences in the PDB are screened, that is, the protein sequences with a similarity greater than a first similarity threshold (e.g., 60%) are removed to obtain the remaining protein sequences (i.e., candidate protein sequences). A protein sequence search module (e.g., hhblits) is used for the remaining protein sequences, and a template search is performed on a second protein structure database (e.g., PDB70 (a collection of protein sequences with a similarity greater than 70%)). If the similarity between the searched template (i.e., the protein sequence in PDB70) and the candidate protein sequence is greater than a second similarity threshold (e.g., 60%), the candidate protein sequence is also removed to obtain the final filtered protein sequence (i.e., the target protein sequence). Finally, a protein structure prediction tool (e.g., tFold) is used for all target protein sequences to generate protein structures, and multiple protein prediction structures of the target protein sequence are obtained, for example, 150 candidate structure models (i.e., protein prediction structures) are generated for each target protein sequence.

[0111] Among them, the sequence search process of hhblits is as follows: each time a multiple sequence alignment is performed (protein sequences are aligned so that the residues at each position are the same or similar, there is a one-to-many relationship at each position, and there are multiple similarity scores at each position), a position-specific matrix (PSSM) is established (representing the similarity between the protein sequence and other protein sequences at each position), and this position-specific matrix is ​​used to search in the protein structure database, and the PSSM is updated in combination with the search results (the searched protein sequences), and the cycle is repeated to finally obtain the final search result (i.e., the protein sequence whose similarity is greater than the second similarity threshold).

[0112] It should be noted that tFold is a protein structure prediction algorithm. The algorithm uses multi-source fusion technology to mine co-evolutionary information in multiple sets of multiple sequence alignments (MSA); then, with the help of the Deep Cross-Attention Residual Network, it can greatly improve the prediction accuracy of some important protein 2D structural information (such as the residue-residue distance and orientation matrix); finally, through a template-assisted free modeling (TBFM, Template-based Free Modeling) method, the structural information in the 3D structural model generated by free modeling (FM, FreeModeling) and template modeling (TBM, Template-based Modeling) is effectively fused, thereby greatly improving the accuracy of the final 3D modeling.

[0113] In step 102, a plurality of predicted protein structures are subjected to anti-clustering processing to obtain a group to which each predicted protein structure belongs, wherein each group includes a plurality of different predicted protein structures.

[0114] For example, the protein prediction structure models generated from the same protein sequence are reversely clustered and grouped so that the protein prediction structure models in the same group are as dissimilar as possible, so that the structure quality prediction is subsequently performed according to the groups to ensure data diversity and robustness during quality prediction. Compared with the related technology that relies on the predicted structures of all proteins, the efficiency of structure quality prediction can be improved while ensuring the accuracy of structure quality prediction.

[0115] See also Figure 4 , Figure 4is an optional flow chart of the compound processing method based on artificial intelligence provided in the embodiment of the present application, Figure 4 Show Figure 3 Step 102 in the method can be implemented through steps 1021 and 1022: in step 1021, a plurality of predicted protein structures are clustered to obtain a category to which each predicted protein structure belongs, wherein the category includes a plurality of different predicted protein structures; in step 1022, a predicted protein structure is sampled from each category, and the set of sampled predicted protein structures is used as a group.

[0116] For example, in order to ensure that each group includes multiple protein prediction structures that are as dissimilar as possible, the multiple protein prediction structures can be clustered first to obtain the category to which each protein prediction structure belongs, so that each category includes multiple protein prediction structures that are as similar as possible, and then a protein prediction structure is sampled from each category, and the sampled protein prediction structures are constructed into a group.

[0117] As an example, when there are K categories, the protein prediction structures of the K categories are traversed in turn, and the first protein prediction structure of the currently traversed category is separated into a new group each time, until all the protein prediction structures of the K categories belong to the new group, so as to achieve anti-clustering grouping, so that the protein prediction structures of the same group are as dissimilar as possible.

[0118] In some embodiments, clustering is performed on multiple predicted protein structures to obtain the category to which each predicted protein structure belongs, including: taking one predicted protein structure as a category; iteratively performing the following processing: performing similarity processing based on the predicted protein structures included in multiple categories to obtain the similarity between categories; merging the categories corresponding to the obtained similarities in descending order of similarity; when the number of predicted protein structures included in the merged category is less than or equal to a quantity threshold, taking the merged category as a new category; when the number of predicted protein structures included in the merged category is greater than the quantity threshold, merging the categories corresponding to the next similarity in descending order of similarity, until the number of predicted protein structures included in all merged categories is greater than the quantity threshold, and the iteration is stopped.

[0119] For example, if the target protein sequence has N predicted protein structures, the maximum number of structural models per group is set to 15 (can be set according to needs), and the maximum number of structural models per class is set to M = ceil (N / 15), that is, the number threshold, where ceil means rounding up, and each protein predicted structure is initialized as a single class. Perform the following iterative process:

[0120] 1. If there are k categories in total, calculate the similarity between any two categories, a total of k*(k-1) / 2 times.

[0121] 2. Obtain the similarity between k*(k-1) / 2 categories, merge the two categories with the highest similarity, and when the number of structural models after merging is not greater than M, merge the two categories and continue with step 1; if the number of predicted protein structures after merging is greater than M, merge the two categories with the second highest similarity, and then determine whether the number of structural models after merging is greater than M, until all categories are traversed.

