T cell receptor sequence and structure co-design method based on diffusion model

Through a diffusion model-based method, combining the sequence and structural information of T cell receptors, an efficient and accurate TCR sequence and structure is designed, solving the problem of lengthy and inefficient design process in the prior art.

CN119049588APending Publication Date: 2024-11-29ZHIFU TECHNOLOGY CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411171993.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The T cell receptor (TCR) design method in the prior art fails to effectively bind sequence information and structural information, resulting in a lengthy and inefficient design process.

Method used

Using a diffusion model-based method, the data set is screened by traversing the VDJdb database, the diffusion model is constructed, the amino acid sequence is generated using U-ViT network and miniMPNN, and affinity prediction is combined with a multi-layer perceptron to deduce the diffusion process to generate TCR sequences and structures.

Benefits of technology

The efficiency and accuracy of TCR design are improved, and by combining sequence and structural information for collaborative design, TCR sequences and structures with specific affinity conditions are generated.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119049588A_ABST
    Figure CN119049588A_ABST
Patent Text Reader

Abstract

The invention provides a T cell receptor sequence and structure co-design method based on a diffusion model, and relates to the technical field of artificial intelligence drug research and development, and the method comprises the steps: traversing a VDJdb database for screening, constructing a plurality of data sets, and generating a noise data set; de-noising by adopting a U-ViT network to generate an amino acid sequence; designing a comprehensive loss function, constructing an expert model, and performing combined affinity prediction through a multilayer perceptron to generate an affinity condition; and generating a TCR sequence and a TCR structure by adopting a random sampler in combination with a guidance gradient. According to the method and the device, the technical problems of lengthy design process and low efficiency caused by separation processing of the TCR sequence and the structure information in the prior art can be solved, the sequence and the structure are jointly designed by combining the TCR sequence and the structure information and utilizing advanced data processing and model construction technologies, and the efficiency and the accuracy of TCR design are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence drug development technology, and in particular to a T cell receptor sequence and structure co-design method based on a diffusion model. Background Art

[0002] T cell receptors (TCRs) play a vital role in the immune system, responsible for identifying and responding to pathogens and malignant cells. TCRs enable T cells to specifically recognize and eliminate infected or cancerous cells, and are key to immune surveillance and treatment. TCRs recognize peptide-major histocompatibility complex (pMHC) through the V(D)J recombination process, generating diverse sequences in the complementarity determining regions (CDRs). Each TCR consists of an α chain and a β chain, each with three CDRs (CDR1, CDR2, and CDR3), of which CDR3B is particularly important in specific recognition. At present, existing TCR design methods often only consider sequence information or structural information, but fail to effectively combine the two, resulting in insufficient performance of the designed TCR in practical applications. There is a lack of effective end-to-end algorithms that can simultaneously consider other parts of the TCR and the CDR region for the coordinated design of sequence and structure.

[0003] In summary, the prior art has a technical problem that the design process is lengthy and inefficient due to the separate processing of TCR sequence and structure information. Summary of the invention

[0004] The purpose of this application is to provide a T cell receptor sequence and structure co-design method based on a diffusion model to solve the technical problem in the prior art that the design process is lengthy and inefficient due to the separate processing of TCR sequence and structure information.

