Methods, devices, media, equipment and products for predicting protein structure
By adding noise to the diffusion model during training and performing multiple denoising processes, the problem of insufficient data in protein structure prediction was solved, achieving high-precision prediction under limited data conditions and enhancing the model's generalization ability.
Patent Information
- Application Number
- CN202510637977.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2026-11-17
Smart Images

Figure CN123120046A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of bioinformatics technology, and in particular to a method, apparatus, computer-readable storage medium, electronic device, and computer program product for predicting the structure of proteins. Background Technology
[0002] Protein structure prediction is a core area of bioinformatics, aiming to infer the three-dimensional structure of proteins from known amino acid sequences. This process is crucial for understanding protein function, mechanisms, and roles within cells, and has profound implications for drug development, disease research, and many other areas. For example, understanding protein structure helps identify potential drug targets and guides structure-based drug design, thereby accelerating the new drug development process.
[0003] Deep learning models for predicting protein structures, provided by related technologies, typically require large-scale protein structure data for training. A lack of sufficiently diverse data can lead to insufficient model generalization ability, affecting the accuracy of protein structure prediction. Summary of the Invention
[0004] This application provides a method, apparatus, computer-readable storage medium, electronic device, and computer program product for predicting the structure of proteins. By introducing a diffusion model to add noise, the training samples are enriched, which can reduce the dependence on experimental structural data while ensuring the structural accuracy of the generated proteins.
[0005] In a first aspect, embodiments of this application provide a method for predicting the structure of a protein. Based on the amino acid sequence of the target protein, a first feature is determined, wherein the first feature is used to determine a condition variable; the condition variable and randomly sampled raw noise are input into a structure prediction model; the structure prediction model performs multiple denoising processes on the raw noise based on the condition variable to obtain the predicted three-dimensional structure of the target protein; wherein the structure prediction model is obtained by training a diffusion model, and the training samples of the diffusion model are obtained by adding artificial noise to the raw samples.
[0006] Secondly, embodiments of this application provide a protein structure prediction device, which includes: a determination module, an input module, and a prediction module.
[0007] The system includes a determination module for determining a first feature based on the amino acid sequence of the target protein, wherein the first feature is used to determine a condition variable; an input module for inputting the condition variable and randomly sampled raw noise into a structure prediction model; and a prediction module for performing multiple denoising processes on the raw noise based on the condition variable using the structure prediction model to obtain the predicted three-dimensional structure of the target protein. The structure prediction model is obtained by training a diffusion model, and the training samples for the diffusion model are obtained by adding artificial noise to the raw samples.
[0008] In an exemplary embodiment, based on the above scheme, the prediction module includes: a first determining unit, a second determining unit, and a processing unit.
[0009] The first processing unit is used to determine the output features of the (i-1)th denoising process to obtain the (i-1)th noisy data representation, where i is an integer greater than 1; the second determining unit is used to determine the input features of the i-th denoising process based on the (i-1)th noisy data representation and the condition variable to obtain the i-th input feature; and the processing unit is used to process the i-th input feature through the denoising module of the structure prediction model to obtain the i-th noisy data representation.
[0010] In an exemplary embodiment, based on the above scheme, the processing unit is specifically configured to: perform noise prediction through the denoising module to determine the estimated noise information in the i-th denoising process based on the i-th input feature, thereby obtaining the i-th estimated noise information; and update the (i-1)-th noisy data representation through the denoising module according to the i-th estimated noise information, thereby obtaining the i-th noisy data representation.
[0011] In an exemplary embodiment, based on the above scheme, the second determining unit is specifically used to: perform feature fusion on the (i-1)th noisy data representation and the condition variable through the feature fusion module of the structure prediction model to obtain the (i-1)th fused feature; and encode the (i-1)th fused feature through the feature encoding module of the structure prediction model to obtain the i-th input feature.
[0012] In an exemplary embodiment, based on the above scheme, the second processing unit includes a fusion subunit and an encoding subunit.
[0013] The fusion subunit is used to: after the input module inputs the condition variable and randomly sampled original noise into the structure prediction model, perform feature fusion on the data representation of the original noise and the condition variable through the feature fusion module of the structure prediction model to obtain a first fused feature; the encoding subunit is used to: encode the first fused feature through the feature encoding module of the structure prediction model to obtain a first encoded feature; the processing unit is specifically used to: process the first encoded feature through the denoising module of the structure prediction model to obtain the output feature of the first denoising process.
[0014] In an exemplary embodiment, based on the above scheme, the prediction module includes a judgment unit and a feedback unit;
[0015] The judgment unit is used to: after processing the i-th input feature through the denoising module of the structure prediction model to obtain the i-th noisy data representation, determine whether the i-th noisy data representation meets a preset condition; and, if the preset condition is met, determine the predicted three-dimensional structure of the target protein based on the i-th noisy data; the feedback unit is used to: generate i-th feedback information if the i-th noisy data does not meet the preset condition, and perform the (i+1)-th denoising process based on the i-th feedback information.
[0016] In an exemplary embodiment, based on the above scheme, the determination unit is specifically used to: determine the confidence level of the i-th noisy data representation; and determine whether the confidence level of the i-th noisy data representation meets a preset confidence threshold.
[0017] The feedback unit is specifically used to: if the i-th confidence level is greater than the first threshold, then the i-th feedback information is used to indicate increasing the denoising step size, where the first threshold is less than the confidence level threshold; if the i-th confidence level is less than the first threshold, then the i-th feedback information is used to indicate decreasing the denoising step size.
[0018] In an exemplary embodiment, based on the above scheme, the diffusion model is a denoised diffusion probability model (DDPM), and the device further includes a training module; the training module includes a first determining unit, a first processing unit, a prediction unit, a construction unit, and an optimization unit.
[0019] The first determining unit is used to control the noise variance at time step t through a noise adjustment strategy during the forward diffusion process of DDPM, and obtain the noise variance at time step t, wherein the noise variance increases nonlinearly with increasing t.
[0020] The first processing unit is used to determine the noisy data representation of the target original sample based on the noise variance of the tth time step and the noisy data representation of the target original sample at the (t-1)th time step, wherein the noisy data representation at the (t-1)th time step is the noisy data representation output at the (t-1)th time step, and wherein the value of t during the forward diffusion process is 1, ..., T-1, T.
[0021] The prediction unit is used to perform feature fusion and feature encoding on the noisy data representation at time step t in the reverse generation process and the sample condition variables corresponding to the target original sample. The encoded features and time step t are input into the denoising module to predict the actual noise added at the corresponding time step in the forward diffusion process, and obtain the predicted noise at time step t. In the reverse generation process, the value of t is T, T-1, ..., 1.
[0022] A construction unit is configured to construct a mean squared error loss function based on the t-th predicted noise and the t-th actual noise; and an optimization unit is configured to optimize the parameters of DDPM based on the mean squared error loss function.
[0023] In an exemplary embodiment, based on the above scheme, the first determining unit is specifically configured to: control the growth rate of the noise variance to be less than a second threshold when t is less than a first threshold, so as to preserve structural features in the early stage of the forward diffusion process; and control the growth rate of the noise variance to be greater than the second threshold when t is greater than the first threshold, so as to make the noise addition time step less than a third threshold.
[0024] In an exemplary embodiment, based on the above scheme, the noise adjustment strategy controls the growth of noise variance by means of a cosine function.
[0025] In an exemplary embodiment, based on the above scheme, the target protein includes a target immunoglobulin; the determination module includes: a vectorization unit, an extraction unit, and a determination unit.
[0026] The vectorization unit is used to vectorize the amino acid sequence of the target immunoglobulin through a pre-trained language model to obtain an initial sequence representation, which includes local chemical environment information and global topological information of the immunoglobulin complementarity determination region (CDR).
[0027] An extraction unit is used to extract features from the initial sequence representation to obtain the CDR-H3 features and / or frame region constraint features of the target immunoglobulin, wherein the CDR-H3 features characterize the local chemical environment information of the CDR-H3 loop, and the frame region constraint features represent the spatial constraint relationship between the frame region and the CDR.
[0028] A determining unit is configured to determine a first feature of the target immunoglobulin based on the CDR-H3 feature and / or the frame region constraint feature.
[0029] In an exemplary embodiment, based on the above scheme, the extraction unit is specifically used to: focus on the local residue relationships within the CDR based on the first type of attention head in the multi-head attention mechanism to obtain CDR-H3 features, wherein the CDR-H3 features reflect the local chemical environment information of the CDR-H3 ring; and identify the long-range dependency relationship between the frame region and the CDR through the second type of attention head in the multi-head attention mechanism to obtain frame region constraint features.
[0030] In an exemplary embodiment, based on the above scheme, the predicted three-dimensional structure includes: the three-dimensional structure of the complex after binding with the target antigen; the determining module is further configured to: determine a second feature based on the amino acid sequence of the target antigen; and determine the condition variable based on the first feature of the target immunoglobulin and the second feature of the target antigen.
[0031] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory. The memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to perform the protein structure prediction method provided in the first aspect.