[0122] In some embodiments, similarity processing is performed based on protein prediction structures included in multiple categories to obtain the similarity between categories, including: performing the following processing on any two categories in the multiple categories: obtaining any first protein prediction structure in one of the two categories, and obtaining multiple target atom pairs in the first protein prediction structure, wherein the two target atoms in the target atom pair belong to different residues and the first Euclidean distance between the two target atoms is less than a distance threshold; taking the set of multiple target atom pairs as the initial set; obtaining any second protein prediction structure in the other of the two categories, and obtaining candidate atom pairs that are the same as the target atom pairs in the second protein prediction structure; when the error between the second Euclidean distance between the two candidate atoms in the candidate atom pair and the first Euclidean distance is less than the error threshold, deleting the target atom pairs corresponding to the candidate atom pairs in the initial set to obtain an updated set; taking the ratio of the number of target atom pairs included in the updated set to the number of target atom pairs included in the initial set as the local distance difference between the first protein prediction structure and the second protein prediction structure; and taking the average value of the multiple local distance differences as the similarity between any two categories.

[0123] For example, if there are k categories, the similarity between any two categories is calculated k*(k-1) / 2 times. The similarity between category A and category B is calculated as Wherein, i and j are index subscripts of protein prediction structures in categories A and B, i.e., the first protein prediction structure and the second protein prediction structure, respectively. lDDT is an index for calculating the similarity between a protein prediction structure model and a label. The present embodiment uses lDDT to calculate the similarity between two protein prediction structures (i.e., the local distance difference between the first protein prediction structure and the second protein prediction structure). The present embodiment is not limited to the calculation formula of clusim(A, B).

[0124] It should be noted that the calculation process of lDDT is as follows: predict structure a for the first protein and predict structure b for the second protein. Extract target atom pairs that meet the following conditions in a: 1. The two atoms do not belong to the same residue; 2. The Euclidean distance between the two atoms is less than the distance threshold R. Set the target atom pairs that meet the above conditions to set L (i.e., the initial set). Query the candidate atom pairs that are the same as set L in b. If the error between the Euclidean distance of the queried candidate atom pairs and the Euclidean distance of the atom pairs in L is within the threshold range T (e.g., T is 0.5 angstroms, 1 angstroms, 2 angstroms, 4 angstroms), then retain the candidate atom pair. Otherwise, remove the candidate atom pair from R. After all atom pairs in L are verified, a new set M (i.e., the updated set) is obtained. Among them, the lDDT single threshold (i.e. 0.5 angstrom, 1 angstrom, 2 angstroms, 4 angstroms) similarity is M / L, and the total lDDT value is the average of the four T threshold similarities, that is, the 0.5 angstrom single threshold similarity, 1 angstrom single threshold similarity, 2 angstrom single threshold similarity, and 4 angstrom single threshold similarity are calculated, and the average of the 0.5 angstrom single threshold similarity, 1 angstrom single threshold similarity, 2 angstrom single threshold similarity, and 4 angstrom single threshold similarity is taken as the similarity between any two categories.

[0125] In step 103, a graph structure construction process is performed based on the spatial structure of the predicted protein structure in each group and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each group.

[0126] By combining the spatial structure of the predicted protein structure and the spatial structure of the target protein sequence to construct a graph structure, the conformation of the predicted protein structure in three-dimensional space is described from multiple angles on the global receptive field of the predicted protein structure, thereby improving the accuracy of protein predicted structure modeling and further improving the accuracy of structural quality prediction.

[0127] See also Figure 5 , Figure 5 is an optional flow chart of the compound processing method based on artificial intelligence provided in the embodiment of the present application, Figure 5 Show Figure 3 Step 103 in the above method can be implemented through steps 1031 to 1034: in step 1031, based on the amino acid residues of the predicted protein structure in each group, the vertices of the graph structure of each predicted protein structure are constructed; in step 1032, based on the spatial structure of each predicted protein structure, the first distance matrix of each predicted protein structure is determined; in step 1033, based on the spatial structure of the target protein sequence, the second distance matrix of the target protein sequence is determined; in step 1034, based on the first distance matrix and the second distance matrix, the edges of the graph structure of the predicted protein structure in each group are constructed.

[0128] For example, after the anti-clustering grouping is completed, a separate graph is constructed for each predicted protein structure in each group (the amino acid sequence is the same, but the 3D conformation is different): the amino acid residues of each predicted protein structure are used as the vertices of the graph structure of the predicted protein structure, and the first distance matrix of each predicted protein structure is determined according to the spatial structure of each predicted protein structure itself. The second distance matrix of the target protein sequence is determined according to the spatial structure of the target protein sequence. The first distance matrix and the second distance matrix are combined to construct the edges of the graph structure of the predicted protein structure in each group.

[0129] In some embodiments, based on the spatial structure of each predicted protein structure, determining a first distance matrix for each predicted protein structure includes: obtaining the Euclidean distance between amino acid residues in each predicted protein structure; and constructing a first distance matrix for each predicted protein structure based on the Euclidean distance between amino acid residues.

[0130] For example, the first distance matrix of the predicted protein structure is extracted based on the structure of the predicted protein structure itself. The amino acid sequence of the predicted protein structure is the same as the target protein sequence, and the Euclidean distances between the amino acid residues in the predicted protein structure are determined. The first distance matrix of the predicted protein structure is constructed based on the Euclidean distances between the amino acid residues. For example, if the amino acid sequence of the predicted protein structure is L in length, the Euclidean distances between the amino acid residues in the predicted protein structure can be calculated, thereby obtaining an L×L distance matrix.

[0131] In some embodiments, based on the spatial structure of the target protein sequence, a second distance matrix of the target protein sequence is determined, including: calling a protein structure prediction model based on the target protein sequence to perform the following processing: performing multi-data source fusion processing on the target protein sequence to obtain co-evolution information of the target protein sequence; performing deep cross-attention processing on the co-evolution information of the target protein sequence to obtain the second distance matrix of the target protein sequence.

[0132] For example, the protein structure prediction model is tFold, which uses the Multi-Source Fusion technology to perform multiple sets of multiple sequence alignments (MSA, Multiple Sequence Alignment) to obtain the co-evolution information of the target protein sequence. Then, with the help of the Deep Cross-Attention Residual Network, the co-evolution information of the target protein sequence is processed by deep cross-attention to obtain the second distance matrix of the target protein sequence (i.e., the residue pair distance and orientation matrix (Rresidue-Residue Distance and Orientation Matrix)). The second distance matrix is ​​used to characterize the Euclidean distance between two amino acid residues in the target protein sequence. tFold can accurately extract the distance matrix of the target protein sequence.

