A method for predicting adaptive immune receptors by integrating gene and sequence information

Abstract

The present invention discloses a method for predicting adaptive immune receptors by integrating gene and sequence information, comprising the following steps: S1, modifying the SC-AIR-BERT model to construct an SC-AIR-BERT-Multi model; the SC-AIR-BERT-Multi model comprises a gene information extraction channel, a sequence information extraction channel, a multimodal fusion module, and two multi-layer perceptrons for multi-task learning; S2, in the gene information extraction channel, using the gene name as input, obtaining the gene feature h of the immune cell receptor gene ; S3, in the sequence information extraction channel, using TCR sequence or BCR sequence as input, obtain the sequence feature h of immune cell receptor seq ; S4, the gene features and sequence features of the V, D, and J gene segments are sent to the multimodal feature fusion module for fusion, and the fused features are generated; S5, the multimodal receptor feature Representation learned in step S4 is mapped to the final TCR or BCR antigen binding specificity prediction and affinity prediction results through two multi-layer perceptrons for prediction.

Description

Technical Field

[0001] The present invention relates to the field of biological technology, and in particular to a method for predicting adaptive immune receptors by integrating gene and sequence information. Background Art

[0002] The human immune system is composed of innate immunity and adaptive immunity. Adaptive immunity is an immune response that recognizes and targets specific pathogens (antigens) after contact with them. T cells and B cells are key components of the adaptive immune system. Antigen recognition is a key factor in T- and B-cell-mediated immunity. T cells and B cells primarily interact with antigens through the T cell receptor (TCR, a protein dimer) and B cell receptor (BCR), respectively. The TCR recognizes and binds antigens presented by the major histocompatibility complex (MHC) on the cell membrane, while the BCR directly binds to specific antigens. Both TCRs and BCRs are composed of two peptide chains (alpha-beta chains or light-heavy chains). Each chain is generated by the rearrangement of a variable (V) gene segment, a diversity (D) gene segment, and a joining (J) gene segment (light chain V, J, heavy chain V, D, J). Genetic rearrangement of the V(D) and J gene segments produces the highly polymorphic adaptive immune receptor gene repertoire. The two chains form a cyclic three-dimensional structure of the complementarity determining region (CDR) (CDR1, CDR2, CDR3) for antigen recognition and binding. When TCR recognizes the MHC-antigen peptide complex, CDR1 and CDR2 recognize and bind to the side walls of the antigen binding groove of the MHC molecule, while CDR3 directly binds to the antigen peptide; moreover, the CDR3 region has the greatest variation, which directly affects the antigen binding specificity of TCR / BCR.

[0003] Predicting antigen binding specificity and affinity based on TCR / BCR sequence information and gene information, and accurately predicting T cell or B cell activation ability based on this, will have a transformative impact on many cutting-edge research fields such as the treatment of infectious diseases and autoimmune diseases, and cancer immune vaccine design.

[0004] The shortcomings of existing technology solutions are as follows: For methods based on manual definitions, such as GLIPH and TCRdist, the manually defined features such as Hamming distance are not learnable and have limitations, and may not fully represent the useful information of TCR / BCR sequences or accurately characterize the distance differences between sequences. Artificial intelligence-based methods such as DeepTCR, TCRAI, and soNNia automatically extract and learn features based on T / B cell receptor sequences (TCR / BCR) and V, D, and J gene information to predict the antigen specificity of TCRs and BCRs. A large amount of data with known antigen binding specificity is required for training to obtain good prediction results. When the data is small, the prediction results are often poor. Moreover, when combining sequence information and gene information, they simply concatenate them without fusing them. The TCR-BERT method uses TCR single-chain data for pre-training, and does not comprehensively consider the information of both chains during the pre-training stage. Its encoding may not fully and comprehensively represent the influence of the two chains of the TCR pair on each other's encoding. It also does not incorporate V(D)J gene information. Moreover, the method is only targeted at TCRs and does not consider BCRs. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting adaptive immune receptors based on fusion gene and sequence information. The method can simultaneously consider TCR and BCR. When using a small amount of labeled data, the V, D, J genes and TCR / BCR sequences are fused together to predict binding affinity and antigen specificity, and the prediction results are accurate.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for predicting adaptive immune receptors that integrates gene and sequence information comprises the following steps:

[0008] S1. Modify the SC-AIR-BERT model to construct the SC-AIR-BERT-Multi model; the SC-AIR-BERT-Multi model contains a gene information extraction channel, a sequence information extraction channel, a multimodal fusion module, and two multilayer perceptrons for multi-task learning;

[0009] S2. In the gene information extraction channel, the gene name is used as input to obtain the gene characteristics h of the immune cell receptor. gene ;

[0010] S3. In the sequence information extraction channel, the TCR sequence or BCR sequence is used as input to obtain the sequence feature h of the immune cell receptor. seq ;

[0011] S4, sending the gene features and sequence features of the V, D, and J gene segments into the multimodal feature fusion module for fusion, and generating fused features;

[0012] S5. Use two multi-layer perceptrons to map the multimodal receptor feature representation learned in step S4 to the final TCR or BCR antigen binding specificity prediction and affinity prediction results to perform antigen binding specificity prediction and affinity prediction.

[0013] Preferably, the modification process of the SC-AIR-BERT model in step S1 is:

[0014] S11. Modify the BERT network: Since the pairing of TCR or BCR chains is relatively random, remove the NSP task of the BERT model and only retain the MLM task;

[0015] S12. Modify the mask strategy: Since kmer sequences are used as input data, there will be two amino acid overlaps between words. When masking a kmer, the two words on the left and right are also masked to prevent the model training from converging too quickly.

[0016] Preferably, the gene information extraction channel in step S2 comprises a gene encoder for learning and encoding the V, D, and J gene information contained in TCR or BCR; the gene encoder takes the gene name as input, and after performing a corresponding digital encoding according to the gene name, fully connects the V, J gene information of the light chain or α chain and the V, D, and J gene information of the heavy chain or β chain of the adaptive immune cell receptor to obtain the gene feature h of the immune cell receptor. gene .

[0017] Preferably, the specific process of step S3 is:

[0018] S31. The CDR3 region sequences of the two chains are processed into kmer sequences respectively, where each kmer represents a word and the entire kmer sequence is considered as a sentence;

[0019] S32: After the two sentences are connected, they are marked, segmented, and positionally encoded;

[0020] S33. Input the encoded TCR or BCR into a 6-layer BERT model. The BERT model is pre-trained with data without epitope information. Through self-supervision, the protein language in TCR or BCR is pre-learned into a model to extract the common features in the TCR or BCR sequence, obtain the encoding of the CDR3 sequence, and obtain the sequence feature h of the immune cell receptor. seq .

[0021] Preferably, the multimodal feature fusion module in step S4 includes a gated attention mechanism and a tensor fusion mechanism, the gated attention mechanism is used to weight and filter the features of each modality, and the tensor fusion mechanism is used to generate the fused overall multimodal features; the specific process of step S4 is:

[0022] S41, based on the gated attention mechanism, the gene features h obtained in step S2 of the immune cell receptor are gene And the sequence feature h obtained in step S3 seq Perform weighted fusion to obtain the immune cell receptor target gene feature h` gene and target sequence features h` seq ;

[0023] S42, target gene feature h' gene With the target sequence feature h` seq Multiply them together to obtain the initial receptor feature hfusion of the immune cell receptor, that is, the process is the immune cell receptor tensor fusion process;

[0024] S43. Perform two full connections (FC1, FC2) on the initial receptor feature hfusion to obtain the receptor feature Representation of the immune cell.

[0025] Preferably, the multilayer perceptron in step S5 is a fully connected network, wherein a classification head is used for antigen binding specificity prediction and a regression head is used for affinity prediction.

[0026] After adopting the above technical solution, the present invention has the following beneficial effects:

[0027] 1. The present invention is a predictive analysis method that is universal for both TCR and BCR, and has a wider range of application scenarios.

[0028] 2. The present invention obtains TCR / BCR encoding based on the SC-AIR-BERT-Multi model, which can better predict affinity and antigen binding specificity even with limited labeled data, and the prediction results are accurate.

[0029] 3. The present invention performs multimodal feature fusion based on the genetic information and sequence information of TCR / BCR to more comprehensively characterize TCR / BCR.

