Classification model construction method based on immune repertoire sequence data, data classification method, MRD long-term monitoring necessity assessment method and storage medium

Abstract

The invention discloses a classification method and system based on immune repertoire sequence data and a storage medium. The method comprises the following steps: carrying out pretreatment of quality control, sequence comparison and clone annotation on immune repertoire sequencing original data; extracting multi-dimensional features including an immune diversity index, immune clonality, a cloning frequency top10, a Gini coefficient, VJ diversity, a 3-mer motif number and a frequency top10; performing Z-score standardization on the characteristics; adopting stratified sampling to divide a training set and a test set; and training the classification model by using a random forest algorithm, and optimizing hyper-parameters through grid search. By establishing a standardized multi-feature fusion analysis process, the technical problem of unstable model performance caused by insufficient feature utilization and non-uniform process in existing immune repertoire data analysis is solved, high-precision classification of samples is realized, and an objective basis is provided for clinical decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and bioinformatics, in particular to a classification model construction method based on immune repertoire sequence data, a data classification method, an MRD long-term monitoring necessity evaluation method and a storage medium. BACKGROUND

[0002] The rapid development of immune repertoire sequencing technology enables comprehensive acquisition of T cell receptor or B cell receptor diversity sequence information in biological samples. Subsequently, there is a technical demand for extracting valuable information from these massive sequence data. In the prior art, there are various schemes for extracting feature indicators from immune repertoire sequences.

[0003] For example, known technical solutions generally include the following processing flow: quality control of raw sequencing data, identification of clones through sequence alignment, and then calculation of certain statistical features such as Shannon entropy and clonality index. These features can reflect the state of the immune system to some extent. In terms of data processing methods, the prior art discloses technical means such as using fastp for quality control, Bowtie2 for sequence alignment, and IMGT / V-QUEST for sequence annotation.

[0004] In terms of model construction, random forest algorithm has been applied to biomedical data analysis due to its ability to handle high-dimensional features and strong resistance to overfitting. In the prior art, there are methods of optimizing model parameters through cross-validation and dividing data sets using stratified sampling.

[0005] Although the prior art provides various immune repertoire data analysis schemes, there are still the following technical defects: Firstly, the existing feature extraction scheme is often limited to a single or a few types of features, and does not fully utilize the multi-dimensional information contained in the immune repertoire data. This single feature analysis method results in limited discrimination ability and insufficient generalization performance of the constructed classification model.

[0006] Secondly, the prior art lacks a unified standard in the data processing flow. In particular, in key links such as clone identification criteria, low-quality data filtering threshold, and feature standardization method, different processing methods will cause significant differences in model performance, reducing the reproducibility and stability of the technical scheme.

[0007] In addition, the prior art lacks systematic optimization of key parameters such as data preprocessing, feature engineering and model training in the model construction process. For example, in the construction of a random forest model, the selection of key hyperparameters such as the number of decision trees and the maximum depth lacks clear technical guidance, resulting in suboptimal model performance.

[0008] In addition, existing technical solutions are often strongly coupled with specific clinical application scenarios. This close connection limits their applicability as a general data analysis method and also affects their ability to seek patent protection.

[0009] Based on the above analysis, the technical problem to be solved by this invention is: how to provide a standardized immune repertoire data analysis scheme that can systematically integrate multi-dimensional features, construct a stable and accurate classification model through optimized data processing flow and model parameter settings, while maintaining its technical attributes as a general data analysis method. Summary of the Invention

[0010] The purpose of this invention is to provide a method for constructing a classification model based on immune repertoire sequence data, a data classification method, a method for assessing the necessity of long-term MRD monitoring, and a storage medium. This addresses the problem mentioned in the background art that the existing technology lacks a standardized immune repertoire data analysis scheme that can systematically integrate multi-dimensional features, construct a stable and accurate classification model through optimized data processing flow and model parameter settings, while maintaining its status as a general data analysis method.

