Data-driven method for predicting the binding affinity of MHC Class II molecules to peptides

Abstract

This invention, based on a data-driven method for predicting the binding affinity of MHC Class II molecules to peptides, belongs to the field of peptide drug development technology and addresses the problem of improving the accuracy of predicting the binding affinity of MHC Class II molecules to peptides. The invention constructs a fused feature vector by extracting features such as multidimensional physicochemical parameter matrices, digitally representing sequence position information and the physicochemical nature of amino acids. This vector is input into a pre-trained prediction model to obtain the predicted binding affinity value. After online purification, the value is transferred by a robotic arm to a detection unit to measure the actual IC50 value. The difference between the measured molecular binding affinity and the model prediction value is calculated, and this difference is used as the trigger criterion for iterative optimization of the model. The measured data is fed back to update the physicochemical matrix weights or model parameters, and the process is iterated until convergence. This invention achieves continuous self-evolution of the model through closed-loop feedback, steadily improving the long-term prediction accuracy of the binding affinity of MHC Class II molecules to peptides.

Description

Technical Field

[0001] This invention belongs to the field of peptide drug development technology, and relates to a data-driven method, system, device and medium for predicting the binding affinity between MHC Class II molecules and peptides. Background Technology

[0002] MHC Class II molecules (such as human HLA-DR, HLA-DP, HLA-DQ, and mouse H-2-IAb) are key molecules in antigen presentation. Their binding affinity to peptides directly determines the initiation strength and specificity of the immune response. Binding affinity is usually expressed as IC50 or Kd. Existing technologies such as IEDB prediction tools, NetMHCIIpan, and the VivaBiotech platform have the following technical challenges:

[0003] (1) Traditional prediction models often use traditional coding methods such as One-hot, BLOSUM or simple sequence features, treating amino acids as discrete symbols and losing the continuous physical and chemical properties of amino acids (such as hydrophobicity, charge, polarity, etc.), resulting in insufficient accuracy and poor correlation between the predicted IC50 / Kd value and the actual experimental value.

[0004] (2) The synthesis, purification and detection of peptides are independent of each other. Experimental data cannot be efficiently and automatically fed back to the prediction model. The model cannot iterate and evolve on its own based on real experimental data and lacks an efficient data closed-loop feedback mechanism.

[0005] (3) Existing automated platforms mostly focus on a single link, such as automated synthesizers or high-throughput screening, and lack an integrated closed-loop system that covers the entire chain of "prediction-synthesis-purification-detection-model iteration", resulting in low R&D efficiency and high costs.

[0006] While existing public datasets for predicting the binding affinity of MHC Class II molecules to peptides (such as IEDB) are rich in content, they still lack high-quality closed-loop data, limiting the application of models in actual immunotherapy and vaccine design. Therefore, it is crucial to improve the accuracy of MHC Class II-peptide binding affinity prediction, making it closer to the actual physicochemical process, and to continuously optimize the performance of the prediction model using newly added experimental data. Summary of the Invention

[0007] The technical problem to be solved by this invention is how to improve the accuracy of predicting the binding force between MHC Class II molecules and peptides.

[0008] The present invention solves the above-mentioned technical problems through the following technical solutions: A data-driven method for predicting the binding affinity between MHC Class II molecules and peptides includes the following steps: S1. Digitally characterize the peptide sequence, use a pre-trained prediction model to predict the binding force of the peptide sequence, obtain the predicted binding force value and sort it, and select the top-K peptide sequences corresponding to the predicted binding force values ​​as candidate peptide sequences. S2, send the candidate peptide sequence to the peptide synthesis and molecular binding force testing platform to obtain the experimentally determined true binding force value; S3, calculate the difference between the predicted binding force value and the actual binding force value. When the difference exceeds the preset threshold, trigger the model iterative optimization, take the candidate peptide sequence and the corresponding actual binding force value as new samples, and perform incremental training and retraining on the prediction model to obtain the updated prediction model.

[0009] Furthermore, S1 includes the following: S1.1, Collect the dataset used to train the prediction model; S1.2, Read the preprocessed sample file and preprocess the sample, including data loading, integrity verification, data cleaning and standardization, and sample quantity safety verification; S1.3, the polypeptide sequence is padded or truncated to a fixed length to fix the sequence to the maximum length L; feature vectors are extracted from each polypeptide sequence from three different dimensions: physicochemical parameters, position and length, and then fused to obtain a fused feature vector; S1.4, a prediction model is constructed based on the XGBoost regression model. The fused feature vector is input into the pre-trained prediction model, and the predicted binding force between the peptide and the target MHC Class II molecule is output. The peptide sequences are sorted according to the predicted binding force and the top-K peptide sequences are selected as candidate peptide sequences.

[0010] Furthermore, S1.3 includes the following: S1.3.1 For the input polypeptide sequence, extract continuous physicochemical parameters for each amino acid residue in the polypeptide sequence, construct a physicochemical parameter matrix and flatten it to obtain the physicochemical parameter feature vector; S1.3.2, perform one-hot position encoding on the input polypeptide sequence to obtain the position encoding feature vector; S1.3.3, normalize the length information of the input polypeptide sequence to obtain the length feature vector; S1.3.4, the physicochemical parameter feature vector, the position encoding feature vector, and the length feature scalar are fused to obtain a fused feature vector.

[0011] Furthermore, the physicochemical parameters include hydrophobicity, isoelectric point, polarity, side chain volume, net charge, number of hydrogen bond donors, number of hydrogen bond acceptors, aromaticity, flexibility index, solvent-accessible surface area tendency, secondary structure tendency, polarizability, aliphatic index contribution, surface exposure tendency, steric hindrance parameter, hydrophilicity, turn tendency, and β-sheet tendency.

[0012] Furthermore, S1.4 includes the following: S1.4.1, randomly divide the fused feature vector and label according to the ratio to obtain the training set and the test set; S1.4.2, Construct a prediction model based on the XGBoost regression model and configure the hyperparameters of the prediction model, including decision tree, learning rate of each tree, maximum depth, sample ratio, feature ratio, L1 regularization, L2 regularization coefficient and objective function. S1.4.3, the gradient boosting framework is used to train the prediction model; S1.4.4 evaluates the prediction model by calculating the mean squared error and the coefficient of determination, and makes predictions on the test set.

