Method and apparatus for generating protein binders, electronic device, and storage medium

Abstract

The present disclosure provides a protein binder generation method and device, an electronic device and a storage medium. The protein binder generation method comprises: obtaining a first protein structure of a plurality of candidate protein binders generated by a target protein; performing helix structure detection on the first protein structure and filtering invalid structures; calling a sequence design component to generate a first protein sequence based on the filtered first protein structure; calling a screening component to perform consistency detection on the first protein sequence and the filtered first protein structure based on a second protein structure corresponding to the target protein, a second protein sequence and first homology information, and determining a target protein binder according to the consistency detection result. The embodiments of the present application can improve the generation efficiency of the protein binder.

Description

Technical Field

[0001] This disclosure relates to the field of bioinformatics, specifically to a method and apparatus for generating protein conjugates, electronic devices, and storage media. Background Technology

[0002] De novo protein design refers to the process of designing protein structures with specific functions and properties from scratch using computational methods and experimental techniques. In recent years, the field of de novo protein design has significantly improved the efficiency of protein design thanks to the rapid development of deep learning and artificial intelligence technologies.

[0003] Although the designed proteins exhibit excellent theoretical performance, their design efficiency in laboratory validation remains low. Therefore, relevant technologies can use screeners to select the most promising protein conjugates from a large number of candidate designs, significantly improving the efficiency and success rate of experimental validation and thus accelerating the practical application of de novo protein conjugate design.

[0004] However, this filter-dependent protein conjugate design process still has low efficiency in protein conjugate generation. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, electronic device, and storage medium for generating protein conjugates, aiming to improve the efficiency of protein conjugate generation.

[0006] To achieve the above objectives, this application proposes a method for generating protein conjugates, applicable to a protein conjugate generation apparatus, the apparatus comprising a sequence design component and a screening component, including:

[0007] Obtain the first protein structure of multiple candidate protein conjugates generated from the target protein;

[0008] The structure of the first protein is analyzed by helical structure detection to filter out invalid structures;

[0009] The sequence design component is invoked to generate a first protein sequence based on the filtered first protein structure.

[0010] The screening component is invoked to perform a consistency test on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence, and the first homology information corresponding to the target protein, and the target protein conjugate is determined based on the consistency test results.

[0011] In some embodiments, the step of detecting the helical structure of the first protein and filtering out invalid structures includes:

[0012] The first protein structure is subjected to helical structure detection, and the first protein structure that includes only a single helical structure or a double helical structure is regarded as an invalid structure and the invalid structure is filtered out.

[0013] In some embodiments, the step of performing a consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence, and the first homology information corresponding to the target protein, and determining the target protein conjugate based on the consistency detection result, includes:

[0014] A complex simulation structure is generated based on the filtered first protein structure and the second protein structure corresponding to the target protein.

[0015] Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex, a predicted structure and prediction confidence of the complex are generated.

[0016] The difference information between the simulated structure of the complex and the predicted structure of the complex is determined, and the target protein conjugate is determined from the plurality of candidate protein conjugates based on the prediction confidence and the difference information.

[0017] In some embodiments, determining the difference information between the simulated structure of the complex and the predicted structure of the complex includes:

[0018] Determine the first difference between the filtered first protein structure and the first predicted structure corresponding to the candidate protein conjugate in the predicted structure of the complex;

[0019] Determine a second difference between the second protein structure and the second predicted structure corresponding to the target protein in the predicted structure of the complex;

[0020] A third difference is determined between the overall simulated structure of the complex and the predicted structure of the complex, and the difference information is determined based on the first difference, the second difference, and the third difference.

[0021] In some embodiments, generating a complex-like structure based on the filtered first protein structure and the second protein structure corresponding to the target protein includes:

[0022] The backbone information in the filtered first protein structure is encoded to obtain a first coordinate matrix, and the backbone information in the second protein structure corresponding to the target protein is encoded to obtain a second coordinate matrix.

[0023] The third coordinate matrix corresponding to the simulated structure of the complex is determined based on the first coordinate matrix and the second coordinate matrix;

[0024] The step of generating a predicted complex structure and prediction confidence based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex includes:

[0025] Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the third coordinate matrix, a predicted structure and prediction confidence of the complex are generated.

[0026] In some embodiments, generating a predicted complex structure and prediction confidence based on the first protein sequence, the second protein sequence corresponding to the target protein, first homology information, and the simulated complex structure includes:

[0027] Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex, a predicted structure of the complex is generated.

[0028] The two-dimensional matrix is ​​divided into blocks based on the regional positions of the candidate protein conjugate and the target protein in the two-dimensional matrix of the predicted structure of the complex.

[0029] Based on the block division results, the overlap matrix of the candidate protein conjugate and the target protein in the two-dimensional matrix is ​​determined, and the prediction confidence is determined based on the value of the overlap matrix.

[0030] Another aspect of this application discloses an apparatus for generating protein conjugates, the apparatus comprising:

[0031] The structure filtering unit is used to obtain the first protein structure of multiple candidate protein conjugates generated from the target protein, and to perform helical structure detection on the first protein structure to filter out invalid structures.

[0032] Sequence design component for generating a first protein sequence based on the filtered first protein structure;

[0033] A screening component is used to perform consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence and the first homology information corresponding to the target protein, and to determine the target protein conjugate based on the consistency detection results.

[0034] Another aspect of this application proposes an electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method described in the first aspect.

[0035] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0036] Another aspect of this application proposes a computer program product comprising a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the methods described above.

[0037] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0038] When generating protein conjugates, the first protein structures of multiple candidate protein conjugates generated from the target protein are first obtained. Then, helical structure detection is performed on each first protein structure to filter out invalid structures. Further, a sequence design component is invoked to generate a corresponding first protein sequence based on the filtered first protein structures. Further, a screening component is invoked to perform a consistency check on the first protein sequence and the filtered first protein structures based on the second protein structure corresponding to the target protein, the second protein sequence, and first homology information, and the target protein conjugate is determined based on the consistency check results. Since this application can perform a first-layer screening of candidate protein conjugates based on the helical structure detection results corresponding to the first protein structure before sequence generation, filtering out unacceptable first protein structures, these filtered first protein structures will not be used for sequence generation, thereby reducing the amount of unnecessary computation in the sequence generation process. Furthermore, this application can perform a consistency check on the second protein sequence and the filtered first protein structure based on the screening component, and determine the target protein conjugate based on the consistency check results. That is, this application can effectively improve the generation efficiency of protein conjugates through a two-layer filtering process constructed by helical structure detection and consistency detection. Attached Figure Description

[0039] The accompanying drawings are provided to further understand the technical solutions of this disclosure and constitute a part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure.