[0133] In some embodiments, based on the spatial structure of the target protein sequence, determining a second distance matrix of the target protein sequence includes: performing a template search process based on the spatial structure of the target protein sequence to obtain a template sequence of the target protein sequence; aligning the target protein sequence and the template sequence to obtain an aligned template sequence; and constructing a second distance matrix of the target protein sequence based on the Euclidean distance between the amino acid residues in the aligned template sequence and the amino acid residues in the target protein sequence.

[0134] For example, the second distance matrix of the target protein sequence is obtained from the template sequence (protein sequence in the protein structure database). Assuming that the length of the target protein sequence is L, 5 template sequences are searched at the same time (i.e., the template search process is performed based on the spatial structure of the target protein sequence to obtain multiple template sequences of the target protein sequence), and the number of these 5 template sequences that can match the target protein sequence is less than or equal to L (individual positions in the template sequence may not match), the Euclidean distances between the L amino acid residues are extracted from the structure of each template sequence, and these Euclidean distances can construct an L×L distance matrix (i.e., the second distance matrix of the target protein sequence). For the case of multiple template sequences, the template sequence with the highest confidence is used as the standard, that is, the distance matrix corresponding to the template sequence with the highest confidence is used.

[0135] In some embodiments, based on the first distance matrix and the second distance matrix, constructing the edge of the graph structure of the protein prediction structure in each group, including: performing the following processing on any two vertices in the protein prediction structure in each group: determining the first Euclidean distance between the two vertices from the first distance matrix; determining the second Euclidean distance between the two vertices from the second distance matrix; performing intersection logic processing on the first Euclidean distance and the second Euclidean distance to obtain the edge of the two vertices.

[0136] For example, the following processing is performed for any two vertices in the predicted protein structure: the first Euclidean distance between the two vertices is determined from the first distance matrix, the second Euclidean distance between the two vertices is determined from the second distance matrix, and the first Euclidean distance and the second Euclidean distance are subjected to intersection logic processing. If the two vertices still meet the edge requirements after the intersection logic processing, an edge between the two vertices is constructed.

[0137] As an example, combined with the intersection logic formula Construct the edge between vertices i and j, C represents the β carbon atom coordinate of the amino acid residue, template represents that the corresponding edge is extracted from the first distance matrix obtained based on the template sequence, prediction represents that the corresponding edge is extracted from the second distance matrix, and decoy represents that the corresponding edge is extracted from the first distance matrix obtained based on the predicted protein structure itself. D is a hyperparameter, and D can be taken as required (for example, 8 angstroms). The embodiments of the present application are not limited to the above intersection logic formula.

[0138] See also Figure 6 , Figure 6 is an optional flow chart of the compound processing method based on artificial intelligence provided in the embodiment of the present application, Figure 6 Show Figure 3 The method further includes steps 105 to 108: in step 105, the similarity between each predicted protein structure and other predicted protein structures in the same group is determined; in step 106, the vertex features of the vertices in the graph structure are determined based on the average value of the similarities and the energy of the vertices in the graph structure; in step 107, the edges of the graph structure are discretized based on the first distance matrix and the second distance matrix to obtain discretized edges; in step 108, the discretized edges are Gaussian distributed to obtain edge features of the edges.

[0139] For example, the lDDT value between each protein predicted structure and other protein predicted structures in the same group is used as the similarity between each protein predicted structure and other protein predicted structures in the same group, and the average lDDT value is used as the consensus feature (i.e., a vertex feature) of the vertex in each protein predicted structure. The first distance matrix is ​​used as a constraint term to calculate the Rosetta (an energy function method used to evaluate whether a protein structure is stable) energy of each vertex, and the Rosetta energy of the vertex is used as a vertex feature.

[0140] For edge feature extraction, the first distance matrix and the second distance matrix are referenced for feature discretization. For example, the edge distance is discretized and divided into 37 categories, where the 2nd to 37th categories represent distances between 2 and 20 angstroms, each 0.5 angstroms is a category, and the single 1st category represents other distances. According to the first distance matrix and the second distance matrix, the probability of each distance classification under the Gaussian distribution can be calculated respectively. Because the first distance matrix and the second distance matrix are involved, the probability of the distance classification under the Gaussian distribution in each distance matrix is ​​averaged to obtain discretized edge features. The embodiment of the present application does not directly use specific numerical values ​​as features, but converts specific numerical values ​​into probability distributions that obey Gauss (feature discretization), thereby solving the situation where there may be errors in the numerical values ​​extracted for individual features.

[0141] In step 104, a structural quality prediction process is performed based on the graph structure of each predicted protein structure to obtain a structural quality index of each predicted protein structure.

[0142] For example, by combining the spatial structure of the predicted protein structure and the spatial structure of the target protein sequence to construct a graph structure, the accuracy of the graph structure can be improved, thereby improving the accuracy of the structural quality prediction, thereby greatly improving the evaluation accuracy of the predicted protein structure through graph learning.

[0143] In some embodiments, a structural quality prediction process is performed based on the graph structure of each predicted protein structure to obtain a structural quality index for each predicted protein structure, including: fusing vertex features and edge features of vertices in the graph structure of each predicted protein structure to obtain fused features of the vertices; performing vertex update processing on the fused features to obtain updated features of the vertices; and performing nonlinear mass mapping processing on the updated features of the vertices to obtain structural quality indexes for the vertices in each predicted protein structure.