[0005] In view of the above problems, the present application provides a T cell receptor sequence and structure co-design method based on a diffusion model, wherein the T cell receptor sequence and structure co-design method based on a diffusion model includes: traversing the VDJdb database for screening, constructing multiple data sets, and the multiple data sets include a training data set, a verification data set, and a test data set; constructing a diffusion model based on the training data set, the verification data set, and the test data set to generate a noise data set; using the U-ViT network to denoise the noise data set, and combining miniMPNN to generate an amino acid sequence; designing a comprehensive loss function based on the amino acid sequence, constructing an expert model, combining the hidden features of miniMPNN and the Embedding of pMHC, and predicting the binding affinity through a multi-layer perceptron to generate affinity conditions; deriving the diffusion process with the affinity conditions through SDEs, and using a random sampler combined with a guiding gradient to generate a TCR sequence and structure.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] By traversing the VDJdb database for screening, multiple data sets are constructed, and the multiple data sets include a training data set, a validation data set, and a test data set; a diffusion model is constructed based on the training data set, the validation data set, and the test data set to generate a noise data set; the noise data set is denoised using the U-ViT network, and an amino acid sequence is generated in combination with miniMPNN; a comprehensive loss function is designed based on the amino acid sequence, an expert model is constructed, and the binding affinity is predicted by a multi-layer perceptron in combination with the hidden features of miniMPNN and the Embedding of pMHC to generate affinity conditions; the diffusion process with the affinity conditions is derived through SDEs, and a random sampler is used in combination with a guiding gradient to generate a TCR sequence and structure. In other words, by combining TCR sequence and structure information, and using advanced data processing and model building techniques, the sequence and structure are jointly designed, thereby improving the efficiency and accuracy of TCR design.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0010] Figure 1 A schematic diagram of a T cell receptor sequence and structure co-design method based on a diffusion model in this application;

[0011] Figure 2 A schematic diagram of the process of generating amino acid sequences in a T cell receptor sequence and structure co-design method based on a diffusion model in the present application. DETAILED DESCRIPTION

[0012] This application provides a T cell receptor sequence and structure co-design method based on a diffusion model, which solves the technical problem in the prior art that the design process is lengthy and inefficient due to the separate processing of TCR sequence and structure information. By combining TCR sequence and structure information and using advanced data processing and model building technology, the sequence and structure are co-designed, which improves the efficiency and accuracy of TCR design.

[0013] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0014] For examples, please see the attached Figure 1 The present application provides a T cell receptor sequence and structure co-design method based on a diffusion model, wherein the T cell receptor sequence and structure co-design method based on a diffusion model specifically comprises the following steps:

[0015] Step 1: Traverse the VDJdb database for screening and construct multiple data sets, which include a training data set, a verification data set, and a test data set.

[0016] Specifically, the VDJdb database was traversed to extract multiple sample data sets containing complete TCR sequences, including CDR3 and V / J gene segments. The classification of species, confidence, and MHC category was determined, and cluster analysis was performed on multiple sample data sets. Similar TCR sequences were grouped into the same cluster to generate multiple sample data clusters. Multiple sample data clusters were divided to determine training data sets, validation data sets, and test data sets. The quality and diversity of the data sets were ensured, which helped to train a model that can accurately predict and design TCR sequences and structures.

[0017] Step 2: construct a diffusion model based on the training data set, the verification data set, and the test data set to generate a noise data set.

[0018] Specifically, a continuous time framework for noise injection and removal is provided through stochastic differential equation modeling. The training data set, validation data set, and test data set are combined for model training, parameter adjustment, and performance evaluation. The diffusion model is a deep learning model that generates data through a reverse process. Its principle is to generate high-quality samples through gradual denoising. In drug design, the diffusion model generates new drug molecules with potential therapeutic effects by learning molecular structures, significantly improving the efficiency and accuracy of drug discovery. Define a forward SDE expression, gradually add noise to the data, and transform it into a simple prior distribution, such as a Gaussian distribution. The diffusion model is simulated through the forward SDE expression, noise is injected and simulated data is generated, protein sequences and structures containing noise are generated, and the noise data set is determined. The noise data set will be used to train denoising models, such as the U-ViT network or other deep learning architectures, to restore the original TCR sequence and structure. The generated noise data set helps improve the training efficiency and performance of the model, enabling it to predict and generate data more accurately.

[0019] Step 3: Use the U-ViT network to denoise the noise data set and combine it with miniMPNN to generate amino acid sequences.

