Cancer auxiliary diagnosis apparatus, training method, electronic device, and storage medium

Abstract

Embodiments of the present application disclose a cancer auxiliary diagnosis device, a training method, an electronic device and a storage medium. In a specific embodiment, the structure comprises: a data input layer configured to receive DNA methylation data, transcriptome data and copy number variation data; a first attention mechanism module configured to fuse features of the DNA methylation data and corresponding transcriptome data to obtain first fused features; a second attention mechanism module configured to fuse features of the DNA methylation data and corresponding copy number variation data to obtain second fused features; a feature splicing module configured to splice the first fused features and the second fused features to obtain spliced features; a fully connected layer configured to reduce the spliced features to an m-dimensional vector; and a classifier configured to map the m-dimensional vector to obtain probability values of m classifications of cancer. The embodiment can reduce the redundancy between features and effectively improve the accuracy of cancer diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology. More particularly, it relates to a cancer auxiliary diagnosis device, a training method, an electronic device and a storage medium. BACKGROUND

[0002] Cancer (malignant tumor) is a major disease that seriously threatens human life and health. The related prediction of cancer is attracting more and more social attention. However, the related prediction of cancer depends on manual work, which has certain subjective factors in judgment, is easy to cause misjudgment, and consumes high human and material resources. SUMMARY

[0003] The present application aims to provide a cancer auxiliary diagnosis device, a training method, an electronic device and a storage medium to solve at least one of the problems in the prior art.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0005] The present application provides a cancer auxiliary diagnosis device in the first aspect, comprising:

[0006] A data input layer for receiving DNA methylation data, transcriptome data and copy number variation data;

[0007] A first attention mechanism module for feature fusion of DNA methylation data and corresponding transcriptome data to obtain first fusion features;

[0008] A second attention mechanism module for feature fusion of DNA methylation data and corresponding copy number variation data to obtain second fusion features;

[0009] A feature fusion module for fusing the first fusion features and the second fusion features to obtain global features;

[0010] A fully connected layer for reducing the global features to m-dimensional vectors;

[0011] A classifier for mapping the m-dimensional vectors to obtain probability values of m classifications of the cancer,

[0012] Wherein, m is a natural number greater than or equal to 2.

[0013] In one embodiment, the classifier is a softmax classifier.

[0014] In one embodiment, the m classifications are cancer types and / or corresponding cancer staging categories or subtype categories.

[0015] In one embodiment,

[0016] The first attention mechanism module fuses the DNA methylation data and the transcriptome data to obtain first fused features, including:

[0017] The attention weight of the transcriptome data A2 on the methylation data A1 is calculated:

[0018]

[0019] wherein Q represents A1, K and V represent A2, d k is the dimension of K.

[0020] In one specific embodiment,

[0021] The first attention mechanism module fuses the DNA methylation data and the transcriptome data to obtain first fused features with the same dimension, including:

[0022] The DNA methylation data and the transcriptome data are represented by using an encoder to obtain representation vectors A1 and A2, respectively;

[0023] The weight w1 of the DNA methylation data relative to the transcriptome data and the weight w2 of the transcriptome data relative to the DNA methylation data are calculated, wherein

[0024]

[0025]

[0026] The weighted methylation data representation vector and the transcriptome data representation vector are calculated according to the weights:

[0027] A′1=w1A1

[0028] A′2=w2A2;

[0029] The first fused features A are obtained by averaging:

[0030]

[0031] In one specific embodiment,

[0032] The second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain second fused features, including:

[0033] The attention weight of the copy number variation data B2 on the methylation data B1 is calculated:

[0034]

[0035] wherein Q represents B1, K and V represent B2, dk is the dimension of K.

[0036] In one specific embodiment,

[0037] The second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain second fused features, including:

[0038] The DNA methylation data and the copy number variation data are represented using an encoder to obtain representation vectors B1 and B2 with the same dimension;

[0039] The weight w3 of the DNA methylation data relative to the copy number variation data and the weight w4 of the copy number variation data relative to the DNA methylation data are calculated, wherein

[0040]

[0041]

[0042] The weighted methylation data representation vector and the copy number variation data representation vector are calculated according to the weights:

[0043] B'1 = w3B1

[0044] B'2 = w4B2

[0045] The second fused features B are obtained by averaging:

[0046]

[0047] The second aspect of the present application provides a training method of the cancer auxiliary diagnosis device of the first aspect, including:

[0048] Obtaining a training data set and a validation set;

[0049] Designing a loss function;

[0050] Inputting the data set into the device for training until a threshold condition is met;

[0051] Using the validation set for verification.