[0032] Fourthly, embodiments of this application provide a chip for implementing the protein structure prediction method provided in the first aspect. Specifically, the chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to execute the protein structure prediction method provided in the first aspect.
[0033] Fifthly, embodiments of this application provide a computer-readable storage medium for storing a computer program that causes a computer to execute the protein structure prediction method provided in the first aspect.
[0034] In a sixth aspect, embodiments of this application provide a computer program product, including computer program instructions that cause a computer to execute the protein structure prediction method provided in the first aspect.
[0035] In a seventh aspect, embodiments of this application provide a computer program that, when run on a computer, causes the computer to execute the protein structure prediction method provided in the first aspect.
[0036] In summary, the solution provided in this application determines conditional variables based on the amino acid sequence of the target protein. These conditional variables and randomly sampled raw noise are then input into the structure prediction model. Furthermore, the denoising module of this structure prediction model performs multiple denoising processes on the raw noise based on the conditional variables to obtain the predicted three-dimensional structure of the target protein. Here, the raw noise serves as the starting point for the reverse generation process, and the conditional variables guide the structure prediction model to progressively denoise the raw noise, generating a structure that conforms to physical rules and has functional specificity. This ensures the accuracy of the generated structure. The structure prediction model is obtained by training a diffusion model, which can reduce its dependence on large-scale raw samples (experimental structural data) through a noise-adding-denoising training paradigm. Specifically, by using artificially constructed noise data during the noise-adding process, the diversity of training samples can be increased. Even with limited experimental structural data, the model can learn to recover the true structure from various noise levels, improving the model's generalization ability. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 A schematic diagram of the system architecture for a molecular property prediction scheme provided in an embodiment of this application;
[0039] Figure 2 This application provides a schematic flowchart of a protein structure prediction method.
[0040] Figure 3 This application provides a schematic diagram of the structure of an antibody language model;
[0041] Figure 4 This application provides a flowchart illustrating a method for determining the first characteristic of a target immunoglobulin.
[0042] Figure 5 This application provides a schematic diagram for feature extraction of an initial sequence.
[0043] Figure 6 A flowchart illustrating a DDPM training method provided in this application embodiment;
[0044] Figure 7 A schematic diagram of a DDPM training process provided in an embodiment of this application;
[0045] Figure 8 A flowchart illustrating a denoising method in a structure prediction process provided in this application embodiment;
[0046] Figure 9 A schematic diagram illustrating a method for predicting the structure of an immunoglobulin-antigen complex provided in an embodiment of this application;
[0047] Figure 10 A schematic block diagram of a protein structure prediction device provided in an embodiment of this application;
[0048] Figure 11 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0050] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein. In embodiments of this application, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that comprises 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 these processes, methods, products, or devices. In the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0051] 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.
[0052] Immunoglobulins (Ig) are a class of proteins that can specifically bind to antigens. Immunoglobulin structure prediction is a core step in immunoglobulin drug development, and the accuracy of this prediction directly impacts immunoglobulin affinity optimization, humanization design, and target specificity. The following section details protein structure prediction schemes provided by relevant technologies, using immunoglobulins as an example.
[0053] Existing technical solutions can be categorized into deep learning-based immunoglobulin design frameworks, traditional computational biology methods, and multimodal fusion strategies. Among these, the deep learning-based immunoglobulin design framework includes the aforementioned H3-OPT scheme. Specifically, the H3-OPT scheme first performs template matching, using a known antibody structure database as templates to find the structure most similar to the target antibody. By comparing the highly variable Complementarity-Determining Regions (CDRs) in these templates, particularly CDR-H3, common features and patterns are extracted, and initial predictions are made based on these patterns. Furthermore, a pre-trained protein language model is introduced, trained on a large dataset of natural immunoglobulin sequences. These models can capture the complex relationships within and between amino acid sequences, thus providing better generalization ability for unseen antigenic epitopes. In H3-OPT, this language model is used to enhance the results of template matching by learning the sequence-structure mapping relationship, further refining the prediction results. Although H3-OPT improves the prediction accuracy of CDR-H3 to some extent, such as reducing its prediction error to approximately However, in order to improve the model's generalization ability and prediction accuracy, a filtered dataset that does not contain duplicate or highly similar structures is needed, i.e., a non-redundant dataset.
[0054] Traditional computational biology methods include the following.
[0055] RosettaAntibody: Grafting template based on the Kabat / Chothia numbering system, non-H3 ring prediction error However, the sampling efficiency for CDR-H3 is low, with a success rate of less than 30%.
[0056] Molecular dynamics simulations can reproduce 98% of the unbound CDR conformation, but cannot handle the dynamic changes in conformational entropy during antigen binding, and have errors in flexible interface prediction.
[0057] The following are some methods for predicting immunoglobulin structure based on multimodal fusion strategies.
[0058] IgFold is an end-to-end prediction framework that combines a pre-trained language model to achieve minute-level predictions, but its prediction error for the nanoimmunoglobulin CDR-H3 reaches [a certain value]. However, the structure of the complex cannot be predicted.
[0059] It is evident that among the solutions provided by the relevant technologies, deep learning methods require training on large-scale immunoglobulin structure databases (e.g., H3-OPT requires a non-redundant dataset), thus the generalization ability of the models needs improvement. Furthermore, models providing high-precision structure predictions rely on experimental validation, while fast prediction tools (such as IgFold) sacrifice structural accuracy, such as the detail accuracy of CDR-H3.
[0060] To address the aforementioned problems in related technologies, the structural prediction model provided in this application is obtained by training a diffusion model. Through the noise-adding process in the diffusion model, different levels of noise can be artificially added to the original structure, thereby creating a large number of variants and obtaining rich training samples. This diversity of training samples allows the model to learn on a broader dataset, rather than relying solely on limited real experimental structural data. Simultaneously, because diverse training samples can be synthesized, the diffusion model reduces its dependence on large-scale experimental validation data. By exposing the model to data with various noise levels during training, the model learns how to recover the correct structure from different starting points (even highly distorted states). This enhances the model's performance when faced with unseen antigen or antibody sequences, i.e., improves its generalization ability.
[0061] The following is through Figure 1 An exemplary system architecture of the embodiments of this application will be described.
[0062] For example, Figure 1 This is a schematic diagram of the system architecture for a protein structure prediction scheme provided in an embodiment of this application. (Reference) Figure 1The system architecture of the implementation environment of this application embodiment may include: electronic device 10, and may also include network, storage device, etc.
[0063] The electronic device 10 is used for a protein structure prediction program. Specifically, the electronic device 10 receives the sequence of the target protein 11 and randomly sampled raw noise 12, executes the protein structure prediction program, and then outputs the three-dimensional structure of the target protein 11.
[0064] refer to Figure 1 In an exemplary embodiment, the protein structure prediction scheme executed by the computing device 10 may include: extracting features from the sequence of the target protein 11 using an antibody language model 110 to determine a condition variable 13. Further, the condition variable 13 and the original noise 12 are input into the structure prediction model 120.
[0065] In an exemplary embodiment, the feature fusion module 121 fuses the condition variable and the original noise 12 to obtain fused features. These fused features are then processed by the feature encoding module 122 and input to the denoising module 123. Specifically, before the first denoising process, to ensure that the structure prediction is optimized in the correct direction from the outset, the original noise 12 and the condition variable 12 are initially fused to form a preliminary coordinate representation. Based on this, the denoising module performs the first noise prediction, predicting how much noise should be removed from the current coordinate representation to make it closer to the true molecular structure.
[0066] refer to Figure 1 The denoising module 123 performs denoising processing multiple times, a preset number of times each time or at each interval, to determine whether the output noisy data representation meets preset requirements (such as confidence level). If the requirements are not met, the noisy data representation is fused and encoded again with the condition variable 13 and then re-enters the denoising module for denoising processing. If the requirements are met, the three-dimensional structure of the target protein 11 is output.
[0067] In the above embodiments, the original noise is used as the starting point for the model's reverse generation process. Conditional variables guide the structure prediction model to progressively denoise the original noise, generating structures that conform to physical rules and have functional specificity. This ensures the accuracy of the generated structures. Since the above structure prediction model is obtained by training a diffusion model, and the diffusion model can be trained using a noise-denoising paradigm, it can reduce the dependence on large-scale original samples (experimental structure data). By using artificially constructed noise data during the noise-adding process, the diversity of training samples can be increased. Even with limited experimental structure data, the model can learn to recover real structures from various noise levels, improving the model's generalization ability.
[0068] It should be noted that, Figure 1The system architecture of the implementation environment of the embodiments of this application is illustrated, but the system construction of the implementation environment of the embodiments of this application is not limited to... Figure 2 As shown.
[0069] The protein structure prediction method of this application will be described in detail below through some embodiments. These embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.
[0070] Figure 2 This application provides a schematic flowchart of a protein structure prediction method P200. Method P200 provides an execution entity that can be, for example... Figure 1 The electronic device 10 shown. (Reference) Figure 2 The method includes S210-S230.