[0020] Specifically, the U-ViT network is used to denoise the noisy dataset. The U-ViT network is a deep learning network that is particularly suitable for image segmentation and denoising tasks. Through multiple layers of self-attention layers and residual noise conditional transformer layers, noise is effectively removed and clear image data is restored. The denoised dataset is passed as input data to the miniMPNN model, which receives these data and extracts protein backbone features from them. miniMPNN is a simplified version of ProteinMPNN that retains the core principles of message passing and autoregressive decoding but reduces computational complexity. The message passing mechanism is used to update the protein backbone features of each node to capture local and global information of the protein structure. Based on the extracted local and global information, the autoregressive decoding method is used to gradually predict the position of each amino acid to generate an amino acid sequence. By using the U-ViT network for denoising and combining the message passing mechanism and autoregressive decoding method of the miniMPNN model, the amino acid sequence is generated more accurately, which helps to improve the efficiency of protein engineering and drug design.

[0021] Step 4: Based on the amino acid sequence, a comprehensive loss function is designed to construct an expert model, and the hidden features of miniMPNN and the Embedding of pMHC are combined to predict the binding affinity through a multi-layer perceptron to generate affinity conditions.

[0022] Specifically, the U-ViT network and miniMPNN model are used to generate amino acid sequences from noisy data. A comprehensive loss function is designed by combining the sequence prediction loss function, the structure denoising score matching loss function, and the binding affinity prediction loss function. It usually involves using deep learning techniques such as multi-layer perceptron (MLP) and graph neural network (ProteinMPNN) to combine amino acid sequence information and TCR structural information to build an expert model. The expert model is a highly specialized model that combines the advantages of multiple models, such as the hidden features of miniMPNN and the embedding of pMHC (peptide-MHC complex). miniMPNN is a graph neural network used to process protein structural data. In the expert model, miniMPNN is used to extract hidden features of proteins, which contain the structural information of proteins. A multi-layer perceptron (MLP) is used to integrate the hidden features of miniMPNN and the embedding of pMHC, and to predict binding affinity. The MLP network is a feed-forward neural network designed to perform regression or classification tasks on the extracted feature vectors to predict binding affinity. The expert model and MLP generate predicted binding affinity between TCR and pMHC. The predicted results of binding affinity generate affinity conditions, which help guide subsequent protein design and optimization. The expert model significantly improves the accuracy of binding affinity prediction and contributes to the advancement of TCR design.

[0023] Step 5: Induce the diffusion process with the affinity conditions through SDEs, use a random sampler combined with a guiding gradient to generate TCR sequences and structures.

[0024] Specifically, the data is gradually transformed into a Gaussian distribution through the forward SDE expression, and a diffusion process with affinity conditions is derived. The generation process is controlled using conditional probability, and the score function of the conditional diffusion is derived through the Bayesian formula to generate affinity conditions. The affinity condition refers to the binding affinity between the TCR and the specific antigen, which is used to guide the derivation of SDEs to ensure that the generated TCR sequences and structures have the required affinity characteristics. Based on the diffusion process of the affinity condition, a random sampler is used for data sampling and combined with a guidance gradient for generation. The random sampler is used to explore different possibilities, while the guidance gradient provides directionality in the generation process. Combined with the guidance gradient, the random sampler adjusts the generation direction during the generation process to generate TCR sequences and structures with specific affinity conditions. Combining the random sampler and the guidance gradient, the generated TCR sequences and structures are more in line with the specific affinity conditions, improving the affinity and specificity of the TCR.

[0025] Furthermore, step one of this application includes:

[0026] Traversing the VDJdb database to extract multiple sample data sets of complete TCR sequences; determining multiple category information based on historical classification data logs; performing cluster analysis on the multiple sample data sets according to the multiple category information to generate multiple sample data clusters, wherein the multiple sample data clusters have a corresponding relationship with the multiple category information; dividing based on the multiple sample data clusters to determine the training data set, the verification data set, and the test data set.