[0052] In one specific embodiment, the inputting the data set into the device for training until the threshold condition is met includes:

[0053] The data input layer receives DNA methylation data, corresponding transcriptome data, and corresponding copy number variation data of a training sample;

[0054] The first attention mechanism module fuses the DNA methylation data and the transcriptome data to obtain first fused features;

[0055] The second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain second fused features;

[0056] The feature splicing module splices the first fused features and the second fused features to obtain spliced features;

[0057] The full connection layer reduces the spliced features to m-dimensional vectors;

[0058] The classifier maps the m-dimensional vectors to obtain probability values of m classifications of the patient with respect to cancer; and

[0059] Supervised training is performed using a cross-entropy loss function, and the training is stopped when a preset threshold is reached.

[0060] In one specific embodiment,

[0061] The first attention mechanism module fuses the DNA methylation data and the transcriptome data to obtain first fused features, including:

[0062] The attention weight of the transcriptome data A2 on the methylation data A1 is calculated:

[0063]

[0064] wherein Q represents A1, K and V represent A2, d k is the dimension of K.

[0065] In one specific embodiment,

[0066] The first attention mechanism module fuses the DNA methylation data and the transcriptome data to obtain first fused features, including:

[0067] The DNA methylation data and the transcriptome data of the sample are represented using an encoder to obtain representation vectors A1 and A2 with the same dimension;

[0068] The weight w1 of the DNA methylation data of the sample relative to the transcriptome data and the weight w2 of the transcriptome data relative to the DNA methylation data are calculated, wherein

[0069]

[0070]

[0071] The weighted methylation data representation vector and the transcriptome data representation vector are calculated according to the weights:

[0072] A'1=w1A1

[0073] A'2 = w2A2;

[0074] The first fusion feature A is obtained by fusing through averaging:

[0075]

[0076] In one specific embodiment,

[0077] The second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain a second fusion feature, comprising:

[0078] The attention weight of the copy number variation data B2 on the methylation data B1 is calculated:

[0079]

[0080] wherein Q represents B1, K and V represent B2, d k is the dimension of K.

[0081] In one specific embodiment,

[0082] The second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain a second fusion feature, comprising:

[0083] The DNA methylation data and the copy number variation data of the sample are represented by using an encoder to obtain representation vectors B1 and B2 with the same dimension;

[0084] The weight w3 of the DNA methylation data of the sample relative to the copy number variation data and the weight w4 of the copy number variation data relative to the DNA methylation data are calculated, wherein

[0085]

[0086]

[0087] The weighted methylation data representation vector and the copy number variation data representation vector are calculated according to the weights:

[0088] B'1 = w3B1

[0089] B'2 = w4B2;

[0090] The second fusion feature B is obtained by fusing through averaging:

[0091]

[0092] In one specific embodiment,

[0093] The data input layer receives DNA methylation data, corresponding transcriptome data and corresponding copy number variation data of the training sample, including:

[0094] Obtaining the DNA methylation data of the training sample from a database containing DNA methylation data;

[0095] Obtaining the transcriptome data under the same sample name as the DNA methylation data of the training sample from a database containing transcriptome data;

[0096] Based on the genomic annotation file containing the mapping relationship between DNA methylation data and copy number variation data, obtaining the copy number variation data corresponding to the DNA methylation data of the training sample from a database containing copy number variation data.

[0097] In one specific embodiment,

[0098] Based on the genomic annotation file containing the mapping relationship between DNA methylation data and copy number variation data, obtaining the copy number variation data corresponding to the DNA methylation data of the training sample from a database containing copy number variation data, including:

[0099] Finding the gene number corresponding to the methylation site from the genomic annotation file to obtain the associated gene;

[0100] Filtering the copy number variation data corresponding to the DNA methylation data sample in the database containing copy number variation data through the gene number.