[0144] For example, based on the graph structure of each protein prediction structure, a message passing neural network (MPNN) is called to perform structural quality prediction (i.e., quality assessment). The forward pass process of MPNN consists of two stages: the message passing stage and the readout stage:

[0145] For the message passing phase, the message function is defined as M t , the vertex update function is U t , t is the current time of the run, and the update process is: in, represents the intermediate information (i.e., fusion features) received by vertex v at time t+1, N(v) represents all adjacent vertices of vertex v, represents the feature vector of vertex v at time t, e vw represents the edge feature between vertices v and w, Represents the updated features of vertex v at time t+1.

[0146] The readout stage uses the read function R to calculate the feature vector of the entire image in, represents the structural quality index of the vertices in the predicted protein structure (i.e., the output of MPNN), and R is the read function.

[0147] The calculation formula of the message function is: Among them, (.,.) represents splicing.

[0148] The calculation formula of the vertex update function is: Among them, σ represents the activation function (sigmoid), deg(v) represents the degree of vertex v, represents a learnable matrix and t represents the time step.

[0149] The read function R converts the states of all hidden layers into Make a connection Among them, f represents a neural network, W t Represents a learnable matrix of readings.

[0150] Finally, MPNN outputs the lDDT value of each amino acid residue in the predicted protein structure (the structural quality index of the vertex in the predicted protein structure), the label data trained by MPNN is the lDDT value of each amino acid residue obtained by comparing the predicted protein structure with the true structure, and the loss function of MPNN is the average absolute error (MAE) between the prediction result of MPNN and the label + the ranking loss function (MarginRankingLoss). It should be noted that the embodiments of the present application are not limited to MPNN.

[0151] The following is an explanation of an exemplary application of the embodiments of the present application in a practical application scenario.

[0152] The embodiments of the present application can be applied to various protein scenarios, for example, in the selection of protein prediction structure models. In the protein tertiary structure prediction process, a certain degree of randomness is introduced into the algorithm process, so that the generated protein prediction structure model contains more possible conformations, and finally a protein prediction structure candidate pool is obtained. Here, the protein prediction structure model is evaluated, and each candidate structure model can be scored, and the model with the highest score is the model with the best structure; it is applied to assist the training of protein structure prediction algorithms under unlabeled data sets. Training the protein structure prediction algorithm model requires real-time knowledge of the accuracy of each prediction output. When the experimental analysis structure of the protein can be obtained, the prediction result can be directly compared with the experimental analysis structure to obtain the prediction accuracy. However, in the entire protein structure database, the number of proteins whose structures have been obtained through experimental analysis is less than 200,000, and the structures of most other protein sequences are unknown. Reasonable use of the scoring model allows the protein structure prediction algorithm to be trained under a larger unlabeled data set.

[0153] Exploring how proteins fold from amino acid sequences into three-dimensional structures is of great significance to modern biology, medicine, and pharmaceuticals. However, when faced with protein sequences without experimental structural analysis, it is impossible to determine whether the structure folded by protein structure prediction technology meets expectations. Therefore, how to evaluate whether a protein prediction structure model is reliable without labels is a relatively important step.

[0154] The embodiment of the present application provides a protein structure prediction quality assessment method based on a message passing network (i.e., an artificial intelligence-based compound processing method). In terms of feature processing, the structure prediction model pool (including multiple prediction structure models) is clustered and grouped in advance to maximize the structural differences in each group, thereby solving the problem of excessive time overhead caused by considering all the structure model pools. In view of the fact that there may be errors in the feature extraction values ​​themselves, the embodiment of the present application does not directly use specific values ​​as features, but converts the specific values ​​into probability distributions that obey Gaussian. In terms of the global receptive field of the predicted structure model, the embodiment of the present application simultaneously refers to the information of multiple distance matrices to describe the conformation of the predicted structure model in three-dimensional space from multiple angles.

[0155] Combine the following Figure 7 The protein structure prediction quality assessment method based on the message passing network provided in the embodiment of the present application is specifically described as follows:

[0156] First, generate a training dataset.

[0157] Collect all protein sequences containing structural tags in the protein structure database (PDB) and screen the protein sequences, that is, filter the protein sequences with a similarity greater than 60%. Use a protein sequence search module (such as hhblits) for the remaining protein sequences, and perform a template search on PDB70 (a collection of protein sequences with a similarity greater than 70%). If the similarity between the searched template and the protein sequence is greater than 60%, the protein sequence is also filtered. Finally, use a protein structure prediction tool (such as tFold) for all filtered protein sequences to perform structure prediction, and generate 150 candidate structure models (i.e., Figure 7 total structural model pool).

[0158] It should be noted that HH-suite is an open source software package for protein sequence search. HH-suite contains programs that can search for similar protein sequences in protein sequence databases. Sequence search is a standard tool in modern biology. Through HH-suite, the function of unknown proteins can be inferred from the functions of proteins with similar sequences. hhsearch and hhblits are the two main programs in HH-suite and are also the entry points for its search function. hhblits has a faster iteration speed than hhsearch.

[0159] The sequence search process of hhblits is as follows: after each multiple sequence alignment (sequence alignment of multiple sequences so that the residues at each position are the same or similar, there is a one-to-many relationship at each position, and there are multiple similarity scores at each position), a position-specific matrix (PSSM) is established (representing the similarity between the target sequence and other sequences in the multiple sequences at each position), and then this position-specific matrix is ​​used to search in the protein structure database, and the PSSM is updated in combination with the statistical significance of the search results. The cycle is repeated to finally obtain the final search results.

[0160] It should be noted that tFold is a protein structure prediction algorithm. The algorithm uses multi-source fusion technology to mine co-evolutionary information in multiple sets of multiple sequence alignments (MSA); then, with the help of the Deep Cross-Attention Residual Network, it can greatly improve the prediction accuracy of some important protein 2D structural information (such as the residue-residue distance and orientation matrix); finally, through a template-assisted free modeling (TBFM, Template-based Free Modeling) method, the structural information in the 3D structural model generated by free modeling (FM, FreeModeling) and template modeling (TBM, Template-based Modeling) is effectively fused, thereby greatly improving the accuracy of the final 3D modeling.

