Filtering methods for second-generation sequencing RNA fusion false positives

Abstract

The present application relates to a filtering method for second-generation sequencing RNA fusion false positives, comprising: integrating fusion gene software results; software result statistics; feature selection; constructing a machine learning model; false positive database filtering; and optionally true positive database backtracking. The present application can effectively filter out false positives from RNA NGS results as much as possible and retain possible driver mutations, thereby effectively reducing the workload of reporters.

Description

TECHNICAL FIELD

[0001] The application relates to a filtering method for false positive RNA fusion of second-generation sequencing, and belongs to the technical field of bioinformatics. BACKGROUND

[0002] Next-Generation Sequencing (NGS), also known as high-throughput sequencing, is mainly characterized by high output and high depth, and can sequence millions of DNA molecules at a time. RNA NGS is a method for obtaining sequence information of original RNA fragments by reverse transcription of RNA fragments into cDNA through a reverse transcriptase and then sequencing the cDNA.

[0003] A fusion gene refers to a chimeric gene formed by the fusion of part of the sequences of two genes; it is generally caused by chromosomal translocation, deletion, inversion, etc.; such chimeric genes usually form abnormal transcripts and proteins in subsequent biological processes, thereby causing the occurrence of tumors.

[0004] Fusion gene detection has been included in the diagnosis and treatment guidelines for lung cancer, colorectal cancer and other cancers by NCCN and WHO guidelines. At present, the main methods for detecting fusion genes are fluorescence in situ hybridization (FISH), immunohistochemistry (IHC), PCR, DNA NGS and RNA NGS. Among them, RNA NGS has obvious advantages over the other several detection methods. On the one hand, NGS can not only detect known specific fusion genes, but also detect unknown fusion genes, and does not depend on the subjective judgment of the detection personnel; on the other hand, compared with DNA NGS, RNA NGS is simpler in probe design, and can improve the fusion detection rate and reduce false negatives. In addition, some gene fusions do not form new fusion transcripts, and these fusions do not appear in RNA NGS detection.

[0005] At present, there is a relatively mature analysis process and several relatively mature analysis software in the field of bioinformatics analysis of RNA NGS detection of fusion genes. After quality control and splicing, the sequencing data is aligned to the reference sequence by STAR, HISAT2 and other software; then a series of relatively mature software can find potential mutation sites from the data: such as StarFusion, FusionCatcher, JAFFA, SOPAfuse, TopHat-Fusion, Arriba and the like. The algorithms of these software are different, and some of them also have their own functions of filtering false positives; finally, many predicted fusion gene results are presented.

[0006] However, in actual case detection, most cases have only one driver fusion gene, and a small number of cases have multiple driver fusion genes. Even if there is a false positive filtering process, due to the limitations of library construction technology and alignment software, there are still many false fusion genes and non-driver fusion genes in the results of these fusion gene searching software.

[0007] At the same time, if the sample quality itself is not good enough, there will be a large number of false positives, which will consume a lot of time for related personnel and be very dependent on the experience of the reporting personnel when detecting large sample quantities. At the same time, due to sample sequencing reasons, some true fusions may have few reads supported and be filtered out by the software itself, which may cause missed detection. In addition, some gene fusions belong to activating fusions, which will increase the expression, so they will often have more read support and be considered as high-confidence true positive fusion genes by the software.

[0008] In summary, there is currently a lack of a technical means that can effectively filter false positives in RNA NGS detection results and find driver mutations.

[0009] Machine learning is a branch of artificial intelligence that aims to develop algorithms and techniques that allow computers to "learn", that is, to automatically improve their performance without explicit programming. It can be mainly divided into supervised learning, unsupervised learning and reinforcement learning; in supervised learning, the algorithm learns through a training data set, which contains input vectors and corresponding target outputs, and the goal of the algorithm is to find the mapping between input and output; unsupervised learning involves cases where there is no labeled data, and the algorithm must discover the structure and rules of the data itself; reinforcement learning is a decision-making process about rewards and punishments, and the algorithm learns how to act in the environment to maximize the cumulative reward obtained.