[0101] In one specific embodiment, the training method further comprises:

[0102] Using cross-entropy loss function for supervised training.

[0103] The third aspect of the present application provides an electronic device comprising a processor and a memory storing a program, wherein the processor implements the training method according to the second aspect when executing the program.

[0104] The fourth aspect of the present application provides a storage medium storing a program, wherein the program is executed to implement the training method according to the second aspect.

[0105] The fifth aspect of the present application provides an electronic device comprising:

[0106] The cancer auxiliary diagnosis device according to the first aspect;

[0107] A user interaction interface for a user to input DNA methylation data, transcriptome data and copy number variation data of a patient, and display probability values of m classifications of cancer suffered by the patient.

[0108] The beneficial effects of the present application are as follows:

[0109] The technical solution described in the present application uses a first attention mechanism module to fuse DNA methylation data and corresponding transcriptome data to obtain first fused features; uses a second attention mechanism module to fuse DNA methylation data and corresponding copy number variation data to obtain second fused features; and fuses the first fused features and the second fused features after fusion, processes the global features obtained after fusion, can reduce the redundancy between features, extracts key information, and effectively improves the accuracy of cancer diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0110] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0111] Figure 1 An exemplary system architecture diagram showing that the method according to one embodiment of the present application can be applied in.

[0112] Figure 2 A schematic diagram of a cancer auxiliary diagnosis device provided by one embodiment of the present application is shown.

[0113] Figure 3 A flowchart of a cancer auxiliary diagnosis method provided by one embodiment of the present application is shown.

[0114] Figure 4 A user interaction interface schematic diagram according to one embodiment of the present application is shown.

[0115] Figure 5 A flowchart of a training method of a cancer auxiliary diagnosis device provided by one embodiment of the present application is shown.

[0116] Figure 6 A database screenshot diagram containing DNA methylation data is shown.

[0117] Figure 7 A database screenshot diagram containing transcriptome data is shown.

[0118] Figure 8 A database screenshot diagram containing copy number variation data is shown.

[0119] Figure 9 A schematic diagram of the gene corresponding to the methylation site under the cg00000029 data sequence is shown.

[0120] Figure 10 A screenshot schematic diagram of the gene corresponding to the chromosome start site is shown.

[0121] Figure 11A structural schematic diagram of a computer system implementing the device provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0122] In order to more clearly illustrate the present application, the present application will be further described below in conjunction with embodiments and drawings. Similar components are denoted by the same reference numerals in the drawings. It should be understood by those skilled in the art that the specific description below is illustrative rather than limiting, and should not limit the scope of protection of the present application.

[0123] Figure 1 An application environment diagram is shown for illustrating the application environment of the cancer auxiliary diagnosis method according to an embodiment of the present application. The cancer auxiliary diagnosis method provided by the present application can be applied in the application environment as shown. Figure 1 The terminal 10 communicates with the server 12 through a network. The terminal 10 can be a desktop device or a mobile terminal, such as a desktop computer, a tablet computer, a smart phone, etc. The server 12 can be a standalone physical server, a cluster of physical servers, or a virtual server.

[0124] The cancer auxiliary diagnosis method according to an embodiment of the present application can run on the server 12. The server 12 receives the DNA methylation data, the transcriptome data and the copy number variation data of a patient from the terminal 10; performs feature fusion on the DNA methylation data and the corresponding transcriptome data to obtain first fusion features; performs feature fusion on the DNA methylation data and the corresponding copy number variation data to obtain second fusion features; splices the first fusion features and the second fusion features to obtain spliced features; reduces the dimensions of the spliced features to m-dimensional vectors; and maps the m-dimensional vectors to obtain probability values of m classifications of the cancer. Further, the server 12 can return the probability values of the m classifications of the cancer to the terminal 10 after obtaining the probability values, so that the terminal 10 can display the probability values of the m classifications of the cancer for reference by medical personnel during diagnosis.

[0125] In order to realize classification prediction, an algorithm model needs to be constructed first.