[0013] Furthermore, S3 includes the following: S3.1 After synthesizing the peptide on the peptide synthesis platform, the actual molecular weight is determined by mass spectrometry. If the absolute value of the difference between the actual molecular weight of the synthesized peptide and the theoretical molecular weight is not higher than 1 Da, the synthesis is considered successful, and S3.2 is executed; otherwise, the current synthesis sample is discarded; where 1 Da represents 1 Dalton. S3.2, The validated peptides are sent into the molecular binding force testing platform to obtain the true pIC50 value of each peptide, denoted as experimental_pIC50; S3.3, Based on the current prediction model, extract the fusion feature vector of the new sequence and predict it to obtain the predicted pIC50 value, denoted as predicted_pIC50; S3.4, calculate the difference between the actual pIC50 value and the predicted pIC50 value: mean_delta = experimental_pIC50 - predicted_pIC50; if If the model iteration optimization is triggered, S3.5 will be executed; otherwise, the new data will only be stored in the historical database, and the model will not be updated for the time being. S3.5: After S3.4 triggers the model iterative optimization, the peptide sequence and the true pIC50 value are added as new samples in the dataset. The prediction model adopts a strategy of combining incremental training and periodic retraining for iterative training. When the termination condition is met, it converges and the updated optimized model is output.

[0014] Furthermore, the strategy of combining incremental training with periodic retraining is specifically as follows: Incremental training: will satisfy The data is used for retraining; Periodic retraining: when the following conditions are met Training will begin after the number of data entries has been accumulated to a preset number.

[0015] Corresponding to the above method, the present invention also provides a data-driven prediction system for the binding affinity between MHC Class II molecules and peptides, comprising: The digital characterization module is used to digitally characterize peptide sequences. It uses a pre-trained prediction model to predict the binding force of peptide sequences, obtains and sorts the predicted binding force values, and selects the top-K peptide sequences corresponding to the predicted binding force values ​​as candidate peptide sequences. The peptide synthesis and molecular binding force testing platform is used to send candidate peptide sequences to the platform to obtain experimentally determined true binding force values. The closed-loop feedback optimization module is used to calculate the difference between the predicted binding force value and the actual binding force value. When the difference exceeds a preset threshold, it triggers model iterative optimization. The candidate peptide sequence and the corresponding actual binding force value are used as new samples to perform incremental training and retraining on the prediction model to obtain the updated prediction model.

[0016] An electronic device includes a memory and a processor, the memory being used to store a program that supports the processor in executing the above-described data-driven method for predicting the binding affinity of MHC Class II molecules and peptides, the processor being configured to execute the program stored in the memory.

[0017] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described data-driven method for predicting the binding affinity between MHC Class II molecules and peptides.

[0018] The advantages of this invention are: This invention triggers a model optimization process based on the difference between measured and predicted binding forces and a threshold comparison, ensuring that the iteration of the prediction model is targeted. It uses a closed-loop feedback optimization interface and SHAP interpretive analysis to locate feature contributions and identify the influence of key physicochemical parameters and sequence positions. During iterative optimization, new closed-loop data is incorporated for incremental retraining to quickly adapt to small sample updates, and historical data is integrated for periodic full-scale retraining to solidify the prediction model's generalization ability. This invention achieves continuous self-evolution of the model through closed-loop feedback, steadily improving the long-term prediction accuracy of MHC Class II molecules and peptide binding forces.

[0019] This invention inputs a polypeptide sequence and a target MHC Class II allele, extracts multidimensional physicochemical parameter matrices and other features to construct a fused feature vector, and digitally represents the positional information of the sequence, the physicochemical nature of the amino acids, etc., and inputs it into a pre-trained prediction model to obtain the predicted binding force value, effectively solving the defects of existing one-hot, BLOSUM and other discrete encoding methods that lose the continuous physical properties of amino acids.

[0020] This invention utilizes an automated platform to synthesize Top-K high-potential sequences. After online purification, these sequences are transferred by a robotic arm to a detection unit to determine their actual IC50 values. The difference between the measured molecular binding forces and the model's predicted values ​​is calculated, and this difference serves as the trigger criterion for iterative model optimization. The measured data is fed back to update the physicochemical matrix weights or model parameters, and this process is iterated until the target affinity threshold is reached or convergence occurs. Through closed-loop feedback, the model can achieve personalized optimization for specific targets or patient HLA types, making it particularly suitable for tumor neoantigen vaccines, therapeutic peptides for autoimmune diseases, and precision vaccine design.

[0021] This invention achieves an unattended closed loop from input sequence prediction to high-affinity candidate peptide production and data feedback, effectively improving R&D efficiency and significantly reducing manpower and reagent costs.

[0022] The prediction method proposed in this invention supports multiple MHC Class II alleles (human HLA-DR / DP / DQ, mouse H-2, etc.), is compatible with non-natural amino acids and modified peptides, and has strong universality. Attached Figure Description

[0023] Figure 1 This is a flowchart of the data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the dataset data storage format according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the code for preprocessing samples according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of digital characterization of a polypeptide sequence according to Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the code for digital characterization of a polypeptide sequence according to Embodiment 1 of the present invention; Figure 6 This is a heat map prediction based on a traditional one-hot encoding model and a multidimensional physicochemical parameter matrix model, according to Embodiment 1 of the present invention. Figure 7 This is a flowchart of the prediction model pre-training process in Embodiment 1 of the present invention; Figure 8This is a schematic diagram of the data set partitioning in Embodiment 1 of the present invention; Figure 9 This is a schematic diagram of the code for constructing the prediction model and training the model according to Embodiment 1 of the present invention; Figure 10 This is the interactive interface of the MHC Class II and peptide binding affinity prediction software in Embodiment 1 of the present invention. Figure 11 This is a schematic diagram of the structure of the polypeptide synthesis platform of Embodiment 1 of the present invention; Figure 12 This is a schematic diagram of the code for model iterative optimization in Embodiment 1 of the present invention; Figure 13 This is a schematic diagram of the prediction results of the software interaction interface in Embodiment 1 of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 1 Specifically, a data-driven method for predicting the binding affinity between MHC Class II molecules and peptides is disclosed, comprising the following steps: S1 involves digitally characterizing the peptide sequence, predicting its binding affinity using a pre-trained prediction model, obtaining and ranking the predicted binding affinity values, and selecting the top-K peptide sequences corresponding to the predicted binding affinity values ​​as candidate peptide sequences. Specifically, S1 includes the following: S1.1, Collect the dataset used to train the prediction model.