[0040] Figure 1 This is a flowchart of a method for generating protein conjugates provided in an embodiment of this application;

[0041] Figure 2This is a flowchart illustrating the determination of a preset invalid spiral structure type provided in an embodiment of this application;

[0042] Figure 3 yes Figure 1 A flowchart of an embodiment of step 104 in the middle;

[0043] Figure 4 yes Figure 4 A flowchart of an embodiment of step 303 in the middle;

[0044] Figure 5 yes Figure 5 A flowchart of an embodiment of step 401;

[0045] Figure 6 yes Figure 4 A flowchart of an embodiment of steps 301 and 302;

[0046] Figure 7 yes Figure 7 A flowchart of an embodiment of step 602 in the middle;

[0047] Figure 8 yes Figure 4 Another embodiment of step 302 in the flowchart;

[0048] Figure 9 yes Figure 4 A flowchart of an embodiment of step 303 in the middle;

[0049] Figure 10 This is a schematic diagram of a method for generating protein conjugates provided in an embodiment of this application;

[0050] Figure 11 This is a schematic diagram of a predicted structure of a complex provided in an embodiment of this application;

[0051] Figure 12 This is a schematic diagram of a protein conjugate generation apparatus provided in an embodiment of this application;

[0052] Figure 13 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0055] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0056] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0057] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0059] First, let's analyze some of the terms used in this application:

[0060] De novo protein design refers to the process of carefully arranging the amino acid sequence of a protein to enable it to spontaneously fold into a desired three-dimensional structure and acquire a specific function. The core of this technology lies in overcoming the limitations of natural proteins, building and designing proteins with entirely new structures and functions from scratch, creating protein structures and functions that have never existed in nature. For example, scientists can use de novo protein design to develop novel enzymes that can catalyze specific chemical reactions for use in drug synthesis, biofuel production, or other industrial processes.

[0061] Protein conjugates are small molecules, peptides, or proteins themselves that can bind to specific proteins. This binding can be non-covalent (e.g., hydrogen bonds, hydrophobic interactions, van der Waals forces) or covalent (e.g., covalent bonds formed through chemical reactions). Protein conjugates play crucial roles in many biological processes, such as signal transduction, enzyme catalysis, immune responses, regulation, and drug design. In drug development, designing small molecules or peptides that can specifically bind to target proteins to block or activate their function is a vital strategy for treating diseases. De novo design of protein conjugates refers to designing new proteins that can specifically bind to the target protein, based on its structural and functional requirements.

[0062] Wet experiments refer to experiments conducted in a laboratory by physically manipulating liquids, biological samples, or chemical reagents. This contrasts with "dry experiments," which primarily rely on computer simulations and data analysis.

[0063] The field of de novo protein design has significantly improved efficiency thanks to the rapid development of deep learning and artificial intelligence. Researchers, leveraging increasingly rich protein databases and fully utilizing existing sequence and structural information from nature, have successfully designed novel proteins with specific functions. From initial stability optimization to the realization of complex functions (such as enzyme activity and signal transduction), this field is gradually moving towards multifunctionality. Furthermore, the application of high-throughput screening technologies and automated experimental equipment has greatly shortened the design and validation cycle, while enhancing the feedback mechanism between computation and experiment. Therefore, de novo design technology for protein conjugates has made significant progress in computational methods, experimental techniques, and design strategies in recent years. Relying on molecular dynamics simulations, machine learning, and high-throughput screening, the design of highly specific and high-affinity conjugates has become possible. These technologies have been widely applied in the development of antiviral and anticancer drugs, synthetic biosensors, and enzyme inhibitors. Despite challenges such as design accuracy, experimental validation efficiency, and immune safety, the field has a very broad application prospect in life sciences and medicine due to multidisciplinary collaboration and technological iteration. Although computationally designed proteins exhibit excellent theoretical performance, their success rate in wet experiments (laboratory validation) remains low. Therefore, before conducting wet experiments on de novo designed proteins, screening of de novo designed proteins can be performed, which can greatly improve the efficiency and success rate of experimental verification, thereby accelerating the practical application of de novo designed proteins.

[0064] Currently, Rosetta, a screening tool based on physical energy functions, can predict the energy state and folding probability of proteins by calculating physical energy functions and using intermolecular interaction energies (such as hydrogen bonds and van der Waals forces) and hydrophobic effects, thereby assessing the folding stability and interaction relationships of de novo designed proteins. However, under complex environments (such as solution effects, pH, and temperature), its energy function's prediction accuracy for the folding probability of de novo designed proteins is insufficient, often leading to significant deviations from actual experimental results. It also incurs extremely high computational costs and has a very low success rate in screening proteins under wet experimental conditions. Furthermore, screening tools combining molecular dynamics simulations, such as GROMACS and AMBER, can simulate conformational changes and stability of proteins under dynamic environments, capturing the dynamic behavior of proteins at the nanosecond to microsecond scale, thereby analyzing the stability of de novo designed proteins and their interaction relationships with target proteins. However, screening tools combining molecular dynamics simulations are computationally very expensive, making it difficult to efficiently process large-scale candidate protein libraries, and their simulation timescales are limited (usually at the nanosecond or microsecond level), potentially failing to capture important interaction conformations between de novo designed proteins and target proteins. Furthermore, statistical and data-driven screeners like ProGen and dl_binder_design can leverage deep learning models to capture potential sequence-structure-function relationships from massive datasets. By combining large-scale statistical associations between the sequence, structure, and function of the target protein, they can predict the folding probability of de novo designed proteins and their interactions with the target protein. However, statistical and data-driven screeners typically use protein monomer structure prediction models as their foundation. When screening large-scale candidate protein libraries, they may not accurately predict the interactions between de novo designed proteins and the target protein, leading to too many false positives. Since protein conjugates can be proteins themselves, the aforementioned screener-dependent protein conjugate design processes still exhibit relatively low efficiency in generating protein conjugates.

[0065] To address the aforementioned issues, this application discloses a method for generating protein conjugates, which can effectively improve the efficiency of protein conjugate generation.

[0066] The method for generating protein conjugates provided in the embodiments of this application will be described below.

[0067] Please see Figure 1 In some embodiments, the method for generating protein conjugates provided in this application includes, but is not limited to, steps 101 to 104.

[0068] Step 101: Obtain the first protein structure of multiple candidate protein conjugates generated from the target protein;

[0069] Step 102: Perform helical structure detection on the first protein structure and filter out invalid structures;

[0070] Step 103: Invoke the sequence design component to generate the first protein sequence based on the filtered first protein structure;

[0071] Step 104: Call the screening component to perform consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence and the first homology information corresponding to the target protein, and determine the target protein conjugate based on the consistency detection results.

[0072] Steps 101 to 104, as illustrated in this embodiment, generate first protein structures of multiple candidate protein conjugates from the target protein. Since some of the generated first protein structures do not meet experimental expression requirements, invalid structures can be filtered out based on the helical structure detection results corresponding to each first protein structure. This structure filtering avoids the time-consuming design process of sequence generation for protein conjugate structures that do not meet the requirements; that is, it first screens out conjugate structures that are compatible with successful wet experiments, thereby improving the generation efficiency of the de novo protein conjugate design process. Subsequently, this application can call the sequence design component to generate the first protein sequence corresponding to the first protein structure, and use the screening component to determine the consistency detection results between each filtered first protein structure and the corresponding first protein sequence. Therefore, the target protein conjugate most likely to succeed in wet experiments can be screened based on the consistency detection results.

[0073] In step 101 of some embodiments, the target protein can be a specific protein to be bound that interacts with the de novo designed protein, i.e., a protein (such as an enzyme, receptor, etc.) to which it needs to be designed or screened for binding. The second protein structure corresponding to the target protein is equivalent to the structure to be bound, which can be the three-dimensional structure of the target protein calculated by experimental methods, including X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy (Cryo-EM), etc. X-ray crystallography can determine the three-dimensional structure of a protein by analyzing the diffraction pattern of a protein crystal with X-rays; nuclear magnetic resonance can study the structure and dynamics of a protein using nuclear magnetic resonance spectroscopy; cryo-electron microscopy can observe the structure of a frozen sample using an electron microscope.

[0074] Binding site information indicates the specific region where the target protein can specifically bind to the protein conjugate. The protein conjugate in this application is also a protein. This application can generate the first protein structure of multiple candidate protein conjugates based on the second protein structure of the target protein and the binding site information.