[0126] Therefore, the server 12 can include a cancer auxiliary diagnosis device 120 according to an embodiment of the present application, as shown. Figure 2 The device 120 includes a data input layer 1200, a first attention mechanism module 1205, a second attention mechanism module 1210, a feature fusion module 1215, a fully connected layer 1220, and a classifier 1225 (equivalent to an application program installed in the server, which includes various functional modules).

[0127] Correspondingly, as shown in Figure 3As shown, the cancer auxiliary diagnosis method according to one embodiment of the present application comprises the following steps:

[0128] S310, receiving DNA methylation data, corresponding transcriptome data and corresponding copy number variation data of the patient.

[0129] In one specific example, the DNA methylation data, corresponding transcriptome data and corresponding copy number variation data of the patient are obtained through various medical examinations of the patient.

[0130] Among them, DNA methylation is a very important epigenetic modification method, which is one of the main mechanisms for regulating gene expression in cell differentiation and development process, participates in the formation of heterochromatin, the silencing of transposons, the regulation of gene expression and the occurrence of imprinting genes. DNA methylation as an epigenetic modification on the genome exists in various organisms. The DNA methylation process will add methyl groups to DNA molecules. Current research shows that the methylation difference of CpG island (the density of CpG is very high, commonly known as CpG island) between tumor samples and normal samples is mostly not in the internal of CpG island but in the CpG island shore. Therefore, DNA methylation data can be applied to infer cancer susceptibility genes, pathogenesis, and the influence of epigenetic modification. The current methods for detecting DNA methylation are: BS method (Bisulfite-conversion), restriction enzyme method, enrichment method.

[0131] The transcriptome data is RNA-seq transcriptome data. RNA transcription is the process of converting genetic information from DNA to RNA. As the first step of protein life synthesis, transcription is the synthesis step of mRNA and non-coding RNA, that is, the process of synthesizing RNA from one strand of double-stranded DNA (template strand for transcription, coding strand not for transcription) as a template, ATP, CTP, GTP, UTP four nucleoside triphosphates as raw materials, and RNA catalytic enzyme as catalyst. Transcriptome data can provide information on what genes are expressed under what conditions, and accordingly infer the function of the corresponding unknown genes and reveal the mechanism of action of specific regulatory genes. Through this molecular marker based on gene expression profile, not only the phenotype attribution of cells can be distinguished, but also the diagnosis of diseases can be used. By comparing the transcriptome differences between normal population and patients, specific expression differences with diagnostic significance related to cancer are screened out.

[0132] Copy number variation (CNV) of DNA fragments is a common form of genomic structural variation and is prevalent in the population. CNV of certain specific genes is considered to be a clinical indicator of tumor progression and prognosis, and has the potential to guide drug use for tumor patients. At present, the commonly used methods for detecting CNV mainly include two categories of experimental methods: low-throughput molecular biology experimental techniques, including chromosome banding, fluorescence in situ hybridization (FISH), droplet digital PCR (ddPCR), etc.; and high-throughput next generation sequencing (NGS) technology, which can detect DNA fragment CNV in the whole genome or target gene interval.

[0133] The medical staff uploads the obtained DNA methylation data, corresponding transcriptome data and corresponding copy number variation data to the server 104 through the human-computer interaction interface (such as the one shown in FIG. 10) on the terminal 102. Figure 4 For example, click the "Upload DNA Methylation Data" button to load the DNA methylation data from the local or remote, such as liver cancer data in CSV, RData, xlsx, etc.

[0134] In the example shown in FIG. 12, step S310 is performed by the data input layer 1200, that is, the data input layer receives the uploaded data. Figure 2

[0135] Step S320, the DNA methylation data and the transcriptome data are fused to obtain the first fused features.

[0136] In one possible implementation, step S120 includes:

[0137] Calculate the attention weight of the transcriptome data A2 on the methylation data A1:

[0138]

[0139] Where Q represents A1, K and V represent A2, and d k is the dimension of K.

[0140] Through the feature fusion of the attention mechanism model, the correlation between the DNA methylation data and the transcriptome data is obtained. However, the above method only calculates the weight influence of the transcriptome data on the DNA methylation data.