[0011] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a classification model based on immune repertoire sequence data includes the following steps: Step S1: Data preprocessing: Obtain raw sequencing data from the immune repertoire and perform quality control, including removing low-quality sequences, adapter sequences, and short sequences to obtain high-quality sequence data; compare the high-quality sequence data with reference sequences, identify and annotate immune receptor genes, set a clone sequence similarity threshold to define the same clone, and remove low-confidence clone sequences. Step S2: Feature Extraction Step: Based on the clone sequence data obtained in the data preprocessing step, calculate multiple features for each sample, including: immune diversity index, immune clonality, and clone frequency. Gini coefficient Diversity ,as well as frequency ; Step S3: Data Standardization Step: Perform Z-score standardization on the extracted features to eliminate the influence of units; Step S4: Dataset partitioning step: Divide the sample dataset labeled with category labels into training set and test set according to a preset ratio, where the category labels are set based on external medical evaluation results; Step S5: Model training step: Using the training set and its corresponding standardized feature data, the random forest algorithm is used to train the model, and a classification model is constructed by optimizing the hyperparameters.

[0012] Furthermore, in the data preprocessing step, the low-quality sequence refers to a sequence in which the proportion of bases with a quality value lower than Q20 exceeds 5%, the short sequence refers to a sequence with a length of less than 200 bp, the low-confidence clone sequence refers to a clone sequence with a sequencing depth of <10×, and the clone sequence similarity threshold is ≥98%.

[0013] Furthermore, the dataset partitioning step employs a stratified sampling strategy, dividing the samples into training and test sets in a 7:3 ratio, and ensuring that the proportion of each category in the training and test sets is consistent with the distribution of the original dataset.

[0014] Furthermore, in the model training step, the basic unit of the random forest algorithm is the CART decision tree, its splitting index is the Gini index, and the stopping conditions include reaching a preset maximum depth or meeting the node purity requirements; multiple decision trees are generated through bootstrap sampling and random feature selection, and the prediction results of each decision tree are integrated through a voting mechanism.

[0015] Furthermore, in the model training step, hyperparameters are optimized through grid search and cross-validation, and these hyperparameters include the number of decision trees. and maximum depth Their optimization ranges are respectively and The training termination condition is: in three consecutive rounds of cross-validation, the F1 value of the validation set fluctuates by ≤0.01, and the difference between the training error and the validation error is ≤0.02.

[0016] A method for classification using a classification model constructed using the method described above, comprising: Obtain the immune repertoire sequencing data of the sample to be tested; The data preprocessing step, feature extraction step, and data standardization step are performed on the immune repertoire sequencing data of the sample to be tested to obtain standardized feature data; The standardized feature data is then input into the trained classification model. Obtain the classification result output by the classification model.

[0017] Further data preprocessing steps include: quality control using FastP software, sequence alignment using Bowtie2, and immune receptor gene annotation using the IMGT / V-QUEST tool.

[0018] Furthermore, in the feature extraction step, the immune diversity index is calculated using Shannon entropy, with the following formula:

[0019] in, Let n be the frequency of the i-th clone, and n be the total number of clones; The immune clonality is calculated using the following formula:

[0020] Where n is the total number of immune cell clones in the sample. It is the frequency of the i-th immune cell clone.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the classification model construction method or a method for classifying data using the method.

[0022] A method for generating reference information for assessing the necessity of long-term MRD surveillance based on immune repertoire sequence data, characterized in that the method described above is applied, wherein: The classification results output by the classification model serve as reference information for assessing the necessity of long-term MRD monitoring for patients. The method includes the step of outputting the reference information.

[0023] Compared with the prior art, the beneficial effects of the present invention are: (1) High prediction accuracy and excellent model performance By integrating seven multidimensional features, including immune diversity index and clonality, and employing the random forest algorithm to construct a prediction model, high-precision classification performance was achieved. Validated on the test set, the model achieved an overall accuracy of 96%, with precision and recall exceeding 96% for the "no monitoring required" category, and precision and recall reaching 92% and 94%, respectively, for the "monitoring required" category. The model's comprehensive performance metrics, including an F1 score of 0.97 (class 0) and 0.93 (class 1) and an AUC of 0.99, indicate excellent discriminative ability and reliability. (2) It has strong technological innovation and breaks through traditional limitations.

[0024] By applying the multidimensional features of immune repertoire sequences (including immune diversity, clonality, Gini coefficient, VJ diversity, motif features, etc.) to the construction of classification models, we have broken through the technical limitations of traditional methods that rely solely on specific tumor gene mutations or protein biomarkers. This provides a new technical approach to solve the problem of inaccurate detection caused by tumor heterogeneity and antigen loss in existing technologies. (3) The technical solution is standardized and has good reproducibility.