[0071] In S210, a first feature is determined based on the amino acid sequence of the target protein, wherein the first feature is used to determine a condition variable.
[0072] The target protein mentioned above refers to any type of protein, such as an immunoglobulin or a non-immunoglobulin. The amino acid sequence of the target protein refers to the linear arrangement of the amino acid units constituting the target protein in a specific order. In this embodiment, the first feature determined based on the amino acid sequence of the target protein is used to guide the structure prediction model to generate the three-dimensional structure of the target protein.
[0073] When the target protein is an immunoglobulin, the complementarity-determining region (CDR) and framework region (FR) information determined by its amino acid sequence can provide guidance for the three-dimensional structure generation process. Therefore, in this case, the first feature mentioned above can include at least one of the features corresponding to the CDR and FFR information. When the target protein is a non-immunoglobulin, the information that can provide guidance for the three-dimensional structure generation process of the structure generation model includes: secondary structure tendency information (e.g., hydrophobic amino acids tend to be located inside α-helices, while hydrophilic amino acids are more likely to appear on β-sheet surfaces), hydrophobic interactions, hydrogen bond networks, etc. Therefore, in this case, the first feature mentioned above can include at least one of the features corresponding to secondary structure tendency information, hydrophobic interactions, and hydrogen bond networks.
[0074] The following example uses the target protein as an immunoglobulin. Figure 3 and Figure 4 This section introduces the specific implementation method for determining the aforementioned first feature.
[0075] Figure 3 This application provides a schematic diagram of the structure of an antibody language model. Figure 4This application provides a flowchart illustrating a method for determining the first characteristic of a target immunoglobulin. Specifically, it will be achieved through methods such as... Figure 3 The antibody language model 110 shown extracts the first feature mentioned above from the amino acid sequence of the target immunoglobulin.
[0076] refer to Figure 3 The antibody language model 110 includes a pre-trained language model 111. Since the pre-trained language model 111 is a model pre-trained on a large-scale dataset, it can provide powerful basic representation capabilities.
[0077] In S410, the sequence of the target immunoglobulin is vectorized by the pre-trained language model 111 to obtain an initial sequence representation, which includes local chemical environment information and global topological information of the immunoglobulin CDR.
[0078] CDRs (Cellular Dependent Residues) are specific segments located within the variable region of immunoglobulins, responsible for direct antigen binding, and are therefore a key focus area for 3D structure prediction. Each immunoglobulin's light and heavy chain variable regions contain three CDRs; specifically, the three CDRs in the light chain are denoted as L1, L2, and L3, and the three CDRs in the heavy chain are denoted as H1, H2, and H3. Local chemical environment information of CDRs describes the specific physical and chemical conditions surrounding amino acid residues within the CDR region, including hydrogen bonding patterns, hydrophobic interactions, and electrostatic interactions. This local chemical environment information relates to the specific interactions of immunoglobulins in three-dimensional space, providing guidance for the 3D structure generation process. Global topological information reflects the 3D structural layout of immunoglobulins, including spatial relationships between different regions, and thus also provides guidance for the 3D structure generation process. Therefore, the initial sequence representation described above is suitable for determining these features, and subsequently for determining the aforementioned conditional variables, to assist in the generation of the structure prediction model.
[0079] In an exemplary embodiment, the pre-trained language model 111 described above can be a Transformer model trained on a large number of unlabeled immunoglobulin sequences (e.g., based on the 558M natural immunoglobulin sequence). Therefore, it can effectively utilize domain-specific knowledge to vectorize the amino acid sequence of the target immunoglobulin, which is beneficial for improving the accuracy of sequence representation. Here, sequence representation refers to the pre-trained language model generating a corresponding vector representation for each amino acid in the input sequence. This vector is obtained through learning the entire sequence and includes the amino acid at that position and its contextual information. Contextual information means that the position of each amino acid not only reflects its own attributes (such as type, charge, etc.) but also its role in the entire sequence and its relationship with other amino acids.
[0080] Multiple sequence alignment (MSA) is used in related technologies to determine initial sequence features. MSA relies on the assumption of evolutionary conservation, extracting co-evolutionary signals by aligning homologous sequences to infer structural features. However, the CDRs of immunoglobulins (especially CDR-H3) are highly variable, with extremely low sequence conservation. Therefore, using MSA in immunoglobulin modeling suffers from high computational complexity; specifically, the time complexity of MSA increases exponentially with the number of sequences, requiring significant computational resources. MSA also suffers from data sparsity issues; for example, CDR sequence diversity reaches trillions, making it difficult for traditional MSA to cover sufficient samples.
[0081] Compared to multiple sequence alignment in related technologies, the pre-trained language model 111 in this embodiment can directly extract the local chemical environment and global topological features of the CDR from the sequence through a representation learning strategy, without relying on direct alignment between sequences. The representation learning strategy refers to the process of automatically discovering effective representations of data. In this embodiment, the sequence embedding vector is directly mapped to structural features, simplifying the conversion process from sequence to structure. Therefore, this embodiment can significantly reduce computation time (e.g., by 93%) and cover a wider range of CDR conformational spaces, thus solving the problems of computational complexity and data sparsity in related technologies.
[0082] The initial sequence features obtained through the pre-trained language model contain rich sequence information. However, directly using these initial sequence features as the first feature to determine the condition variables may introduce redundant information into the 3D structure prediction process, leading to model overfitting. Therefore, in order to ensure that the determined condition variables can efficiently guide the 3D result prediction of the structure prediction model, it is necessary to further extract targeted features from the initial sequence representation to obtain features that are more instructive for the 3D structure generation task.
[0083] refer to Figure 3 The antibody language model 110 may also include a feature extraction module 112. The feature extraction module 112 is used to extract specific features useful for protein structure prediction from the initial sequence representation.
[0084] In S420, the initial sequence representation is subjected to feature extraction by the feature extraction module 112 to obtain the CDR-H3 feature of the target immunoglobulin and / or the frame region constraint feature.
[0085] CDR-H3 is the third CDR in the immunoglobulin heavy chain and belongs to one of the most diverse parts of the variable region of the antibody heavy chain, located at the center of the antigen-binding site. Therefore, CDR-H3 plays a crucial role in the specific recognition and binding of antigens. However, due to the large variation in length and highly diverse amino acid composition of CDR-H3, predicting the structure of the CDR-H3 region is more challenging. Accurate prediction of the CDR-H3 region helps in understanding how antibodies interact with specific antigens; therefore, the structure of the CDR-H3 region needs to be a key focus during the three-dimensional structure generation process. Therefore, in this embodiment, features related to CDR-H3 (denoted as CDR-H3 features) need to be extracted from the initial sequence representation to determine the first feature, which is ultimately used to determine the aforementioned conditional variables to guide the three-dimensional structure generation process.
[0086] For example, since the initial sequence features mentioned above contain local chemical environment information of the CDR, the CDR-H3 features obtained after feature extraction can characterize the local chemical environment information of the CDR-H3 ring. For instance, the CDR-H3 features can be features representing the polar / hydrophobic distribution pattern of key residues in the CDR-H3 ring.
[0087] The framework region refers to the portion of the antibody variable region excluding the CDRs, primarily composed of β-sheets, forming the basic structural framework of the antibody. Each antibody's light and heavy chain variable regions contain four framework regions (FR1, FR2, FR3, and FR4), interconnected by linking sequences and surrounding three CDRs. Thus, the framework region provides structural support for the CDRs and ensures the stability of the overall three-dimensional structure by maintaining the antibody's basic folding pattern. Therefore, in this embodiment, features related to the framework region (denoted as framework region constraint features) need to be extracted from the initial sequence representation to determine the first feature, which is ultimately used to determine the aforementioned conditional variables to guide the three-dimensional structure generation process.
[0088] For example, since the initial sequence features mentioned above contain global topological information about immunoglobulins, the frame region constraint features obtained after feature extraction can represent the spatial constraint relationship between the frame region and the CDR. For instance, the frame region constraint features can be used to reflect the spatial constraint effect of the heavy chain and light chain frame regions on the CDR conformation.
[0089] Specifically, the input to the feature extraction module 112 is the initial sequence features output by the pre-trained language model 111, and the output information is the feature representation extracted from the initial sequence features, such as CDR-H3 features and / or frame region constraint features. These features can more accurately reflect the physicochemical properties of the antibody and its potential spatial conformation.
[0090] For example, the embodiments of this application utilize a multi-head attention mechanism to help focus on the properties of specific parts or patterns in a sequence. The above feature extraction process can be implemented through the multi-head attention mechanism in the Transformer model, thereby improving the relevance and accuracy of the extracted features. Figure 5 This application provides a schematic diagram of feature extraction of initial sequence features.
[0091] refer to Figure 5 For example, a multi-head attention layer includes a first type of attention head (X1-X4) and a second type of attention head (Y1-Y3). The first type of attention head focuses on capturing the details inside CDR-H3 to accurately extract the CDR-H3 features, while the second type of attention head may focus on long-range dependencies across chains to accurately extract the upper frame constraint features.