[0141] Therefore, another possible implementation is provided, and thus step S120 includes:

[0142] Use the encoder to perform feature representation on the DNA methylation data and the transcriptome data of the patient to obtain the same dimension representation vectors A1 and A2, respectively;​

[0143] computing a weight w1 of the DNA methylation data relative to the transcriptome data and a weight w2 of the transcriptome data relative to the DNA methylation data, wherein

[0144]

[0145]

[0146] computing a weighted methylation data feature vector and a weighted transcriptome data feature vector according to the weights:

[0147] A'1=w1A1

[0148] A'2=w2A2;

[0149] fusing by averaging to obtain a first fused feature A:

[0150]

[0151] In this embodiment, the attention weights of the methylation data and the transcriptome data are calculated from each other, which can further mine the correlation between them, and since both types of omics data are given weights, the interpretability of the classification result is improved.

[0152] In Figure 2 In the example shown in FIG. 12, step S320 is performed by the first attention mechanism module 1205.

[0153] The inventors have found that methylation is the binding site of the transcription factor in the promoter, and there are usually a large number of CpG islands in the promoter, which are in a highly methylated state. Methylation can change the original configuration of the promoter in the gene, thereby interfering with the binding of specific transcription factors to specific recognition sites on the promoter, and thereby affecting the transcriptional regulation of downstream genes and the occurrence and development of tumors. Analyzing the relationship between methylation and transcriptome data can effectively improve the prediction efficiency and accuracy of tumors.

[0154] Step S330, fusing the DNA methylation data and the copy number variation data to obtain a second fused feature.

[0155] In one possible implementation, step S330 includes:

[0156] computing an attention weight of the copy number variation data B2 on the methylation data B1:

[0157]

[0158] wherein Q represents B1, K and V represent B2, and d k is the dimension of K.

[0159] The correlation between DNA methylation data and copy number variation data was obtained through feature fusion using an attention mechanism model. However, the above method only calculates the weight of copy number variation data on DNA methylation data.

[0160] Therefore, another possible implementation is provided, and thus, the steps are as follows:

[0161] The encoder was used to perform feature representation on the patient's DNA methylation data and copy number variation data, resulting in representation vectors B1 and B2 of the same dimension;

[0162] Calculate the weights w3 and w4 of DNA methylation data relative to copy number variation data, where...

[0163]

[0164]

[0165] Calculate the weighted methylation data representation vector and copy number variation data representation vector based on the aforementioned weights:

[0166] B′1=w3B1

[0167] B′2=w4B2;

[0168] The second fusion feature B is obtained by averaging and fusion.

[0169]

[0170] In this implementation, attention weights are calculated between methylation data and copy number variation data to further explore their correlations. Furthermore, since both types of omics data are assigned weights, the interpretability of the classification results is improved.

[0171] exist Figure 2 In the example shown, step S330 is performed by the second attention mechanism module 1210.

[0172] The inventors discovered that gene methylation regulates gene expression, and copy number variations can lead to different degrees of gene expression differences, which play a role in the formation of normal phenotypes and the occurrence and development of diseases. Integrating and analyzing methylation and copy number variation data can effectively improve the accuracy of tumor detection.

[0173] It should be noted that the tumor feature information with greater attention weight has greater influence on the prediction result, and vice versa. In this embodiment, the transmission of the attention weight is used to introduce the weight w1 of the DNA methylation data relative to the transcriptome data, the weight w2 of the transcriptome data relative to the DNA methylation data, the weight w3 of the DNA methylation data relative to the copy number variation data, and the weight w4 of the copy number variation data relative to the DNA methylation data, so that more useful features for prediction can be extracted from the acquired tumor data, which provides data support for the accuracy of the prediction result.

[0174] S340, fusing the first fusion feature and the second fusion feature to obtain a global feature.

[0175] In one specific example, the feature fusion is performed in a splicing manner; the feature fusion can also be performed in a weighted summation manner, and the feature fusion can also be performed in a mapping manner.

[0176] The feature fusion in the splicing manner refers to splicing the feature data of each feature type to obtain the global object feature. The weighted summation manner refers to weighted summation of the feature data of each feature type to obtain the global object feature. In the weighted summation process, the weight of the feature data of each feature type can be set to the same value or different values. The feature fusion in the mapping manner can be feature fusion of the feature data by a trained neural network model to obtain the global object feature. Each sample of the trained neural network model can include feature data of at least two feature types and an expected global object feature.