[0161] Then, the structural model pool is clustered for preprocessing.

[0162] First, the 150 protein prediction structure models generated from the same protein sequence are clustered and grouped. Each group contains a maximum of 15 prediction structure models. The ultimate goal of grouping is to make the prediction structure models in the same group as dissimilar as possible. The clustering process is as follows:

[0163] 1) If there are N prediction structural models in the same sequence, set the maximum number of structural models in each group M = ceil (N / 15), where ceil means rounding up, and initialize each prediction structural model as a single category.

[0164] 2) If there are k categories in total, calculate the similarity between any two categories, a total of k*(k-1) / 2 times.

[0165] The similarity calculation between category A and category B is shown in formula (1):

[0166]

[0167] Among them, i and j are the index subscripts of the predicted structural models under categories A and B respectively. lDDT is an indicator for calculating the similarity between the predicted protein structural model and the label. Here, lDDT is used to calculate the similarity between the two predicted structural models.

[0168] It should be noted that the calculation process of lDDT is as follows: for the predicted structure b and the real structure a. Extract the atom pairs that meet the following conditions in a: 1. The two atoms do not belong to the same residue; 2. The Euclidean distance between the two atoms is less than the threshold R. Set the atom pairs that meet the above conditions as set L. Query the atom pairs that are the same as set L in b. If the error of the Euclidean distance of the queried atom pair compared with the atom pair in L is within the threshold range T (T is 0.5 angstroms, 1 angstroms, 2 angstroms, 4 angstroms), then retain the atom pair, otherwise remove the atom pair from R, and obtain a new set M after all atom pairs in L are verified. The lDDT single threshold similarity is M / L, and the total lDDT value is the average of the 4 T threshold similarities.

[0169] 3) Obtain the calculation results of k*(k-1) / 2 times of category similarity, merge the two categories corresponding to the highest category similarity, and when the number of structural models after the merger is not greater than M, merge the two categories and continue to step 2); if the number of structural models after the merger is greater than M, merge the two categories with the second highest similarity, and then determine whether the number of structural models after the merger is greater than M. If all are traversed and still cannot be merged, proceed to step 4).

[0170] 4) If there are K categories left after the merging is completed, the prediction structural models of the K categories are traversed in turn, and the first prediction structural model of the currently traversed category is separated into a new set each time. The new set obtained by traversing once contains at most 15 prediction structural models, until all the structural models of the K categories belong to the new set to achieve clustering and grouping, so that the prediction structural models of the same set are as dissimilar as possible.

[0171] After clustering is completed, a separate map is built for each predicted structural model (same amino acid sequence but different 3D conformation):

[0172] First, use hhblits to parse the target sequence ( Figure 7 The template search is performed using the input sequence in the template, and the template sequences with the top 5 template scores are extracted. According to the multiple sequence alignment (MSA, Multiple Sequence Alignment) of the template sequences, the distance matrix (first distance matrix) of the β carbon (GLY uses α carbon) at the corresponding residue position in the template sequence structure is obtained. At the same time, the distance matrix (third distance matrix) of the predicted structural model is extracted, and the distance matrix (fourth distance matrix) of the target sequence is predicted using the structure prediction algorithm tFold. The distance matrix (first distance matrix) in each group of predicted structural models and the distance matrix (fourth distance matrix) of the predicted target sequence are the same. Among them, the second distance matrix includes at least one of the third distance matrix and the fourth distance matrix, and the consensus feature of each predicted structural model is the average value of the lDDT value between the predicted structural model and all other predicted structural models in the group.

[0173] It should be noted that the distance matrix of the β carbon at the corresponding residue position is obtained from the template sequence. Figure 7 The length of the input sequence is L. At the same time, assume that 5 template sequences are searched, and the number of these 5 template sequences that can match the input sequence is less than or equal to L (some positions in the template sequence may not match). The Euclidean distances between the L residues are extracted from the structure of each template sequence, and these Euclidean distances can be used to construct an L×L distance matrix. In the case of multiple template sequences, the template sequence with the highest confidence will prevail.

[0174] It should be noted that the distance matrix of the predicted structural model is extracted based on the structure of the predicted structural model itself. Figure 7 The input sequence is the same. When the sequence length is L, the Euclidean distance between each pair of residues in the predicted structural model to be evaluated can be calculated to obtain an L×L distance matrix.

[0175] Taking amino acid residues as vertices, the edge set is combined with the above three distance matrices to construct a graph. The calculation formula is shown in formula (2):

[0176]

[0177] Where i and j represent the index of the amino acid residue, C represents the β-carbon atom coordinate of the amino acid residue, and v i Represents vertex i, template indicates that the corresponding edge is extracted from the template distance matrix (i.e., the first distance matrix), prediction indicates that the corresponding edge is extracted from the self-tfold predicted distance matrix (i.e., the fourth distance matrix), and decoy indicates that the corresponding edge is extracted from the third distance matrix. D is a hyperparameter, and D can take a value of 8 angstroms.

[0178] During the graph construction process, the vertex features and edge features of the graph are extracted. For vertex features, the predicted distance matrix is ​​used as a constraint to calculate the Rosetta (an energy function method used to evaluate whether a protein structure is stable) energy of each residue. For edge features, the distance matrix predicted by the reference prediction algorithm (the fourth distance matrix), the distance matrix calculated based on the template structure (i.e., the template sequence) (the first distance matrix), and the distance matrix calculated based on the predicted structure model itself (the third distance matrix) are also discretized. First, the distance needs to be discretized and divided into 37 categories, of which categories 2 to 37 represent distances between 2 and 20 angstroms, with each 0.5 angstrom being a category, and the single category 1 represents other distances. Based on the distance matrix, the probability of each distance classification under the Gaussian distribution can be calculated. Since a total of three distance matrices are involved, the result needs to average the probability of each distance classification under the Gaussian distribution.