[0092] refer to Figure 5 In the initial sequence features, C1-C5 represent features related to the CDR region, and F1-F5 represent features related to the frame region. Specifically, in the multi-head attention mechanism, the first type of attention head focuses on the local residue interactions within the CDR region to extract CDR-H3 features. Figure 5 In the diagram, "C1→C3" indicates that the first type of attention head X1 calculates a higher attention weight between this pair of residues, therefore focusing on the interaction between C1 and C3. Similarly, "C4→C1" indicates that the first type of attention head X3 calculates a higher attention weight between this pair of residues, therefore focusing on the interaction between C4 and C1. As can be seen, the arrows corresponding to the first type of attention head in the diagram show that each first type of attention head mainly focuses on the local interactions between different residues within the CDR region, thereby capturing local details and short-range dependencies. The second type of attention head identifies the long-range dependencies between the frame region and the CDR to extract the spatial constraint features of the frame region on the CDR conformation. Figure 5 In the diagram, "C2→F3" indicates that the second type of attention head Y1 calculates a higher attention weight between this pair of residues, therefore the second type of attention head Y1 focuses on the interaction between C2 and F3. As can be seen, the arrows corresponding to the second type of attention heads in the diagram indicate that each second type of attention head focuses on the interaction between different residues in the CDR region and the frame interval, thus enabling the identification of dependencies between distant residues.
[0093] As can be seen, the pre-trained language model 111 provides powerful sequence representation capabilities for subsequent tasks, while the feature extraction module 112, based on these representations, focuses on extracting and emphasizing key information for achieving specific goals (such as accurate antibody structure prediction). The two modules work together to improve the accuracy and reliability of the final structure prediction.
[0094] refer to Figure 3 The antibody language model 110 may also include a sequence feature embedding module 113.
[0095] In S430, the first feature of the target immunoglobulin is determined by the sequence feature embedding module based on the CDR-H3 feature and / or the frame region constraint feature.
[0096] The sequence feature embedding module 113 is used to transform the CDR-H3 features and / or the frame region constraint features into a form suitable for downstream tasks (such as 3D structure prediction). Specifically, the sequence feature embedding module 113 can denoise the CDR-H3 features and / or the frame region constraint features through linear transformation, nonlinear transformation, and / or attention mechanisms to filter out information irrelevant to or interfering with the 3D structure prediction task. This also reduces the dimensionality of the features, making model training more efficient and reducing the risk of overfitting. Furthermore, the sequence feature embedding module 113 can enhance information relevant to the 3D structure prediction task. For example, in antibody 3D structure prediction, some features in the CDR-H3 region may be more important than others, requiring enhancement of these features.
[0097] After introducing the above-described embodiments for determining the first feature, embodiments for determining condition variables based on the first feature are described.
[0098] In this embodiment, a condition variable is used as the core input of the structure prediction model to guide the denoising process of the structure prediction model. By constraining the biological rationality of the structure, it promotes the generation of structures that conform to physical rules and have functional specificity, thereby improving the prediction accuracy of immunoglobulin structures.
[0099] The information contained in the condition variables is related to the type of prediction task. Specifically, the embodiments of this application can be used for various types of three-dimensional structure prediction tasks: protein structure data prediction and antigen-immunoglobulin complex structure data prediction. Individual antibody structure prediction refers to the prediction of the three-dimensional structure of the antibody itself. The aim is to determine the spatial arrangement of the various regions of the antibody, especially the CDRs in the variable regions. Complex structure prediction refers to the prediction of the structure when the antibody binds to a specific antigen. Complex structure prediction needs to consider the structural characteristics of the antibody itself, as well as the structural information of the antigen and the interaction mode between the two. This prediction is more complex because it needs to consider the physicochemical interactions between the antibody and the antigen simultaneously, including hydrogen bonds, hydrophobic interactions, van der Waals forces, etc. For example, the amino acid sequence of an immunoglobulin (antibody) is known (e.g., obtained through phage display, single B cell sequencing, etc.), but its three-dimensional structure (especially the complex structure when bound to a specific antigen) is unknown. The complex structure prediction task provided in the embodiments of this application can then be used to predict the spatial conformation of the antibody after binding to its target antigen.
[0100] The condition variables differ depending on the type of prediction task. Specifically, for a prediction task focusing solely on protein structure, the condition variables only need to include a first feature determined based on the protein's amino acid sequence. For example, when the protein is an immunoglobulin and the prediction task is to predict only the immunoglobulin's protein structure, the condition variables only need to include a first feature determined based on the immunoglobulin's amino acid sequence. When the protein is an immunoglobulin and the prediction task is to predict the structure of an immunoglobulin-antigen complex, the condition variables include a first feature determined based on the immunoglobulin's amino acid sequence and a second feature determined based on the antigen's amino acid sequence.
[0101] When the prediction task involves protein structure data, the first feature 14 output by the sequence feature embedding module 113 can be directly used as the condition variable 13. In other embodiments, to provide more precise constraints on the protein structure generation process, the first feature 14 output by the sequence feature embedding module 113 can be combined with other features of the target protein to determine the condition variable 13. The richer the content of the condition variable, the more prior knowledge it can provide to the model, making the generation process more precisely constrained. The other features of the target protein can be at least one of the following: the amino acid sequence of the target protein, the initial sequence representation output by the pre-trained language model 111, etc.
[0102] When predicting the structural data of an antigen-immunoglobulin complex, it is necessary to determine the characteristics (i.e., secondary characteristics) of the target antigen in order to determine the condition variables. In the structural prediction model provided in this application, the condition variables guide the generation process. By using information specific to both immunoglobulins and antigens as condition variables, the model can "know" what kind of complex structure it is attempting to build during the generation process. This provides guidance, helping the model avoid errors caused by randomness and improving the accuracy and reliability of the final prediction results. Furthermore, the more comprehensive information contained in the condition variables enhances the model's generalization ability; that is, using condition variables allows the model to better handle novel antigens or antibodies. Even without directly relevant experimental structural data, as long as sufficient amino acid sequences and characteristic information are provided, the model can make reasonable inferences using existing knowledge.
[0103] Specifically, in the process of predicting the structure of immunoglobulin-antigen complexes, the antigen information used to determine the conditional variables requires at least the amino acid sequence of the antigen, as it is the basis for generating the structural model. If the conditional variables further include the three-dimensional structural information of the antigen, the prediction accuracy can be significantly improved. Furthermore, antigenic epitopes are the specific sites in the antigen molecule that specifically recognize and bind to immunoglobulins; therefore, antigenic epitopes are crucial for accurately predicting the binding interface. To ensure accurate prediction, antigenic epitopes can be included as part of the aforementioned second feature. The epitope information of the antigen can be predicted using computational methods. Therefore, the second feature of the antigen used to determine the conditional variables at least includes the amino acid sequence of the antigen, and can also include the structural information of the antigenic epitopes and antigen, depending on actual needs. Further, the first feature of the target immunoglobulin and the second feature of the target antigen are used as conditional variables for predicting the structural data of the antigen-immunoglobulin complex.
[0104] After determining the condition variables through the above embodiments, structural prediction can be performed based on the trained diffusion model (i.e., the structural prediction model). This involves executing steps S220 and S230. Before detailing the 3D structural prediction using the structural prediction model, this paper first introduces an embodiment of the model training.
[0105] In some embodiments, the above diffusion model is a denoising diffusion probabilistic model (DDPM). The training process of DDPM will be described in detail below.
[0106] Figure 6 This is a flowchart illustrating a DDPM training method provided in an embodiment of this application.
[0107] For example, for a task predicting protein structure, the original samples mentioned above are structural data of proteins, with different original samples representing structural data of different proteins. For a task predicting the structure of immunoglobulin-antigen complexes, the original samples mentioned above are structural data of immunoglobulin-antigen complexes, with different original samples representing structural data of different complexes.
[0108] refer to Figure 6 The DDPM process for any original sample in the two prediction tasks described above is divided into two main processes: forward diffusion and reverse generation. In the forward diffusion process, DDPM gradually adds noise to the original sample over T time steps (S610-S620), ultimately reducing the experimental structural data of the original sample to pure noise. Then, in the reverse generation process, DDPM gradually learns how to reverse the process (S630) to generate / predict the structural data of the original sample through T time steps of denoising.
[0109] The forward diffusion process refers to the gradual addition of Gaussian noise to disrupt the original sample (such as the structure of the antigen-immunoglobulin complex) until the data is transformed into a pure noise distribution. Its mathematical form can be expressed as:
[0110]
[0111] Where, β t x represents the noise variance at time step t, used to adjust the amount of noise added at time step t. t Let I represent the noisy data output at time step t, where I is the identity matrix, indicating that the noise is isotropic. Formula (1) represents the noise at the given previous time step x. t-1 In the case of current time step x t The distribution of follows a Gaussian distribution (normal distribution), with a mean of . Covariance is β t I. That is to say, during the forward diffusion process, for x t-1 Adding a certain amount of noise to obtain x t .