[0030] 4. The downstream of the present invention adopts multi-task learning, which takes into account multiple tasks such as antigen-specific recognition and binding affinity prediction, while improving the robustness and generalization ability of the model itself. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2 This is a framework diagram of the SC-AIR-BERT-Multi model of the present invention;

[0033] Figure 3 Flowchart of the SC-AIR-BERT-Multi model of the present invention;

[0034] Figure 4 Flowchart of the SC-AIR-BERT model of the present invention;

[0035] Figure 5 This is a diagram showing the results of TCR antigen-specific recognition of the present invention;

[0036] Figure 6 This is a diagram showing the BCR antigen-specific recognition results of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] Example

[0039] like Figures 1 to 6 As shown, a method for predicting adaptive immune receptors by integrating gene and sequence information comprises the following steps:

[0040] S1. Modify the SC-AIR-BERT model to construct the SC-AIR-BERT-Multi model; the SC-AIR-BERT-Multi model contains a gene information extraction channel, a sequence information extraction channel, a multimodal fusion module, and two multilayer perceptrons for multi-task learning;

[0041] The modification process of the SC-AIR-BERT model in step S1 is as follows:

[0042] S11. Modify the BERT network: Since the pairing of TCR or BCR chains is relatively random, remove the NSP task of the BERT model and only retain the MLM task;

[0043] S12. Modify the mask strategy: Since kmer sequences are used as input data, there will be two amino acid overlaps between words. When masking a kmer, the two words on the left and right are also masked to prevent the model training from converging too quickly.

[0044] S2. In the gene information extraction channel, the gene name is used as input to obtain the gene characteristics h of the immune cell receptor. gene;

[0045] The gene information extraction channel in step S2 includes a gene encoder for learning and encoding the V, D, and J gene information contained in TCR or BCR; the gene encoder takes the gene name as input, and after encoding the corresponding number according to the gene name, fully connects the V, J gene information of the light chain or α chain and the V, D, and J gene information of the heavy chain or β chain of the adaptive immune cell receptor to obtain the gene feature h of the immune cell receptor. gene ;

[0046] S3. In the sequence information extraction channel, the TCR sequence or BCR sequence is used as input to obtain the sequence feature h of the immune cell receptor. seq ;

[0047] The specific process of step S3 is:

[0048] S31. The CDR3 region sequences of the two chains are processed into kmer sequences respectively, where each kmer represents a word and the entire kmer sequence is considered as a sentence;

[0049] S32: After the two sentences are connected, they are marked, segmented, and positionally encoded;

[0050] S33. Input the encoded TCR or BCR into a 6-layer BERT model. The BERT model is pre-trained with a large dataset without epitope information. Through self-supervision, the protein language in the TCR or BCR is pre-learned into a model to extract the common features in the TCR or BCR sequence, obtain the encoding of the CDR3 sequence, and obtain the sequence feature h of the immune cell receptor. seq ;

[0051] S4, sending the gene features and sequence features of the V, D, and J gene segments into the multimodal feature fusion module for fusion, and generating fused features;

[0052] The multimodal feature fusion module in step S4 includes a gated attention mechanism and a tensor fusion mechanism. The gated attention mechanism is used to weight and filter the features of each modality, and the tensor fusion mechanism is used to generate the fused overall multimodal features. The specific process of step S4 is as follows:

[0053] S41, based on the gated attention mechanism, the gene features h obtained in step S2 of the immune cell receptor are gene And the sequence feature h obtained in step S3 seq Perform weighted fusion to obtain the immune cell receptor target gene feature h` gene and target sequence features h` seq ;

[0054] S42, target gene feature h'gene With the target sequence feature h` seq Multiply them together to obtain the initial receptor feature hfusion of the immune cell receptor, that is, the process is the immune cell receptor tensor fusion process;

[0055] S43, perform two full connections (FC1, FC2) on the initial receptor feature hfusion to obtain the receptor feature representation of the immune cell;

[0056] S5. Mapping the multimodal receptor feature representation learned in step S4 to the final TCR or BCR antigen binding specificity prediction and affinity prediction results through two multi-layer perceptrons to perform antigen binding specificity prediction and affinity prediction;

[0057] The multilayer perceptron in step S5 is a fully connected network, wherein a classification head is used for antigen binding specificity prediction and a regression head is used for affinity prediction.