[0026] In this embodiment, the dataset is prepared by synthesizing peptides using a peptide synthesis and molecular binding force testing platform and detecting molecular binding force using an ELISA reader. The measured data is then presented as follows: Figure 2 The format is saved in a CSV file.

[0027] like Figure 2As shown in the figure, MHC represents the molecular type, length represents the number of amino acids in the polypeptide, peptide represents the polypeptide sequence, eq represents the sign relationship, ic50 is the core indicator for measuring binding force, the smaller the value, the stronger the binding force, and pIC50 is the negative logarithm of IC50. In this embodiment, pIC50 is used as the label for regression training.

[0028] In this embodiment, the peptide synthesis and molecular binding force testing platform can be an automated editing and synthesis peptide platform disclosed in the invention patent application number 202510283249.6.

[0029] S1.2, Read the preprocessed sample file and preprocess the sample, including data loading, integrity verification, data cleaning and standardization, and sample quantity safety verification.

[0030] Preferably, in this embodiment, data loading is performed first. The resolve_existing_file method is called to locate the input file, with the default file name being mhc_processed_smoke_new.csv. The pandas.read_csv interface is used to read the IEDBMHC Class II binding force dataset, which includes three core fields: peptide (peptide sequence), length (sequence length), and pIC50 (binding force tag), totaling 692 original records.

[0031] Secondly, perform integrity checks: check whether the data contains the required columns (corresponding to the core fields mentioned above), and throw an exception if they are missing to ensure the integrity of the input data.

[0032] Next, data cleaning and standardization were performed: Dropna was used to remove rows containing missing values; the peptide field was uniformly converted to string type and leading and trailing spaces were removed, and empty sequences were filtered out; the length and pIC50 fields were converted to numeric type using pd.to_numeric, and values ​​that failed to be converted were cleaned; finally, a clean and uniformly formatted valid sample dataset was obtained.

[0033] Finally, a sample quantity safety check is performed: if the number of valid samples after cleaning is less than 5, the program will actively throw an exception to avoid unstable model training due to insufficient samples.

[0034] like Figure 3 The following is the program content corresponding to step S1.2, which involves preprocessing the sample.

[0035] S1.3, Digital characterization of peptide sequences: The peptide sequences are padded or truncated to a fixed length, fixing the sequence to a maximum length L; feature vectors are extracted from each peptide sequence from three different dimensions: physicochemical parameters, position, and length, and then fused to obtain a fused feature vector; for example... Figure 4 As shown, the specific steps include: S1.3.1 For the input polypeptide sequence, extract 18-dimensional continuous physicochemical parameters for each amino acid residue in the polypeptide sequence, including hydrophobicity, isoelectric point, polarity, side chain volume, net charge, number of hydrogen bond donors, number of hydrogen bond acceptors, aromaticity, flexibility index, solvent-accessible surface area tendency, secondary structure tendency, polarizability, aliphatic index contribution, surface exposure tendency, steric hindrance parameter, hydrophilicity, turn tendency, and β-sheet tendency. After fixing the sequence to the maximum length L, construct an L×18 physicochemical parameter matrix and flatten it to obtain the physicochemical parameter feature vector.

[0036] Preferably, in this embodiment, the maximum sequence length L is set to 15. In this embodiment, the continuous physicochemical parameters are obtained by referring to the AAindex database, the Kyte-Doolittle hydrophobicity scale, the Grantham polarity value, and standard biochemical literature. Each amino acid residue corresponds to an 18-dimensional continuous vector. The specific parameter types are shown in Table 1 below: Table 1. Names and descriptions of continuous physicochemical parameters

[0037] Preferably, in this embodiment, the physicochemical parameter features are extracted using the extract_chem_features() function. Each amino acid is mapped to an 18-dimensional continuous physicochemical parameter vector, and 15 positions in the sequence are extracted separately to form a 15×18=270-dimensional flattened matrix. The physicochemical parameter feature matrix is ​​used to give the model physicochemical intuition, allowing the model to understand the properties of amino acids, such as the property that hydrophobic residues are conducive to binding.

[0038] S1.3.2, perform one-hot position encoding on the input polypeptide sequence to obtain the position encoded feature vector.

[0039] Preferably, in this embodiment, positional encoding feature extraction is achieved using the onehot_encode() function. The known sequence is fixed at length 15, and each position is represented by a 20-dimensional one-hot vector (corresponding to 20 amino acids), forming a 15×20=300-dimensional matrix. The positional encoding feature vector is used to accurately record which amino acid corresponds to which position, preserving position-specific information.

[0040] S1.3.3, normalize the length information of the input polypeptide sequence to obtain the length feature vector.

[0041] In this embodiment, each amino acid is assigned a position, for a total of 20 amino acids.

[0042] Preferably, the length feature vector is obtained by normalizing the peptide sequence length by dividing it by 20. Since the binding behavior of long peptides and short peptides may be different, the normalized length feature vector is used to provide global scale information, and the dimension of the length feature vector is 1.

[0043] S1.3.4, the physicochemical parameter feature vector, the position encoding feature vector, and the length feature scalar are fused to obtain a fused feature vector.

[0044] In this embodiment, numpy.hstack is used to horizontally concatenate the three features obtained in S1.3.1 to S1.3.4 above to obtain a 571-dimensional high-dimensional feature vector X, which is the complete multi-dimensional digital representation of each sample.

[0045] The fused matrix not only preserves the positional information of the sequence but also incorporates the physicochemical properties of the amino acids. This enables the model to understand important principles such as "hydrophobic residues tend to enter the MHC binding pocket" and "charged residues affect electrostatic complementarity," significantly improving the accuracy and generalization ability of the predictions. After the above multi-dimensional feature extraction and fusion, the original sequence data is transformed into a unified high-dimensional numerical feature matrix, providing multi-dimensional complementary information such as positional information, physicochemical properties, and global scale for subsequent XGBoost regression model training.

[0046] like Figure 5 The image shows part of the core code corresponding to step S1.3, which involves digitally characterizing the polypeptide sequence.