[0112] In related technologies, the forward diffusion process of DDPM uses a linear increase in noise, resulting in a large total number of noise addition time steps. Correspondingly, the denoising time steps in the generation stage are also relatively large. To improve model training / inference efficiency, this application provides an improved noise adjustment strategy that controls the noise addition rate by non-linearly scheduling the amount of noise added, effectively improving model inference efficiency. For example, the noise adjustment strategy of this application can compress the 1000-step denoising process of DDPM in related technologies to 200 steps.
[0113] refer to Figure 6 Execute S610: During the forward diffusion process, the noise variance at time step t is controlled by a noise adjustment strategy to obtain the noise variance at time step t, wherein the noise variance increases non-linearly with increasing t.
[0114] In the forward diffusion stage of DDPM, the value of t is 1, ..., T-1, T.
[0115] In an exemplary embodiment, this application further subdivides the loading stage into molecular stages, with different growth rates of noise variance in different sub-stages. Specifically, in S610-1, through a noise adjustment strategy, when t is less than a first threshold, the growth rate of the noise variance is relatively small (e.g., less than a second threshold), in order to preserve structural features in the early stages of the forward diffusion process. It can be seen that in the early stages of the forward diffusion process (when t is small), the improved scheduling strategy makes β... t The smaller the noise, the less noise is added, which helps to preserve more structural information of the original sample. For example, it enables the model to retain the local chemical environment and global topological features of key functional regions during the noise addition process.
[0116] In S610-2, through the noise adjustment strategy, when t is greater than the first threshold, the growth rate of the noise variance is greater than the second threshold, so that the noise addition time step is less than the third threshold. It can be seen that in the later stage of the forward diffusion process (when t is larger), the improved scheduling strategy makes β... t The larger the noise, the faster the data can be converted into pure noise, thus speeding up the process.
[0117] In an exemplary embodiment, the noise adjustment strategy can be expressed as formula (2).
[0118]
[0119] Where, β t β represents the noise variance at time step t; min This represents the minimum noise variance during the entire forward diffusion process. By setting a lower limit for the noise variance, it can be ensured that even in the initial steps of the forward diffusion process, some noise is added; β max This represents the maximum noise variance during the entire forward diffusion process. By setting an upper limit for the noise variance, we can ensure that enough noise is introduced in the last few steps of the forward diffusion process to completely cover the original sample; T represents the total number of time steps.
[0120] Referring to formula (2), this embodiment uses a cosine function instead of a linear function, making the noise addition rate change more smoothly at different time steps and conforming to the distribution characteristics of natural signals. The scheduling strategy shown in formula (2) can control the early addition of noise in the noise addition stage to be slower, which helps to retain more structural details of the original samples, so that the high-frequency details in the noisy data representation can be slowly removed, and avoid being quickly erased by noise. Since the goal of the reverse generation process of DDPM is to reconstruct the original samples step by step from the noise, slow noise addition in the early stage of the forward diffusion stage is beneficial to make the reverse generation process more likely to learn how to recover the structure from the "slightly perturbed" data, thereby improving the overall training stability and generation quality. The scheduling strategy shown in formula (2) can control the later addition of noise in the noise addition stage to be faster, that is, by accelerating the addition of noise, it helps to compress the number of steps. In summary, through the above noise adjustment method, the time steps can be reduced and the training / model inference efficiency can be improved while ensuring the model prediction accuracy.
[0121] In S620, the noisy data representation of the target original sample is determined based on the t-th noise variance and the (t-1)-th noisy data representation of the target original sample, wherein the (t-1)-th noisy data representation is the noisy data representation output at the (t-1)-th time step.
[0122] Referring to formula (1), the original target sample can be represented as x0. In the t-th time step of noise addition, according to the t-th noise variance β t And the noisy data representation x at t-1 t-1 Determine the noisy data representation x of the t-th bit of the original sample. t .
[0123] After adding noise for T time steps of forward diffusion, we obtain x T Approximately a standard normal distribution, it can be called the final noisy data representation or pure noise representation.
[0124] In related technologies, the reverse generation process of DDPM essentially involves transforming the final noisy data representation x T The model performs stepwise denoising over T time steps using a parameterized neural network (e.g., U-Net) to recover the data distribution of the original target sample x0. Specifically, the denoising process relies on the model's predictive ability; at each time step t, the model needs to estimate or predict the noise component ∈ in the current coordinate representation based on the current time step and the input data.
[0125] In this embodiment, the structure prediction process is based on condition variables. Therefore, at time step T of the training phase (the values of t in the reverse generation phase are T, T-1, ..., 1), the x obtained from the forward diffusion needs to be adjusted. TAfter being fused and encoded with the corresponding conditional variables, noise prediction is then performed, specifically predicting how much noise was added at time step T of the forward diffusion stage. It should be noted that for protein structure prediction tasks, the above conditional variables include the first feature. For immunoglobulin-antigen complex structure prediction tasks, the above conditional variables include both the first and second features.
[0126] In some embodiments, a set of training samples containing the original sample includes the original sample x0 and its corresponding condition variable C0. In this case, the condition variable C0 can be determined based on the original sample x0 through the embodiment corresponding to S210. In this case, the parameters of the antibody language model do not need to be optimized during the training process of DDPM, i.e., the parameters of the pre-trained language model can be used, and the model training speed is faster. In some embodiments, a set of training samples containing the original sample includes the original sample x0 and the amino acid sequence of the associated immunoglobulin. In this case, the parameters of the antibody language model will be optimized during the training process of DDPM, which is suitable for more accurate predictions.
[0127] refer to Figure 7 Taking a set of training samples containing the original samples, including the original sample x0 and its corresponding condition variable c, as an example, we will illustrate the reverse generation process of DDPM.
[0128] In S630, at time step t in the reverse generation process, feature fusion and feature encoding are performed on the sample condition variables corresponding to the (t+1)th noisy data representation and the target original sample. The encoded features and time step t are input into the denoising module to predict the tth actual noise added at the corresponding time step in the forward diffusion process, and thus obtain the tth predicted noise.
[0129] refer to Figure 7 At time step T of the reverse generation process, the final data representation x of the target original sample is generated. T The condition variable c corresponding to the sample is input into the feature fusion module 121, and the sample fusion feature is obtained after fusion processing.
[0130] The above sample fusion features are input into the feature encoding module 122, and the sample encoded features y are obtained after feature encoding. T .
[0131] Encode the sample feature y T And the current time step T input denoising module 123, after denoising processing, yields the Tth prediction noise at time step T. Furthermore, based on the T-th prediction noise and the final data representation x T It can be determined that the denoised data representation z of the Tth denoising layer is... T .
[0132] refer to Figure 7 In the reverse generation process, at time step t (t takes values T, T-1, ..., 1), the denoised data representation z from the previous time step is... t+1 The condition variable c corresponding to the sample is input into the feature fusion module 121, and the sample fusion feature is obtained after fusion processing.
[0133] The above sample fusion features are input into the feature encoding module 122, and the sample encoded features y are obtained after feature encoding. t .
[0134] Encode the sample feature y Tt The current time step t is input to the denoising module 123, and after denoising processing, the t-th prediction noise at time step t is obtained. Furthermore, z is represented based on the t-th prediction noise and the denoised data from the previous time step. t+1 It can be determined that the denoised data representation z at time t is... t .
[0135] Refer again Figure 6 In S640, a mean squared error loss function is constructed based on the t-th predicted noise and the t-th actual noise; in S650, the parameters of DDPM are optimized based on the mean squared error loss function.
[0136] The loss function mentioned above can be expressed as formula (3).
[0137]
[0138] in, This indicates that for time step t, the original sample And the expectation of noise ∈; Indicates cumulative product. Represents all α values from the start to time step t. s The product of β s This represents the noise variance at each time step.
[0139] The objective of the loss function described above is to minimize the noise in the model's predictions. The difference between the noise and the actual noise added. In this way, the model learns how to accurately estimate the added noise based on the input data and time steps, and then learns to reverse the entire noise addition process, ultimately generating an output close to the original sample without noise.
[0140] The above embodiments combined Figure 6 and Figure 7The training process of the structure prediction model is introduced, specifically including parameter optimization of the feature fusion module, feature encoding module, and denoising module. In the following examples, S220 and S230 illustrate an implementation of structure prediction for immune proteins based on the optimized structure prediction model.
[0141] In S220, the conditional variable and randomly sampled raw noise are input into the structure prediction model. In S230, the denoising module of the structure prediction model performs multiple denoising processes on the raw noise based on the conditional variable to obtain the predicted three-dimensional structure of the target immunoglobulin.
[0142] Figure 8 This is a flowchart illustrating a denoising method in the structure prediction process provided in this application embodiment. The denoising method shown in the figure can be applied to both protein structure prediction tasks and immunoglobulin-antigen complex structure prediction tasks. Figure 8 The embodiment shown takes the i-th (i is an integer greater than 1) denoising process as an example to illustrate the specific implementation of the denoising process.
[0143] In S810, the output features of the (i-1)th denoising process are determined to obtain the (i-1)th noisy data representation.