[0025] The specific parameters and standards for each key step in the technical solution were clarified, including the quality filtering standard (Q20, 200bp) and clone identification standard (similarity ≥98%, sequencing depth ≥10×) in data preprocessing, and the hyperparameter combination in model training (n_estimators=100, max_depth=10). Through 10 repeated validations, the model performance standard deviation was ≤0.02, demonstrating the good stability and reproducibility of the solution, laying the foundation for industrial application. (4) It has significant clinical practical value and clear resource optimization effect.

[0026] Classification methods can provide objective evidence for clinical decision-making and help distinguish between patient groups that require long-term monitoring and those that do not. The disclosure document demonstrates through clinical data comparison that, after optimizing the monitoring strategy using the method of this invention, the average monitoring period for patients in all age groups is significantly shortened, effectively avoiding the problems of over-monitoring and under-monitoring, thereby improving the efficiency of medical resource utilization and reducing the burden on patients. Attached Figure Description

[0027] Figure 1 This is a flowchart of a classification model construction method based on immune repertoire sequence data according to the present invention; Figure 2 A visualization of the confusion matrix of the model's prediction results; Figure 3 The model's ROC curve (AUC=0.99); Figure 4 A flowchart illustrating the method for classifying data using the classification model constructed by the aforementioned classification model construction method. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0029] Please see Figure 1 This invention provides a technical solution: a method for constructing a classification model based on immune repertoire sequence data, comprising the following steps: Step S1: Data preprocessing: Obtain raw sequencing data from the immune repertoire and perform quality control, including removing low-quality sequences, adapter sequences, and short sequences to obtain high-quality sequence data; compare the high-quality sequence data with reference sequences, identify and annotate immune receptor genes, set a clone sequence similarity threshold to define the same clone, and remove low-confidence clone sequences. Step S2: Feature extraction step: Based on the clone sequence data obtained in the data preprocessing step, calculate multiple features for each sample, including: immune diversity index, immune clonality, top 10 clone frequencies, Gini coefficient, VJ diversity, motif count (3-mer), and top 10 motif frequencies (3-mer). The specific meanings of the above indicators are as follows: Immune diversity: used to measure the diversity of immune cell receptor sequences. In this invention, Shannon entropy is used for calculation, and the formula is:

[0030] in, Let n be the frequency of the i-th clone, and n be the total number of clones. A higher value indicates richer immune diversity.

[0031] Immune clonality: refers to the degree of homogeneity of an immune cell clonal population, quantified by the clonality index. A higher index indicates a greater likelihood of the presence of a dominant clone. The formula is: , Where n is the total number of immune cell clones in the sample (i.e., the number of unique clone sequences identified by immune repertoire sequencing, and low-confidence clones with sequencing depth <10× need to be removed). The frequency of the i-th immune cell clone The log-weighted sum of cloning frequencies (reflecting the dispersion of the cloning distribution; the smaller the value, the more concentrated the distribution). This is the reference value for maximum dispersion (when all cloning frequencies are exactly equal). ,at this time Clonal index = 0).

[0032] Top 10 Clonal Frequencies: This study selected the frequencies of the top 10 most abundant immune cell clones in the sample to analyze their dominant role in the overall immune response from the perspective of dominant clones. Dynamically monitoring the changing trends of these high-frequency clones not only allows for effective tracking of disease progression but also assists in evaluating treatment efficacy. For example, significant changes in the frequency of dominant clones during immunotherapy often indicate the effectiveness of the treatment regimen.

[0033] The Gini coefficient, a classic statistic in economics, is applied to immunological research after being mathematically derived from the Lorenz curve. It is used to quantify the degree of imbalance in the frequency of immune cell clones. Its formula is:

[0034] in, Represents the Gini coefficient. For the sample size, This represents the mean frequency of all immune cell clones. and They represent the first The and the first The values ​​for each sample range from 0 to 1. A value closer to 1 indicates a more concentrated clonal distribution, suggesting a possible abnormal immune response; conversely, a value closer to 0 indicates a more uniform clonal distribution and a normal physiological state of the immune system. For example, if the frequencies of five immune cell clones in a sample are 0.3, 0.25, 0.2, 0.15, and 0.1, with a mean of... =0.2, and substituting into the formula, we get the Gini coefficient G=0.18, indicating that the clone distribution is relatively uniform.

