Deep learning-based prediction method for hla class i binding to tcr

Abstract

A deep learning-based method for predicting HLA-I binding to TCRs is proposed. This method collects comprehensive peptide-TCR binding records from four databases: IEDB, VDJdb, PIRD, and McPas-TCR, forming a dataset. The steps include: preprocessing the data before inputting it into the model to obtain a one-hot encoding matrix; inputting the one-hot encoding matrix into a deep learning model for learning, and predicting the binding probability. This invention can efficiently and accurately predict the binding affinity between TCRs and HLA-I complexes and requires only the CDR3β sequence.

Description

Technical Field

[0001] This invention relates to a technology at the intersection of computer science and genetic engineering, specifically a deep learning-based method for predicting the binding of major histocompatibility complex HLA-I class to TCR (PanTCR). Background Technology

[0002] The T-cell receptor (TCR) is a molecule located on the surface of T cells that recognizes and binds to antigens, initiating an immune response in T cells. The TCR consists of two polypeptide chains: an α-chain and a β-chain. The HLA-I molecule, on the other hand, is found on the surface of almost all nucleated cells and can bind to and present antigens to T cells. The HLA-I molecule also consists of two polypeptide chains: an α-chain and a β2-microglobulin chain. When the TCR binds to the HLA-I complex, the complementarity-determining regions (CDRs) on the α- and β-chains of the TCR bind to the peptide presented on the HLA-I molecule, triggering an immune response in T cells. This binding is highly specific. This specific binding is a crucial step in antigen recognition and immune response in the immune system. The binding of the TCR to the HLA-I complex is an important aspect of the immune response and can be used to develop novel vaccines, diagnostics, and immunotherapies.

[0003] Existing techniques include clustering TCRs and decoding their antigen-specific binding patterns; employing peptide-specific TCR binding prediction models; and using peptide-TCR binding prediction models that are not limited to specific peptides, but require known binding TCRs for model training. However, these existing computational tools perform poorly in predicting TCR binding to neoantigens or exogenous peptides because they cannot generalize to peptides that are unseen or not perceived by the immune system. These peptides may be generated by mutations or viral infections and are important for immunological research and immunotherapy. Summary of the Invention

[0004] This invention addresses the limitations of machine learning computational tools in predicting TCR binding to neoantigens or exogenous peptides, which cannot be extended to peptides that are unseen or not perceived by the immune system and require both CDR3β and CDR3α sequences. It proposes a deep learning-based method for predicting HLA-I class TCR binding, which can efficiently and accurately predict the binding affinity between TCR and HLA-I complexes and only requires the CDR3β sequence.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to a deep learning-based prediction method combining HLA-I class and TCR, comprising the following steps:

[0007] Step 1: Collect comprehensive peptide-TCR binding records from four databases: IEDB, VDJdb, PIRD, and McPas-TCR, and construct a dataset.

[0008] In the dataset, each data sample consists of a peptide sequence and a TCR sequence.

[0009] The dataset was specifically constructed by collecting a large number of TCRs from the peripheral blood of 587 healthy volunteers, randomly selecting 14,230 data points as non-binding TCRs for use as controls, and constructing a dataset containing 28,463 data samples with a positive to negative sample ratio of 1:1.

[0010] Step 2: Preprocess the data before it enters the model to obtain the one-hot encoding matrix, specifically including:

[0011] 2.1) Merge the peptide sequence and TCR sequence of each data sample, and retain the original peptide sequence and TCR sequence to obtain the peptide sequence, TCR sequence and merged sequence.

[0012] 2.2) Perform one-hot encoding on the sequences obtained in step 2.1 to obtain three one-hot encoding matrices.

[0013] Step 3: Input the one-hot encoding matrix obtained in Step 2 into the deep learning model for learning, and predict the combination probability, specifically including:

[0014] 3.1) The one-hot encoding matrix of the TCR sequence is decoded by the CNN-LSTM module to obtain the decoded features. The one-hot encoding matrices of peptide sequences and merged sequences are transformed from two-dimensional matrices into one-dimensional features.

[0015] 3.2) After concatenating the decoded features obtained in step 3.1) with the two types of one-dimensional features into a fused feature matrix, the matrix is ​​input into the fully connected network of the deep learning model to map the feature space to the label space, thereby enabling the prediction of the binding probability of antigen peptides to TCR.