[0144] When determining the input information for the current denoising process, it is necessary to first obtain the output features of the previous denoising process. Furthermore, the input information for the current denoising process is determined by combining the output features of the previous denoising process with condition variables. Thus, in each denoising process, the amount of denoising is guided by condition variables. This allows the model to generate structural data that conforms to the aforementioned condition variables during inference, leading to more accurate generation of the three-dimensional structure of complexes or proteins.
[0145] In an exemplary embodiment, Figure 9 This is a schematic diagram of a method for predicting the structure of an immunoglobulin-antigen complex provided in an embodiment of this application.
[0146] refer to Figure 9 The amino acid sequence of the target immunoglobulin is input into the antibody language model 110, and after processing, the corresponding first feature is obtained. This feature will be used as a conditional variable. In the first denoising process, the amino acid sequence 91 of the antigen and the original noise 12 are processed by the Combineinputts module 72. The purpose of the Combineinputts module 72 is to integrate all relevant input information and provide a unified representation for subsequent processing steps.
[0147] Furthermore, the first feature and the output feature of the CombineInputs module 92 (containing the aforementioned second feature and the original noise data representation) are input to the feature fusion module 121 to perform feature fusion on the input information. The feature fusion module 121 is responsible for integrating feature information from different sources to form a unified feature representation, ensuring that all relevant information is considered and avoiding omissions. The fused features are then input to the feature encoding module 122, where feature encoding further extracts deeper feature representations, enabling the model to better understand and utilize this information, thereby improving the accuracy and reliability of the final prediction. In this embodiment, through the coordinated work of feature fusion and feature encoding, the system can effectively handle complex immunoglobulin-antigen complex structure prediction tasks, which is beneficial for generating high-quality prediction results.
[0148] In the first denoising process, the features output by the feature encoding module 122 (the first encoded features) are input to the denoising module 123. The denoising module 123 performs noise prediction based on the first encoded features to determine the estimated noise information in the first denoising process. When the denoising module 123 performs noise prediction, the model utilizes trained parameters, which are learned during the training phase by minimizing a loss function (such as the mean squared error loss function shown in formula (3)), thereby ensuring prediction accuracy. Further, the denoising module 123 updates the first encoded features based on the estimated noise information to obtain a first noisy data representation.
[0149] Therefore, in the i-th denoising process, the noisy data representation of the (i-1)th time can be used as the starting point.
[0150] In S820, the input features for the i-th denoising process are determined based on the (i-1)-th noisy data representation and the condition variable, thus obtaining the i-th input feature.
[0151] For example, to ensure that the condition variable efficiently guides the model's structured data generation process, this application embodiment divides the combination of the output features from the previous denoising process and the condition variable into two processing stages. In the first stage, feature fusion is performed to integrate the two sets of information into a unified feature representation, i.e., the fused feature. In the second stage, feature encoding is performed to encode the fused feature, which can further extract deeper feature representations, enabling the model to better understand and utilize the information contained in the features.
[0152] Continue to refer to Figure 9In S820-1, the feature fusion module 121 of the structural prediction model fuses the (i-1)th noisy data representation with the condition variable to obtain the (i-1)th fused feature. This integrates the condition variable with the noisy data representation output from the previous step, resulting in a unified feature representation, the (i-1)th fused feature. In S820-2, the feature encoding module 122 encodes the (i-1)th fused feature to obtain the ith input feature. Encoding the (i-1)th fused feature allows for further extraction of deeper feature representations, enabling the model to better understand and utilize the information contained in the features based on the ith input feature, thus improving the accuracy and reliability of the final prediction result.
[0153] Furthermore, the denoising module 123 processes the i-th input feature to obtain the i-th noisy data representation. Specifically, this includes steps S830 and S840.
[0154] In S830, noise prediction is performed by the denoising module to determine the estimated noise information in the i-th denoising process based on the i-th input feature, thereby obtaining the i-th estimated noise information. In S840, the denoising module updates the (i-1)-th noisy data representation according to the i-th estimated noise information, thereby obtaining the i-th noisy data representation.
[0155] During the inference phase of the aforementioned structure prediction model, denoising is performed at preset intervals (e.g., once, three times, etc.) to determine whether the latest noisy data representation meets preset conditions, such as whether the confidence level reaches a confidence threshold, in order to evaluate the quality and reliability of the currently generated noisy data representation. If the preset conditions are met, the latest noisy data representation can be used as the predicted 3D structure for the current prediction task. That is, when the prediction task is a complex structure prediction, the predicted 3D structure of the target immunoglobulin-antigen complex is obtained; or when the prediction task is a protein structure prediction, the predicted 3D structure of the target protein is obtained. Because the latest noisy data representation has a high confidence level, the predicted 3D structure has high accuracy and reliability.
[0156] If the latest noisy data does not meet the preset conditions, feedback information is determined based on the current confidence level, and the next denoising process is performed based on the feedback information to achieve dynamic adjustment of the generation process, thereby improving the accuracy and efficiency of the final prediction result.
[0157] As can be seen, in this embodiment of the application, even when the latest noisy data representation does not meet the preset conditions, feedback information can still be provided for subsequent denoising steps, thereby helping the model dynamically adjust its denoising strategy based on the quality of the currently generated structure. For example, the denoising step size can be dynamically adjusted.
[0158] Specifically, if the confidence level of the i-th noisy data representation, although not meeting the confidence threshold, is relatively large (e.g., greater than the first threshold), it indicates that the current structure is close to the true structure. In this case, the denoising step size can be increased (i.e., the number of denoising processes is reduced) to accelerate convergence without sacrificing accuracy. Therefore, the i-th feedback information generated in this case is used to indicate increasing the denoising step size. Specifically, in the (i+1)-th denoising process, if the estimated noise amount obtained in the noise prediction stage is X and the original denoising step size is 1, where the original denoising step size of 1 indicates that the noise removal amount in the noisy data update stage is X, increasing the denoising step size will indicate that the noise removal amount in the noisy data update stage is greater than X. It can be seen that increasing the denoising step size means removing more noise to improve the model inference efficiency.
[0159] If the confidence level of the i-th noisy data representation is very low (e.g., less than the first threshold mentioned above), it indicates that there is still a significant gap between the current structure and the target structure. In this case, the iteration step size can be reduced for more refined adjustments. Therefore, the i-th feedback information generated in this situation is used to indicate a reduction in the denoising step size. Specifically, in the (i+1)-th denoising process, if the estimated noise level obtained in the noise prediction stage is X and the original denoising step size is 1, reducing the denoising step size will indicate that the noise removal amount in the noisy data update stage is less than X. It is evident that reducing the denoising step size means removing smaller noise to more precisely approximate the true structure.
[0160] In an exemplary embodiment, if the confidence level of the i-th noisy data representation is very low (e.g., less than the first threshold mentioned above), the feedback information may further include reminders to update condition variables and / or reminders to introduce additional constraints. Updating condition variables (such as characteristics of antibodies and antigens) ensures that they correctly guide the model to learn in the right direction. Introducing additional physical or chemical constraints forces the generated structure to meet certain biological rationality requirements, such as interatomic distance restrictions, stereochemical rules, etc.
[0161] As can be seen, in the model inference stage, the embodiments of this application can achieve efficient generation of antigen-immunoglobulin interface conformation by dynamically adjusting the strategy.
[0162] This application provides a method for predicting immunoglobulin structures based on a generative deep learning model. By constructing a three-dimensional conformational generation model, it achieves high-precision and low-computational-cost prediction of unknown immunoglobulin structures. Specifically, the structure prediction model is obtained by training a diffusion model. The diffusion model can reduce its dependence on large-scale original samples (experimental structural data) through a noise-adding and denoising training paradigm. Specifically, the noise-adding process uses artificially constructed noise data to increase the diversity of training samples. Even with limited experimental structural data, the model can learn to recover the true structure from various noise levels, improving the model's generalization ability. For example, it can support the prediction of immunoglobulin binding interfaces for novel antigens (such as undisclosed epitopes).
[0163] Simultaneously, a deep learning-driven diffusion model is employed, based on conditional variables (such as the spatial characteristics of antigen epitopes and the sequence-structure coupling information of immunoglobulin complementarity-determining regions), to generate a three-dimensional conformation of the antigen-immunoglobulin binding interface. The prediction accuracy is significantly improved compared to traditional homology modeling methods and deep learning methods, with an improvement of approximately 30%. Furthermore, the embodiments of this application can extend the application of immunoglobulin drug development. For example, the prediction results can be combined with immunoglobulin humanization design to output candidate structures that can guide the engineering modification of bispecific immunoglobulins, shortening the immunoglobulin drug discovery cycle by 40%-60%.
[0164] In the process of determining conditional variables, introducing an antibody language model that includes a pre-trained language model can effectively utilize knowledge within the domain.
[0165] This application provides an end-to-end generative model that, by utilizing a diffusion model step compression method, reduces the single-structure prediction time to the minute level, meeting the requirements for screening millions of candidate libraries. It supports batch input of immunoglobulin sequences, achieving a single-structure prediction time of 3 minutes based on the end-to-end generative model (approximately 200 times faster than the AlphaFold-Multimer approach in related technologies), thus meeting the industrial-grade requirements for screening millions of candidate libraries.