[0035] VJ diversity: This index focuses on the sequence diversity resulting from recombination of the variable region (V) and linker region (J) of immune receptors, revealing the richness of the immune repertoire at the molecular level. Random recombination of VJ gene fragments is the basis for the generation of diverse antigen receptors. The diversity of VJ recombination sequences is calculated using Shannon entropy, with the same formula as the immune diversity index, where is the frequency of the i-th VJ recombination sequence and n is the total number of VJ recombination sequences. This effectively assesses the immune system's ability to generate effective antigen receptors and provides an important reference for evaluating immune function.

[0036] Motif count (3-mer): As an important component of the morphological characteristics of the CDR3 region, this indicator counts the number of occurrences of a characteristic motif with a length of 3 amino acids. The CDR3 region is a key region for direct binding between the antigen receptor and the antigen, and its structural complexity directly affects the antigen recognition specificity. The frequency and distribution pattern of the 3-mer motif can effectively reflect the structural characteristics and antigen recognition potential of the CDR3 region.

[0037] Top 10 3-mer motif frequencies: This study extracted the 10 most frequently occurring 3-mer motifs and their proportions. By capturing high-frequency structural patterns in the CDR3 region, it provides direct clues for analyzing antigen recognition specificity. Different 3-mer motifs may correspond to specific antigenic epitopes. Tracking the dynamic changes of these high-frequency motifs helps to deepen the understanding of the immune system's response mechanisms to antigens.

[0038] Step S3: Data Standardization Step: Perform Z-score standardization on the extracted features to eliminate the influence of units; Step S4: Dataset partitioning step: Divide the sample dataset labeled with category labels into training set and test set according to a preset ratio, where the category labels are set based on external medical evaluation results; Step S5: Model training step: Using the training set and its corresponding standardized feature data, the random forest algorithm is used to train the model, and a classification model is constructed by optimizing the hyperparameters.

[0039] In the feature extraction step, the immune diversity index is calculated using Shannon entropy, with the following formula:

[0040] in, Let n be the frequency of the i-th clone, and n be the total number of clones; The immune clonality (referring to the degree of homogeneity of an immune cell clonal population, quantified by a clonality index; a higher index indicates a greater likelihood of the presence of a dominant clone) is calculated using the following formula:

[0041] Where n is the total number of immune cell clones in the sample. It is the frequency of the i-th immune cell clone.

[0042] See Figure 4 This application also provides a method for classification using a classification model constructed by the described method, comprising: The process involves acquiring the sequencing data of the immune repertoire of the sample to be tested; performing the data preprocessing, feature extraction, and data standardization steps on the sequencing data of the immune repertoire of the sample to obtain standardized feature data; inputting the standardized feature data into the trained classification model; and obtaining the classification result output by the classification model. Further, the data preprocessing steps include: quality control using FastP software, sequence alignment using Bowtie2, and immune receptor gene annotation using the IMGT / V-QUEST tool.

[0043] Specifically, it includes the following steps: The specific process of this method is as follows: 1. Data Acquisition and Preprocessing Raw sequencing data from patients' immune repertoires were collected, and the raw data underwent quality control using FASTP software. First, low-quality sequences generated during sequencing were removed (sequences with a quality value below Q20 exceeding 5% were defined as low-quality sequences). Then, sequencing adapter sequences and short sequences less than 200 bp were removed, ultimately yielding high-quality immune cell receptor sequence data. The quality-filtered sequence data was compared with reference sequences from the International Immunogene Information System (IMGT) or other authoritative immune repertoires using alignment tools such as Bowtie2. Clonal sequences were identified and extracted, and immune receptor gene annotation was performed using the IMGT / V-QUEST tool. Clonal sequences with a similarity ≥98% were considered the same clone, and low-confidence clones with a sequencing depth <10× were removed, providing accurate data for subsequent immune diversity index and clonality analysis.

[0044] 2. Feature Extraction

[0045] To construct a classification model for predicting the necessity of long-term MRD monitoring in patients, 745 clinical samples were first labeled. Of these, 496 samples with negative long-term MRD monitoring results were labeled as category 0, representing patients who do not require continuous monitoring; 249 positive samples were labeled as category 1, corresponding to patients requiring long-term monitoring. This classification method is closely related to clinical decision-making needs, providing a clear target variable for subsequent model training. The sample collection process strictly followed standardized operating procedures to ensure the representativeness and consistency of the samples.