[0010] Machine learning has a wide range of applications in many fields, such as identifying disease patterns in medical diagnosis and predicting stock prices in financial markets. With the continuous advancement of technology and the increasing amount of available data, machine learning is becoming increasingly important.

[0011] Support Vector Machine (SVM) is a very popular machine learning algorithm that can be mainly used for pattern recognition, classification and regression analysis.

[0012] The core idea of SVM is to find the optimal decision boundary in the feature space, i.e. the so-called maximum margin hyperplane, which can separate different classes with the largest margin. In binary classification problem, this hyperplane is a line (in 2D space) or a plane (in 3D space), and a hyperplane in higher dimensional space. The so-called "support vectors" are the data points closest to this decision boundary, which play a decisive role in defining the maximum margin. The optimization goal of SVM is to maximize the minimum distance of these support vectors to the decision boundary, and then construct a classifier with the best generalization ability.

[0013] SVM algorithm has shown superior performance on many problems, especially on classification problems with small samples. When choosing the appropriate kernel function and adjusting the parameters (such as soft margin parameter), SVM can exhibit extremely high precision and excellent control of generalization error. Although the calculation of SVM on large-scale data sets may be relatively time-consuming and sensitive to parameter selection, it is still one of the indispensable machine learning tools in many fields. SUMMARY

[0014] The main purpose of the present application is to overcome the problems existing in the prior art based on the prior art conditions, to provide a filtering method for false positives of RNA fusion of second-generation sequencing, which can effectively filter out false positives from RNA NGS results as much as possible and retain possible driver mutations, thereby effectively reducing the workload of reporters.

[0015] The technical solution of the present application to solve its technical problems is as follows:

[0016] A filtering method for false positives of RNA fusion of second-generation sequencing, comprising the following steps:

[0017] First step, using the obtained RNA NGS data; using Arriba software to process RNA NGS data and obtaining the first prediction result; using StarFusion software to process RNA NGS data and obtaining the second prediction result, or / and using FusionCatcher software to process RNA NGS data and obtaining the third prediction result; in the first prediction result, selecting the prediction results appearing in the second prediction result at the same time or / and appearing in the third prediction result at the same time as the preliminary screening result.

[0018] Second step, count the true positives and false positives in the preliminary screening result, and mark them as the basis for model training and testing.

[0019] Third step, the results obtained in the second step are verified by correlation test, features determined as having high correlation with false positives according to preset standards are reserved, data of the features are collected and labeled as training and test data.

[0020] Fourth step, a support vector machine (SVM) in machine learning is selected as a classification model, a preset proportion of data in the data set obtained in the third step is randomly selected as a training set, the remaining data is used as a verification set, and the classification effect is evaluated, the model parameters are adjusted, and the sensitivity is preferentially ensured so that a driver fusion cannot be labeled as a false positive.

[0021] Fifth step, a false positive database constructed by known false positive fusion genes is used; the results labeled as true positives in the data processed by the classification model obtained in the fourth step are filtered by the false positive database, and finally fusion gene prediction results are obtained.

[0022] Further improved technical solutions of the application are as follows:

[0023] Preferably, the filtering method further comprises: sixth step, a known driver fusion gene database is used; fusion genes filtered out in the fifth step due to not reaching a threshold value of support reads but existing in the known driver fusion gene database are put into a candidate list; when a driver fusion gene is found in the fusion gene prediction results obtained in the fifth step, the candidate list is not considered; when no driver fusion gene is found in the fusion gene prediction results obtained in the fifth step, whether the fusion genes in the candidate list are driver fusion genes is verified.

[0024] Preferably, in the third step, the features having high correlation with false positives are selected from: site1, site2, type, split_reads1, split_reads2, coverage1, coverage2, confidence, reading_frame, and tags.