[0166] The above text combined Figures 2 to 9 This document describes embodiments of the protein structure prediction method of this application. The following section, in conjunction with… Figure 10 Examples of the protein structure prediction device of this application are described.
[0167] Figure 10 This is a schematic block diagram of a protein structure prediction device 1000 provided in an embodiment of this application, wherein the device 1000 can be configured in an electronic device.
[0168] refer to Figure 10The protein structure prediction device 1000 includes: a determination module 1001, an input module 1002, and a prediction module 1003.
[0169] The determination module 1001 is used to determine a first feature based on the amino acid sequence of the target protein, wherein the first feature is used to determine a condition variable; the input module 1002 is used to input the condition variable and randomly sampled raw noise into the structure prediction model; the prediction module 1003 is used to perform multiple denoising processes on the raw noise based on the condition variable through the structure prediction model to obtain the predicted three-dimensional structure of the target protein; wherein the structure prediction model is obtained by training a diffusion model, and the training samples of the diffusion model are obtained by adding artificial noise to the raw samples.
[0170] In an exemplary embodiment, based on the above scheme, the prediction module 1003 includes: a first determining unit, a second determining unit, and a processing unit.
[0171] The first processing unit is used to determine the output features of the (i-1)th denoising process to obtain the (i-1)th noisy data representation, where i is an integer greater than 1; the second determining unit is used to determine the input features of the i-th denoising process based on the (i-1)th noisy data representation and the condition variable to obtain the i-th input feature; and the processing unit is used to process the i-th input feature through the denoising module of the structure prediction model to obtain the i-th noisy data representation.
[0172] In an exemplary embodiment, based on the above scheme, the processing unit is specifically configured to: perform noise prediction through the denoising module to determine the estimated noise information in the i-th denoising process based on the i-th input feature, thereby obtaining the i-th estimated noise information; and update the (i-1)-th noisy data representation through the denoising module according to the i-th estimated noise information, thereby obtaining the i-th noisy data representation.
[0173] In an exemplary embodiment, based on the above scheme, the second determining unit is specifically used to: perform feature fusion on the (i-1)th noisy data representation and the condition variable through the feature fusion module of the structure prediction model to obtain the (i-1)th fused feature; and encode the (i-1)th fused feature through the feature encoding module of the structure prediction model to obtain the i-th input feature.
[0174] In an exemplary embodiment, based on the above scheme, the second processing unit includes a fusion subunit and an encoding subunit.
[0175] The fusion subunit is used to: after the input module 1002 inputs the condition variable and the randomly sampled original noise into the structure prediction model, perform feature fusion on the data representation of the original noise and the condition variable through the feature fusion module of the structure prediction model to obtain a first fused feature; the encoding subunit is used to: encode the first fused feature through the feature encoding module of the structure prediction model to obtain a first encoded feature; the processing unit is specifically used to: process the first encoded feature through the denoising module of the structure prediction model to obtain the output feature of the first denoising process.
[0176] In an exemplary embodiment, based on the above scheme, the prediction module 1003 includes a judgment unit and a feedback unit;
[0177] The judgment unit is used to: after processing the i-th input feature through the denoising module of the structure prediction model to obtain the i-th noisy data representation, determine whether the i-th noisy data representation meets a preset condition; and, if the preset condition is met, determine the predicted three-dimensional structure of the target protein based on the i-th noisy data; the feedback unit is used to: generate i-th feedback information if the i-th noisy data does not meet the preset condition, and perform the (i+1)-th denoising process based on the i-th feedback information.
[0178] In an exemplary embodiment, based on the above scheme, the determination unit is specifically used to: determine the confidence level of the i-th noisy data representation; and determine whether the confidence level of the i-th noisy data representation meets a preset confidence threshold.
[0179] The feedback unit is specifically used to: if the i-th confidence level is greater than the first threshold, then the i-th feedback information is used to indicate increasing the denoising step size, where the first threshold is less than the confidence level threshold; if the i-th confidence level is less than the first threshold, then the i-th feedback information is used to indicate decreasing the denoising step size.
[0180] In an exemplary embodiment, based on the above scheme, the diffusion model is a denoised diffusion probability model (DDPM), and the device further includes a training module; the training module includes a first determining unit, a first processing unit, a prediction unit, a construction unit, and an optimization unit.
[0181] The first determining unit is used to control the noise variance at time step t through a noise adjustment strategy during the forward diffusion process of DDPM, and obtain the noise variance at time step t, wherein the noise variance increases nonlinearly with increasing t.
[0182] The first processing unit is used to determine the noisy data representation of the target original sample based on the noise variance of the tth time step and the noisy data representation of the target original sample at the (t-1)th time step, wherein the noisy data representation at the (t-1)th time step is the noisy data representation output at the (t-1)th time step, and wherein the value of t during the forward diffusion process is 1, ..., T-1, T.
[0183] The prediction unit is used to perform feature fusion and feature encoding on the noisy data representation at time step t in the reverse generation process and the sample condition variables corresponding to the target original sample. The encoded features and time step t are input into the denoising module to predict the actual noise added at the corresponding time step in the forward diffusion process, and obtain the predicted noise at time step t. In the reverse generation process, the value of t is T, T-1, ..., 1.
[0184] A construction unit is configured to construct a mean squared error loss function based on the t-th predicted noise and the t-th actual noise; and an optimization unit is configured to optimize the parameters of DDPM based on the mean squared error loss function.
[0185] In an exemplary embodiment, based on the above scheme, the first determining unit is specifically configured to: control the growth rate of the noise variance to be less than a second threshold when t is less than a first threshold, so as to preserve structural features in the early stage of the forward diffusion process; and control the growth rate of the noise variance to be greater than the second threshold when t is greater than the first threshold, so as to make the noise addition time step less than a third threshold.
[0186] In an exemplary embodiment, based on the above scheme, the noise adjustment strategy controls the growth of noise variance by means of a cosine function.
[0187] In an exemplary embodiment, based on the above scheme, the target protein includes a target immunoglobulin; the determination module 1001 includes: a vectorization unit, an extraction unit, and a determination unit.
[0188] The vectorization unit is used to vectorize the amino acid sequence of the target immunoglobulin through a pre-trained language model to obtain an initial sequence representation, which includes local chemical environment information and global topological information of the immunoglobulin complementarity determination region (CDR).
[0189] An extraction unit is used to extract features from the initial sequence representation to obtain the CDR-H3 features and / or frame region constraint features of the target immunoglobulin, wherein the CDR-H3 features characterize the local chemical environment information of the CDR-H3 loop, and the frame region constraint features represent the spatial constraint relationship between the frame region and the CDR.
[0190] A determining unit is configured to determine a first feature of the target immunoglobulin based on the CDR-H3 feature and / or the frame region constraint feature.
[0191] In an exemplary embodiment, based on the above scheme, the extraction unit is specifically used to: focus on the local residue relationships within the CDR based on the first type of attention head in the multi-head attention mechanism to obtain CDR-H3 features, wherein the CDR-H3 features reflect the local chemical environment information of the CDR-H3 ring; and identify the long-range dependency relationship between the frame region and the CDR through the second type of attention head in the multi-head attention mechanism to obtain frame region constraint features.
[0192] In an exemplary embodiment, based on the above scheme, the predicted three-dimensional structure includes: the three-dimensional structure of the complex after binding with the target antigen; the determining module is further configured to: determine a second feature based on the amino acid sequence of the target antigen; and determine the condition variable based on the first feature of the target immunoglobulin and the second feature of the target antigen.
[0193] It should be understood that, as Figure 10 The embodiment of the protein structure prediction device shown corresponds to the protein structure prediction method embodiment described above, and a similar description can be found in the method embodiment. To avoid repetition, it will not be repeated here. Specifically, through... Figure 10 The information interaction between the various modules in the protein structure prediction device shown can execute the embodiments of the protein structure prediction method described above, through methods such as... Figure 10 The information interaction between the various modules in the protein structure prediction device shown can execute the embodiments of the protein structure prediction method described above. For the sake of brevity, the method embodiments corresponding to the aforementioned and other operations and / or functions of each module in the device will not be described again here.
[0194] The above description, in conjunction with the accompanying drawings, describes the operation and maintenance related apparatus of the software agent according to the embodiments of this application from the perspective of functional modules. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by the integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.
[0195] This application also provides an electronic device.
[0196] Figure 11 This is a schematic block diagram of an electronic device 1100 provided in an embodiment of this application. As described above, the operation and maintenance related devices of the software agent can be deployed in, for example... Figure 11 The electronic device shown can therefore be used to perform the aforementioned protein structure prediction method.
[0197] like Figure 11 As shown, the electronic device 1100 may include:
[0198] The system includes a memory 1110 and a processor 1120. The memory 1110 stores a computer program 1130 and transfers the program code 1130 to the processor 1120. In other words, the processor 1120 can call and run the computer program 1130 from the memory 1110 to implement the methods in the embodiments of this application.