[0179] like Figure 7 As shown in the figure, after the features are discretized, the message passing neural network (MPNN) is used for quality assessment, where the message passing step size is set to 8, the hidden layer dimension of the vertex and edge is 96, and the data label is the local lDDT of each residue in the predicted structure. When calculating the loss, in addition to calculating the absolute average error of the local lDDT, the ranking loss function (MarginRankingLoss) is also required.

[0180] The input of MPNN is a definition graph G(v,e), and the feature vector of vertex v is x v (including Rosetta energy and consensus features), edge e vw Linking vertices v and w, the edge features include distance discretization features, and the forward transmission process of MPNN includes two stages: message passing stage and readout stage.

[0181] For the message passing phase, the message function is defined as M t , the vertex update function is U t , t is the current time of the operation, and the update process is as shown in formula (3):

[0182]

[0183] in, represents the information received by vertex v at time t+1, N(v) represents all adjacent vertices of vertex v, represents the feature vector of vertex v at time t, e vw Represents the edge features between vertices v and w.

[0184] In the readout stage, the read function R is used to calculate the feature vector of the entire image. The calculation process is shown in formula (4):

[0185]

[0186] in, represents the final output vector (i.e., the output of MPNN), and R is the reading function.

[0187] The calculation formula of the message function is shown in formula (5):

[0188]

[0189] Among them, (.,.) represents splicing.

[0190] The calculation formula of the vertex update function is shown in formula (6):

[0191]

[0192] Among them, σ represents the activation function (sigmoid), deg(v) represents the degree of vertex v, represents a learnable matrix and t represents the time step.

[0193] The read function R converts the states of all hidden layers into Make the connection as shown in formula (7):

[0194]

[0195] Among them, f represents a neural network, W t Represents a learnable matrix of readings.

[0196] Finally, MPNN outputs the lDDT value of each residue in the predicted structural model. The trained label data is the lDDT value of each residue obtained by comparing the predicted structure with the true structure. The loss function is the mean absolute error (MAE) between the predicted result and the label + the ranking loss function (MarginRankingLoss).

[0197] The present application example uses data from the CAMEO platform (protein structure prediction automatic evaluation platform) to evaluate the effect of this solution. If the local IDDT score of an amino acid residue is greater than 0.6 during the evaluation process, the quality of the prediction is judged to be good, otherwise it is judged to be poor. The solution proposed in the present application example significantly exceeds other technologies in the main evaluation index ROC-AUC during the evaluation process.

[0198] In summary, the embodiment of the present application pre-clustered and grouped the structural model pool, discretized the numerical features, and then integrated the structural information of the protein amino acid residues into a graph with reference to the multi-scale spatial structure information, and finally greatly improved the evaluation accuracy of the protein prediction structure model through graph learning.

[0199] So far, the compound processing method based on artificial intelligence provided by the embodiment of the present application has been described in combination with the exemplary application and implementation of the server provided by the embodiment of the present application. The embodiment of the present application also provides a compound processing device. In actual applications, each functional module in the compound processing device can be implemented by hardware resources of electronic devices (such as terminal devices, servers or server clusters), such as computing resources such as processors, communication resources (such as for supporting various communication methods such as optical cables and cellular), and memory. Figure 2 A compound processing device 555 stored in a memory 550 is shown, which may be software in the form of programs and plug-ins, for example, software modules designed in programming languages ​​such as C / C++ and Java, application software designed in programming languages ​​such as C / C++ and Java, or dedicated software modules in large software systems, application programming interfaces, plug-ins, cloud services, and other implementations.

[0200] The compound processing device 555 includes a series of modules, including an acquisition module 5551, an anti-clustering module 5552, a construction module 5553, and a prediction module 5554. The following further describes how the various modules in the compound processing device 555 provided in the embodiment of the present application cooperate to implement the compound processing solution.

[0201] An acquisition module 5551 is used to acquire multiple protein prediction structures of a target protein sequence; an anti-clustering module 5552 is used to perform anti-clustering processing on the multiple protein prediction structures to obtain a group to which each of the protein prediction structures belongs, wherein each of the groups includes multiple different protein prediction structures; a construction module 5553 is used to perform graph structure construction processing based on the spatial structure of the protein prediction structure in each of the groups and the spatial structure of the target protein sequence to obtain a graph structure of the protein prediction structure in each of the groups; a prediction module 5554 is used to perform structure quality prediction processing based on the graph structure of each of the protein prediction structures to obtain a structure quality index for each of the protein prediction structures.

[0202] In some embodiments, the anti-clustering module 5552 is further used to perform clustering processing on the multiple protein prediction structures to obtain the category to which each of the protein prediction structures belongs, wherein the category includes multiple different protein prediction structures; sample a protein prediction structure from each of the categories, and use the set of sampled protein prediction structures as a group.

[0203] In some embodiments, the anti-clustering module 5552 is also used to take one of the protein prediction structures as a category; iteratively perform the following processing: perform similarity processing based on the protein prediction structures included in multiple categories to obtain the similarity between the categories; merge the categories corresponding to the obtained similarities in order from high to low; when the number of protein prediction structures included in the merged category is less than or equal to the quantity threshold, use the merged category as the new category; when the number of protein prediction structures included in the merged category is greater than the quantity threshold, merge the categories corresponding to the next similarity of the similarity in order from high to low, until the number of protein prediction structures included in all the merged categories is greater than the quantity threshold, and stop the iteration.