[0047] This invention proposes a digital characterization method based on a multidimensional physicochemical parameter matrix, upgrading the traditional discrete amino acid sequence encoding into a continuous vector matrix with physicochemical meaning (digital twin encoding). This effectively solves the defect of existing discrete encoding methods such as One-hot and BLOSUM that lose the continuous physical properties of amino acids. Traditional One-hot encoding treats amino acids as isolated symbols, losing the physical similarity between molecules, while the continuous physicochemical parameter matrix used in this invention endows the model with "physicochemical intuition".

[0048] like Figure 6 As shown, Figure 6 The middle left image shows hotspot predictions using a one-hot encoding model. Figure 6The right-middle figure shows the hotspot prediction using the multidimensional physicochemical parameter matrix model based on this invention. Experimental results show that on the same dataset, after adding the multidimensional physicochemical parameter matrix, the model's R² score increased from 0.23 with the basic one-hot encoding to over 0.57, and the correlation between prediction accuracy and experimental measurements was significantly improved. The matrix representation method proposed in step S1 has good scalability, allowing for the convenient addition of more physicochemical dimensions (such as topologically polar surfaces (TPSA), local charge distribution, etc.), or adaptive adjustment of parameter weights for specific MHC alleles, providing a solid foundation for subsequent closed-loop feedback optimization.

[0049] S1.4, a prediction model is constructed based on the XGBoost regression model. The fused feature vector is input into the pre-trained prediction model, which outputs the predicted binding force between the peptide and the target MHC Class II molecule. The peptide sequences are then sorted according to the predicted binding force values, and the top-K peptide sequences are selected as candidate peptide sequences. The predicted binding force value is represented by the pIC50 value. Specifically, as shown... Figure 7 As shown, S1.4 includes the following: S1.4.1, Dataset Partitioning: The `sklearn.model_selection.train_test_split` function was used to randomly split the fused feature matrix X (571 dimensions) and the label y (pIC50) in an 80:20 ratio, resulting in a training set (approximately 553 samples) and a test set (approximately 139 samples). The random seed was fixed at 42 to ensure the reproducibility of the experiment.

[0050] S1.4.2, Constructing a prediction model based on the XGBoost regression model: Preferably, this embodiment uses an XGBoost regressor (XGBRegressor) to construct the prediction model. The main hyperparameters of the prediction model are configured as follows: decision trees n_estimators=700, learning rate of each tree learning_rate=0.025, maximum depth of each tree max_depth=7, sample ratio and feature ratio used for each tree are 0.85 and 0.8 respectively, L1 regularization and L2 regularization coefficients are configured as 0.1 and 1.0 respectively, and a regression task objective function is adopted.

[0051] S1.4.3, Model Training: Preferably, offline training is performed using `model.fit(X_train, y_train)`. In this embodiment, the XGBoost regression model employs a gradient boosting framework to automatically learn the complex interactions and importance weights among the dimensions of the 571-dimensional features.

[0052] S1.4.4, Model Evaluation: Preferably, in this embodiment, predictions are made on a test set, the mean squared error (MSE) and the coefficient of determination (R²) are calculated, and a scatter plot of the true values ​​and predicted values ​​is drawn for visual verification.

[0053] like Figure 8 , Figure 9 The following is a portion of the core code for steps S1.4.1 to S1.4.4.

[0054] In this embodiment, the XGBoost regression model is trained by instantiating an XGBRegressor and configuring its hyperparameters, and then completing the offline training phase by calling model.fit(X_train, y_train). The XGBoost regression model automatically constructs multiple decision trees and continuously fits the residuals through gradient boosting to ultimately obtain a strong regression model.

[0055] Furthermore, this invention reserves a closed-loop feedback optimization interface, `feedback_optimize()`, after the model's offline training is completed. Based on the candidate peptide sequences selected in S1, the peptides are synthesized, purified, and their binding forces are measured using a peptide synthesis and molecular binding force testing platform. The experimentally measured pIC50 is then compared with the predicted pIC50 of the prediction model. When the difference exceeds a preset threshold, the `feedback_optimize()` interface is called to perform SHAP interpretive analysis, adding the new validation data to the training set, re-extracting the fusion feature vector, and incrementally / retraining the XGBoost model to achieve iterative optimization of the prediction model's hyperparameters.

[0056] The closed-loop mechanism proposed in this invention enables the prediction model to continuously self-evolve based on real experiments, significantly improving the long-term accuracy and industrial applicability of the model in predicting the binding affinity of MHC Class II molecules to peptides. For example... Figure 10 The image shows the interface between MHCClass II and the peptide binding affinity prediction software.

[0057] S2, sending the candidate polypeptide sequence to the polypeptide synthesis and molecular binding force testing platform to obtain the experimentally determined actual binding force value. Specifically, S2 includes the following: like Figure 11As shown, the entire peptide synthesis platform communicates remotely with each station via an industrial control computer according to different protocols. The user sets the types of peptides to be synthesized based on the candidate peptide sequences. Then, the industrial control computer receives JSON format instructions from the upper-level scheduling platform via remote communication, parses the instructions into control instructions for the corresponding stations, and sends them to each workstation within the peptide synthesis platform. After receiving the corresponding instructions, each station operates the corresponding consumable storage to start the corresponding work. The consumable storage includes: centrifuge tubes, dialysis cards, concentration tubes, and ELISA plates.

[0058] The peptide synthesis platform includes an industrial control computer, and a robotic arm, a ground rail, a hard rock control board, a peptide synthesizer, a nitrogen blowing and a peptide injection station, a liquid filling and capping station, a liquid chromatography purification station, a liquid nitrogen freezing station, a shaker, and a centrifuge, all connected to the industrial control computer. It also includes a single-axis guide rail, an electromagnetic system, and a peristaltic pump; after receiving instructions from the scheduling platform, the industrial control computer sends some control instructions to the Hardstone industrial control board, which drives the single-axis guide rail, the electromagnetic system, and the peristaltic pump according to the instructions.

[0059] In this embodiment, the working stations of the polypeptide synthesis platform, such as the polypeptide synthesizer, nitrogen blowing and aspiration liquid and polypeptide injection station, liquid filling and capping station, liquid chromatography purification station, liquid nitrogen freezing station, shaker, centrifuge and enzyme-linked immunosorbent assay (ELISA) reader, can adopt the same structure and connection relationship as those disclosed in the invention patent application number 202510283249.6.