[0025] More preferably, in the third step, site1 is the position information of the first breakpoint; site2 is the position information of the second breakpoint; type indicates the type of the event that generates the fusion; split_read1 indicates the number of supporting reads of the left breakpoint with anchor support, and wherein the repeated reads are not included; split_read2 indicates the number of supporting reads of the right breakpoint with anchor support, and wherein the repeated reads are not included; coverage1 indicates the number of fragments retained on one side of the left breakpoint, and wherein the repeated reads are included; coverage2 indicates the number of fragments retained on one side of the right breakpoint, and wherein the repeated reads are included; confidence indicates the confidence of the fusion gene assessed by the software, selected from one of high, medium, and low; reading_frame indicates whether the reading frame of the 3' end gene of the fusion gene is complete, selected from one of complete, incomplete, and unable to judge; and tags indicate whether there is a record in the database, which includes the Mitelman database.

[0026] Preferably, in the second step, the data in the preliminary screening results are labeled as true positives and false positives according to known true positive sample data and false positive sample data as markers, and statistics are performed at the same time.

[0027] Preferably, in the third step, the statistical software package is used to analyze the correlation between the features and the labels of the results obtained in the second step; and the preset standard is that if the correlation index is greater than 0.7, it is determined that the correlation with the false positives is high.

[0028] Preferably, in the fourth step, the machine learning framework scikit-learn is used to train and evaluate the SVM classification model; a weighted evaluation standard including Fβ-measure is used, wherein the β parameter is set to 5; and when evaluating the classification effect, the performance of the SVM classification model is verified by referring to the confusion matrix and the ROC curve, respectively.

[0029] Preferably, in the fifth step, the false positive database is updated according to the changes of the known false positive fusion genes.

[0030] Preferably, in the sixth step, the data sources of the known driver fusion gene database include cosmic and ChimerDB.

[0031] By using the above preferred schemes, the specific technical features of each step can be further optimized.

[0032] Compared with the prior art, the method of the present application filters out false positives from the results of RNA NGS as much as possible and retains possible driver mutations, thereby effectively reducing the workload of the reporting personnel. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The specific implementation flowchart of embodiment 1 of the present application. DETAILED DESCRIPTION

[0034] The filtering method for false positives of second-generation sequencing RNA fusion of the present application comprises the following steps:

[0035] Step 1: Use the obtained RNA NGS data; use Arriba software to process the RNA NGS data and obtain the first prediction result; use StarFusion software to process the RNA NGS data and obtain the second prediction result, or / and use FusionCatcher software to process the RNA NGS data and obtain the third prediction result; in the first prediction result, select the prediction results that appear simultaneously in the second prediction result and / or appear simultaneously in the third prediction result as the preliminary screening result.

[0036] Note: Step 1 is the fusion gene software result integration step. In order to improve the accuracy of the fusion gene software prediction, the present application selects three excellent fusion gene software: StarFusion, FusionCatcher and Arriba; among them, the prediction result of Arriba software is better; in the prediction result of Arriba software, select the prediction that also appears in at least one of the other two software as the preliminary screening result.

[0037] Step 2: Count and mark the true positives and false positives in the preliminary screening result as the basis for model training and testing.

[0038] Note: Step 2 is the software result statistics step.

[0039] Step 3: Verify the results obtained in Step 2 through correlation test, retain the features that are judged to have high correlation with false positives according to the preset standard, collect the data of these features and add labels as training and testing data.

[0040] Note: The third step is the feature selection step. Among them, the features with high false positive correlation are, for example: site1, site2, type, split_reads1, split_reads2, coverage1, coverage2, confidence, reading_frame, tags. site1 is the location information of the first breakpoint (such as UTR, CDS, intron, exon, splice-site, etc.); site2 is the location information of the second breakpoint; type indicates the type of event that generates the fusion (such as translocation, inversion, duplication, deletion, etc.); split_read1 indicates the number of supporting reads of the left breakpoint with anchor support, and does not include repeated reads; split_read2 indicates the number of supporting reads of the right breakpoint with anchor support, and does not include repeated reads; coverage1 indicates the number of fragments retained on one side of the left breakpoint, and includes repeated reads; coverage2 indicates the number of fragments retained on one side of the right breakpoint, and includes repeated reads; confidence indicates the confidence of the software evaluation of the fusion gene, selected from one of high, medium, and low; reading_frame indicates whether the 3' end gene reading frame of the fusion gene is complete, selected from one of complete, incomplete, and unable to judge; tags indicate whether there is a record in the database, which includes the Mitelman database.