[0016] The deep learning model includes a CNN-LSTM module and a fully connected network. The CNN-LSTM module consists of 64 one-dimensional convolutional kernels of size 3 and a single-layer bidirectional LSTM. The matrix is ​​decoded by the convolutional kernels and then subjected to ReLU activation and pooling operations. The temporal features of the sequence are then extracted by the 64-channel bidirectional LSTM layer. The fully connected network module uses a four-layer perceptron for dimensionality reduction mapping, which maps the features from the feature space to the label space.

[0017] The deep learning model described above is trained and validated using a five-fold cross-validation method to ensure that there is no overfitting.

[0018] Technical effect

[0019] This invention employs a CNN-LSTM module to decode the one-hot encoding matrix of TCR sequences, extracting high-order features while preserving the low-order features of the original one-hot encoding of the merged sequence and peptide sequence. Then, by fusing these decoded features and learning through a four-layer perceptron, both high- and low-order features of the sequence are processed simultaneously to predict the binding probability of the antigen peptide to the TCR. This unique and innovative technique provides a novel and effective approach for exploring the binding probability of HLA-I class complexes to TCRs. Compared to existing technologies, this invention utilizes only the CDR3β information of the HLA-I class complex peptide sequence and TCR sequence, without requiring additional feature extraction or feature engineering; it is applicable to various HLA-I class complexes, and even to neoantigen peptides of HLA-I class complexes without training data. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention;

[0021] Figure 2 This is a schematic diagram of the test set (PR-AUC) during the five-fold cross-validation training process;

[0022] Figure 3 A schematic diagram showing the results of the test set comparison (ROC-AUC and PR-AUC) with the test set set at zero.

[0023] Figure 4 A schematic diagram showing the comparison of experimental results (ROC-AUC and PR-AUC) on the normal test set;

[0024] Figure 5 A schematic diagram of the CNN-LSTM module processing TCR sequences in this invention. Detailed Implementation

[0025] like Figure 1 As shown, this embodiment specifically includes:

[0026] Step 1) Data Collection and Preparation: Peptide-TCR binding data were collected from four databases: IEDB, VDJdb, PIRD, and McPas-TCR. Unbound TCRs were obtained from peripheral blood of healthy volunteers as controls. Data sets were constructed, with each data example including both the peptide sequence and the TCR sequence.

[0027] Step 2) Data Preprocessing: Merge the peptide and TCR sequences of each data sample, retaining the original one-hot encoding. One-hot encoding is performed on these three types of sequences (peptide sequences, TCR sequences, and merged sequences), specifically including:

[0028] 2.1) Merging the TCR sequence and the peptide sequence of the HLA-I complex, that is, splicing the novel antigen peptide sequence presented on the HLA molecule for TCR recognition with the CDR3β sequence, specifically including:

[0029] a) The TCR sequence and the HLA-I complex sequence are merged. This process focuses on the CDR3β (Complementary Determining Region 3β) sequence of the TCR. Variations in the CDR3β sequence determine the T cell's ability to recognize specific antigens selectively.

[0030] b) The neoantigen peptide sequence presented to the TCR for recognition on the HLA molecule is spliced ​​with the CDR3β sequence. Specifically, first, the neoantigen peptide sequence on the HLA-I molecule is filled with the symbol '-' when it is less than 11 amino acid residues long, and the CDR3β sequence is filled with the symbol '-' when it is less than 24 amino acid residues long. Then, the TCR's CDR3b sequence (24 amino acid residues long) is spliced ​​after the neoantigen peptide sequence to form a sequence of 35 amino acid residues long.

[0031] For example, the neoantigen peptide sequence 'NLVPMVATV' and the TCR sequence 'CASSHPRPRVDSYNEQFF' are spliced ​​together to form:

[0032] The binding sequence of 'NLVPMVATV--CASSHPRPRVDSYNEQFF------', while preserving 'NLVPMVATV--' and

[0033] 'CASSHPRPRVDSYNEQFF------'.

[0034] 2.2) The binding sequence, neoantigen peptide sequence, and TCR CDR3b sequence are encoded using a one-hot matrix to obtain the encoding matrix: each amino acid residue is encoded into a vector, the length of which is equal to the number of amino acid residue types. In this vector, only the position corresponding to the amino acid is 1, and all other positions are 0. In this way, each binding sequence can be converted into a numerical matrix, facilitating the next step of calculation.

