2019-nCov case classification and clinical result prediction method based on machine learning
Through machine learning, combining RNA-seq and immune library characteristics, the problem of time-consuming and false negative nucleic acid detection is solved, and the accurate case classification and early infection detection of 2019-nCov was achieved, which improves the accuracy and sensitivity of the detection. Provides scientific basis for treatment and monitoring.
Patent Information
- Application Number
- CN202310869920.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-17
- Publication Date
- 2025-01-14
AI Technical Summary
The existing nucleic acid test takes a long time and has false negatives, making it difficult to accurately detect patients with early infection and asymptomatic infection, and cannot predict the development of the disease.
Using a machine learning-based approach, combining RNA-seq data and immune library characteristics, biomarkers were screened out through differential expression gene analysis and logistic regression model for 2019-nCov case classification and clinical outcome prediction.
Accurately distinguishing the 2019-nCov infected population has been achieved, the accuracy and sensitivity of detection has been improved, the infection can be detected early, the infection status and immunity can be judged, the false alarm rate and the missed alarm rate are reduced, and the scientific basis for treatment and monitoring is provided.
Smart Images
Figure CN119314662A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of biotechnology, and more specifically discloses a 2019-nCov case classification and clinical outcome prediction method based on machine learning. Background Art
[0002] For the treatment, prevention and monitoring of 2019-nCov, it is necessary not only to diagnose and classify cases quickly and accurately, but also to accurately predict the clinical development of 2019-nCov. At present, nucleic acid testing is used as the test result of the new coronavirus, and nucleic acid testing is also regarded as the gold standard for the detection of this disease, but nucleic acid testing is time-consuming and has a certain false negative, and usually multiple tests are required to ensure the correct test results. In addition, nucleic acid testing cannot accurately detect patients with early infection and asymptomatic infection, and cannot predict the development of the patient's condition. Summary of the invention
[0003] The main technical problem solved by the present invention is to provide a 2019-nCov case classification and clinical outcome prediction method based on machine learning, which can solve the problem that nucleic acid testing is time-consuming and has a certain amount of false negatives, and usually requires multiple tests to ensure that the test results are correct. In addition, nucleic acid testing cannot accurately detect patients with early infection and asymptomatic infection, and cannot predict the progression of the patient's condition.
[0004] To solve the above technical problems, according to one aspect of the present invention, more specifically, a 2019-nCov case classification and clinical outcome prediction method based on machine learning includes the following steps:
[0005] S1. Obtain RNA-seq data of PBMCs from 194 2019-nCov recovered patients, acute patients, and healthy subjects. Use Fastp to clean the raw RNA-seq data, and then use HI SAT38 to map to the human genome (GRCh38).
[0006] S2, FeatureCounts tool of Subread package was used to count the number of reads aligned to genes;
[0007] S3. Then, the differentially expressed genes were analyzed using the pyDESeq2 package. The differentially expressed genes were divided according to the following criteria: up-regulated genes were divided according to the criteria of adjusted p-value (pAdj) < 0.05 and log2FoldChange ≥ 1, and down-regulated genes were divided according to the criteria of adjusted p-value (pAdj) < 0.05 and log2FoldChange ≤ -1. Gene counts were normalized using pyDESeq2 and converted to log2 (normalized counts + 1) format, and then the batch effect was removed using the limma R package. The result was the normalized gene expression level.
[0008] S4. Use MiXCR to extract the clonal types of BCR and TCR, IGH, IGL, IGK, TRA, and TRB chains, and then analyze the number of clones (richness), diversity (diversity), and clonality (clonality) of each chain of IGH, IGL, IGK, TRA, and TRB;
[0009] S5. The Python module Sklearn, which integrates classic machine learning algorithms, was used to implement the above model. In order to evaluate the performance achieved by each feature combination, a triple cross-validation was performed. The samples were randomly and evenly divided into three groups, each group was called a fold. For each fold, one of them was used as a test set, and the remaining two folds were regarded as training sets. Then, this process was repeated until all folds were used as test sets. The performance of the model was evaluated by the average accuracy and average AUC of the prediction results of the three fold test sets. The top 500 combinations were sorted from large to small according to the AUC value, and the most common parameters were selected from the combinations with the top AUC. These parameters were used as the input combination of the model for training, verification, and prediction to obtain the final model and results. Finally, in order to determine the reliability of these biomarkers, a confusion matrix was established to show that the use of these biomarkers can distinguish the various disease courses of these samples with ideal accuracy.