[0041] The fourth step is to select a support vector machine (SVM) classification model in machine learning, randomly select a predetermined proportion of data from the data set obtained in the third step as a training set, use the remaining data as a validation set, and evaluate the classification effect, adjust the model parameters, and prioritize sensitivity to ensure that the driver fusion cannot be labeled as false positive, and finally obtain the prediction result.

[0042] Note: The fourth step is the step of constructing a machine learning model. Among them, an example of a random selection ratio is: randomly select 70% from the data set as a training set, and 30% as a validation set.

[0043] The fifth step is to use a false positive database constructed from known false positive fusion genes; filter the results labeled as true in the prediction results obtained in the fourth step with the false positive database, and finally obtain the fusion gene prediction results.

[0044] Note: The fifth step is the false positive database filtering step. Among them, the known false positive fusion genes used to construct the false positive database can be false positive fusion genes commonly collected by the laboratory / project team, can be some common read-through fusions, can be some false fusions that often occur during library construction, and can be some circular RNAs, some passenger rearrangements, etc.

[0045] Step 6, use known driver fusion gene database; put the fusion genes filtered out in step 5 due to the number of supporting reads not reaching the threshold but existing in the known driver fusion gene database into a candidate list; when a driver fusion gene is found in the fusion gene prediction results obtained in step 5, the candidate list is not considered; when no driver fusion gene is found in the fusion gene prediction results obtained in step 5, it is verified whether the fusion gene in the candidate list is a driver fusion gene.

[0046] Note: Step 6 is a true positive database backtracking step. Due to the fact that in a few cases, the true positive fusion may have very few supporting reads due to sequencing and sample quality problems in the database construction, the number of supporting reads may not reach the threshold and the fusion may be filtered out by the software. Therefore, a database of known driver fusion genes is constructed according to the fusions in the known database and the accumulated data in the laboratory, and the fusions existing in the database and filtered out by the software due to too few supporting reads are listed separately in a candidate list, when a driver fusion is found, the list does not need to be considered, and when no driver fusion is found, it can be verified whether the fusion in the list is true.

[0047] The application will be further described in detail below with reference to the accompanying drawings and in conjunction with the embodiments. However, the application is not limited to the examples given.

[0048] Example 1

[0049] The specific implementation process of this embodiment is shown in Figure 1 .

[0050] The specific implementation steps of this embodiment are as follows.

[0051] 1. Data preparation

[0052] 1.1 Obtain software preliminary screening results

[0053] 1.1.1 Run StarFusion, FusionCatcher, and Arriba software in a parallel computing environment by writing an automated script to ensure efficient and accurate data processing.

[0054] 1.1.2 Use a Python script to automatically import the result files of the three software: star-fusion.fusion_predictions.abridged.tsv, summary_candidate_fusions.txt, and fusions.tsv, parse the results, and obtain fusion gene prediction lists of different software.

[0055] 1.1.3 Apply multiple matching algorithm, consider gene name standardization and approximate matching technique (e.g. use a range of breakpoint tolerance) to bridge the difference in gene identification.

[0056] 1.1.4 Integrate fusion event results, for Arriba results, only keep those fusion predictions that are also identified at least once by StarFusion or FusionCatcher to reduce false positive rate.

[0057] 1.1.5 Save the filtered results in a data file in a unified table format for easy follow-up processing.

[0058] 1.2 Label the data

[0059] 1.2.1 Collect positive and false positive sample data from clinical reports and lab logs.

[0060] 1.2.2 Label the data as true positive and false positive according to the information collected in 1.2.1, then organize them in a unified format and export, for feature selection and model training steps, the total of 1140 fusion predictions, among which there are 97 positive results and 1043 false positive results.

[0061] 1.3 Feature selection

[0062] 1.3.1 Use statistical software package (R's corrplot package) to analyze the correlation between features and labels.

[0063] 1.3.2 For high correlation between features, use model-based feature selection to remove redundant features.

