Method for constructing prediction model

Abstract

The present invention provides a method for constructing a prediction model, which obtains synthetic sample data by means of mixing sequencing data of positive samples and sequencing data of negative samples. Therefore, the efficiency of model construction is improved.

Description

A method for building predictive models Technical Field

[0001] This application relates to the field of bioinformatics and to a method for constructing predictive models. Background Technology

[0002] Existing prediction models may suffer from problems such as overfitting, poor robustness, and weak generalization ability, resulting in low sensitivity and poor accuracy in the final prediction model.

[0003] There has always been a need in this field for prediction models with high sensitivity and high accuracy. Summary of the Invention

[0004] To address the aforementioned technical problems, the inventors have developed a highly sensitive and accurate prediction model by training the model using positive, synthetic, and negative samples. The inventors discovered that sampling and mixing sequencing data from positive samples (e.g., late-stage / high-tumor-percentage positive samples) with negative samples can simulate the signal of early-stage / low-tumor-percentage positive samples. This method also provides a larger number of training samples, ultimately enabling data augmentation. Training the machine learning model with this augmented data yields a highly sensitive and accurate prediction model.

[0005] In one aspect, the present invention provides a method for constructing a predictive model, comprising: obtaining sequencing data of positive samples and sequencing data of negative samples; sampling and mixing the sequencing data of the positive samples and the sequencing data of the negative samples to obtain synthetic sample data; using the positive samples and synthetic sample data as the experimental group and the negative sample data as the control group to train a machine learning model.

[0006] In one aspect, the present invention also provides a prediction model constructed using the method of the present invention.

[0007] In one aspect, the present invention also provides a data augmentation method for improving the performance of a predictive model, comprising: obtaining sequencing data of positive samples and sequencing data of negative samples, sampling and mixing the sequencing data of positive samples and sequencing data of negative samples to obtain synthetic sample data, and using the synthetic sample data as training data for machine learning.

[0008] In some implementations, the positive sample is a late positive sample.

[0009] In some implementations, the synthetic sample data is obtained by sampling and mixing the sequencing data of a positive sample with the sequencing data of multiple negative samples. Preferably, the sequencing data of a positive sample is sampled and mixed with the sequencing data of 2, 3, 4, 5, 6, 7, 8, 9 or 10 negative samples.

[0010] In some implementations, a positive sample is paired with multiple negative samples.

[0011] In some implementations, sequencing data from a positive sample are mixed proportionally with sequencing data from multiple negative samples.

[0012] In some implementations, sequencing data from a positive sample are mixed with sequencing data from multiple negative samples according to a fixed target tumor percentage gradient.

[0013] In some implementations, sequencing includes targeted methylation sequencing and / or whole-genome sequencing (e.g., low-depth whole-genome sequencing).

[0014] In some implementations, the tumor percentage of each sample is obtained based on whole-genome sequencing data analysis. Negative samples whose tumor percentage falls within the 95th percentile of the negative population are defined as negative samples, and positive samples whose tumor percentage falls outside the 95th percentile of the negative population are defined as positive samples.

[0015] In some implementations, the sequencing depth is 0.1X, 0.2X, 0.3X, 0.4X, 0.5X, 0.6X, 0.7X, 0.8X, 0.9X, 1.0X, 2.0X, 4.0X, 5.0X, 6.0X, 7.0X, 8.0X, 9.0X, 10.0X, 20.0X, 30.0X, 500X, 1000X, 1500X, 2000X, or higher, or any depth between the aforementioned values.

[0016] In some implementations, obtaining sequencing data for positive samples refers to obtaining sequencing data for late-stage positive samples.

[0017] In some implementations, the model is a machine learning model. In other implementations, the machine learning model is a deep learning model.

[0018] In some implementations, the machine learning model is selected from: Transformer, xgboost, random forest, support vector machine, and logistic regression. Attached Figure Description

[0019] Figure 1: Model building flowchart.

[0020] Figure 2: ROC curves of machine learning detection models. Top: ROC curve for CRC validation set; Bottom: ROC curve for HCC validation set. Detailed Implementation

[0021] This invention provides a novel method for constructing predictive models. Synthetic samples are obtained by mixing sequencing data from positive samples (e.g., positive samples with a high tumor percentage) and negative samples to simulate positive samples with a low tumor percentage, thereby increasing the number of training samples and achieving data augmentation. By using positive and synthetic samples as the experimental group and negative samples as the control group for model training, this method can obtain predictive models with high sensitivity and accuracy even with a small number of samples (especially a small number of early-stage positive samples).