[0027] Specifically, the VDJdb database was traversed to extract multiple sample data sets containing complete TCR sequences, including CDR3 (complementarity determining region 3) and V / J (variable and joining) gene segments. VDJdb is a data resource library dedicated to immunogenomics research, which mainly collects and organizes information about the sequences of T cell receptors (TCRs) and B cell receptors (BCRs) and their corresponding antigen specificity. Historical classification data logs contain previous TCR studies, known TCR-antigen interaction data, or other related bioinformatics data. These historical classification data logs were analyzed to determine multiple category information, namely data classification criteria, including species (human), confidence (≥0, indicating data reliability) and MHC class (MHCI, i.e. major histocompatibility complex class I).

[0028] According to the determined multiple category information, clustering algorithms (such as hierarchical clustering, K-means clustering, etc.) are used to group the sample data sets so that similar TCR sequences are classified into the same cluster. Using the CDR regions of TCRA and TCRB (complementary determining regions of TCRα and β), an identity threshold of 0.7 is used, that is, samples with a sequence similarity of more than 70% are classified into the same category. Cluster analysis is an unsupervised learning method used to group similar data samples together, which helps to identify and remove sequence redundancy while ensuring the diversity of the data set. The result of the cluster analysis is multiple data clusters, each cluster containing similar TCR sequences. These clusters correspond to the previously determined category information, which means that each cluster represents a specific biological characteristic or function.

[0029] The multiple sample data clusters obtained by clustering are divided into training data sets (for model training), validation data sets (for model parameter adjustment) and test data sets (for model performance evaluation). Usually, the samples in the cluster are allocated according to a certain ratio (such as 80% training, 10% validation, 10% testing) to ensure that each data set is representative. At the same time, the training data set is large enough to train the model, and the validation and test data sets are used to evaluate the performance of the model. High-quality and reasonably classified data sets help improve the training efficiency and performance of the model, enabling it to more accurately predict the binding of TCR to antigens.

[0030] Further, step 2 of this application includes:

[0031] The diffusion model is constructed based on stochastic differential equations in combination with the training data set, the validation data set, and the test data set; a forward SDE expression is defined: dx=f(x,t)dt+g(t)dw; wherein t is time, x is the state of TCR, f(x,t) is a drift term used to describe deterministic changes, g(t) is a diffusion coefficient used to adjust randomness, and dw is an increment of the Wiener process, reflecting randomness; the diffusion model is simulated by the forward SDE expression to inject simulated noise and remove simulated data generation processes, generate protein sequences and structures, and determine the noise data set.

[0032] Specifically, through modeling with stochastic differential equations (SDEs), a continuous-time framework for noise injection and removal is provided to generate new protein sequences and structures. SDEs are a mathematical tool used to describe random processes that change over time and simulate the generation process of TCR sequences and structures. The forward SDE expression is defined as dx = f(x, t) dt + g(t) dw, where t is time, x is the state of the TCR, f(x, t) is the drift term used to describe deterministic changes, g(t) is the diffusion coefficient used to adjust the randomness, and dw is the increment of the Wiener process, which reflects the randomness. The diffusion model is simulated by the forward SDE expression, noise is injected, and simulated data is generated. Noise is added to the TCR sequence and structure, and then this perturbation is reversed to generate new sequences and structures. The generated protein sequence and structure containing noise are used as a noise dataset to train a model to remove noise, thereby generating clear protein sequences and structures. The noise dataset will be used to train a denoising model, such as the U-ViT network or other deep learning architectures, to recover the original TCR sequence and structure. By simulating the dynamic changes of TCR sequences and structures while taking into account the randomness in biological systems, we combine determinism and randomness to generate more diverse and realistic TCR sequences and structures.

[0033] Further, as attached Figure 2 As shown, step three of this application includes:

[0034] The U-ViT network is used to denoise the noise data set to generate a denoised data set; the denoised data set is used as input data and synchronized to the miniMPNN model to obtain protein skeleton features; the protein skeleton features of each node are updated through the message passing mechanism of the miniMPNN model to capture local and global information of the protein structure; based on the local information and the global information, the position of each amino acid is gradually predicted using an autoregressive decoding method to generate the amino acid sequence.