[0010] Furthermore, the clonal richness is used to represent the number of unique clonal types of each sample to indicate the richness; unique clonal types are strictly defined by the exact match of receptor nucleotide sequences (CDR3 sequence, V / J fragment and hypermutation).
[0011] Furthermore, the diversity is expressed using the Shannon index:
[0012]
[0013] Where N represents the total number of clonal types in sample X, and p xiis the frequency of clonotype i. Generally speaking, more unique clonotypes and more uniform clonotype distribution will increase the diversity of a given library.
[0014] Uniformity:
[0015] E=H / H max
[0016] Among them, H is the Shannon index, H max is H when the probability of each clone appearing is the same; is the standardized Shannon index. The greater the uniformity, the better the diversity.
[0017] Furthermore, the clonality:
[0018] Clonality=1-E
[0019] The differentially expressed genes obtained above were merged with the immune repertoire index data, and the top differentially expressed genes and 15 immune repertoire index data of richness, diversity, and clonality of 5 chains, including IGH, IGL, IGK, TRA, and TRB, were selected as candidate markers. These markers were randomly combined to obtain a data set;
[0020] Using the logistic regression model as the classifier, the logistic regression function is expressed as follows:
[0021] logit=b0+b1f1+b2f2+...+b n f n
[0022] Among them, f1 is f n Indicates different elements; b0 represents the bias, b1 is from b n represents the weights of these features; the sigmoid function is applied to convert the logit into a score between 0 and 1. The sigmoid function is expressed as follows:
[0023] Score=1 / (1+e -logit ).
[0024] Furthermore, in S5, for the model training in this step, multivariate logistic regression (LR) is performed.
[0025] The beneficial effects of the 2019-nCov case classification and clinical outcome prediction method based on machine learning of the present invention are as follows: the present invention innovatively combines immune repertoire features with RNA-seq, and uses machine learning methods to screen a group of biomarkers, through which the 2019-nCov infected population can be predicted and the acute phase patients, the recovery phase patients and the unexposed healthy persons can be distinguished. The present invention has high accuracy. In this model, the AUC values of the acute phase, the recovery phase and the healthy population all reach 0.99, which can accurately distinguish the 2019-nCov infected population. The present invention can predict the early infected population. Other detection technologies usually need to detect the new coronavirus several days after infection, so they cannot detect the early infected population. At the same time, the gold standard for the detection of the new coronavirus, fluorescent quantitative PCR, cannot distinguish the different disease states (acute phase, recovery phase and healthy persons) after the new coronavirus infection, because the virus content in the recovery phase patients is extremely low, and its genes are difficult to be effectively detected. The present invention combines immune repertoire features with RNA-seq, which can not only detect the current infection status, but also determine whether the patient has been infected with the new coronavirus or has immunity to the new coronavirus. In particular, it has high accuracy even for mutant strains. Through the above advantages, the present invention can not only improve the accuracy and sensitivity of detection, reduce the false alarm rate and missed alarm rate, but also provide a scientific basis for the treatment and monitoring of patients. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0027] Figure 1 is a schematic diagram of the method flow chart;
[0028] Figure 2 It is the cluster heat map of differentially expressed genes;
[0029] Figure 3 This is a comparison of the B cell immune repertoire characteristics of the acute phase, recovery phase and healthy subjects (represented by IGH);
[0030] Figure 4 This is a comparison of the T cell immune repertoire characteristics of the acute phase, recovery phase and healthy subjects (represented by TRB);
[0031] Figure 5 It is the ROC curve diagram;
[0032] Figure 6 This is the confusion matrix result diagram. DETAILED DESCRIPTION
[0033] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.