[0075] Candidate protein conjugates can refer to proteins that may interact with a specific target protein. In this application, the framework structure of the protein conjugate can be designed de novo, meaning that proteins selected from the de novo protein design process are those that need to bind to the specific target protein. The first protein structure can be a three-dimensional structure generated by folding the protein sequence corresponding to the de novo protein conjugate; that is, amino acids are linked together by peptide bonds to form polypeptide chains, and these chains are then folded into specific three-dimensional shapes through various interactions (such as hydrogen bonds, hydrophobic interactions, van der Waals forces, etc.).

[0076] The first protein structure can be a three-dimensional structure of multiple potential conjugates generated using molecular docking software (such as AutoDock or Rosetta) based on the second protein structure of the target protein and binding site information. The binding site information is used to constrain the spatial conformation of the protein conjugate, ensuring that the scaffold structure of the generated conjugate matches the target site.

[0077] The first protein structure can also be a three-dimensional structure of multiple potential conjugates generated using a deep learning model, based on the second protein structure of the target protein and binding site information. For example, a protein structure generation model based on a diffusion model (RFdiffusion) can be used as the core module to construct a backbone structure design component. The second protein structure and the specified binding sites of the target protein (i.e., binding site information) are used as inputs to the RFdiffusion algorithm. Starting from the protein conjugate backbone structure represented by random noise after diffusion processing, through multiple denoising iterations, the randomly noise-inducing protein conjugate backbone structure is gradually optimized into a structure conforming to the natural protein conformation, until a theoretically valid de novo protein conjugate backbone structure (i.e., the first protein structure) that can bind to the target protein is generated. The RFdiffusion algorithm is a deep learning-based protein design method that combines a random field model and a diffusion process to predict and design the first protein structure. Its core technologies include target protein structure processing, probabilistic modeling, diffusion simulation, and structure optimization. It can generate high-precision three-dimensional structures of protein Ca atoms, especially performing well when detailed structural information is lacking. Furthermore, embodiments of this application may also employ other algorithms to generate protein conjugate structures, such as the AlphaFold algorithm, without limitation or elaboration. Thus, by rapidly generating multiple first protein structures, the probability of identifying the target protein conjugate most likely to succeed in wet experiments can be increased.

[0078] In step 102 of some embodiments, helical structure detection refers to the process of identifying the helical structure type corresponding to the candidate structure of the protein-protein complex using a secondary structure analysis algorithm. This secondary structure analysis algorithm may include the Definition of Secondary Structure of Proteins (DSSP) algorithm, the STRuctural ID Entification (STRIDE) algorithm, etc. The DSSP algorithm is a standardized algorithm for classifying the secondary structure conformations of amino acid residues in a second protein structure. It includes supporting software and a database for parsing secondary structure information in the second protein sequence. The DSSP algorithm can identify secondary structures such as α-helices and β-sheets by analyzing hydrogen bond patterns and geometric angles in the three-dimensional structure of the protein. Specifically, it can identify eight types of secondary structure conformations, each with its own identifier. These eight conformations can be 310 helices, α-helices, and π-helices, identified by G, H, and I, respectively. STRIDE can be seen as an improvement on DSSP. STRIDE statistically analyzes the distribution of dihedral angles of helical and lamellar residue backbones on the Laplace diagram and then excludes residues with abnormal backbone dihedral angles, even if these residues form hydrogen bond patterns corresponding to helical and lamellar structures. However, STRIDE focuses on the local geometric features of each residue and does not consider how to make the overall secondary structure fragments more consistent.

[0079] The filtered first protein structure is the remaining first protein structure after filtering multiple first protein structures based on helical structure detection results. This filtered first protein structure better meets the requirements for success in a specific experiment. For example, in a wet experimental scenario, most of the generated first protein structures contain single-helical or double-helical structures, which are unlikely to be expressed in solution, representing invalid designs in a wet experimental scenario. These invalid backbone structure designs increase the computational cost and time required for subsequent sequence design of the backbone structure, reducing the computational efficiency of the de novo protein conjugate design process. Therefore, to address this issue, this application first performs helical structure detection on the first protein structure of candidate protein conjugates and filters out backbone structures that clearly do not meet experimental requirements based on the helical structure detection results. This avoids the time-consuming design process of generating sequences from unsuitable protein conjugate structures, thereby improving the generation efficiency of the de novo protein conjugate design process.

[0080] A protein sequence is the amino acid sequence that makes up a protein, which determines the protein's structure and function. The protein sequence corresponding to the first protein structure of the filtered candidate protein conjugate can be called the first protein sequence, and the protein sequence corresponding to the protein structure of the target protein can be called the second protein sequence.

[0081] Homology information can be genetic information derived from a common ancestor in biology, providing additional sequence and structural data that helps improve and validate the structure of a second protein obtained through experimental methods. In the embodiments of this application, the homology information of the target protein can be referred to as the first homology information.

[0082] In one specific embodiment, step 102 may specifically include: performing helical structure detection on the first protein structure, identifying first protein structures that only include single-helical or double-helical structures as invalid structures, and filtering out invalid structures. That is, in a wet experimental setting, since single-helical and double-helical structures are unlikely to be expressed in solution, first protein structures that only include single-helical or double-helical structures can be considered invalid structural designs.

[0083] In the above embodiments, this application filters out invalid design structures that do not meet the design requirements of specific experimental scenarios by matching the type of the conjugate structure with the preset invalid helical structure type (such as single helical structure or double helical structure). This prevents these invalid design structures from being passed to the downstream de novo protein conjugate sequence design and only performs sequence design on the valid protein conjugate structure, thereby reducing the number of subsequent sequence generation and improving the computational efficiency of the de novo protein conjugate design process.

[0084] Please see Figure 2 In some embodiments, prior to step 102, the present application embodiments further include a step of determining a preset invalid spiral structure type, and the process of determining the preset invalid spiral structure type includes, but is not limited to, steps 201 to 202.

[0085] Step 201: Obtain information on the design process and structural design requirements for protein conjugates;

[0086] Step 202: Obtain multiple candidate spiral structure types, and determine the preset invalid spiral structure type from the multiple candidate spiral structure types based on the design operation process information and structural design requirement information.

[0087] In step 201 of some embodiments, the design operation process information refers to the specific experimental steps, conditions, and parameters required when designing protein conjugates, such as polymerase chain reaction (PCR), mass spectrometry analysis, protein purification methods, binding energy determination techniques, cell culture, etc. The structural design requirements information indicates the design requirements for the designed protein conjugate, including binding site characteristics (such as hydrophobic pockets, hydrogen bond networks), functional requirements (such as affinity, specificity), etc.

[0088] In step 202 of some embodiments, the candidate helical structure type refers to a potential helical structure type obtained through calculation or experiment. This application can determine candidate helical structure types using bioinformatics tools (such as DSSP, STRIDE) or experimental methods (such as nuclear magnetic resonance). Preset invalid helical structure types refer to candidate helical structure types that do not meet the experimental design conditions, determined based on design operation process information and structural design requirements information.

[0089] In the above embodiments, this application can determine the preset invalid helical structure type from multiple candidate helical structure types by designing operation process information and structural design requirement information, and use it to detect whether the generated protein conjugate structure type is a valid structure type. In this way, invalid helical structure types can be eliminated in advance, reducing the cost of subsequent simulation and experimentation, thereby improving the efficiency and accuracy of determining the target protein conjugate.