[0035] Furthermore, the present application also includes the following steps:

[0036] The U-ViT network consists of multiple layers of self-attention layers and residual noise conditional transformer layers.

[0037] Specifically, the U-ViT network includes multiple layers of self-attention layers and residual noise conditional transformer layers. The U-ViT network is a complex model designed to process noisy protein data with high accuracy. U-ViT has a hidden dimension of 256, and the internal representation of each processing unit in the network has 256 features. The network consists of 6 residual noise conditional transformer layers, each of which contains 8 self-attention layers, and the dimension of each attention head is 32. The self-attention layer allows the network to dynamically assign weights between different positions, capturing long-range dependencies in the data. By using multi-head attention, the network is able to process information in different representation subspaces in parallel, thereby improving the expressiveness of the model. The U-ViT network is used to denoise the noisy protein data, that is, to recover a clean signal from the protein data containing noise. The U-shaped structure of U-ViT helps to maintain spatial information while denoising, which is crucial for recovering the original structure of the protein. The denoised data forms a denoised dataset, which contains the protein structure information recovered from the noise. The denoised dataset contains the recovered protein structure information, usually in the form of atomic coordinates, interatomic distances, backbone dihedrals, etc.

[0038] The denoised dataset is synchronized as input data to the miniMPNN model. miniMPNN is a graph neural network that is specifically designed to process protein structure data, receive this data, and extract protein skeleton features from it. The miniMPNN model updates the protein skeleton features of each node through its message passing mechanism, taking into account the interactions and dependencies between nodes, thereby capturing local information (e.g., the chemical properties of individual amino acids) and global information (e.g., the folding pattern of the entire protein chain) of the protein structure. Based on the extracted local and global information, the position of each amino acid is gradually predicted using the autoregressive decoding method to generate an amino acid sequence. The autoregressive decoding method is a sequence generation technique that generates the entire sequence by gradually predicting each element in the sequence. In protein structure prediction, it is used to predict amino acid sequences. In the process of generating amino acid sequences, protein skeleton features such as interatomic distances and backbone dihedral angles are combined. These features provide detailed information on the protein structure and help to predict amino acid sequences more accurately. The U-ViT network first removes the noise in the data and provides a clear input for the miniMPNN model. The miniMPNN model uses its message passing mechanism to understand and simulate the structure of proteins, and finally generates accurate amino acid sequences through an autoregressive decoder. The generated amino acid sequences have higher accuracy and reliability, which helps to improve the efficiency of protein engineering and drug design.

[0039] Further, step 4 of this application includes:

[0040] The structure denoising score matching loss function is: Among them, D θ is the scoring network, σ t is the noise level, x t is a noise structure, x0 is a denoising structure; the sequence prediction loss function is: Among them, ψ θ is the sequence prediction model, r i is the ith amino acid, N res is the number of amino acids; the binding affinity prediction loss function is: Among them, V θ To predict affinity, is the true affinity; the comprehensive loss function is: in, is the comprehensive loss function, is the structure denoising score matching loss function, is the sequence prediction loss function, Predict the loss function for the binding affinity.

[0041] Specifically, the structure denoising score matching loss function It is used to measure the difference between the denoised structure and the original structure. θ is the scoring network, σ t is the noise level, x t is the noise structure, and x0 is the denoising structure. Sequence prediction loss function Aims to optimize the prediction of amino acid sequence from protein structure, where ψ θ is the sequence prediction model, r i is the ith amino acid, N res is the number of amino acids. Binding affinity prediction loss function Evaluate the difference between the predicted affinity and the true affinity, where V θ To predict affinity, is the true affinity. The comprehensive loss function is the sum of the three loss functions: structure denoising score matching loss function, sequence prediction loss function, and affinity prediction loss function. It comprehensively evaluates the performance of the model in structure denoising, sequence prediction, and affinity prediction.