[0022] The method of this invention obtains synthetic sample data that can simulate early-stage / low-tumor-percentage cancer by mixing sequencing data of positive samples (e.g., positive samples with late-stage / high tumor percentage) and sequencing data of negative samples, thereby achieving data augmentation and improving the predictive sensitivity and accuracy of cancer prediction models for cancer (especially early-stage cancer), while also helping to improve the generalization of the model.

[0023] Throughout this specification, the terms "an embodiment," "one implementation," "some embodiments," "implementation," "specific embodiment," "related embodiment," "a particular embodiment," "another embodiment," or "further embodiment," or combinations thereof, refer to a specific feature, structure, or characteristic described in conjunction with that embodiment being included in at least one embodiment of the invention. Therefore, the appearance of these phrases in various places throughout this specification does not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0024] As used in this article, the terms "optional" or "optionally" include both selection and non-selection. For example, "optional modification" includes both being modified and not being modified.

[0025] As used herein, the terms “a” and “the” are generally interpreted to encompass both singular and plural forms.

[0026] As used herein, the term "comprising" means the phrase "including (but not limited to)" and may be used interchangeably with it. The term "including" as used herein means the phrase "including (but not limited to)" and may be used interchangeably with it. The technical solutions using "comprising" or "including" in this patent may be further defined as "comprising" or "forming".

[0027] As used herein, the term "late-stage positive sample" refers to a sample from advanced-stage cancer. Those skilled in the art can distinguish between early-stage and late-stage cancer based on medical knowledge. For example, in some embodiments, stage 0 and 1-3 cancers are considered early-stage cancers, and stage 4 cancers are considered late-stage cancers. In some embodiments, stage 0 and 1 cancers are considered early-stage cancers, and stages 2-4 cancers are considered late-stage cancers.

[0028] In some implementations, the positive samples described herein are from cancer samples, and the negative samples described herein are from healthy samples.

[0029] Synthetic sample data

[0030] The synthetic sample data described in this invention is obtained by mixing sequencing data from positive and negative samples. The sequencing data from positive and negative samples can be mixed according to a ratio or a fixed gradient representing the proportion of the target tumor.

[0031] In some embodiments, the synthetic sample data is obtained by sampling and mixing the sequencing data of one positive sample with the sequencing data of multiple negative samples. Preferably, in some embodiments, the sequencing data of one positive sample is sampled and mixed with 2, 3, 4, 5, 6, 7, 8, 9 or 10 negative samples.

[0032] In some implementations, the positive sample is paired with multiple negative samples. The term "paired" means that a positive sample is diluted using sequencing data from a fixed set of negative samples at different gradients or ratios.

[0033] The term "proportional mixing" as used in this article refers to sampling and mixing the sequencing data of positive samples and negative samples according to a specific preset ratio (gradient).

[0034] The "mixing according to a fixed target tumor proportion gradient" mentioned in this article refers to pre-setting the tumor proportion (target tumor proportion) gradient of the synthetic samples, and then calculating the mixing ratio of the sequencing data of a single positive sample with the sequencing data of a negative sample based on the tumor proportion of a single positive sample and the target tumor proportion.

[0035] In some implementations, the positive sample is mixed with multiple negative samples in a proportional manner.

[0036] In some implementations, the positive sample is mixed with multiple negative samples according to a fixed target tumor proportion gradient.

[0037] Methods for building models

[0038] This invention provides a method for constructing a cancer prediction model, comprising:

[0039] Obtain sequencing data for both positive and negative samples;

[0040] Synthetic sample data is obtained by sampling and mixing sequencing data from positive and negative samples.

[0041] The machine learning model was trained using positive and synthetic sample data as the experimental group and negative sample data as the control group.

[0042] In some implementations, the number of positive samples is greater than 20, greater than 30, greater than 50, greater than 200, greater than 300, greater than 400, greater than 500, greater than 600, greater than 700, greater than 800, greater than 900, greater than 1000 or more.

[0043] In some implementations, the number of negative samples is greater than 200, greater than 300, greater than 400, greater than 500, greater than 600, greater than 700, greater than 800, greater than 900, greater than 1000 or more.

[0044] Machine learning models

[0045] Those skilled in the art will understand that any machine learning model in the art can be used in the methods described herein.

[0046] In some implementations, the machine learning model is a deep learning model.

[0047] In some implementations, the machine learning model is a supervised learning or unsupervised learning machine learning model.

[0048] In some implementations, the machine learning model is selected from: Transformer, xgboost, random forest, support vector machine, and logistic regression.

[0049] Free DNA

[0050] In some embodiments, cell-free DNA is dispersed, floating DNA obtained from an organism, generally referring to DNA present in body fluids, free outside cells or vesicles. Cell-free DNA is well known in the art and generally refers to DNA that floats freely in body fluids.