[0060] The molecular binding force testing platform includes an ELISA reader and analysis software; the ELISA reader reads the original OD value or fluorescence value on the ELISA plate, exports the data as a CSV file, and the host computer program calls the analysis software to perform nonlinear regression fitting and calculate the pIC50 value; in this embodiment, GraphPad Prism analysis software is preferred.

[0061] Furthermore, a detailed description of each workstation in the peptide synthesis platform is provided: The peptide synthesizer used is a CEM Liberty Blue 2.0, which can automatically inject multiple amino acids into the synthesis chamber on demand for synthesis. Resin powder is added to DMF solution for swelling, and then a peristaltic pump is used to pump the DMF resin solution into the synthesis chamber. The bottom solenoid valve of the synthesis chamber is then opened, and the peristaltic pump drains the DMF solution from the chamber. The solenoid valve is then closed, and the synthesizer is started to inject amino acids to begin synthesis. After synthesis, a cleavage reagent is injected into the synthesis chamber to cleave the synthesized peptides from the resin. Then, sufficient DMF solution is injected into the synthesis chamber to swell the resin, rinse it clean, and drain it. This process achieves automatic resin addition and synthesis chamber rinsing.

[0062] The nitrogen blowing and aspiration liquid and peptide injection station moves the bottle to the catheter via the x-axis and z-axis to complete the nitrogen blowing and aspiration liquid and peptide injection operations. Compared to the conventional method that requires three catheters, this embodiment selects to reuse one catheter for nitrogen blowing and aspiration liquid. A two-position three-way solenoid valve is used to select which functional catheter is connected to the bottle, and a separate catheter is used for peptide injection to prevent cross-contamination.

[0063] The liquid filling and capping station uses a peristaltic pump to control the pumping of liquid, and the slider is controlled by a linear motor module.

[0064] The liquid chromatography purification station is composed of a Shimadzu liquid chromatograph, a three-axis moving module, and a multi-channel injection pump. The reagent to be purified is injected into the liquid chromatography quantitative loop, and the pump on the liquid chromatograph pumps the reagent from the quantitative loop into the detection cell for detection. When a peak appears in the chromatographic host computer interface, the signal is transmitted to the three-axis moving module. The three-axis moving module moves to the designated bottle opening and selects whether to discharge to the waste liquid tank or the receiving tube through a solenoid valve. Based on the signal automatically sent by the liquid chromatography software, the three-axis moves to the designated position to receive the reagent.

[0065] The liquid nitrogen freezing station includes a solid-state relay, a liquid nitrogen pump, and a freezing basin. Liquid nitrogen is pumped into the freezing basin by the liquid nitrogen pump, and then a robotic arm picks up the bottles to be frozen and transfers them to the liquid nitrogen basin. The solid-state relay is controlled by the digital pins of the industrial control board, thereby controlling the start and stop of the liquid nitrogen pump. The amount of liquid nitrogen injected is roughly controlled according to the start time.

[0066] The freeze dryer and freezing chamber include a freeze dryer, a vacuum chamber, a manifold, and a solenoid valve. The freeze dryer is started and stopped by a robotic arm simulating manual operation. The ground rail moves to a designated position, and the robotic arm clicks the button to start freeze drying and vacuuming. As vacuuming begins, the lid of the vacuum chamber is firmly held in place under negative pressure by the pressure applied by the robotic arm. After the freeze drying time is up, the robotic arm shuts off the freeze dryer and vacuuming. The industrial control board activates the solenoid valve to open the pressure relief valve, restoring the pressure inside the chamber to atmospheric pressure. The robotic arm then opens its grippers to push the lid open.

[0067] The shaker is controlled by a stepper motor (FOC) and works in conjunction with a robotic arm to accurately grip the bottles inside the shaker.

[0068] The clutch includes a clutch, a stepper motor, a three-phase asynchronous motor, and a frequency converter. The industrial control computer sends control commands to the centrifuge control board via 485 communication, setting the speed, centrifuge time, positioning hole position, door opening command, and door closing command. The centrifuge control board forwards the control commands. For example, when positioning, it closes the clutch, then connects the stepper motor and the three-phase asynchronous motor together to rotate the turntable to the designated position. If it receives a start centrifuge command, it opens the clutch, directly driving the frequency converter to drive the rotating turntable to start centrifuging. In addition, it is equipped with a serial port screen to view the set parameters in real time and to set corresponding control commands.

[0069] The overall control process for automated peptide synthesis and molecular binding force testing in this embodiment is as follows: The peptide synthesis platform uses the HTTP communication protocol and sends JSON-formatted commands to the industrial control computer via POST. The host URL is: http: / / 192.168.1.89:7080 / data. The control commands are as follows: S2.1, Take a bottle and place it at the receiving station for the cutting reagent, and unscrew the cap to prepare for receiving; take a centrifuge tube and place it at the num position of the liquid phase sample collection; receive the cut peptide solution.

[0070] In this embodiment, the software automatically generates the polypeptide sequence to be synthesized from the host computer. Then, DMF resin solution is injected through the polypeptide synthesizer. After the DMF is discharged, the remaining resin hangs on the filter screen. Various amino acids are then injected. After the synthesis is completed, a cleavage reagent is injected into the synthesis chamber to cut off the polypeptide hanging on the resin. Afterward, DMF solution is repeatedly circulated to rinse the resin clean for the next synthesis.

[0071] After the robotic arm is equipped with a rotary electric cylinder, it drives the guide rail to move to the centrifuge tube storage area to pick up a bottle, transfers it to the sample receiving station, and then moves the bottle on the station directly below the sample receiving port through a single-axis guide rail to start the sample receiving process.

[0072] S2.2, Transfer the received peptide solution to the nitrogen blowing station and start nitrogen blowing.

[0073] In this embodiment, the bottle receiving the peptide solution is transferred to the nitrogen blowing station by the cooperation of the ground rail and the robotic arm to accelerate the evaporation of the cutting reagent by nitrogen blowing. The nitrogen blowing or liquid suction function is selected by switching the solenoid valve. The nitrogen blowing function is controlled by the first solenoid valve and the nitrogen blowing pipeline is selected by the second solenoid valve.