[0035] Step 3) Construct and train a deep learning model: Use a CNN-LSTM module to decode the one-hot encoding matrix of the TCR sequence to obtain high-order features while retaining the original low-order features of the one-hot encoding. Fuse these features into a comprehensive feature matrix, and then input the fused feature matrix into a fully connected network module of a four-layer perceptron. Through neural network training, both high-order and low-order features are learned simultaneously to predict the binding probability of the antigen peptide to the TCR. Finally, a five-fold cross-validation method is used for model training and validation to ensure the model's performance, generalization ability, and robustness.

[0036] The deep learning model includes: CNN, LSTM and perceptron, wherein: CNN performs convolution processing based on the time series information of the CDR3b sequence of TCR to obtain high-order feature results; LSTM extracts temporal features based on the high-order feature information decoded by CNN to obtain high-order temporal features; and the perceptron performs dimensionality reduction processing based on the fusion information of high-order temporal features and low-order sequence features, maps it to the label space and performs probability prediction.

[0037] The convolution calculation in the CNN is specifically as follows: Where: y i It is the i-th element of the output, σ is the activation function, and W j X is the weight of the j-th convolutional kernel. i+j is the (i+j)th element of the input sequence, and b is the bias.

[0038] The calculation of the LSTM long short-term memory network is specifically as follows: f t =σ(W f ·[h t-1 x t ]+b f ), i t =σ(W i ·[h t-1 x t ]+b i ), o t =σ(W o ·[h t-1 x t ]+bo), h t =o t ·tanh(C t ), where: f t It's the Gate of Oblivion, i t It's an input gate, o t It's an output gate. It is the candidate unit state, C t It is the unit state, h t This is the output.

[0039] The perceptron calculation specifically involves: given an input feature vector x = (x1, x2, ..., x3) and a corresponding weight vector w = (w1, w2, ..., w...). n Given a threshold b, the perceptron makes classification predictions in the following way: Here, f is the activation function, and this model uses the ReLU activation function.

[0040] The five-fold cross-validation method refers to: splitting the integrated dataset into 5 parts, with each part consisting of a training set and a test set. Each part is used as the test set once, and each part is used as the training set 4 times. Five-fold cross-validation is used to train and evaluate the model's performance. Then, the model with the smallest loss on the test set is selected as the best model. The output results of the five models obtained from the five cross-validations are averaged to obtain the final model output result.

[0041] The training method described employs the ADAM optimizer and cosine annealing strategy to train the model. The model's parameters are updated using the cross-entropy loss function and the backpropagation algorithm.

[0042] In the actual experiment, the specific environment was set up within the PyTorch framework. An initial learning rate of 0.001 and a batch size of 1000 were chosen, and a total of 50 training epochs were performed. The experimental results are as follows: Figure 2 As shown, the ROC-AUC of the test set changes during the five-fold cross-validation training process. After 50 rounds of training, the ROC-AUC value of the validation set gradually increases, eventually stabilizing at approximately 0.83. The final ROC-AUC of the five-fold model is 0.828 ± 0.014, indicating that the training method achieves good performance in this experiment.

[0043] In addition, comprehensive peptide-TCR binding records were collected from four databases (including IEDB, VDJdb, PIRD, and McPas-TCR) to form a dataset; this dataset retains human TCR records that describe HLA alleles belonging to HLA-I; in VDJdb, records with a confidence score of 0 were further excluded and only high-confidence binding TCR records from the PIRD dataset were retained; only records containing the CDR3 beta chain were retained.

[0044] The dataset is divided into a training set, a normal test set, and a zero-case test set. Peptides in the zero-case test set are not visible in any training set of the model of this invention or any other existing tool, while peptides in the normal test set are allowed to appear in the training set. Finally, the model of this invention is compared with other well-known models on these two test sets.

[0045] like Figure 3 As shown, the results of the zero-case test set comparison experiment demonstrate that the present invention (PanTCR) significantly outperforms other methods in the zero-case scenario, achieving a ROC-AUC of 0.774 and a PR-AUC of 0.817 (pMTnet's ROC-AUC was 0.563 and PR-AUC was 0.555; ERGO2's ROC-AUC was 0.496 and PR-AUC was 0.542; DLpTCR's ROC-AUC was 0.517 and PR-AUC was 0.488; and PanPep's ROC-AUC was 0.744 and PR-AUC was 0.754). Therefore, PanTCR can predict the binding of unseen peptide-specific TCRs, demonstrating great potential for recognizing exogenous or neoantigens in various immunological studies and clinical applications.