[0090] In step 103 of some embodiments, the first protein sequence is the sequence of a de novo designed protein conjugate generated based on the filtered first protein structure. This application can generate a sequence from the filtered first protein structure using a second protein sequence design algorithm to obtain the first protein sequence corresponding to the effective structure. The second protein sequence design algorithm may include, but is not limited to, a Protein Message Passing Neural Network (ProteinMPNN), an Early Maladaptive Schema Inventory (ESM-IF), a Ligand-based Message Passing Neural Network (LigandMPNN), a Solvent-Modulated Message Passing Neural Network (SolMPNN), etc.

[0091] For example, when using the ProteinMPNN algorithm to generate the first protein sequence, the effectively filtered first protein structure can be used as input and represented as a graph. Nodes in the graph represent the amino acid residue positions of the backbone structure, and edges represent the spatial relationships between residues in the backbone structure. Then, a message-passing mechanism is used to capture the geometric information of the backbone and the interaction information of adjacent residues, and the optimal amino acid type at each residue position is predicted step by step, thereby generating a first protein sequence with high energy stability and functional adaptability to the filtered first protein structure.

[0092] In step 104 of some embodiments, because the de novo designed protein conjugate backbone structure and the de novo designed first protein sequence (i.e., the first protein sequence and its corresponding protein structure) are not necessarily consistent, the generated first protein sequence may not fold into the structure of the corresponding protein conjugate. Therefore, it is necessary to screen out sequences from multiple first protein sequences that are consistent with the filtered first protein structure. Thus, this application can perform consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence, the first homology information, the first protein sequence, and the filtered first protein structure to screen out target protein conjugates with consistency. This allows for accurate prediction of the interaction between the de novo designed protein conjugate and the target protein, thereby increasing the probability of experimental success.

[0093] The protein conjugate generation device provided in this application includes a backbone structure design component (equivalent to the structure filtering unit of this application), a sequence design component, and a screening component, which will be described in subsequent embodiments and will not be repeated here.

[0094] Please see Figure 3 In some embodiments, step 104 specifically includes, but is not limited to, steps 301 to 303.

[0095] Step 301: Generate a complex simulation structure based on the filtered first protein structure and the second protein structure corresponding to the target protein;

[0096] Step 302: Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex, generate the predicted structure and prediction confidence of the complex.

[0097] Step 303: Determine the difference information between the simulated structure and the predicted structure of the complex, and determine the target protein conjugate from multiple candidate protein conjugates based on the prediction confidence and the difference information.

[0098] In step 301 of some embodiments, the simulated complex structure may refer to the three-dimensional structure of the complex formed by the expected interaction between biomolecules such as proteins and proteins and proteins and ligands, simulated by computational methods. Here, it refers to the complex structure under the interaction between the three-dimensional structure of the filtered candidate protein conjugate (i.e., the filtered first protein structure) and the laboratory structure of the specific target protein (i.e., the second protein structure).

[0099] In step 302 of some embodiments, the predicted complex structure may be a three-dimensional structure constructed based on the binding sites and functional regions of the two proteins indicated by the second and first protein sequences, combined with the sequence and structural information of other proteins similar to the second protein sequence, including first homology information, and simulating the expected interactions indicated by the complex structure. Prediction confidence can be an assessment of the degree of certainty regarding the predicted complex structure generated by the deep learning model; for example, it can refer to the model's confidence in its prediction results.

[0100] In some embodiments, step 302 may be performed as follows: determining the complex sequence and pairing result based on the first protein sequence and the second protein sequence; generating alignment result representation information based on multiple sequence alignment in the complex sequence and the first homology information, and generating pairing representation information based on the pairing result, the simulated structure of the complex, and the homology template in the first homology information; and generating the predicted structure and prediction confidence of the complex based on the alignment result representation information and the pairing representation information.

[0101] The complex sequence can refer to the set of amino acid sequences of all protein chains that make up a protein complex. That is, in a protein complex, multiple protein molecules may interact to form a stable complex structure. Each protein molecule has its own amino acid sequence, and these sequences together determine the overall structure and function of the complex.

[0102] Pairing results can be the analysis of possible interactions and binding patterns between two molecules. By comparing the first and second protein sequences, their similarities and differences can be identified, indicating the possible interaction regions between the two proteins.

[0103] Multiple sequence alignment (MSA) is the alignment of three or more biological sequences (such as protein sequences, DNA sequences, or RNA sequences). Its purpose is to align identical bases or amino acids from different sequences in the same column as much as possible without altering the sequence order, assuming that sequences in the same column are evolutionarily homologous. The embodiments in this application illustrate the similarities and differences between protein sequences and homologous sequences.

[0104] The alignment results can represent information that, after homology alignment of protein sequences in a complex, the alignment results are associated with other parts of the complex sequence (i.e., the first protein sequence). They can provide complete information on the structure and function of the complex and be presented in a standardized, visualized, or computable form.

[0105] Homologous templates can refer to the three-dimensional structure of a protein that is homologous (sequence similarity) to the second protein sequence of the target protein and whose three-dimensional structure is known.

[0106] Pairing representation information refers to a detailed description of the interaction relationships between individual protein monomers in a protein complex. It can include not only which residues interact, but also the type, strength, and spatial location of these interactions. In other words, pairing representation information is generated based on pairing results, simulated complex structures, and homologous templates to describe the interaction patterns between monomers in the complex. Alignment result representation information and pairing representation information provide rich sequence and structural details, enabling more accurate capture of the interaction patterns between monomers in the protein complex, generating predicted complex structures and prediction confidence levels. The representation information format is typically a simplified and abstract data representation method that highlights key information and removes unnecessary details, allowing for more efficient data processing and reduced computational complexity.

[0107] In some embodiments, step 302 may also be performed by: performing structural prediction on the first protein sequence, the second protein sequence, the first homology information and the simulated structure of the complex based on the protein structure prediction model, and obtaining the predicted structure of the complex and the prediction confidence corresponding to the predicted structure of the complex.

[0108] In some embodiments, the model parameters of the protein structure prediction model are updated using the following method:

[0109] Obtain the second homology information corresponding to the protein sequence of the first sample, and predict the second homology information based on the protein structure prediction model to obtain the predicted structure of the sample;

[0110] Determine the comparison value between the predicted structure of the sample and the structure of the reference complex;

[0111] The parameters of the protein structure prediction model are updated based on the difference between the comparison value and the reference local distance difference test score and the reference alignment error.

[0112] Among them, protein structure prediction models can be tools for inferring the three-dimensional spatial structure of proteins from amino acid sequences. They can integrate evolutionary information (such as homologous information) and physicochemical laws (such as interatomic interactions indicated by protein sequences and complex simulated structures) and use machine learning algorithms to predict the conformation of proteins after folding.

[0113] The first sample protein sequence can refer to a specific amino acid sequence used to validate or adjust the performance of a protein structure prediction model during training or optimization. It can be a known experimentally resolved structure (such as a reference structure obtained by X-ray crystallography or cryo-electron microscopy) or a protein pre-labeled as having a specific function.

[0114] The reference complex structure can refer to the three-dimensional structure of a protein complex that has been resolved by known experiments and serves as a benchmark or validation standard during protein structure prediction or model training.

[0115] The comparison value can be an indicator that quantifies the similarity or difference between the predicted structure and the reference structure. It can be determined by calculating the geometric deviation or local residue matching degree between the predicted structure and the reference structure.

[0116] The Local Distance Difference Test score (LDDT-score) is a non-additive score used to evaluate the local distance differences among all atoms in the model. It does not rely on global alignment but directly compares the distance differences between corresponding atoms in the predicted and reference structures. The LDDT-score ranges from 0 to 1; a score closer to 1 indicates better local consistency between the predicted and reference structures.