[0204] In some embodiments, the anti-clustering module 5552 is further used to perform the following processing on any two of the multiple categories: obtain any first protein prediction structure in one of the two categories, and obtain multiple target atom pairs in the first protein prediction structure, wherein the two target atoms in the target atom pairs belong to different residues and the first Euclidean distance between the two target atoms is less than a distance threshold; use the set of the multiple target atom pairs as the initial set; obtain any second protein prediction structure in the other of the two categories, and obtain candidate atom pairs that are the same as the target atom pairs in the second protein prediction structure; when the error between the second Euclidean distance between the two candidate atoms in the candidate atom pairs and the first Euclidean distance is less than the error threshold, delete the target atom pairs corresponding to the candidate atom pairs from the initial set to obtain an updated set; use the ratio of the number of target atom pairs included in the updated set to the number of target atom pairs included in the initial set as the local distance difference between the first protein prediction structure and the second protein prediction structure; and use the average value of the multiple local distance differences as the similarity between any two categories.

[0205] In some embodiments, the construction module 5553 is also used to construct the vertices of the graph structure of each predicted protein structure based on the amino acid residues of the predicted protein structure in each of the groups; determine the first distance matrix of each predicted protein structure based on the spatial structure of each predicted protein structure; determine the second distance matrix of the target protein sequence based on the spatial structure of the target protein sequence; and construct the edges of the graph structure of the predicted protein structure in each of the groups based on the first distance matrix and the second distance matrix.

[0206] In some embodiments, the construction module 5553 is further used to obtain the Euclidean distance between the amino acid residues in each of the predicted protein structures; and construct a first distance matrix for each of the predicted protein structures based on the Euclidean distance between the amino acid residues.

[0207] In some embodiments, the construction module 5553 is also used to call a protein structure prediction model based on the target protein sequence to perform the following processing: performing multi-data source fusion processing on the target protein sequence to obtain the co-evolution information of the target protein sequence; performing deep cross-attention processing on the co-evolution information of the target protein sequence to obtain a second distance matrix of the target protein sequence.

[0208] In some embodiments, the construction module 5553 is also used to perform template search processing based on the spatial structure of the target protein sequence to obtain the template sequence of the target protein sequence; align the target protein sequence and the template sequence to obtain the aligned template sequence; and construct a second distance matrix of the target protein sequence based on the Euclidean distance between the amino acid residues in the aligned template sequence and the amino acid residues in the target protein sequence.

[0209] In some embodiments, the construction module 5553 is also used to perform the following processing on any two vertices in the predicted protein structure in each of the groups: determine the first Euclidean distance between the two vertices from the first distance matrix; determine the second Euclidean distance between the two vertices from the second distance matrix; perform intersection logic processing on the first Euclidean distance and the second Euclidean distance to obtain the edge of the two vertices.

[0210] In some embodiments, the construction module 5553 is also used to determine the similarity between each of the predicted protein structures and other predicted protein structures in the same group; determine the vertex features of the vertices in the graph structure based on the average value of the similarities and the energy of the vertices in the graph structure; perform distance discretization processing on the edges of the graph structure based on the first distance matrix and the second distance matrix to obtain discretized edges; perform Gaussian distribution processing on the discretized edges to obtain edge features of the edges.

[0211] In some embodiments, the prediction module 5554 is also used to fuse the vertex features and edge features of the vertices in the graph structure of each of the protein prediction structures to obtain the fused features of the vertices; perform vertex update processing on the fused features to obtain the updated features of the vertices; perform nonlinear mass mapping processing on the updated features of the vertices to obtain the structural quality index of the vertices in each of the protein prediction structures.

[0212] In some embodiments, the acquisition module 5551 is also used to screen the protein sequences in the first protein structure database based on the similarity between protein sequences to obtain candidate protein sequences; perform template search on the candidate protein sequences based on the template sequences in the second protein structure database to obtain the target protein sequence; perform protein structure generation processing on the target protein sequence to obtain multiple predicted protein structures of the target protein sequence.

[0213] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned artificial intelligence-based compound processing method of the embodiment of the present application.

[0214] The present application embodiment provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will be caused to execute the artificial intelligence-based compound processing method provided by the present application embodiment, for example, Figure 3-Figure 6 An artificial intelligence-based compound processing method is shown.

[0215] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0216] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0217] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0218] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0219] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A compound processing method based on artificial intelligence, characterized in that: The method comprises: Obtain multiple protein predicted structures of the target protein sequence; Performing anti-clustering processing on the multiple predicted protein structures to obtain a group to which each predicted protein structure belongs, wherein each group includes multiple different predicted protein structures; Performing graph structure construction processing based on the spatial structure of the predicted protein structure in each of the groups and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each of the groups; A structural quality prediction process is performed based on the graph structure of each of the predicted protein structures to obtain a structural quality index of each of the predicted protein structures.

2. The method according to claim 1, characterized in that The performing anti-clustering processing on the plurality of predicted protein structures to obtain the group to which each predicted protein structure belongs includes: Performing clustering processing on the multiple predicted protein structures to obtain a category to which each predicted protein structure belongs, wherein the category includes multiple different predicted protein structures; A predicted protein structure is sampled from each of the categories, and a set of the sampled predicted protein structures is used as a group.

3. The method according to claim 2, characterized in that The clustering process is performed on the plurality of predicted protein structures to obtain the category to which each predicted protein structure belongs, including: Taking a predicted protein structure as one of the categories; The following processing is performed iteratively: Performing similarity processing based on predicted protein structures included in the plurality of categories to obtain similarities between the categories; According to the order of the similarities from high to low, the categories corresponding to the obtained similarities are merged; When the number of predicted protein structures included in the merged category is less than or equal to the quantity threshold, taking the merged category as the new category; When the number of predicted protein structures included in the merged categories is greater than the quantity threshold, the categories corresponding to the next similarity of the similarity are merged in descending order of the similarity until the number of predicted protein structures included in all merged categories is greater than the quantity threshold, and the iteration is stopped, wherein the next similarity is a similarity that is adjacent to the similarity in the descending order of the similarity and is less than the similarity.