[0051] In some embodiments, the cell-free DNA is derived from blood or plasma. In other embodiments, the cell-free DNA is derived from normal cells such as peripheral blood cells.

[0052] sample

[0053] The DNA (such as cell-free DNA) used for sequencing and analysis in this invention can be extracted from a sample. In some embodiments, the sample is a bodily fluid. In some embodiments, the sample is blood. In some embodiments, the sample is peripheral blood. In some embodiments, the sample is plasma, such as peripheral blood plasma.

[0054] In some embodiments, the volume of the sample used in the method of the present invention is 1.0, 1.5, 2, 2.5, 3, 3.5, 4, 4.5, 5, 5.5, 6, 6.5, 7, 7.5, 8, 8.5, 9, 9.5, 10, 20, 30, 50 ml or any value between the above. In some embodiments, the volume of the sample used in the method of the present invention is at least 0.5 ml. In some embodiments, the volume of the sample used in the method of the present invention is at most 10 ml.

[0055] All samples described in this invention can be obtained using standard techniques in the art. For example, blood samples can be obtained using standard techniques, such as using needles and syringes. In some embodiments, the blood sample is a peripheral blood sample. Alternatively, the blood sample can be a separated portion of peripheral blood, such as a plasma sample. In some embodiments, free DNA is extracted from the sample, for example, from nucleosomes in the sample. In some embodiments, after obtaining the blood sample, total DNA can be extracted from the sample using standard techniques known to those skilled in the art. In some embodiments, intact cells are removed prior to DNA extraction, thereby extracting only freely floating DNA. Intact cells can be removed by any method known in the art, such as by centrifugation or by gradient separation, such as by Ficol gradient separation, as a non-limiting example. A non-limiting example of DNA extraction is the FlexiGene DNA Kit (QIAGEN). Standard techniques for receiving cell-free DNA extraction are known to those skilled in the art, and a non-limiting example is the QIAamp Circulating Nucleic Acid Kit (QIAGEN).

[0056] Library construction and sequencing

[0057] Those skilled in the art will understand that library construction and sequencing can be performed using any suitable method in the art.

[0058] In some implementations, the positive and negative samples are obtained by sequencing cell-free DNA.

[0059] In some implementations, the sequencing method is a whole-genome sequencing method. In some implementations, the sequencing method is low-depth whole-genome sequencing.

[0060] In some implementations, the sequencing method is targeted methylation sequencing.

[0061] In some embodiments, the amount of cell-free DNA used to build the library is 0.01, 0.05, 0.1, 0.5, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, 200, 250, 300, 400, or 500 ng, or any value between these values. In some embodiments, the amount of cell-free DNA used to build the library is 10 to 100 ng.

[0062] In some embodiments, the sequencing depth is 1500X, 1000X, 500X, 300X, 200X, 150X, 100X, 50X, 40X, 30X, 20X, 10X, 9X, 8X, 7X, 6X, 5X, 4X, 3X, 2X, 1X, or any of the values ​​above. In some embodiments, the sequencing depth is 1000X to 30X. In some embodiments, the sequencing depth is 10X.

[0063] In some embodiments, sequencing is next-generation sequencing. Next-generation sequencing, also known as high-throughput sequencing or massively parallel sequencing, is any sequencing method that enables rapid, high-throughput sequencing of base pairs from DNA or RNA samples. In some embodiments, sequencing is high-throughput sequencing. In some embodiments, sequencing is massively parallel sequencing. Such sequencing is well known in the art and may include the use of Illumina sequencers, MGI sequencers, nanopore sequencers, and ion-based semiconductor sequencers as non-limiting examples. For non-limiting examples, sequencing machines such as the Illumina Nextseq 500 machine can be used, and the Illumina 500 / 550V2 kit can be used for processing. In some embodiments, sequencing is whole-genome sequencing. In some embodiments, only a portion of the genome is sequenced.

[0064] Application of the method of the present invention

[0065] In one aspect, the present invention provides the use of the method of the present invention or the model constructed therefrom in one or more of the following: early diagnosis or screening of cancer, evaluation of cancer treatment plans, detection of cancer progression, detection of residual cancer, or other applications.

[0066] The information generated by the method of this invention is an intermediate result for diagnosis and cannot be directly used as a basis for determining whether a person has cancer.

[0067] In some embodiments, the model constructed by the method described in this invention can be used for cancer risk assessment. In some embodiments, the method described in this invention can be used to determine cancer progression or cancer stage.

[0068] The beneficial effects of the present invention will be illustrated below through examples. Those skilled in the art will recognize that these examples are illustrative and not restrictive. These examples do not limit the scope of the invention in any way. The experimental methods described in the following examples are conventional methods unless otherwise stated; reagents and materials are commercially available unless otherwise stated.