[0034] According to one aspect of the present invention, Figure 1 As shown in the figure, a method for classifying 2019-nCov cases and predicting clinical outcomes based on machine learning is provided, including the following steps: a total of 194 2019-nCov recovered, acute patients and healthy PBMC RNA-seq data were obtained through RNA transcriptome sequencing experiments and public databases (68 recovery period sample data were obtained by laboratory collection sample library construction sequencing, 62 acute phase samples (30 of which were samples of patients infected with the variant strain-Omicron (Table 1)) and 64 healthy person data (Table 2) were obtained by SRA (sequence Read The RNA-seq raw data were downloaded from the R-Archive database, and Fastp was used to clean the RNA-seq data, and then mapped to the human genome (GRCh38) using HISAT38. The FeatureCounts tool of the Subread package was used to calculate the number of reads aligned to the gene, and then pyDESeq2 was used to analyze the differentially expressed genes. The differentially expressed genes were divided according to the following criteria: upregulated genes were divided according to the criteria of adjusted p value (pAdj) < 0.05 and log2FoldChange ≥ 1, and downregulated genes were divided according to the criteria of adjusted p value (pAdj) < 0.05 and log2FoldChange ≤ -1. Gene counts were normalized by pyDESeq2 and converted to log2 (normalized counts + 1) format. The batch effect was then removed by the limma R package, and the result was the normalized gene expression level ( Figure 2 ).
[0035] Table 1 Run number of RNA-seq data of PBMC in acute phase samples
[0036]
[0037]
[0038] Note: Acute1-30 is the acute phase data of Omicron
[0039] Table 2 Run number of RNA-seq data of PBMC samples from healthy subjects
[0040]
[0041]
[0042] MiXCR was used to extract the clonal types of BCR and TCR, IGH, IGL, IGK, TRA, and TRB chains, and then the richness, diversity, and clonality of each chain of IGH, IGL, IGK, TRA, and TRB were analyzed. Figure 3 , Figure 4 ). The differentially expressed genes obtained above were merged with the immune repertoire index data, and the 8 differentially expressed gene data with the top |log2FoldChange| values and 15 immune repertoire index data of richness, diversity, and clonality of 5 chains, including IGH, IGL, IGK, TRA, and TRB, were selected as candidate markers. These markers were randomly combined by selecting 1-5 each time to obtain a data set.
[0043] Use Python module Sklearn to implement logistic regression model. Randomly divide the samples into three groups, use these data sets for cross validation, sort the prediction results by AUC value from large to small, take the top 500 combinations, and find that the four gene expression parameters ENSG00000211896, ENSG00000143546, ENSG00000125538, and ENSG00000197223 appear in the top 80% of the combinations with AUC. At the same time, the three parameters IGK_clon ality, IGH_clonality, and IGL_clonality also appear in the top combinations. Therefore, these seven parameters are used as the input combination of the model for training, verification, and prediction to obtain the final model and results. In this model, the AUC of the acute phase, recovery phase, and healthy subjects all reach above 0.99 ( Figure 5 In the figure, the horizontal axis is the false positive rate (1-specificity), and the vertical axis is the true positive rate (sensitivity).
[0044] In order to evaluate the performance of the model, multiple evaluation indicators such as precision, sensitivity, specificity, and accuracy are used to evaluate the model. There are four situations in the test results: the true category of the sample is a positive example, and the result predicted by the model is also a positive example, that is, a true positive (Ture Positive TP); the true category of the sample is a negative example, and the model predicts it as a negative example, that is, a true negative (Ture Negative TN); the true category of the sample is a negative example, but the model predicts it as a positive example, that is, a false positive (False Positive FP); the true category of the sample is a positive example, but the model predicts it as a negative example, that is, a false negative (False Negative FN).
[0045] Precision refers to the ratio of correctly predicted positive data to predicted positive data. Precision reflects the model's ability to distinguish positive samples, the higher the better. The formula for calculating precision is:
[0046] Precision = TP / (TP+FP)
[0047] Sensitivity, also called recall, reflects the model's coverage of positive samples. The higher the better. The calculation formula for sensitivity is:
[0048] Sensitivity = TP / (TP+FN)
[0049] Specificity reflects the ability of a classifier or model to correctly predict the full range of negative samples, and increases the proportion of negative samples predicted as negative samples, that is, the proportion of negative samples predicted as negative samples to the total negative samples. The larger the TNR value, the better the performance. The specificity calculation formula is:
[0050] Specificity = FN / (TP+TN)
[0051] Accuracy is the most common evaluation indicator, and it is easy to understand. It is the percentage of correct predictions in the total number of samples, which is the number of correctly classified samples divided by the total number of samples. Generally speaking, the higher the accuracy, the better the classifier. Its calculation formula is:
[0052] Accuracy = (TP+TN) / (TP+TN+FP+FN)
[0053] To determine the reliability of this set of biomarkers, a confusion matrix was established ( Table 3 ), and the total number of test samples selected was 194, with an overall accuracy of 98.0%, indicating that the use of these biomarkers could distinguish various disease courses in these samples with ideal accuracy.