[0074] S2.3, transfer to the ether injection station, start the ether injection, and cause the peptide to precipitate.

[0075] In this embodiment, a ground rail and a robotic arm work together to transfer the peptide solution from which the cutting reagent has been removed by nitrogen blowing to the ether dosing station, and the amount of ether dosing is controlled by a peristaltic pump.

[0076] S2.4, transfer the bottle from the ether station to the nitrogen blowing station, and then transfer the bottle from the nitrogen blowing station to the ether station for subsequent capping. After opening the centrifuge door, position it at the ID number hole.

[0077] In this embodiment, a ground rail and a robotic arm work together to transfer the ether-injected peptide solution to a nitrogen blowing station for mixing the ether and peptide solution.

[0078] S2.5, Transfer the bottles from the ether station to the centrifuge for centrifugation.

[0079] In this embodiment, a ground rail and a robotic arm work together to transfer the nitrogen-blown peptide solution to the capping station, where the cap is screwed on. Then, the centrifuge door is opened and rotated to the designated position, transferring the bottle into the centrifuge. Next, a balancing command is used to balance the centrifuge. Specifically, the centrifuge turntable rotates and positions itself symmetrically, then an empty bottle is picked up, the cap is opened, the appropriate amount of water is added, the cap is screwed on, and the bottle is transferred to the centrifuge for balancing. After these actions are completed, the centrifuge door is closed, the centrifugation speed and time are set, and centrifugation is started.

[0080] S2.6 After transferring the bottles in the centrifuge to the liquid suction station, start the liquid suction and discharge the supernatant.

[0081] In this embodiment, the centrifuge door is opened and rotated to the hole for placing peptide bottles. The bottles containing peptide solutions are transferred to the aspiration / nitrogen blowing combined station by the cooperation of the ground rail and the robotic arm. Then, the centrifuge is rotated to the balancing hole, the robotic arm throws the balancing bottles into the recovery port, and finally the aspiration is started.

[0082] S2.7, transfer the bottle to the water-adding station, add ml of water, and then transfer it to the shaker to dissolve.

[0083] In this embodiment, the peptide bottle after liquid aspiration is transferred to the water injection station, where a specified amount of water is added using a peristaltic pump, and then transferred to a high-precision positionable shaker for dissolution.

[0084] S2.8, after the polypeptide solution has been dissolved, transfer the vial to the liquid chromatography injection station.

[0085] S2.9, Start the liquid chromatography purification workstation to inject the sample and send the liquid chromatography command.

[0086] S2.10, transfer num bottles purified by liquid chromatography to the liquid nitrogen freezing station, starting from start_num bottles and transferring a preset number and freezing for a preset time.

[0087] In this embodiment, the purified bottles are transferred to liquid nitrogen freezing, and the liquid nitrogen pump is started and stopped by controlling a solid-state relay to pump liquid nitrogen into the liquid nitrogen basin. Considering that the number of purified bottles may exceed the number of stations in the freezing basin, this can be achieved by freezing in batches.

[0088] S2.11, Place the num bottles in the liquid nitrogen basin into the freeze-drying chamber and freeze for time minutes.

[0089] In this embodiment, after liquid nitrogen freezing, the sample needs to be transferred to a freeze dryer for freeze drying. This embodiment features a customized vacuum chamber for the freeze dryer to maintain the temperature and vacuum level within it. The vacuum chamber has two gas pipe interfaces: one connects to the freeze dryer's manifold for vacuuming and supplying the freezing atmosphere, and the other connects to a solenoid valve for pressure relief. The solenoid valve is controlled to select whether to connect to the outside environment. When vacuuming is closed, opening the solenoid valve allows for pressure relief, and then the vacuum chamber door can be opened to place or remove vials containing peptides.

[0090] S2.12, add ml of solvent to each of the num bottles in the freeze-drying chamber, transfer them to a shaker to assist in dissolution, and shake for time minutes.

[0091] S2.13, transfer num dialysis cards to the shaker, transfer the solutions from num bottles to the dialysis cards and dialyze for the preset time, open the centrifuge door and position it at num.

[0092] In this embodiment, the robotic arm and the ground rail work together to transfer the dialysis card from the storage to the dialysis stand. Then, the robotic arm changes the end effector to a pipette to transfer the solution dissolved in the centrifuge tube on the shaker to the dialysis card and use the shaker to assist dialysis.

[0093] S2.14, transfer the solution from dialysis card number num to concentration tube number num, tighten the cap, and then transfer it to the centrifuge port directly opposite the inlet. Place dialysis card number num and centrifuge tube number num in the recovery port, set the centrifuge speed and time, and then start the centrifuge.

[0094] In this embodiment, the robotic arm, in coordination with the ground rail, moves to a designated position to grip a concentration tube and places it at the capping station. The end effector is then replaced with a large-scale gripper. The cap is then unscrewed and placed at the designated position. The end effector is then replaced with a pipette. After inserting a tip, the solution in the dialysis card is transferred to the concentration tube. The end effector is replaced with the large-scale gripper again. The cap of the concentration tube is then screwed on, the centrifuge door is opened, and the turntable is rotated to port 0. The concentration tube is transferred to the centrifuge, and the centrifuge turntable is rotated to port 3, the symmetrical port of port 0, for centrifuge balancing. An empty centrifuge tube is gripped and transferred to the capping and water filling station. The cap is unscrewed, and water is added. The centrifuge tube with water added is then transferred to port 3 of the centrifuge. The centrifuge door is closed, and centrifugation and concentration are started.

[0095] S2.15, transfer the concentration tube from the well facing the window of the centrifuge to the clamping station, open the cap, then transfer the reagent to well number num of the ELISA plate, start the ELISA reader, and the robotic arm clamps the ELISA plate into the ELISA reader for testing.

[0096] In this embodiment, the centrifuge is positioned at well 0 and the hatch is opened. The robotic arm and ground rail work together to transfer the concentrated tube from well 0 to the capping station. After replacing the cap with a large-scale gripper, the cap is unscrewed. Then, the end effector is converted to a pipette to add the solution in the concentrated tube to the first well of the ELISA plate. The robotic arm then closes the concentrated tube and discards it into the recycling port. The centrifuge turntable rotates to well 3, the robotic arm changes the end effector to a large-scale gripper, and then removes the balanced centrifuge tube from the centrifuge. After reading the plate, the ELISA reader software can directly export a CSV file, and the host computer program can automatically call Prism to directly output the IC50 value, achieving full automation of the process.