[0199] For example, the processor 1120 can be used to execute the steps in the protein structure prediction method according to the instructions in the computer program 1130, or to execute the steps in the protein structure prediction method.
[0200] In some embodiments of this application, the processor 1120 may include, but is not limited to:
[0201] General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
[0202] In some embodiments of this application, the memory 1110 includes, but is not limited to:
[0203] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).
[0204] In some embodiments of this application, the computer program 1130 may be divided into one or more modules, which are stored in the memory 1110 and executed by the processor 1120 to complete the protein structure prediction method provided in this application, or to complete the steps in the protein structure prediction method described above. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 1130 in the electronic device.
[0205] like Figure 11As shown, the electronic device 1100 may further include:
[0206] Transceiver 1140, which can be connected to processor 1120 or memory 1110.
[0207] The processor 1120 can control the transceiver 1140 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 1140 may include a transmitter and a receiver. The transceiver 1140 may further include antennas, and the number of antennas may be one or more.
[0208] It should be understood that the various components in the electronic device 1100 are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.
[0209] According to one aspect of this application, a computer storage medium is provided that stores a computer program thereon, which, when executed by a computer, enables the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.
[0210] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method described in the above-described method embodiments.
[0211] In other words, when implemented using software, it can be implemented wholly or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., Digital Video Disc (DVD)), or a semiconductor medium (e.g., Solid State Disk (SSD)).
[0212] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0213] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules 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; the indirect coupling or communication connection between devices or modules may be electrical, mechanical, or other forms.
[0214] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0215] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the structure of a protein, characterized in that, The method includes: Based on the amino acid sequence of the target protein, a first feature is determined, wherein the first feature is used to determine conditional variables; The conditional variables and randomly sampled raw noise are input into the structural prediction model; The original noise is denoised multiple times using the structure prediction model based on the condition variables to obtain the predicted three-dimensional structure of the target protein. The structure prediction model is obtained by training a diffusion model, and the training samples of the diffusion model are obtained by adding artificial noise to the original samples.
2. The method according to claim 1, characterized in that, The original noise is denoised multiple times using the structure prediction model based on the conditional variables to obtain the predicted three-dimensional structure of the target protein, including: Determine the output features of the (i-1)th denoising process to obtain the (i-1)th noisy data representation, where i is an integer greater than 1; Based on the (i-1)th noisy data representation and the condition variable, determine the input features for the i-th denoising process to obtain the i-th input feature; The denoising module of the structure prediction model processes the i-th input feature to obtain the i-th noisy data representation.
3. The method according to claim 2, characterized in that, The denoising module of the structure prediction model processes the i-th input feature to obtain the i-th noisy data representation, including: The noise prediction is performed by the denoising module to determine the estimated noise information in the i-th denoising process based on the i-th input feature, thereby obtaining the i-th estimated noise information; The denoising module updates the (i-1)th noisy data representation based on the i-th estimated noise information to obtain the i-th noisy data representation.
4. The method according to claim 2, characterized in that, Based on the (i-1)th noisy data representation and the condition variable, the input features for the i-th denoising process are determined, resulting in the i-th input features, including: The feature fusion module of the structural prediction model performs feature fusion on the (i-1)th noisy data representation and the condition variable to obtain the (i-1)th fused feature. The i-th fused feature is encoded by the feature encoding module of the structure prediction model to obtain the i-th input feature.
5. The method according to claim 2, characterized in that, After inputting the conditional variables and randomly sampled raw noise into the structural prediction model, the method further includes: The feature fusion module of the structural prediction model performs feature fusion on the data representation of the original noise and the condition variables to obtain the first fused feature. The first fused feature is encoded by the feature encoding module of the structure prediction model to obtain the first encoded feature; The first encoded feature is processed by the denoising module of the structure prediction model to obtain the output feature of the first denoising process.
6. The method according to claim 2, characterized in that, After processing the i-th input feature through the denoising module of the structure prediction model to obtain the i-th noisy data representation, the method further includes: Determine whether the i-th noisy data representation meets the preset conditions; Under the condition that the preset conditions are met, the predicted three-dimensional structure of the target protein is determined based on the i-th noisy data; If the i-th noisy data does not meet the preset conditions, generate the i-th feedback information and perform the (i+1)-th denoising process based on the i-th feedback information.
7. The method according to claim 6, characterized in that, Determining whether the i-th noisy data representation meets preset conditions includes: Determine the confidence level of the i-th noisy data representation; Determine whether the confidence level of the i-th noisy data representation meets the preset confidence threshold; If the i-th noisy data does not meet the preset conditions, generate the i-th feedback information, including: If the i-th confidence level is greater than the first threshold, then the i-th feedback information is used to indicate increasing the denoising step size, and the first threshold is less than the confidence level threshold. If the i-th confidence level is less than the first threshold, then the i-th feedback information is used to indicate a reduction in the denoising step size.
8. The method according to claim 1, characterized in that, The diffusion model is the Denoising Diffusion Probability Model (DDPM), and the method further includes: During the forward diffusion process of DDPM, the noise variance at time step t is controlled by a noise adjustment strategy to obtain the noise variance at time step t, wherein the noise variance increases non-linearly with the increase of t. Based on the noise variance at time t and the noisy data representation at time t-1 of the original target sample, the noisy data representation at time t of the original target sample is determined, wherein the noisy data representation at time t-1 is the noisy data representation output at time t-1, wherein the value of t during the forward diffusion process is 1, ..., T-1, T. At time step t in the reverse generation process, feature fusion and feature encoding are performed on the noisy data representation at time step t+1 and the sample condition variables corresponding to the target original sample. The encoded features and time step t are input into the denoising module to predict the actual noise added at time step t in the forward diffusion process, and the predicted noise at time step t is obtained. In the reverse generation process, the value of t is T, T-1, ..., 1. Based on the predicted noise at the t-th digit and the actual noise at the t-th digit, a mean squared error loss function is constructed; The parameters of DDPM are optimized based on the mean squared error loss function.
9. The method according to claim 8, characterized in that, The noise variance at time step t is controlled through a noise adjustment strategy, including: By employing a noise adjustment strategy, when t is less than a first threshold, the growth rate of the noise variance is less than a second threshold, so as to preserve structural features in the early stages of the forward diffusion process. By employing the noise adjustment strategy, when t is greater than the first threshold, the growth rate of the noise variance is greater than the second threshold, so that the noise addition time step is less than the third threshold.
10. The method according to claim 9, characterized in that, In the noise adjustment strategy, the growth of noise variance is controlled by using a cosine function.
11. The method according to any one of claims 1 to 10, characterized in that, The target protein includes a target immunoglobulin; based on the amino acid sequence of the target protein, a first characteristic is determined, including: The amino acid sequence of the target immunoglobulin is vectorized using a pre-trained language model to obtain an initial sequence representation, which includes local chemical environment information and global topological information of the immunoglobulin complementarity-determining region (CDR). Feature extraction is performed on the initial sequence representation to obtain the CDR-H3 feature and / or frame region constraint feature of the target immunoglobulin, wherein the CDR-H3 feature characterizes the local chemical environment information of the CDR-H3 loop, and the frame region constraint feature represents the spatial constraint relationship between the frame region and the CDR. The first feature of the target immunoglobulin is determined based on the CDR-H3 feature and / or the framework region constraint feature.
12. The method according to claim 11, characterized in that, The feature extraction of the initial sequence representation includes: Based on the first type of attention head in the multi-head attention mechanism, the local residue relationships within the CDR are focused to obtain CDR-H3 features, which reflect the local chemical environment information of the CDR-H3 ring; By using the second type of attention head in the multi-head attention mechanism, the long-range dependency between the frame region and the CDR is identified to obtain the frame region constraint features.
13. The method according to claim 11, characterized in that, The predicted three-dimensional structure includes: the three-dimensional structure of the complex after binding with the target antigen; the method further includes: The second characteristic is determined based on the amino acid sequence of the target antigen; The conditional variable is determined based on the first characteristic of the target immunoglobulin and the second characteristic of the target antigen.
14. A protein structure prediction device, characterized in that, The device includes: A determination module is used to determine a first feature based on the amino acid sequence of the target protein, wherein the first feature is used to determine a condition variable; The input module is used to input the conditional variables and randomly sampled raw noise into the structural prediction model; The prediction module is used to perform multiple denoising processes on the original noise based on the condition variables using the structure prediction model to obtain the predicted three-dimensional structure of the target protein. The structure prediction model is obtained by training a diffusion model, and the training samples of the diffusion model are obtained by adding artificial noise to the original samples.
15. A computer-readable storage medium, characterized in that, Used to store computer programs; The computer program causes the computer to perform the protein structure prediction method as described in any one of claims 1 to 13.
16. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the protein structure prediction method as described in any one of claims 1 to 13.
17. A computer program product, characterized in that, It includes computer program instructions that cause a computer to perform the protein structure prediction method as described in any one of claims 1 to 13.