4. The method according to claim 3, characterized in that: The similarity processing is performed based on the predicted protein structures included in the plurality of categories to obtain the similarity between the categories, including: The following processing is performed for any two of the multiple categories: Obtain any first protein predicted structure in one of any two of the categories, and obtain a plurality of target atom pairs in the first protein predicted structure, wherein two target atoms in the target atom pairs belong to different residues, and a first Euclidean distance between the two target atoms is less than a distance threshold; Taking the set of the plurality of target atom pairs as an initial set; Obtain any second protein predicted structure in another category of any two of the categories, and obtain a candidate atom pair in the second protein predicted structure that is identical to the target atom pair; When the error between the second Euclidean distance between two candidate atoms in the candidate atom pair and the first Euclidean distance is greater than or equal to an error threshold, deleting the target atom pair corresponding to the candidate atom pair from the initial set to obtain an updated set; Using the ratio of the number of target atom pairs included in the updated set to the number of target atom pairs included in the initial set as the local distance difference between the first protein predicted structure and the second protein predicted structure; An average value of the plurality of local distance differences is taken as the similarity between any two of the categories.

5. The method according to claim 1, characterized in that The graph structure construction process is performed based on the spatial structure of the predicted protein structure in each of the groups and the spatial structure of the target protein sequence to obtain the graph structure of the predicted protein structure in each of the groups, including: Based on the amino acid residues of the predicted protein structure in each of the groups, constructing the vertices of the graph structure of each of the predicted protein structures; Determining a first distance matrix for each of the predicted protein structures based on the spatial structure of each of the predicted protein structures; Determining a second distance matrix of the target protein sequence based on the spatial structure of the target protein sequence; Based on the first distance matrix and the second distance matrix, edges of a graph structure of predicted protein structures in each of the groups are constructed.

6. The method according to claim 5, characterized in that Determining a first distance matrix of each predicted protein structure based on the spatial structure of each predicted protein structure comprises: Obtaining the Euclidean distance between the amino acid residues in each predicted structure of the protein; Based on the Euclidean distances between the amino acid residues, a first distance matrix is ​​constructed for each of the predicted protein structures.

7. The method according to claim 5, characterized in that The step of determining a second distance matrix of the target protein sequence based on the spatial structure of the target protein sequence comprises: Based on the target protein sequence, the protein structure prediction model is called to perform the following processing: Performing multi-data source fusion processing on the target protein sequence to obtain co-evolution information of the target protein sequence; Deep cross-attention processing is performed on the co-evolution information of the target protein sequence to obtain a second distance matrix of the target protein sequence.

8. The method according to claim 5, characterized in that The step of determining a second distance matrix of the target protein sequence based on the spatial structure of the target protein sequence comprises: Performing template search processing based on the spatial structure of the target protein sequence to obtain a template sequence of the target protein sequence; Aligning the target protein sequence and the template sequence to obtain the aligned template sequence; Based on the aligned Euclidean distances between the amino acid residues in the template sequence and the amino acid residues in the target protein sequence, a second distance matrix of the target protein sequence is constructed.

9. The method according to claim 5, characterized in that The step of constructing edges of a graph structure of predicted protein structures in each of the groups based on the first distance matrix and the second distance matrix comprises: The following processing is performed for any two vertices in the predicted protein structure in each of the groups: Determining a third Euclidean distance between the two vertices from the first distance matrix; Determining a fourth Euclidean distance between the two vertices from the second distance matrix; An intersection logic process is performed on the third Euclidean distance and the fourth Euclidean distance to obtain an edge between the two vertices.

10. The method according to claim 5, characterized in that The method further comprises: Determining the similarity between each of the predicted protein structures and other predicted protein structures in the same group; Determining vertex features of vertices in the graph structure based on an average value of the similarities and energies of the vertices in the graph structure; Performing distance discretization processing on the edges of the graph structure based on the first distance matrix and the second distance matrix to obtain discretized edges; Gaussian distribution processing is performed on the discretized edges to obtain edge features of the edges.

11. The method according to claim 1, characterized in that: The step of performing structural quality prediction processing based on the graph structure of each of the predicted protein structures to obtain a structural quality index of each of the predicted protein structures comprises: fusing vertex features and edge features of vertices in the graph structure of each protein prediction structure to obtain fused features of the vertices; Performing vertex update processing on the fused features to obtain updated features of the vertices; The updated features of the vertices are subjected to nonlinear quality mapping processing to obtain a structural quality index of each vertex in the predicted protein structure.

12. The method according to claim 1, characterized in that The step of obtaining a plurality of predicted protein structures of a target protein sequence comprises: Based on the similarity between protein sequences, protein sequences in the first protein structure database are screened to obtain candidate protein sequences; Based on the template sequence in the second protein structure database, a template search is performed on the candidate protein sequence to obtain the target protein sequence; The target protein sequence is processed for protein structure generation to obtain a plurality of predicted protein structures of the target protein sequence.

13. A compound processing device based on artificial intelligence, characterized in that: The device comprises: An acquisition module, used to obtain multiple protein predicted structures of a target protein sequence; an anti-clustering module, used for performing anti-clustering processing on the plurality of predicted protein structures to obtain a group to which each predicted protein structure belongs, wherein each group includes a plurality of different predicted protein structures; A construction module, used for performing graph structure construction processing based on the spatial structure of the predicted protein structure in each of the groups and the spatial structure of the target protein sequence, to obtain the graph structure of the predicted protein structure in each of the groups; The prediction module is used to perform structure quality prediction processing based on the graph structure of each of the predicted protein structures to obtain a structure quality index for each of the predicted protein structures.

14. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the artificial intelligence-based compound processing method according to any one of claims 1 to 12 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium, characterized in that: Executable instructions are stored, which are used to implement the artificial intelligence-based compound processing method described in any one of claims 1 to 12 when executed by a processor.

16. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the artificial intelligence-based compound processing method according to any one of claims 1 to 12 is implemented.

Images