[0042] Further, step five of this application includes:

[0043] The data is gradually transformed into a Gaussian distribution through the forward SDE expression, and the diffusion process with the affinity condition is derived; the generation process is controlled by conditional probability, and the score function of the conditional diffusion is derived through the Bayesian formula: in, is the logarithmic probability gradient under different noise levels t, γ is a hyperparameter used for the influence of conditional information, and y is the affinity condition; the affinity condition of the TCR sequence generated by the conditional diffusion score function is derived using the Bayesian formula; based on the diffusion process of the affinity condition, the random sampler is used to perform data sampling, and the generation direction is adjusted in combination with the guidance gradient of the expert model to generate the TCR sequence and structure.

[0044] Specifically, noise is gradually added to the data through forward SDE, transforming it into a simple prior distribution, such as a Gaussian distribution, and deriving SDEs that can conditionally guide the diffusion process. The conditional diffusion model aims to generate samples that meet certain conditions, such as binding affinity to the target in the pMHC complex. The generation process is controlled using conditional probabilities, and the score function of conditional diffusion is derived through the Bayesian formula: in, is the logarithmic probability gradient under different noise levels t, describing how the probability distribution of x changes over time under specific conditions. γ is a hyperparameter used to control the influence of conditional information on the gradient. When γ = 0, the conditional information is ignored. y is the affinity condition. The conditional diffusion score function derived from the Bayesian formula is used to generate the affinity condition of the TCR sequence, and a new TCR sequence is generated according to the score function. The TCR sequence has a specific affinity condition. A random sampler is used for data sampling, which integrates the guidance gradient while optionally injecting noise at each step. By adding noise steps to a higher noise level and then taking a larger denoising step to remove the added noise, it helps the model to generate diverse and innovative TCR sequences and structures while maintaining the affinity condition. By combining the guidance of the Bayesian formula, the random sampler, and the expert model, through the steps of noise injection and removal, TCR sequences and structures with specific affinity conditions are generated, which helps to improve the efficacy of immunotherapy.

[0045] Furthermore, the present application also includes the following steps:

[0046] Retrieve the Cα root mean square deviation expression: Among them, x i and i are the coordinates of the i-th Cα atom in the generated and target structures, respectively, and N is the number of Cα atoms; based on the alignment of the TCR structure with the original target TCR structure, the Cα atom is used as a comparison benchmark, and the Cα root mean square deviation of the TCR structure and the original target TCR structure is calculated by the Cα root mean square deviation expression to obtain the Cα root mean square deviation of the original structure; the amino acid recovery rate expression is retrieved: Among them, Number of correctly predicted amino acids is the number of correctly predicted amino acids, Total number of amino acids is the total number of amino acids, and AAR is the amino acid recovery rate; the amino acid sequence is compared with the target sequence bit by bit, and the amino acid recovery rate is obtained by calculating the amino acid recovery rate expression; the TCR sequence and structure are evaluated by the Cα root mean square deviation of the original structure and the amino acid recovery rate, and the TCR sequence and structure are optimized according to the evaluation results to generate a TCR optimized sequence and structure.