[0058] Performance testing:

[0059] The SC-AIR-BERT-Multi model of the present invention and the traditional SC-AIR-BERT model, TCRAL model, DeepTCR model, TCR-BERT model and SoNNia model were used to predict TCR and BCR antigen specific recognition. The results of TCR antigen specific recognition are shown in Figure 2. Figure 5 As shown, the BCR antigen specific recognition results are as follows Figure 6 As shown in the figure, the AUC value shows that the SC-AIR-BERT-Multi model used in the present invention has better recognition and prediction results than other existing methods such as the SC-AIR-BERT model, TCRAL model, DeepTCR model, TCR-BERT model and SoNNia model.

[0060] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for predicting adaptive immune receptors by integrating gene and sequence information, characterized in that: The following steps are involved: S1. Modify the SC-AIR-BERT model to construct the SC-AIR-BERT-Multi model; the SC-AIR-BERT-Multi model contains a gene information extraction channel, a sequence information extraction channel, a multimodal fusion module, and two multilayer perceptrons for multi-task learning; The modification process of the SC-AIR-BERT model in step S1 is as follows: S11. Modify the BERT network: Since the pairing of TCR or BCR chains is relatively random, remove the NSP task of the BERT model and only retain the MLM task; S12. Modify the mask strategy: Since kmer sequences are used as input data, there will be two amino acid overlaps between words. When masking a kmer, the two words on the left and right are also masked to prevent the model training from converging too quickly. S2. In the gene information extraction channel, the gene name is used as input to obtain the gene characteristics h of the immune cell receptor. gene ; S3. In the sequence information extraction channel, the TCR sequence or BCR sequence is used as input to obtain the sequence feature h of the immune cell receptor. seq ; S4, sending the gene features and sequence features of the V, D, and J gene segments into the multimodal feature fusion module for fusion, and generating fused features; The multimodal feature fusion module in step S4 includes a gated attention mechanism and a tensor fusion mechanism. The gated attention mechanism is used to weight and filter the features of each modality, and the tensor fusion mechanism is used to generate the fused overall multimodal features. The specific process of step S4 is as follows: S41, based on the gated attention mechanism, the gene features h obtained in step S2 of the immune cell receptor are gene And the sequence feature h obtained in step S3 seq Perform weighted fusion to obtain the immune cell receptor target gene feature h gene and target sequence feature hˋ seq ; S42, target gene feature h gene and the target sequence feature hˋ seq Multiply them together to obtain the initial receptor feature hfusion of the immune cell receptor, that is, the process is the immune cell receptor tensor fusion process; S43, perform two full connections (FC1, FC2) on the initial receptor feature hfusion to obtain the receptor feature representation of the immune cell; S5. Use two multi-layer perceptrons to map the multimodal receptor feature representation learned in step S4 to the final TCR or BCR antigen binding specificity prediction and affinity prediction results to perform antigen binding specificity prediction and affinity prediction.

2. The method for predicting adaptive immune receptors based on fusion gene and sequence information according to claim 1, wherein: The gene information extraction channel in step S2 includes a gene encoder for learning and encoding the V, D, and J gene information contained in TCR or BCR; the gene encoder takes the gene name as input, and after encoding the corresponding number according to the gene name, fully connects the V, J gene information of the light chain or α chain and the V, D, and J gene information of the heavy chain or β chain of the adaptive immune cell receptor to obtain the gene feature h of the immune cell receptor. gene .

3. The method for predicting adaptive immune receptors based on fusion gene and sequence information according to claim 2, wherein: The specific process of step S3 is: S31. The CDR3 region sequences of the two chains are processed into kmer sequences respectively, where each kmer represents a word and the entire kmer sequence is considered as a sentence; S32: After the two sentences are connected, they are marked, segmented, and positionally encoded; S33. Input the encoded TCR or BCR into a 6-layer BERT model. The BERT model is pre-trained with data without epitope information. Through self-supervision, the protein language in TCR or BCR is pre-learned into a model to extract the common features in the TCR or BCR sequence, obtain the encoding of the CDR3 sequence, and obtain the sequence feature h of the immune cell receptor. seq .

4. The method for predicting adaptive immune receptors based on fusion gene and sequence information according to claim 3, wherein: The multilayer perceptron in step S5 is a fully connected network, wherein a classification head is used for antigen binding specificity prediction and a regression head is used for affinity prediction.

Images