[0097] S2.16, take bottle number num, add ml of water, balance the mixture, and transfer it to the centrifuge. Position the centrifuge in advance. Take a bottle from the centrifuge tube storage location and place it at the num position for liquid phase sample collection. Discard the bottle from the balanced centrifuge into the recycling port.

[0098] In this embodiment, the robotic arm transfers the concentrated sample to the ELISA plate, the ELISA reader reads the data, and the host computer automatically calls the analysis software (such as Prism) to calculate the pIC50 value, and the experimentally determined pIC50 value is used as the true binding force value.

[0099] S3, calculate the difference between the predicted binding force value and the actual binding force value. When the difference exceeds the preset threshold, trigger the model iterative optimization, take the candidate peptide sequence and the corresponding actual binding force value as new samples, and perform incremental training and retraining on the prediction model to obtain the updated prediction model.

[0100] This embodiment proposes a closed-loop feedback iterative optimization algorithm based on the difference between the actual measured molecular binding force and the model prediction value, which enables the prediction model to continuously self-evolve based on the difference in actual binding force.

[0101] After the automated peptide synthesis platform completes the solid-phase synthesis (SPPS) of the target peptide, it first verifies the consistency between the actual molecular weight and the theoretical molecular weight using mass spectrometry (MS) or liquid chromatography-mass spectrometry (LC-MS), retaining only the correctly synthesized peptides. Next, the molecular binding force testing platform measures the actual binding force of the verified peptides to obtain the experimental pIC50 value. Then, the system compares the experimentally measured actual pIC50 with the pIC50 predicted by the current model, and automatically triggers iterative optimization of the model based on the degree of difference.

[0102] In this embodiment, the input to the iterative optimization algorithm includes: The current prediction model, the set of newly synthesized and validated peptide sequences (new_peptides), the corresponding experimentally determined true binding affinity (experimental_pIC50), historical training data, and the difference threshold (threshold).

[0103] In this embodiment, the output of the iterative optimization algorithm includes the updated optimized model `updated_model`. The specific algorithm content is as follows: S3.1, Verification of Synthesis Correctness: After synthesizing the peptide on the peptide synthesis platform, the actual molecular weight is determined by mass spectrometry. If the absolute value of the difference between the actual molecular weight of the synthesized peptide and the theoretical molecular weight is not higher than 1 Da, the synthesis is considered successful, and S3.2 is executed; otherwise, the current synthesis sample is discarded; where 1 Da represents 1 Dalton, approximately equal to 1 g / mol.

[0104] In this embodiment, some impurities are generated during the peptide synthesis process. In order to obtain the target peptide, the pure peptide separated by liquid chromatography needs to be transferred to mass spectrometry for molecular weight identification (i.e., verification of the correctness of synthesis). The peptide with the molecular weight matching the target peptide is then used for subsequent operations, which effectively improves the accuracy of binding force determination.

[0105] S3.2, Determination of true binding force: The validated peptides are sent into the molecular binding force testing platform to obtain the true pIC50 value of each peptide, denoted as experimental_pIC50.

[0106] S3.3, Difference Calculation: Based on the current prediction model, extract the fused feature vector of the new sequence and predict it to obtain the predicted pIC50 value, denoted as predicted_pIC50.

[0107] like Figure 13As shown, with the peptide sequence NVKYLAKILCLKTEI as input, the prediction model outputs a predicted pIC50 value of 17450.29 nM, and gives the evaluation result that the MHC Class II molecule and the peptide are weakly bound.

[0108] S3.4, Optimize trigger judgment: Calculate the difference between the actual pIC50 value and the predicted pIC50 value, mean_delta = experimental_pIC50 - predicted_pIC50; if If the model is updated, iterative optimization will be triggered, and S3.5 will be executed; otherwise, the new data will only be stored in the historical database, and the model will not be updated for the time being.

[0109] Preferably, the difference threshold is set to 0.5 based on experience.

[0110] S3.5: After S3.4 triggers the model iterative optimization, the peptide sequence and the true pIC50 value are added as new samples to the dataset. The prediction model is trained iteratively using a strategy combining incremental training and periodic retraining. It converges when the termination condition is met, and the updated optimized model `updated_model` is output. The specific code is as follows: Figure 12 As shown.

[0111] In this embodiment, the strategy of combining incremental training with periodic retraining is specifically as follows: Incremental training: will satisfy The data is used for retraining; Periodic retraining: when the following conditions are met The data entries are accumulated to a preset number before training is performed; preferably, in this embodiment, the preset number is 500.

[0112] In this embodiment, the updated optimized model `updated_model` is saved, named `model_v{version}_iter{iteration}.json`. The mean delta, R² improvement, and number of new samples in this optimization iteration are recorded. Furthermore, this embodiment can also use SHAP values ​​to analyze which features contribute most to the new data, helping to explain the reasons for the optimization.

[0113] In this embodiment, the termination condition includes: The R² improvement was less than 0.01 for three consecutive test sets. The total number of iterations reaches a preset limit, preferably 50 rounds; The platform considers the number of verification samples to have reached a preset scale, preferably when at least 200 new samples are added.

[0114] In a preferred embodiment, after multiple rounds of optimization iterations, the mean squared error (MSE) of the initial 100 rounds of data was 0.4900 and the R² score was 0.5368, which was improved to 0.3995 and the R² score was 0.6223. The accuracy and realism of the model prediction were improved, and the predicted value obtained after the model iteration optimization was less than 1 nm from the actual measured value.

[0115] Example 2 An apparatus includes a memory and a processor, the memory being used to store a program that enables the processor to execute the data-driven MHC Class II molecule-peptide binding affinity prediction method of Embodiment 1, the processor being configured to execute the program stored in the memory.

[0116] Example 3 A storage medium storing a computer program, which, when executed by a processor, performs the steps of the data-driven method for predicting the binding affinity between MHC Class II molecules and peptides in Example 1.