[0117] Predicted Alignment Error (PAE) measures the relative position and orientation error between two residues in a predicted structure. A higher PAE value indicates a higher prediction error and therefore lower confidence. PAE can be represented as a matrix, where each element corresponds to the prediction error between two residues. During parameter updates in protein structure prediction models, LDDT-score and PAE can be used to assess the accuracy of the predicted structure and adjust model parameters accordingly. For example, if the predicted structure's local distance difference test score is lower than the reference score or its alignment error is higher, it indicates a significant difference between the predicted and reference structures, and the model parameters may require further optimization. By comparing the LDDT-score and PAE of the predicted structure with the reference values, the model can accurately pinpoint the sources of error in the prediction, allowing for targeted parameter adjustments to reduce errors and improve the overall accuracy of the predicted structure.

[0118] In step 303 of some embodiments, the difference information can be a quantitative indicator measuring the magnitude of the difference between two structures. In protein complex structure prediction, the difference information can be a comparison of the similarity or difference between the simulated structure of the complex and the predicted structure, or it can be the calculation of the distance between corresponding atoms in the two structures. The smaller the difference information, the more similar the predicted structure of the complex is to the simulated structure, and the higher the accuracy of the prediction. The confidence index indicates the model's level of confidence in its prediction results. A high prediction confidence indicates that the predicted structure of the complex is more consistent with the expected structure, which provides an important positive signal for screening candidate protein conjugates. Combining the prediction accuracy indicated by the difference information with a more comprehensive assessment of whether the candidate protein conjugates meet the expected structural requirements, this assessment result can be used as the screening result. Proteins that meet the expectations can be more accurately screened from multiple proteins to be screened for wet experiments, improving the efficiency of wet experiments.

[0119] In the above embodiments, this application generates a simulated complex structure by detecting the first protein structure after filtering using helical structure detection and the second protein structure of the target protein. This simulates the complex that the corresponding candidate protein conjugate and the target protein may form in the natural environment, providing a representation of protein-protein interactions. Then, by combining the second protein sequences of the candidate protein conjugate and the target protein, the first homology information of the target protein, and the simulated complex structure, a predicted complex structure is generated. The confidence level of the generated predicted complex structure is used as the prediction confidence level. The difference between the simulated complex structure and the predicted complex structure, combined with the prediction confidence level, can determine the screening results of the candidate protein conjugate. Furthermore, by comparing the similarity between the predicted structure and the simulated structure, as well as the prediction confidence level, the protein that best matches the expectation in structure can be identified, thereby screening the candidate protein conjugate most likely to succeed in wet experiments.

[0120] Please see Figure 4 In some embodiments, step 303 specifically includes, but is not limited to, steps 401 to 403.

[0121] Step 401: Determine the first difference between the filtered first protein structure and the first predicted structure corresponding to the candidate protein conjugate in the predicted structure of the complex;

[0122] Step 402: Determine the second difference between the second protein structure and the second predicted structure corresponding to the target protein in the predicted structure of the complex;

[0123] Step 403: Determine the third difference between the overall simulated structure and the predicted structure of the complex, and determine the difference information based on the first difference, the second difference, and the third difference.

[0124] In step 401 of some embodiments, the first difference may be the difference between the filtered first protein structure in the local differences between the simulated complex structure and the predicted complex structure and the first predicted structure corresponding to the candidate protein conjugate in the predicted complex structure. This can be calculated by measuring the distance between corresponding atoms in the two structures, such as Euclidean distance. The predicted complex structure consists of a candidate protein predicted structure and a target protein predicted structure, and the first predicted structure may specifically refer to the candidate protein predicted structure.

[0125] In step 402 of some embodiments, the second difference may be the difference between the second protein structure and the second predicted structure corresponding to the target protein in the local difference between the simulated structure and the predicted structure of the complex, or the difference in the straight-line distance between the two structures in three-dimensional space.

[0126] In step 403 of some embodiments, the third difference can be determined by comparing the overall difference between the simulated complex structure and the predicted complex structure. This can be achieved by calculating the overall difference, allowing for a comprehensive evaluation of the predicted complex structure. For example, it can assess the difference between the interaction relationships between two proteins within the structure and the simulated complex structure, identifying potential structural deviations. The difference information between the simulated and predicted complex structures can be determined by both the local differences (first and second differences) and the overall difference (third difference). This approach not only assesses the accuracy of individual protein structures but also evaluates the structure of the protein complex as a whole, improving the accuracy of the difference between the simulated and predicted complex structures based on the differences in each component and their interactions.

[0127] Please see Figure 5 In some embodiments, step 401 specifically includes, but is not limited to, steps 501 to 505.

[0128] Step 501: Spatially align the first predicted structure with the filtered first protein structure;

[0129] Step 502: Determine the fourth difference between the spatially aligned first predicted structure and the filtered first protein structure;

[0130] Step 503: Spatial alignment of the second predicted structure with the second protein structure;

[0131] Step 504: Determine the fifth difference between the spatially aligned first predicted structure and the filtered first protein structure;

[0132] Step 505: Determine the first difference based on the fourth and fifth differences.

[0133] In this embodiment, spatial alignment can be a process of matching two or more molecular structures in three-dimensional space. Specifically, it can involve finding corresponding parts within the molecules and calculating the rotation and translation operations required to align one molecule with another, ensuring that the two structures are based on the same coordinate system. It is understood that before determining the first difference between the filtered first protein structure and the first predicted structure corresponding to the candidate protein conjugate in the predicted complex structure, a difference comparison can be performed based on the spatially aligned structures. The filtered first protein structure can be directly spatially aligned with the first predicted structure corresponding to the candidate protein conjugate in the predicted complex structure, ensuring that the corresponding parts of the filtered first protein structure and the first predicted structure are accurately matched in three-dimensional space. This allows the difference between the two structures to be calculated as a fourth difference, making the difference calculation more accurate.

[0134] Before determining the difference between the first predicted structure and the filtered first protein structure, besides focusing solely on the alignment of the candidate protein's individual structure, we can also consider the interaction of the entire complex and then examine the candidate protein conjugate. For example, we can align the coordinate system of the second predicted structure corresponding to the target protein in the complex's predicted structure to the coordinate system of the second protein structure in the complex's simulated structure. In this case, the filtered first protein structure in the complex's simulated structure, based on its interaction with the second protein structure, is in the same coordinate system as the first predicted structure. The difference between this first predicted structure and the filtered first protein structure includes the difference in their interaction relationships, which can be called the fifth difference. Therefore, determining the first difference between the filtered first protein structure and the first predicted structure based on the fourth and fifth differences can provide more information about protein interactions beyond simply evaluating whether the candidate protein conjugate structure meets expectations. This helps to more comprehensively understand whether the function and stability of the complex meet expectations.

[0135] It is understandable that the first and second predicted structures in the complex prediction structure can be distinguished by identifiers.

[0136] An identifier can be a string of characters or numbers used to uniquely identify a molecule, protein, gene, or other biological entity, serving as a label to identify specific protein structures. When processing large amounts of data, identifiers can be used to quickly and accurately locate corresponding specific protein structures. That is, different identifiers in the simulated complex structure represent the filtered first and second protein structures, respectively, and similarly, different identifiers in the predicted complex structure represent the first and second predicted structures. Therefore, when determining the difference between the first predicted structure and the filtered first protein structure, the first predicted structure can be distinguished first by the identifiers in the predicted complex structure, and then the first difference between the filtered first protein structure and the first predicted structure can be determined. Identifiers ensure accurate identification and differentiation of different protein structures in the simulated and predicted complex structures, simplifying the data retrieval and processing process.