[0046] like Figure 4 The results, shown below, are comparative experiments on a normal test set, demonstrating that the present invention can easily generalize to normal conditions and exhibits relatively high performance even under normal circumstances. PanPep and PanTCR were selected for comparison under normal conditions. First, the present invention was trained using the prepared training set, while the weights trained on the same dataset for PanPep were downloaded, resulting in models that can be fairly compared. Next, both models were tested using the same normal test set. The experimental results show that the ROC-AUC of the present invention reached 0.855, while that of PanPep was 0.792. Further testing using the PR-AUC metric showed that the PR-AUC of the present invention reached 0.874, while that of PanPep was 0.796.

[0047] Compared to existing technologies, this invention employs a CNN-LSTM module to decode the one-hot encoding matrix of the TCR sequence, extracting high-order features while preserving the low-order features of the original one-hot encoding of the merged sequence and peptide sequence. Then, by fusing these decoded features and learning through a four-layer perceptron, both high- and low-order features of the sequence are processed simultaneously to predict the binding probability of the antigen peptide to the TCR. This rapid, accurate, and reliable prediction of TCR-complex binding can be widely applied in bioinformatics, medicine, and clinical practice, providing strong support and guidance for new drug development and personalized medicine, as well as offering important tools and methods for the research and development of gene immunotherapy.

[0048] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A prediction method combining HLA-I class and TCR based on deep learning, characterized in that, Includes the following steps: Step 1: Collect comprehensive peptide-TCR binding records from four databases: IEDB, VDJdb, PIRD, and McPas-TCR, and construct a dataset; Step 2: Preprocess the data before it enters the model to obtain the one-hot encoding matrix, specifically including: 2.1) The peptide sequence and TCR sequence of each data sample were merged, and the original peptide sequence and TCR sequence were retained to obtain the peptide sequence, TCR sequence and merged sequence. Specifically, the following steps were taken: First, the length of the neoantigen peptide sequence on the HLA-I molecule was less than 11 amino acid residues after filling it with the symbol '-', and the length of the CDR3β sequence was less than 24 amino acid residues after filling it with the symbol '-'. Then, the CDR3b sequence of the TCR was spliced ​​after the neoantigen peptide sequence to form a sequence with a length of 35 amino acid residues. 2.2) Perform one-hot encoding on the sequences obtained in step 2.1 to obtain three one-hot encoding matrices; Step 3: Input the one-hot encoding matrix obtained in Step 2 into the deep learning model for learning, and predict the combination probability, specifically including: 3.1) The one-hot encoding matrix of the TCR sequence is decoded by the CNN-LSTM module to obtain the decoded features. The one-hot encoding matrices of peptide sequences and merged sequences are transformed from two-dimensional matrices into one-dimensional features. 3.2) After concatenating the decoded features obtained in step 3.1) with the two types of one-dimensional features into a fused feature matrix, the matrix is ​​input into the fully connected network of the deep learning model to map the feature space to the label space, thereby enabling the prediction of the binding probability of antigen peptides to TCR. The deep learning model includes a CNN-LSTM module and a fully connected network. The CNN-LSTM module consists of 64 one-dimensional convolutional kernels of size 3 and a single-layer bidirectional LSTM. The matrix is ​​decoded by the convolutional kernels and then subjected to ReLU activation and pooling operations. The temporal features of the sequence are then extracted by the 64-channel bidirectional LSTM layer. The fully connected network module uses a four-layer perceptron for dimensionality reduction mapping, which maps the features from the feature space to the label space.

2. The deep learning-based prediction method combining HLA-I class and TCR as described in claim 1, characterized in that, In the dataset, each data sample consists of a peptide sequence and a TCR sequence.

3. The deep learning-based prediction method combining HLA-I class and TCR as described in claim 1, characterized in that, The dataset was specifically constructed by collecting a large number of TCRs from the peripheral blood of 587 healthy volunteers, randomly selecting 14,230 data points as non-binding TCRs for use as controls, and constructing a dataset containing 28,463 data samples with a positive to negative sample ratio of 1:

1.

4. The deep learning-based prediction method combining HLA-I class and TCR as described in claim 1, characterized in that, The deep learning model described above is trained and validated using a five-fold cross-validation method to ensure that there is no overfitting.

Images