[0054] Table 3 Confusion matrix results
[0055]
[0056] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention also fall within the protection scope of the present invention.
Claims
1. A 2019-nCov case classification and clinical outcome prediction method based on machine learning, characterized in that: The following steps are involved: S1. Obtain RNA-seq data of PBMCs from 194 2019-nCov recovered patients, acute patients, and healthy subjects. Use Fastp to clean the raw RNA-seq data, and then use HI SAT38 to map to the human genome (GRCh38). S2, FeatureCounts tool of Subread package was used to count the number of reads aligned to genes; S3, then pyDESeq2 was used to analyze the differentially expressed genes; Differentially expressed genes were divided according to the following criteria: up-regulated genes were divided according to the criteria of adjusted p-value (pAdj) < 0.05 and log2FoldChange ≥ 1, and down-regulated genes were divided according to the criteria of adjusted p-value (pAdj) < 0.05 and log2FoldChange ≤ -1; gene counts were normalized by pyDESeq2 and converted to log2 (normalized counts + 1) format, and then the batch effect was removed by the limma R package, and the result was the normalized gene expression level; S4. Use MiXCR to extract the clonal types of BCR and TCR, IGH, IGL, IGK, TRA, and TRB chains, and then analyze the number of clones (richness), diversity (diversity), and clonality (clonality) of each chain of IGH, IGL, IGK, TRA, and TRB; S5. The Python module Sklearn, which integrates classic machine learning algorithms, was used to implement the above model. In order to evaluate the performance achieved by each feature combination, a triple cross-validation was performed. The samples were randomly and evenly divided into three groups, each group was called a fold. For each fold, one of them was used as a test set, and the remaining two folds were regarded as training sets. Then, this process was repeated until all folds were used as test sets. The performance of the model was evaluated by the average accuracy and average AUC of the test set predictions in the three folds. The top 500 combinations were sorted from large to small according to the AUC value, and the most common parameters were selected from the combinations with the top AUCs. These parameters were used as the input combination of the model for training, verification, and prediction to obtain the final model and results. Finally, in order to determine the reliability of these biomarkers, a confusion matrix was established to show that the use of these biomarkers can distinguish the various disease courses of these samples with ideal accuracy.
2. The 2019-nCov case classification and clinical outcome prediction method based on machine learning according to claim 1, characterized in that: The clonal richness is used to represent the number of unique clonotypes per sample to indicate richness; unique clonotypes are strictly defined by exact matches of receptor nucleotide sequences (CDR3 sequences, V / J segments, and hypermutations).
3. The 2019-nCov case classification and clinical outcome prediction method based on machine learning according to claim 1, characterized in that: The diversity is expressed using the Shannon index: Where N represents the total number of clonal types in sample X, and p xi is the frequency of clonotype i. Generally speaking, more unique clonotypes and more uniform clonotype distribution will increase the diversity of a given library. Uniformity: E=H / H max Among them, H is the Shannon index, H max is H when the probability of each clone appearing is the same; is the standardized Shannon index. The greater the uniformity, the better the diversity.
4. The 2019-nCov case classification and clinical outcome prediction method based on machine learning according to claim 1, characterized in that: The clonality: Clonality=1-E The differentially expressed genes obtained above were merged with the immune repertoire index data, and the top differentially expressed genes and 15 immune repertoire index data of richness, diversity, and clonality of 5 chains, including IGH, IGL, IGK, TRA, and TRB, were selected as candidate markers. These markers were randomly combined to obtain a data set; Using the logistic regression model as the classifier, the logistic regression function is expressed as follows: logit=b0+b1f1+b2f2+...+b n f n Among them, f1 is f n Indicates different elements; b0 represents the bias, b1 is from b n represents the weights of these features; the sigmoid function is applied to convert the logit into a score between 0 and 1. The sigmoid function is expressed as follows: Score=1 / (1+e -logit )。 5. The 2019-nCov case classification and clinical outcome prediction method based on machine learning according to claim 1, characterized in that: In S5, for the model training in this step, multivariate logistic regression (LR) is performed.