[0046] For each sample, seven feature data categories were calculated: immune diversity, immune clonality, top 10 clone frequencies, Gini coefficient, VJ diversity, 3-mer motif count, and top 10 3-mer motif frequencies. The original values ​​exhibited significant dimensional differences (e.g., different units for immune cell proportions and gene fragment lengths). This study employed Z-score normalization to dimensionlessly process all feature data. This method maps all feature data to the same scale, effectively eliminating the potential impact of dimensions on model training and improving model convergence speed and prediction accuracy. In practice, the normalized parameters (mean and standard deviation) were calculated based on the training set and applied to the test set data transformation to avoid data leakage.

[0047] 3. Dataset partitioning

[0048] To ensure the constructed predictive model has good generalization ability and can stably perform predictive work on unknown clinical samples, this study adopted a stratified sampling strategy, dividing 745 labeled samples into a training set and a test set in a 7:3 ratio. The training set contained 521 samples, mainly used for learning and optimizing model parameters and building the model; the test set contained 224 samples, completely independent of the training process, and used solely to evaluate the model's predictive performance on unknown data and verify its practicality.

[0049] The sampling process was implemented using the StratifiedShuffleSplit module from the Scikit-learn library in Python. This module strictly controls the distribution ratio of positive and negative samples in the two datasets during the sampling process, ensuring that the ratio of negative samples (class 0) to positive samples (class 1) in both the training and test sets is approximately 496:249, similar to the original dataset. Specifically, the training set contains approximately 346 negative samples and 175 positive samples; the test set contains approximately 150 negative samples and 74 positive samples. This stratified partitioning strategy effectively avoids model bias caused by data imbalance, making the model training results more reliable and representative. Furthermore, the chi-square test was used to verify the inter-group distribution of the split training and test sets. The results showed P>0.05, proving that there was no significant difference in the distribution of the two sets of data, meeting the requirements for model training and evaluation.

[0050] 4. Model Training

[0051] This study employs the random forest algorithm to construct a predictive model for the necessity of long-term monitoring of minimal residual disease (MRD). Random forest, as an ensemble learning algorithm based on the Bagging strategy, integrates the prediction results of each base learner by constructing multiple decision tree classifiers and using a voting mechanism, demonstrating excellent performance in handling high-dimensional nonlinear data. This algorithm can effectively resolve the complex mapping relationship between immune cell characteristics (including immune diversity index, clonality, etc.) and MRD monitoring results, perfectly meeting the prediction needs of this study.

[0052] Let the training data be ,in , This application contains a total of 521 samples, each with 7 standardized characteristics (such as immune diversity index, clonality, etc.).

[0053] Tag definition: Category 0: Negative long-term MRD monitoring results indicate a patient group that does not require continuous monitoring; Category 1: Patients with positive long-term MRD monitoring results, corresponding to the patient group requiring long-term monitoring.

[0054] (1) Learning a single decision tree (CART algorithm)

[0055] The basic unit of a random forest is the CART decision tree. The partitioning process: CART uses a recursive binary split, continuously dividing the sample space into two child nodes based on the selected features of the samples, until a stopping condition is met (such as the tree reaching its maximum depth or the number of samples in a node being too small).

[0056] Split Indicators: In classification tasks, CART uses the Gini index to measure node purity.

[0057] in, denoted as the proportion of samples in node t that belong to category k (i.e., category 0 or category 1).

[0058] When all samples in a node belong to the same class, the Gini index is 0 (the "purest").

[0059] CART selects features and split points that minimize the Gini index at each step to divide the samples.

[0060] Stopping condition: In this application, the maximum depth is set to 10. When the depth of the tree reaches 10 or the node is "pure", the partitioning will stop.

[0061] The decision tree obtained through this training can progressively classify patient samples into "Class 0 (no monitoring required)" or "Class 1 (long-term monitoring required)".

[0062] (2) Generation of multiple decision trees

[0063] To improve prediction performance, random forests generate multiple differentiated decision trees in the following way: Bootstrap sampling (Bagging): Samples are drawn with replacement from the training set to form multiple training subsets.