[0064] 1.3.3 Keep features with high correlation (>0.7) with positive labels, the remaining features are: site1, site2, type, split_reads1, split_reads2, coverage1, coverage2, confidence, reading_frame, tags. To improve credibility and interpretability, the following table shows the potential relationship between these features and false positives:

[0065]

[0066]

[0067] 2. Establishment and evaluation of classification model

[0068] 2.1 Training set and test set selection

[0069] 2.1.1 Due to the imbalance of data, stratified sampling method is used to separate the test set and training set from the unbalanced dataset, 30% of the data is extracted as the test set and 70% as the training set.

[0070] 2.1.2 Preprocess the training set, encode the label and standardize the data to meet the input requirements of the SVM model.

[0071] 2.2 Model selection and evaluation

[0072] 2.2.1 Apply the machine learning framework scikit-learn to train and evaluate the SVM model.

[0073] 2.2.2 Due to the class imbalance, weighted evaluation criteria including Fβ-measure are used. Considering that finding the significance of driving fusion in detection is very important, false negatives should be avoided as much as possible, so the β parameter is set to 5 to increase the weight of true positive prediction accuracy.

[0074] 2.2.3 In addition to Fβ-measure, multiple indicators such as confusion matrix, ROC curve, etc. are used to verify the performance of the model from multiple aspects. Pay special attention to the false negative rate and the false positive rate (false positives can also be removed by manual screening at the end), because it is related to the diagnosis and subsequent treatment decision of the patient.

[0075] 2.3 The confusion matrix of the prediction results of the test set is as follows:

[0076] Predicted true positives Predicted false positives Actual true positives 68 0 Actual false positives 153 577

[0077] The accuracy of the prediction results is 30.77%, the recall rate is 100%, and the F5-measure is 0.9204.

[0078] 3. Database-based filtering

[0079] 3.1 Filtering of false positive database

[0080] 3.1.1 Due to the need to ensure high recall rate in step 2.2, a certain number of false positives will still be retained. In particular, some experimental processes will appear fusion, which often has some laboratory specificity, so building an updateable false positive database based on laboratory and clinical data results can further filter out these false positives to a large extent.

[0081] 3.1.2 Write a Python script to perform database screening on each predicted fusion event based on the above database, and remove known false positive events. The results of the false positive database after filtering are as follows:

[0082]

[0083] 3.2 Retrospective of true positive database

[0084] 3.2.1 Further set an updateable true positive database based on step 3.1, including but not limited to cosmic, ChimerDB and other database sources and laboratory collected true positive information, in case of potential true positive fusion events that may exist but are filtered out by software due to too few supporting reads.

[0085] 3.2.2 Write a retrospective query script to scan the filter list, identify potential positive fusion events, and record them for backup.

[0086] 3.2.3 Create a comprehensive database management system to facilitate information exchange and synchronization between different databases, ensure data traceability and analysis accuracy.

[0087] Based on all the above embodiments, the key technical points of the present application are as follows:

[0088] 1. Feature selection for building machine learning model, select and judge some feature combinations with higher false positive correlation and smaller correlation between each feature from the results of Arriba software: site1, site2, type, split_reads1, split_reads2, coverage1, coverage2, confidence, reading_frame, tags, including the support vector machine model built later.

[0089] 2. Database additional filtering step, combined with machine learning model can filter out false positives in various possible situations.

[0090] 3. Model evaluation index, due to the imbalance between negative and positive samples, and in detection, it is unacceptable to treat true positives as false positives, so the evaluation index takes F5 value, although there will be more false positives, but can avoid false negatives.

[0091] In summary, the method of the present application combines the results of multiple fusion gene analysis software, extracts features related to false positives to build a feature matrix, and then builds a prediction model through support vector machine algorithm (SVM) in machine learning to filter the fusion gene results predicted by software, thereby marking the false positives in the results. The present application can effectively filter out false positives from RNA NGS results and retain possible driver mutations, thereby effectively reducing the workload of the reporting personnel.

[0092] In addition to the embodiments described above, the present application can have other implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the protection scope of the present application.