[0117] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein can be considered as a ordered list of executable instructions for implementing logical functions, which can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-based system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0118] For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples of computer-readable media (a non-exhaustive list) include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CD-ROM). Additionally, a computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optical scanning of the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0119] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0120] In the description of this specification, references to terms such as "in a preferred embodiment," "preferred," "in this embodiment," "specific," or "furthermore," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0121] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven method for predicting the binding affinity between MHC Class II molecules and peptides, characterized in that, Includes the following steps: S1. Digitally characterize the peptide sequence, use a pre-trained prediction model to predict the binding force of the peptide sequence, obtain the predicted binding force value and sort it, and select the top-K peptide sequences corresponding to the predicted binding force values ​​as candidate peptide sequences. S2, send the candidate peptide sequence to the peptide synthesis and molecular binding force testing platform to obtain the experimentally determined true binding force value; S3, calculate the difference between the predicted binding force value and the actual binding force value. When the difference exceeds the preset threshold, trigger the model iterative optimization, take the candidate peptide sequence and the corresponding actual binding force value as new samples, and perform incremental training and retraining on the prediction model to obtain the updated prediction model.

2. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 1, characterized in that, S1 includes the following: S1.1, Collect the dataset used to train the prediction model; S1.2, Read the preprocessed sample file and preprocess the sample, including data loading, integrity verification, data cleaning and standardization, and sample quantity safety verification; S1.3, the polypeptide sequence is padded or truncated to a fixed length to fix the sequence to the maximum length L; feature vectors are extracted from each polypeptide sequence from three different dimensions: physicochemical parameters, position and length, and then fused to obtain a fused feature vector; S1.4, a prediction model is constructed based on the XGBoost regression model. The fused feature vector is input into the pre-trained prediction model, and the predicted binding force between the peptide and the target MHC Class II molecule is output. The peptide sequences are sorted according to the predicted binding force and the top-K peptide sequences are selected as candidate peptide sequences.

3. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 2, characterized in that, S1.3 includes the following: S1.3.1 For the input polypeptide sequence, extract continuous physicochemical parameters for each amino acid residue in the polypeptide sequence, construct a physicochemical parameter matrix and flatten it to obtain the physicochemical parameter feature vector; S1.3.2, perform one-hot position encoding on the input polypeptide sequence to obtain the position encoding feature vector; S1.3.3, normalize the length information of the input polypeptide sequence to obtain the length feature vector; S1.3.4, the physicochemical parameter feature vector, the position encoding feature vector, and the length feature scalar are fused to obtain a fused feature vector.

4. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 3, characterized in that, The physicochemical parameters include hydrophobicity, isoelectric point, polarity, side chain volume, net charge, number of hydrogen bond donors, number of hydrogen bond acceptors, aromaticity, flexibility index, solvent-accessible surface area tendency, secondary structure tendency, polarizability, aliphatic index contribution, surface exposure tendency, steric hindrance parameters, hydrophilicity, turn tendency, and β-sheet tendency.

5. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 2, characterized in that, S1.4 includes the following: S1.4.1, randomly divide the fused feature vector and label according to the ratio to obtain the training set and the test set; S1.4.2, Construct a prediction model based on the XGBoost regression model and configure the hyperparameters of the prediction model, including decision tree, learning rate of each tree, maximum depth, sample ratio, feature ratio, L1 regularization, L2 regularization coefficient and objective function. S1.4.3, the gradient boosting framework is used to train the prediction model; S1.4.4 evaluates the prediction model by calculating the mean squared error and the coefficient of determination, and makes predictions on the test set.

6. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 1, characterized in that, S3 includes the following: S3.1 After synthesizing the peptide on the peptide synthesis platform, the actual molecular weight is determined by mass spectrometry. If the absolute value of the difference between the actual molecular weight of the synthesized peptide and the theoretical molecular weight is not higher than 1 Da, the synthesis is considered successful, and S3.2 is executed; otherwise, the current synthesis sample is discarded; where 1 Da represents 1 Dalton. S3.2, The validated peptides are sent into the molecular binding force testing platform to obtain the true pIC50 value of each peptide, denoted as experimental_pIC50; S3.3, Based on the current prediction model, extract the fusion feature vector of the new sequence and predict it to obtain the predicted pIC50 value, denoted as predicted_pIC50; S3.4, calculate the difference between the actual pIC50 value and the predicted pIC50 value: mean_delta = experimental_pIC50 - predicted_pIC50; if If the model iteration optimization is triggered, S3.5 will be executed; otherwise, the new data will only be stored in the historical database, and the model will not be updated for the time being. S3.5: After S3.4 triggers the model iterative optimization, the peptide sequence and the true pIC50 value are added as new samples in the dataset. The prediction model adopts a strategy of combining incremental training and periodic retraining for iterative training. When the termination condition is met, it converges and the updated optimized model is output.

7. The data-driven method for predicting the binding affinity between MHC Class II molecules and peptides according to claim 6, characterized in that, The strategy of combining incremental training with periodic retraining is as follows: Incremental training: will satisfy The data is used for retraining; Periodic retraining: when the following conditions are met Training will begin after the number of data entries has been accumulated to a preset number.

8. A data-driven prediction system for the binding affinity between MHC Class II molecules and peptides, characterized in that, include: The digital characterization module is used to digitally characterize peptide sequences. It uses a pre-trained prediction model to predict the binding force of peptide sequences, obtains and sorts the predicted binding force values, and selects the top-K peptide sequences corresponding to the predicted binding force values ​​as candidate peptide sequences. The peptide synthesis and molecular binding force testing platform is used to send candidate peptide sequences to the platform to obtain experimentally determined true binding force values. The closed-loop feedback optimization module is used to calculate the difference between the predicted binding force value and the actual binding force value. When the difference exceeds a preset threshold, it triggers model iterative optimization. The candidate peptide sequence and the corresponding actual binding force value are used as new samples to perform incremental training and retraining on the prediction model to obtain the updated prediction model.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the data-driven MHC Class II molecule-peptide binding affinity prediction method according to any one of claims 1 to 7, the processor being configured to execute the program stored in the memory.

10. A storage medium storing a computer program, characterized in that, When the computer program is run by the processor, it performs the steps of the data-driven method for predicting the binding affinity of MHC Class II molecules and peptides as described in any one of claims 1 to 7.

Images