[0137] Please see Figure 6 In some embodiments, steps 301 and 302 include, but are not limited to, steps 601 to 603.

[0138] Step 601: Encode the backbone information in the filtered first protein structure to obtain a first coordinate matrix; encode the backbone information in the second protein structure corresponding to the target protein to obtain a second coordinate matrix.

[0139] Step 602: Determine the third coordinate matrix corresponding to the simulated structure of the complex based on the first and second coordinate matrices;

[0140] Step 603: Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the third coordinate matrix, generate the predicted structure and prediction confidence of the complex.

[0141] In this embodiment, the backbone information can refer to the atomic coordinates that constitute the basic structure of a protein. This information describes the position of each atom in the protein in three-dimensional space. The position of each atom in three-dimensional space can be represented in the form of X, Y, and Z coordinates.

[0142] The first coordinate matrix can be a matrix obtained by converting the backbone information in the filtered first protein structure into a numerical form. It can contain the spatial coordinate information of all atoms that make up the protein backbone, usually including coordinate values ​​in three dimensions: X, Y, and Z.

[0143] The second coordinate matrix can be obtained by converting the backbone information in the second protein structure into a numerical form, thus providing structural information about the target protein.

[0144] The third coordinate matrix can be a matrix that integrates information from the first and second coordinate matrices. It describes the predicted complex structure and contains the spatial coordinates of all atoms in the complex. The interaction relationships between the first and second coordinate matrices can guide the adjustment of atomic coordinates in these matrices to find a stable structure for the third coordinate matrix. Therefore, the structural information of the simulated complex structure can be used to predict the predicted complex structure. By using a precise coordinate matrix, the interactions and binding patterns between proteins can be simulated more accurately, thereby improving the accuracy of the predicted complex structure.

[0145] Please see Figure 7 In some embodiments, step 602 includes, but is not limited to, steps 701 to 702.

[0146] Step 701: Spatially align the first coordinate matrix and the second coordinate matrix;

[0147] Step 702: Based on the binding sites of the second protein structure and the first protein structure, merge the spatially aligned first coordinate matrix and the second coordinate matrix to obtain the third coordinate matrix.

[0148] In this embodiment, a specific algorithm (such as the least squares method) is used to perform translation, rotation and other operations on the two coordinate matrices to achieve spatial alignment, so that the two protein structures overlap as much as possible in space, thereby placing them in a unified spatial reference frame.

[0149] Protein binding sites are key regions where proteins interact with other molecules (such as another protein, ligands, etc.). After spatial alignment, the two coordinate matrices are integrated based on the binding site information of the first and second proteins. Specifically, the coordinate information of atoms in the two protein structures is combined according to the binding site priority rule to form a new coordinate matrix, which can be called the third coordinate matrix. The third coordinate matrix contains all the atomic coordinate information of the two protein structures and considers the binding sites, reflecting the overall spatial structural characteristics of the two proteins in their interaction state.

[0150] Please see Figure 8 In some embodiments, step 302 includes, but is not limited to, steps 801 to 803.

[0151] Step 801: Generate the predicted structure of the complex based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex;

[0152] Step 802: Divide the two-dimensional matrix into blocks based on the regional positions of candidate protein conjugates and target proteins in the two-dimensional matrix of the predicted complex structure;

[0153] Step 803: Determine the overlap matrix of the candidate protein conjugate and the target protein in the two-dimensional matrix based on the block results, and determine the prediction confidence based on the value of the overlap matrix.

[0154] In this embodiment, the predicted structure of the complex can be represented as a two-dimensional matrix. Each element in the matrix can represent a specific attribute of a location in the protein structure, such as interatomic distance or interaction energy. In the two-dimensional matrix, the candidate protein conjugate and the target protein each occupy a certain region, and the positions of these regions can be obtained through analysis of the predicted structure of the complex. For example, the row and column ranges corresponding to the protein in the two-dimensional matrix can be determined based on the amino acid residue numbers or domain information of the protein.

[0155] The block partitioning operation can divide a two-dimensional matrix into different sub-matrices based on the determined regional locations of candidate protein conjugates and target proteins.

[0156] An overlap matrix is ​​the overlapping portion of the candidate protein-conjugate and target protein regions within a segmented two-dimensional matrix. The elements in the overlap matrix reflect the interaction information between the candidate protein-conjugate and the target protein within the overlapping regions. The values ​​in the overlap matrix can represent various information, such as the strength and probability of the interaction. Statistical analysis and evaluation of the values ​​in the overlap matrix can determine the confidence level of the predicted complex structure. For example, if most elements in the overlap matrix indicate strong interactions, the predicted complex structure can be considered to have high confidence; conversely, if the values ​​in the overlap matrix show weak or unstable interactions, the prediction confidence is low. Segmentation allows for more targeted analysis of the interaction relationships between candidate protein-conjugates and target proteins, as well as their structural characteristics within the complex, quantifying prediction quality and improving analytical efficiency.

[0157] Please see Figure 9 In some embodiments, step 303 includes, but is not limited to, steps 901 to 902.

[0158] Step 901: When the prediction confidence and difference information meet the preset index threshold conditions, the candidate protein conjugate is determined as the target protein conjugate.

[0159] Step 902: When the prediction confidence and difference information do not meet the preset index threshold conditions, the candidate protein conjugate is determined to be an invalid protein conjugate.

[0160] In this implementation, the preset threshold conditions can be a set of judgment criteria set based on prediction confidence and difference information. When the prediction confidence reaches the preset confidence level and the difference information reaches the preset difference information, the candidate protein conjugate can be determined to be a protein conjugate that meets the expected structure, and can be further subjected to laboratory verification procedures, i.e., it can become the target protein conjugate. For example, if the prediction confidence level of a candidate protein conjugate is 95 and the root mean square deviation (RMSD) is... (unit of distance) If the parameters (nanometers) meet the threshold conditions, the protein is identified as a target protein conjugate and further experimental verification is performed.

[0161] If the prediction confidence level does not reach the preset confidence level or the difference information does not reach the preset difference information, the candidate protein conjugate can be considered as not meeting expectations. In this case, it can be removed from the set of protein conjugates in the laboratory validation process and considered an invalid protein conjugate. Therefore, by setting reasonable prediction confidence and difference information thresholds, candidate protein conjugates with high confidence and low difference can be screened, thereby improving the reliability of the prediction results.

[0162] The method for generating protein conjugates provided in this application will be described in detail below with reference to a specific embodiment.

[0163] In one specific embodiment, the protein conjugate generation process provided in this application may include three different components, such as... Figure 10 As shown, these are the de novo protein conjugate scaffold structure design component 1010 (equivalent to the structure filtering unit in the protein conjugate generation device of this application, which will be described in subsequent embodiments), the de novo protein conjugate sequence design component 1020, and the de novo protein conjugate screening component 1030.

[0164] The generation process can include three different input sources 1040: the experimental structure of the target protein to be bound 1041, the binding site information corresponding to the experimental structure 1042, and the second protein sequence of the target protein 1043. These three input sources are input into the backbone structure design component 1010 for backbone structure design, resulting in first protein structures of multiple candidate protein conjugates. Then, helical structure detection can be performed on each first protein structure, and based on the helical structure detection results, first protein structures containing only single-helical or double-helical structures can be filtered from the multiple first protein structures to obtain the filtered first protein structure.