[0064] Random feature selection: At each split node, the best split point is found only among a randomly selected subset of features, rather than considering all features.

[0065] In this application, grid search + 5-fold cross-validation is used to optimize hyperparameters: Number of decision trees Finally selected Maximum depth Finally selected Because when n_estimators=100 and max_depth=10, the average F1 score of the model's 5-fold cross-validation reaches 0.95, which is better than n_estimators=50 (F1=0.91), n_estimators=150 (F1=0.94) and max_depth=8 (F1=0.92) and max_depth=12 (F1=0.93), the optimal hyperparameter combination is determined.

[0066] (3) Integration results

[0067] Once all decision trees are trained, the random forest uses a voting mechanism to provide the final prediction: Each tree independently predicts one category (class 0 or class 1); The final output is the majority vote result:

[0068] in: Patients predicted to not require long-term monitoring; Patients predicted to require long-term monitoring.

[0069] (4) Training terminated

[0070] During training, we monitor the model's training error (i.e., the prediction error on the training set) and validation error (the prediction error during cross-validation) in real time. If the F1 score on the validation set fluctuates by ≤0.01 in three consecutive rounds of cross-validation, and the difference between the training error and the validation error is ≤0.02, it indicates that the model has reached its optimal training state. At this point, training is stopped to avoid overfitting.

[0071] 5. Model Evaluation

[0072] The trained random forest model was used to predict the results of 224 samples in the test set. The model was input with the 7-class standardized feature data of the samples and output the predicted label (0 or 1) for each sample, thus obtaining the complete prediction results. Figure 2 The prediction process employs batch processing to improve computational efficiency.

[0073] This application uses a confusion matrix as the core evaluation metric to comprehensively evaluate the classification performance of the model. The confusion matrix displays the matching between the model's predictions and the actual labels from four dimensions:

[0074] Based on the aforementioned confusion matrix, further derived metrics such as precision, recall, and F1 score are calculated to evaluate the model's accuracy and reliability in identifying negative samples (no long-term monitoring required) and positive samples (long-term monitoring required) from multiple dimensions, providing a quantitative basis for the model's practical application value. The calculation formulas and meanings of each metric are as follows:

[0075] Accuracy: This indicates the proportion of samples that were predicted to be positive by the model but were actually positive, reflecting the accuracy of the model's positive predictions and avoiding over-monitoring. Recall rate: This represents the proportion of samples that were actually positive but were correctly predicted as positive by the model, reflecting the model's ability to identify high-risk patients and avoiding under-monitoring. F1 value: Precision is the harmonic mean of precision and recall, which comprehensively reflects the overall classification performance of the model and avoids the limitations of evaluating a single metric.

[0076] To evaluate the stability of the model, this study repeated the data splitting (using stratified sampling each time, maintaining a 7:3 split ratio) and model training 10 times. After each training, the precision, recall, and F1 score of the model on the test set were calculated. Finally, the mean and standard deviation of the 10 results were taken to ensure the stability of the model performance.

[0077] At the same time, the ROC curve (Receiving Operator Characteristic curve) of the model was plotted. Figure 3 The AUC value (area under the curve) is calculated to visually demonstrate the model's classification performance at different prediction thresholds. The ROC curve uses the false positive rate (FPR) as the x-axis and the true positive rate (TPR, i.e., recall) as the y-axis. The closer the curve is to the upper left corner, the better the model's classification performance. The AUC value ranges from 0 to 1; the closer the AUC value is to 1, the stronger the model's discriminative ability.

[0078] Through comprehensive analysis of confusion matrix, ROC curve, and AUC value, the random forest model constructed in this invention demonstrates excellent accuracy, sensitivity, and specificity in predicting the necessity of long-term monitoring of minimal residual disease (MRD) in patients (validated on the test set, the model accuracy reaches 96%, and the AUC value reaches 0.99). In practical applications, the model evaluation results can be comprehensively interpreted in conjunction with clinical expert opinions to ensure that the evaluation conclusions meet actual clinical needs. Through the above comprehensive and rigorous model construction and evaluation process, the method for predicting the necessity of long-term monitoring of minimal residual disease based on immune cell receptor sequence characteristics proposed in this invention is expected to provide clinicians with objective and accurate decision support for MRD monitoring, effectively improve the current unreasonable monitoring strategies, and promote technological progress in the field of tumor diagnosis and treatment.