Claims

1. A method of filtering second-generation sequencing RNA fusion false positives, characterized in that, The method comprises the following steps: Step 1: using the obtained RNA NGS data; using Arriba software to process the RNA NGS data and obtaining a first prediction result; using StarFusion software to process the RNA NGS data and obtaining a second prediction result, or / and using FusionCatcher software to process the RNA NGS data and obtaining a third prediction result; in the first prediction result, selecting the prediction results appearing in the second prediction result at the same time or / and appearing in the third prediction result at the same time as the preliminary screening results; Step 2: counting and marking the true positives and false positives in the preliminary screening results as the basis for model training and testing; Step 3: verifying the results obtained in Step 2 through correlation testing, retaining features with high correlation degrees with false positives according to preset standards, collecting data of the features and adding labels as training and testing data; Step 4: selecting a support vector machine (SVM) in machine learning as a classification model, randomly selecting a preset proportion of data from the data set obtained in Step 3 as a training set, using the remaining data as a validation set, and evaluating the classification effect, adjusting the model parameters, and prioritizing sensitivity to ensure that the driver fusion cannot be labeled as false positive; Step 5: using a false positive database constructed with known false positive fusion genes; filtering the results labeled as true positives in the data processed by the classification model obtained in Step 4 with the false positive database, and finally obtaining fusion gene prediction results.

2. The method of claim 1, wherein the method further comprises, The filtering method further comprises Step 6: using a known driver fusion gene database; placing fusion genes that are filtered out in Step 5 due to support read numbers not reaching a threshold but exist in the known driver fusion gene database into a candidate list; when a driver fusion gene is found in the fusion gene prediction results obtained in Step 5, the candidate list is not considered; when no driver fusion gene is found in the fusion gene prediction results obtained in Step 5, it is verified whether the fusion genes in the candidate list are driver fusion genes.

3. The filtering method for false positives of RNA fusion of second-generation sequencing according to claim 1, wherein the features with high correlation degrees with false positives in Step 3 are selected from site1, site2, type, split_reads1, split_reads2, coverage1, coverage2, confidence, reading_frame, and tags.

4. The method of claim 3, wherein the method further comprises, In the third step, site1 is the position information of the first breakpoint; site2 is the position information of the second breakpoint; type indicates the type of the event that generates the fusion; split_read1 indicates the number of supporting reads of the left breakpoint with anchor support, and does not include duplicate reads; split_read2 indicates the number of supporting reads of the right breakpoint with anchor support, and does not include duplicate reads; coverage1 indicates the number of fragments retained on one side of the left breakpoint, and includes duplicate reads; coverage2 indicates the number of fragments retained on one side of the right breakpoint, and includes duplicate reads; confidence indicates the confidence of the fusion gene evaluated by the software, selected from one of high, medium, and low; reading_frame indicates whether the reading frame of the 3' end of the fusion gene is complete, selected from one of complete, incomplete, and unable to judge; and tags indicate whether there is a record in the database, which includes the Mitelman database.

5. The method for filtering false positives of second-generation sequencing RNA fusion according to claim 1, characterized in that, In the second step, the data in the preliminary screening results are labeled as true positives and false positives according to known true positive sample data and false positive sample data as markers, and statistics are performed at the same time.

6. The method of claim 5, wherein the method further comprises: In the third step, a statistical software package is used to analyze the correlation between the characteristics and the labels of the results obtained in the second step; the preset standard is that if the correlation index is greater than 0.7, it is determined that the correlation with the false positive is high.

7. The method of claim 6, wherein the method further comprises, In the fourth step, a machine learning framework scikit-learn is used to train and evaluate the SVM classification model; a weighted evaluation standard including Fβ-measure is used, wherein the β parameter is set to 5; when evaluating the classification effect, the SVM classification model is verified for performance by referring to the confusion matrix and the ROC curve, respectively.

8. The method of claim 7, wherein the method further comprises, In the fifth step, the false positive database is updated according to the known false positive fusion gene changes.

9. The method of claim 2, wherein the method further comprises, In the sixth step, the data sources of the known driver fusion gene database include cosmic and ChimerDB.

Images