[0165] Further, the obtained multiple filtered first protein structures are input into the sequence design component 1020 to obtain the first protein sequence corresponding to the filtered first protein structure. Then, the first protein sequence is input into the screening component 1030 (also called a filter). The screening component 1030 can process and predict the structure of the complex design structure formed by the input first protein sequence of the de novo designed protein conjugate, the second protein sequence of the target protein to be bound, and the experimental structure of the filtered first protein structure and the target protein, resulting in a complex prediction structure 1050. Simultaneously, three prediction values ​​1060 (pAE_interaction, plddt_complex, and pDockQ) are obtained from the output file (the specific content of these prediction values ​​will be explained later). Using the complex prediction structure 1050 and the complex design structure, four root mean square deviation values ​​1070 (target_aligned_target_r.msd, binder_aligned_binder_r.msd, target_aligned_binder_r.msd, and all_aligned_complex_r.msd) are calculated (the specific content of these root mean square deviation values ​​will be explained later). Finally, based on the set threshold values, de novo protein conjugates that meet the threshold values ​​are selected as target protein conjugates 1080 for wet experiments.

[0166] The filtering process of the filtering component provided in this application will be described in detail below with reference to a specific embodiment.

[0167] First, train the filter. For example, the AlphaFold-Multimer prediction model can be used as the basis for the selection filter. The homology information of the first sample protein sequence is used as input, and the three-dimensional coordinates of the experimental complex structure, the local distance difference test score (lddt-score) between the predicted structure and the experimental structure, and the alignment error (AE) are used as labels to train AlphaFold-Multimer to obtain the target filter.

[0168] Furthermore, the predicted structure of the complex is obtained based on the target filter. Specifically, the predicted structure of the complex can be obtained by predicting the structure of the complex designed with the target protein in the set of de novo designed proteins (i.e., the first protein sequence in the above embodiment), the target protein to be bound (i.e., the second protein sequence in the above embodiment), the homology information of the target protein, the design backbone structure of the de novo designed protein conjugate (i.e., the filtered first protein structure in the above embodiment), and the experimental structure of the target protein. For example, one can refer to... Figure 11 The schematic diagram of the predicted structure of the complex shown is based on the input of the test protein sequence 1101 (i.e., the first protein sequence) of the de novo designed protein, the target protein sequence 1102 of the target protein, the design backbone structure 1103 of the de novo designed protein conjugate, and the experimental structure 1041 of the target protein. Input processing can be performed first.

[0169] In addition, the specific processing flow can be as follows: generate a complex sequence 1104 based on the test protein sequence 1101 and the target protein sequence 1102; perform a gene database search based on the target protein sequence 1102 to obtain multiple sequence alignment (MSA) 1105 from the homology information; pair the test protein sequence 1101 and the target protein sequence 1102 to obtain the pairing result 1106; determine the complex design structure 1107 based on the designed backbone structure 1103 and the experimental structure 1041, which can be based on the interface information of the complex design structure constructed from the coordinate matrix corresponding to the designed backbone structure 1103 and the experimental structure 1041, or it can be based directly on the distance matrix extracted from the complex design structure; and perform a structure database search based on the target protein sequence 1102 to obtain the homology template 1108 from the homology information.

[0170] The target filter can determine multiple sequence alignment expression information 1109 based on the complex sequence 1104 and multiple sequence alignment 1105, and determine the first paired expression information 1110 based on the pairing result 1106, the complex design structure 1107, and the homologous template 1108. The evolution module 1111 can extract evolutionary and conserved information from the multiple sequence alignment expression information 1109 based on a self-attention mechanism to obtain single expression information 1112, and update the intrastructural relationships in the first paired expression information 1110 based on a self-attention mechanism to obtain the second paired expression information 1113. The structure module 1114 can predict the three-dimensional structure of the protein complex based on the single expression information 1112 and the second paired expression information 1113. To improve the accuracy of the prediction, multiple cyclic predictions (e.g., three cycles) can be performed based on the single expression information 1112, the second paired expression information 1113, and the predicted three-dimensional structure obtained each time. The structural representation information is continuously updated and optimized to obtain the final predicted complex structure 1050.

[0171] Once the target screener predicts the predicted structure of the complex of the de novo designed protein and the target protein, the following steps can be performed sequentially:

[0172] Calculate the root mean square deviation between the predicted target protein structure and the experimental target protein structure after alignment in the predicted target protein structure of the complex, and denote this value as target_aligned_target_r.msd;

[0173] Calculate the root mean square deviation between the predicted de novo protein structure and the designed backbone structure of the de novo protein conjugate after alignment in the predicted structure of the complex, and denote this value as binder_aligned_binder_r.msd;

[0174] Calculate the root mean square deviation between the de novo designed protein prediction structure and the designed backbone structure of the de novo designed protein conjugate after aligning the target protein prediction structure with the target protein experimental structure in the complex prediction structure, and denote this value as target_aligned_binder_r.msd;

[0175] Calculate the root mean square deviation between the predicted and designed structures of the complex after aligning them, and denote this value as all_aligned_complex_r.msd;

[0176] In the output feature file generated along with the target filter when predicting the predicted structure of the de novo protein conjugate complex with the target protein, the predicted alignment error of interaction (pAE_interaction), the predicted local distance difference test of complex (plddt_complex), and the predicted docking quality (pDockQ) are extracted. Among these, plddt_complex and pDockQ can be obtained directly from the AlphaFold-Multimer output file.

[0177] `pAE_interaction` is obtained "indirectly" from the output file of AlphaFold-Multimer. The original file is a two-dimensional matrix. This matrix can be divided into blocks based on the interval positions of the target protein prediction structure and the de novo protein prediction structure within the entire two-dimensional matrix. This yields the values ​​of the overlapping blocks of the target and de novo protein prediction structures. `pAE_interaction` represents the error of the interaction between the two parts of the predicted structure; a smaller difference indicates a stronger interaction. `plddt_complex` represents the confidence level of the predicted complex structure; a higher value indicates a more reliable predicted complex, indirectly reflecting the correctness of both the target and de novo protein prediction structures. `pDockQ` represents the accuracy of the interface (interaction surface) between the target and de novo protein prediction structures; a higher value indicates a more accurate interface.

[0178] When all the values ​​simultaneously meet the following conditions (i.e., some set indicator thresholds), the corresponding de novo designed protein conjugate is considered to meet the design requirements and can be used as the target protein conjugate for wet experiments. For example, plddt_complex>90.0, pDockQ>0.8.

[0179] In addition, to verify the beneficial effects of the optimization portion of the protein conjugate design process provided in the embodiments of this application, this application can be verified in two parts:

[0180] (1) Regarding the efficiency of de novo protein conjugate structure and sequence generation: To verify the beneficialness of backbone structure filtering between the de novo protein conjugate backbone structure design component and the de novo protein conjugate sequence design component in this invention, this application can use the target protein IL7Ra and its corresponding binding site as test data for the process efficiency gain effect. Specifically, this application can use the backbone structure design component of the above embodiments (such as those built based on RFdiffusion) to design 10,000 backbone structures for the target protein IL7Ra. After backbone structure filtering, only 2,007 non-single-helix or non-double-helix de novo protein conjugate backbone structures are retained, and these protein conjugate backbone structures are handed over to the sequence design component (such as those built based on ProteinMPNN) for sequence generation. Compared with the process without backbone structure filtering, the backbone structure filtering process of this application embodiment can improve the generation efficiency of de novo protein conjugate backbone structures and de novo protein conjugate sequences by 80%.