[0069] The experimental techniques and methods used in this embodiment are conventional unless otherwise specified. For example, experimental methods in the following embodiments that do not specify specific conditions are generally performed under conventional conditions or according to the conditions recommended by the manufacturer. Unless otherwise specified, the materials and reagents used in the embodiments can be obtained through legitimate commercial channels.

[0070] Example

[0071] Based on the tumor percentage of cell-free DNA, a method of in silico data dilution was used to synthesize positive samples simulating early-stage / low-tumor percentage by diluting positive (cancer) samples with a high tumor percentage from negative (healthy) samples. This increased the amount of data, performed data augmentation, created a better training dataset for the model, and helped improve model performance.

[0072] 1. First, low-pass whole-genome sequencing (WGS) and targeted methylation sequencing (TBS) were performed on clinical samples (colorectal cancer (CRC), liver cancer (HCC), and negative samples). The tumor fraction (TF) of the samples was estimated from the low-pass whole-genome sequencing data.

[0073] 2. Because the tumor percentage estimation in samples with low tumor percentage is inaccurate, and negative samples also have a certain tumor percentage, the 95th percentile of the tumor percentage in negative samples is used as the value. Negative samples with a tumor percentage below this value and tumor samples with a tumor percentage above the threshold are selected for subsequent analysis.

[0074] 3. Sample dilution / data synthesis:

[0075] Each late-stage positive sample was diluted using pool reads consisting of a fixed pair of n (2-10) negative samples (within the 95th percentile of tumor prevalence) (all TBS data).

[0076] Dilution to a fixed tumor fraction (TF) gradient:

[0077] CRC: 0.05, 0.03, 0.02, 0.01, 0.007, 0.005, 0.003.

[0078] HCC: 0.05, 0.03, 0.02, 0.01, 0.007, 0.005, 0.003.

[0079] (All samples were diluted only to a TF gradient lower than the proportion of tumors in the sample itself)

[0080] Data processing:

[0081] For TBS sample data, sequencing data (reads) are randomly extracted from the fastq / bam file according to a calculated proportion to synthesize new sample data. The methylation characteristics of the samples are obtained through a TBS methylation analysis workflow.

[0082] Tumor diagnostic models were trained using methylation feature data from positive, synthetic, and negative samples.

[0083] 4. Model Training

[0084] Table 1: Performance comparison of CRC and HCC models trained with xgboost.

[0085] Results: Modeling by adding diluted samples (synthetic data) improves performance compared to modeling using late samples alone, achieving similar performance to modeling using both early and late samples (Table 2, Figure 2).

[0086] Table 2 Performance of CRC and HCC models

Claims

1. A method for constructing a predictive model, comprising: Determine the sequencing data of positive and negative samples; Synthetic sample data is obtained by sampling and mixing sequencing data from positive and negative samples. The prediction model was trained by using positive and synthetic sample data as the experimental group and negative sample data as the control group.

2. The method according to claim 1, wherein the synthetic sample data is obtained by sampling and mixing the sequencing data of a positive sample with the sequencing data of multiple negative samples.

3. The method of claim 2, wherein the positive sample is paired with a plurality of negative samples.

4. The method according to claim 2 or 3, wherein the sequencing data of the positive sample is mixed with the sequencing data of the plurality of negative samples in a proportional manner.

5. The method according to claim 2 or 3, wherein the sequencing data of a positive sample is mixed with the sequencing data of multiple negative samples according to a fixed target tumor proportion gradient.

6. The method according to any one of the preceding claims, wherein the sequencing comprises targeted methylation sequencing and / or whole genome sequencing (e.g., low-depth whole genome sequencing).

7. The method according to any one of the preceding claims, wherein the tumor percentage of each sample is obtained by analyzing whole-genome sequencing data, and negative samples whose tumor percentage falls within the 95th percentile of negative samples are defined as negative samples, and positive samples whose tumor percentage falls outside the 95th percentile of negative samples are defined as positive samples.

8. The method according to any one of the preceding claims, wherein the sequencing depth is 0.1X, 0.2X, 0.3X, 0.4X, 0.5X, 0.6X, 0.7X, 0.8X, 0.9X, 1.0X, 2.0X, 4.0X, 5.0X, 6.0X, 7.0X, 8.0X, 9.0X, 10.0X, 20.0X, 30.0X, 500X, 1000X, 1500X or higher, or any depth between the aforementioned values.

9. The method of any of the preceding claims, wherein obtaining sequencing data of a positive sample refers to obtaining sequencing data of a late positive sample.

10. A prediction model constructed by the method according to any one of the preceding claims.

Images