[0079] This invention, through systematic analysis of immune cell receptor sequence characteristics and combined with advanced machine learning algorithms, breaks through the limitations of traditional MRD monitoring that relies solely on tumor-related genes or protein biomarkers. It can provide clinicians with a more scientific and accurate basis for long-term patient MRD monitoring decisions, effectively improving the current polarized status quo of monitoring strategies.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a classification model based on immune repertoire sequence data, characterized in that, The method comprises the following steps: Step S1: data preprocessing: obtaining immune repertoire sequencing raw data, performing quality control, including removing low-quality sequences, adapter sequences and short sequences to obtain high-quality sequence data; aligning the high-quality sequence data with the reference sequence, identifying and annotating immune receptor genes, setting a clonal sequence similarity threshold to define the same clone, and removing low-confidence clonal sequences; Step S2: feature extraction step: based on the clone sequence data obtained by the data preprocessing step, a plurality of features of each sample are calculated, including: immune diversity index, immune clonality, clone frequency Gini coefficient, diversity, and frequency ; Step S3: data standardization: Z-score standardization processing is performed on the extracted plurality of features to eliminate the influence of dimension; Step S4: data set division: dividing the sample data set labeled with a class label into a training set and a test set according to a preset ratio, wherein the class label is set based on external medical evaluation results; Step S5: model training: using the training set and its corresponding standardized feature data, a random forest algorithm is used for model training, and a classification model is constructed by optimizing hyperparameters.

2. The method of claim 1, wherein, In the data preprocessing step, the low-quality sequence refers to a sequence with a base ratio of quality value less than Q20 exceeding 5%, the short sequence refers to a sequence with a length less than 200 bp, the low-confidence clonal sequence refers to a clonal sequence with a sequencing depth <10x, and the clonal sequence similarity threshold is ≥98%.

3. The method of claim 1, wherein, The data set division step adopts a stratified sampling strategy, divides the samples into a training set and a test set according to a ratio of 7:3, and ensures that the proportions of samples of various categories in the training set and the test set are consistent with the distribution of the original data set.

4. The method of claim 1, wherein, In the model training step, the basic unit of the random forest algorithm is a CART decision tree, the split index is the Gini index, and the stop conditions include reaching a preset maximum depth or the node purity meeting the requirements; a plurality of decision trees are generated through Bootstrap sampling and random feature selection, and the prediction results of the decision trees are integrated through a voting mechanism.

5. The method according to claim 1 or 4, characterized in that, In the model training step, hyperparameters including the number of decision trees and the maximum depth are optimized through grid search and cross-validation, and the optimization ranges are 10-100 and 5-20, respectively ; and ;​​ The training termination condition is that in the last 3 rounds of cross-validation, the F1 value of the validation set fluctuates ≤0.01, and the difference between the training error and the validation error is ≤0.

02.

6. A method for data classification using a classification model constructed by the method of any one of claims 1-5, wherein, It comprises: obtaining immune repertoire sequencing data of a to-be-tested sample; performing the data preprocessing step, the feature extraction step and the data standardization step on the immune repertoire sequencing data of the to-be-tested sample to obtain standardized feature data; inputting the standardized feature data into the trained classification model; obtaining the classification result output by the classification model.

7. The method of claim 6, wherein, The data preprocessing step comprises: using fastp software for quality control, using Bowtie2 for sequence alignment, and using IMGT / V-QUEST tool for immune receptor gene annotation.

8. The method of claim 6, wherein, In the feature extraction step, the immune diversity index is calculated by Shannon entropy, and the formula is: wherein, fi is the frequency of the ith clone, n is the total number of clones; The immune clonality is calculated by the following formula: where n is the total number of immune cell clones in the sample, is the frequency of the i-th immune cell clone.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the classification model construction method of any one of claims 1-5, or to implement the method of classifying of any one of claims 6-8.

10. A method for generating reference information for MRD long-term monitoring necessity assessment based on immunorepertoire sequence data, characterized by, The method of any one of claims 6 to 8 is applied, wherein: the classification result output by the classification model serves as reference information for evaluating the necessity of long-term MRD monitoring of a patient; The method comprises the step of outputting the reference information.

Images