[0047] Specifically, the Cα root mean square deviation expression is called Used to evaluate the similarity between two protein structures, where N is the number of Cα atoms and x iand yi are the coordinates of the i-th Cα atom in the generated structure and the target structure, respectively. RMSD is the root mean square deviation, which is used to measure the difference between two structures. A lower RMSD value indicates that the predicted structure is very similar to the target structure. Place the TCR structure to be evaluated and the TCR structure of the original target in the same coordinate system to ensure that the two structures are in similar conformations when compared so that the difference between the two can be accurately calculated. RMSD is a measure of the deviation between the structures by calculating the coordinate difference of the Cα atoms in the two structures. After the alignment is completed, calculate the coordinate difference of each Cα atom in the two structures. Obtain the Cα root mean square deviation of the original structure, that is, the average deviation of the original target structure at the Cα atom level compared to the generated structure. Retrieve the amino acid recovery rate expression, It indicates the ratio of the number of correctly predicted amino acids to the total number of amino acids in the predicted structure. Among them, AAR is the amino acid recovery rate, Number of correctly predicted amino acids is the number of correctly predicted amino acids, and Total number of amino acids is the total number of amino acids. The generated amino acid sequence is aligned with the target sequence bit by bit to calculate the amino acid recovery rate. The amino acid recovery rate is the ratio of correct amino acids in the predicted sequence. The higher the amino acid recovery rate, the more similar the predicted sequence is to the target sequence. The Cα root mean square deviation is used to evaluate the difference between the predicted structure and the target result. The lower the RMSD value, the more similar the predicted structure is to the target structure. According to the evaluation results of the amino acid recovery rate and RMSD, the TCR sequence and structure are optimized, and the amino acid sequence or structure parameters are adjusted to improve the similarity with the target sequence and structure. After the optimization process, the optimized sequence and structure of the TCR are generated. These optimized sequences and structures have improvements in amino acid recovery rate and structural similarity. The evaluation and optimization process may require multiple iterations to gradually improve the quality and accuracy of the generated TCR sequence and structure. After multiple evaluations and optimizations, an optimized TCR sequence and structure that is highly similar to the target sequence and structure is generated.

[0048] In summary, the T cell receptor sequence and structure co-design method based on the diffusion model provided in this application has the following technical effects:

[0049] By traversing the VDJdb database for screening, multiple data sets are constructed, and the multiple data sets include a training data set, a validation data set, and a test data set; a diffusion model is constructed based on the training data set, the validation data set, and the test data set to generate a noise data set; the noise data set is denoised using the U-ViT network, and an amino acid sequence is generated in combination with miniMPNN; a comprehensive loss function is designed based on the amino acid sequence, an expert model is constructed, and the binding affinity is predicted by a multi-layer perceptron in combination with the hidden features of miniMPNN and the Embedding of pMHC to generate affinity conditions; the diffusion process with the affinity conditions is derived through SDEs, and a random sampler is used in combination with a guiding gradient to generate a TCR sequence and structure. In other words, by combining TCR sequence and structure information, and using advanced data processing and model building techniques, the sequence and structure are jointly designed, thereby improving the efficiency and accuracy of TCR design.

[0050] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0051] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A T cell receptor sequence and structure co-design method based on a diffusion model, characterized in that: include: Traversing the VDJdb database for screening, and constructing multiple data sets, wherein the multiple data sets include a training data set, a verification data set, and a test data set; Building a diffusion model based on the training data set, the validation data set, and the test data set to generate a noise data set; The U-ViT network was used to denoise the noise dataset and combined with miniMPNN to generate amino acid sequences; Based on the amino acid sequence, a comprehensive loss function is designed, an expert model is constructed, and the binding affinity prediction is performed through a multi-layer perceptron to generate affinity conditions by combining the hidden features of miniMPNN and the Embedding of pMHC; TCR sequences and structures were generated by inducing diffusion processes with described affinity conditions through SDEs, using a random sampler combined with a guiding gradient.

2. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 1, characterized in that: Traverse the VDJdb database for screening and construct multiple data sets, including: Traversing the VDJdb database to extract multiple sample data sets of complete TCR sequences; determining a plurality of category information based on historical classification data logs; Performing cluster analysis on the plurality of sample data sets according to the plurality of category information to generate a plurality of sample data clusters, wherein the plurality of sample data clusters have a corresponding relationship with the plurality of category information; The training data set, the validation data set, and the test data set are determined based on the division of the multiple sample data clusters.

3. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 1, characterized in that: Building a diffusion model based on the training data set, the verification data set, and the test data set to generate a noise data set includes: Constructing the diffusion model based on stochastic differential equations in combination with the training data set, the validation data set, and the test data set; Define the forward SDE expression: dx=f(x,t)dt+g(t)dw; Where t is time, x is the state of TCR, f(x,t) is the drift term used to describe deterministic changes, g(t) is the diffusion coefficient used to adjust randomness, and dw is the increment of the Wiener process, reflecting randomness; The diffusion model is subjected to a simulated noise injection and removal simulation data generation process through the forward SDE expression to generate protein sequences and structures and determine the noise data set.

4. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 1, characterized in that: The U-ViT network is used to denoise the noise dataset and combined with miniMPNN to generate amino acid sequences, including: Using the U-ViT network to denoise the noise data set to generate a denoised data set; The denoised data set is used as input data and synchronized to the miniMPNN model to obtain protein skeleton features; Updating the protein skeleton features of each node through the message passing mechanism of the miniMPNN model to capture local and global information of the protein structure; The amino acid sequence is generated by gradually predicting the position of each amino acid using an autoregressive decoding method based on the local information and the global information.

5. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 4, characterized in that: The U-ViT network consists of multiple layers of self-attention layers and residual noise conditional transformer layers.

6. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 1, characterized in that: A comprehensive loss function is designed based on the amino acid sequence, including: The structure denoising score matching loss function is: Among them, D θ is the scoring network, σ t is the noise level, x t is the noise structure, x0 is the denoising structure; The sequence prediction loss function is: Among them, ψ θ is the sequence prediction model, r i is the ith amino acid, N res is the number of amino acids; The binding affinity prediction loss function is: Among them, V θ To predict affinity, For real affinity; The comprehensive loss function is: in, is the comprehensive loss function, is the structure denoising score matching loss function, is the sequence prediction loss function, Predict the loss function for the binding affinity.

7. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 3, characterized in that: The diffusion process with the described affinity conditions is derived through SDEs, using a random sampler combined with a guiding gradient to generate TCR sequences and structures, including: The data are gradually converted into a Gaussian distribution through the forward SDE expression, and the diffusion process with the affinity condition is derived; Use conditional probability to control the generation process, and derive the score function of conditional diffusion through the Bayesian formula: in, is the log probability gradient under different noise levels t, γ is a hyperparameter for the influence of conditional information, and y is the affinity condition; Derivation of the affinity condition of the TCR sequence generated by the conditional diffusion score function using the Bayesian formula; Based on the diffusion process of the affinity condition, the random sampler is used to perform data sampling, and the generation direction is adjusted in combination with the guidance gradient of the expert model to generate the TCR sequence and structure.

8. A T cell receptor sequence and structure co-design method based on a diffusion model as claimed in claim 1, characterized in that: Also includes: Retrieve the Cα root mean square deviation expression: Among them, x i and i are the coordinates of the i-th Cα atom in the generated and target structures, respectively, and N is the number of Cα atoms; Based on the alignment of the TCR structure with the original target TCR structure, the Cα atom is used as a comparison reference, and the Cα root mean square deviation of the TCR structure and the original target TCR structure is calculated by the Cα root mean square deviation expression to obtain the Cα root mean square deviation of the original structure; Retrieve the amino acid recovery rate expression: Among them, Number of correctly predicted amino acids is the number of correctly predicted amino acids, Total number of amino acids is the total number of amino acids, and AAR is the amino acid recovery rate; The amino acid sequence is compared with the target sequence bit by bit, and the amino acid recovery rate is calculated by the amino acid recovery rate expression; The TCR sequence and structure are evaluated by the Cα root mean square deviation of the original structure and the amino acid recovery rate, and the TCR sequence and structure are optimized according to the evaluation results to generate a TCR optimized sequence and structure.

Citation Information

Cited By

  • Molecular optimization method, system and device based on neural network and storage medium

    CN121237260A

  • Molecular generation method, system, medium and device for drug discovery

    CN121281695A