[0181] (2) For the wet experimental candidate screening of de novo designed protein conjugate backbone structures and de novo designed protein conjugate sequences. To verify that the innovative screener in this invention is beneficial in improving the success rate of wet experimental candidate screening, this application can select publicly disclosed de novo designed protein conjugate sequences and de novo designed protein conjugate backbone structures that have been successfully screened in wet experiments as test data for the gain effect of wet experimental candidate screening in the process of this invention, and screen protein conjugates that may be successful in wet experiments through the above technical solution. As shown in Table 1 below, the wet experimental screening success rate of the screener in this application on de novo designed protein datasets of four different target proteins is compared with the wet experimental screening success rate of the dl_binder_design screener on the corresponding datasets.

[0182] Table 1. Comparison of wet test screening success rates between the filter in this application and the dl_binder_design filter.

[0183]

[0184] Specifically, for the target proteins IL7Ra and TrkA, the wet test success rate of the filter provided in this application is 6.3% and 2.4% higher than that of the dl_binder_design filter, respectively. For the Insulin Receptor, although the wet test success rate of the filter provided in this application is not significantly improved, it is still 0.3% higher than that of the dl_binder_design filter. For Pdl1, the wet test success rate of the filter provided in this application is almost twice that of the dl_binder_design filter, which is significantly better than the dl_binder_design filter by 12.4%.

[0185] In the above embodiments, the embodiments of this application can first screen candidate protein conjugates based on the detection results of the helical structure corresponding to the protein conjugate structure, avoiding the design time consumption when generating sequences of protein conjugate structures that do not meet the requirements, and can effectively improve the generation efficiency of protein conjugates. Furthermore, the prediction confidence reflects the prediction accuracy of the complex predicted structure, and the difference information can quantify the similarity between the predicted complex structure and the simulated complex structure. By combining the analysis of prediction accuracy and similarity, the accuracy of screening whether the predicted complex structure meets the expected results can be improved, so as to realize the second screening of protein conjugate structures, thereby accurately determining the target protein conjugate that binds to the target protein at the binding site, which can effectively improve the design accuracy of protein conjugates.

[0186] Reference Figure 12 This application also provides an apparatus for generating protein conjugates, the apparatus comprising:

[0187] The structure filtering unit 1210 is used to acquire the first protein structure of multiple candidate protein conjugates generated from the target protein, and to perform helical structure detection on the first protein structure to filter out invalid structures.

[0188] Sequence design component 1020 is used to generate a first protein sequence based on the filtered first protein structure;

[0189] The screening component 1030 is used to perform consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence and the first homology information corresponding to the target protein, and to determine the target protein conjugate based on the consistency detection results.

[0190] It is evident that the content of the above-described protein conjugate generation method embodiments is applicable to the embodiments of this protein conjugate generation device. The specific functions implemented by this protein conjugate generation device embodiment are the same as those of the above-described protein conjugate generation method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-described protein conjugate generation method embodiments.

[0191] Please see Figure 13 , Figure 13 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0192] The processor 1301 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0193] The memory 1302 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1302 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1302 and is called and executed by the processor 1301 to execute the protein conjugate generation method of the embodiments of this application.

[0194] The input / output interface 1303 is used to implement information input and output;

[0195] The communication interface 1304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0196] Bus 1305 transmits information between various components of the device (e.g., processor 1301, memory 1302, input / output interface 1303, and communication interface 1304);

[0197] The processor 1301, memory 1302, input / output interface 1303 and communication interface 1304 are connected to each other within the device via bus 1305.

[0198] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for generating protein conjugates.

[0199] This application also provides a computer program product comprising a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the method for generating the protein conjugate described above.

[0200] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this disclosure and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “including,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0201] It should be understood that in this disclosure, "at least one item" means one or more, and "more than one" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0202] It should be understood that in the description of the embodiments of this application, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0203] In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0204] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0205] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0206] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0207] It should also be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0208] The above is a detailed description of the embodiments of this disclosure. However, this disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this disclosure. All such equivalent modifications or substitutions are included within the scope defined by the claims of this disclosure.

Claims

1. A method for generating a protein conjugate, characterized in that, An apparatus for generating protein conjugates, the apparatus comprising a sequence design component and a screening component, the method comprising: Obtain the first protein structure of multiple candidate protein conjugates generated from the target protein; The structure of the first protein is analyzed by helical structure detection to filter out invalid structures; The sequence design component is invoked to generate a first protein sequence based on the filtered first protein structure. The screening component is invoked to perform a consistency test on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence, and the first homology information corresponding to the target protein, and the target protein conjugate is determined based on the consistency test results.

2. The method according to claim 1, characterized in that, The step of detecting the helical structure of the first protein and filtering out invalid structures includes: The first protein structure is subjected to helical structure detection, and the first protein structure that includes only a single helical structure or a double helical structure is regarded as an invalid structure and the invalid structure is filtered out.

3. The method according to claim 1, characterized in that, The step of performing a consistency test on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence, and the first homology information corresponding to the target protein, and determining the target protein conjugate based on the consistency test results, includes: A complex simulation structure is generated based on the filtered first protein structure and the second protein structure corresponding to the target protein. Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex, a predicted structure and prediction confidence of the complex are generated. The difference information between the simulated structure of the complex and the predicted structure of the complex is determined, and the target protein conjugate is determined from the plurality of candidate protein conjugates based on the prediction confidence and the difference information.

4. The method according to claim 3, characterized in that, The determination of the difference information between the simulated structure and the predicted structure of the complex includes: Determine the first difference between the filtered first protein structure and the first predicted structure corresponding to the candidate protein conjugate in the predicted structure of the complex; Determine a second difference between the second protein structure and the second predicted structure corresponding to the target protein in the predicted structure of the complex; A third difference is determined between the overall simulated structure of the complex and the predicted structure of the complex, and the difference information is determined based on the first difference, the second difference, and the third difference.

5. The method according to claim 3, characterized in that, The step of generating a complex-like structure based on the filtered first protein structure and the second protein structure corresponding to the target protein includes: The backbone information in the filtered first protein structure is encoded to obtain a first coordinate matrix, and the backbone information in the second protein structure corresponding to the target protein is encoded to obtain a second coordinate matrix. The third coordinate matrix corresponding to the simulated structure of the complex is determined based on the first coordinate matrix and the second coordinate matrix; The step of generating a predicted complex structure and prediction confidence based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex includes: Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the third coordinate matrix, a predicted structure and prediction confidence of the complex are generated.

6. The method according to claim 3, characterized in that, The step of generating a predicted complex structure and prediction confidence based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex includes: Based on the first protein sequence, the second protein sequence corresponding to the target protein, the first homology information, and the simulated structure of the complex, a predicted structure of the complex is generated. The two-dimensional matrix is ​​divided into blocks based on the regional positions of the candidate protein conjugate and the target protein in the two-dimensional matrix of the predicted structure of the complex. Based on the block division results, the overlap matrix of the candidate protein conjugate and the target protein in the two-dimensional matrix is ​​determined, and the prediction confidence is determined based on the value of the overlap matrix.

7. An apparatus for generating protein conjugates, characterized in that, The device includes: The structure filtering unit is used to obtain the first protein structure of multiple candidate protein conjugates generated from the target protein, and to perform helical structure detection on the first protein structure to filter out invalid structures. Sequence design component for generating a first protein sequence based on the filtered first protein structure; A screening component is used to perform consistency detection on the first protein sequence and the filtered first protein structure based on the second protein structure, the second protein sequence and the first homology information corresponding to the target protein, and to determine the target protein conjugate based on the consistency detection results.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the method described in any one of claims 1 to 6.

Images