[0177] In Figure 2 In the example shown, step S340 is performed by the feature fusion module 1215.

[0178] S350, dimensionally reducing the global feature into an m-dimensional vector.

[0179] The value of m is the number of categories obtained by the cancer prediction method according to the present application. For example, the prediction method of the present application can predict the stage, corresponding cancer or subtype of cancer (for example, the stage of cancer is divided into early stage, middle stage and late stage; the classification subtype is divided into type I, type II and type III), which depends on the training samples used for training the model as described later.

[0180] In Figure 2 In the example shown, step S350 is performed by the fully connected layer 1220.

[0181] S360, mapping the m-dimensional vector to obtain probability values of m categories of cancer of the patient.

[0182] The mapping process of the m-dimensional vector can be mapping by a mapping function such as a softmax classifier, or mapping by a trained neural network model. In the process of mapping by a mapping function, the mapping function used can be set according to the relevance of the object features and the disease attributes of the disease. In the training process of the trained neural network model, each sample used includes object features and expected classification results.

[0183] In Figure 2 In the example shown, step S360 is performed by the classifier 1225.

[0184] The present application uses an attention mechanism model to analyze the correlation between RNA expression and DNA methylation modification and the correlation between DNA methylation and copy number variation in samples obtained by two groups of omics integration, which overcomes the one-sidedness and inaccuracy of partial data in single omics data analysis. In addition, using the attention mechanism model, each part of the input is given different weights, and the tumor feature information with a larger attention weight has a stronger impact on the prediction result, and vice versa. In this embodiment, by using the transmission of attention weights, the weights w1 of DNA methylation data relative to transcriptome data, the weights w2 of transcriptome data relative to DNA methylation data, the weights w3 of DNA methylation data relative to copy number variation data, and the weights w4 of copy number variation data relative to DNA methylation data are introduced, which can extract more useful features for prediction from the obtained tumor data, which provides data support for the accuracy of the prediction result.

[0185] Those skilled in the art can understand that before using the above algorithm model for cancer auxiliary diagnosis, the model needs to be trained.

[0186] The design idea of the training method of the above-mentioned cancer auxiliary diagnosis device is as follows, as shown in Figure 5

[0187] S510, data set making: collect DNA methylation data, transcriptome data and copy number variation data of various cancer diseases in the database and corresponding cancer disease result data as positive samples, and DNA methylation data, transcriptome data and copy number variation data of other normal population and normal population results as negative samples, classify according to various early, medium and late (or type I, type II and type III) stages of cancer, and randomly divide into training set and validation set according to the ratio of 0.8:0.2.

[0188] ​A cancer-aided diagnostic device is constructed, consisting of a data input layer, a first attention mechanism module, a second attention mechanism module, a feature fusion module, a fully connected layer, and a classifier.

[0189] S520. Design the loss function: Use the cross-entropy loss (CE Loss) function, which is responsible for supervising the classification process.

[0190] S530, Model Training: Input the training set into the cancer auxiliary diagnostic device and use a loss function for supervised training until the threshold condition is met.

[0191] S540. Input the validation set into the trained model for validation.

[0192] The training step is similar to the steps described above, with the main difference being that it requires obtaining training and test samples from a known database.

[0193] Specifically, in a particular example, step S510 includes:

[0194] Step S5101: Obtain DNA methylation data of training samples from a database containing DNA methylation data.

[0195] like Figure 6 The image shows DNA methylation data obtained from a database containing DNA methylation data. The first row contains sample names, representing different case names; the first column contains methylation sites. Figure 6 The values ​​located at a sample name in a row and a methylation site in a column represent the percentage of methylation signal intensity, calculated using the following formula:

[0196] β = M ÷ (M + U + offset)

[0197] Where U represents the unmethylated signal intensity, M represents the methylated signal intensity, and offset represents the offset amount.

[0198] Step S5102: Obtain transcriptome data from a database containing transcriptome data under the same sample name as the DNA methylation data of the training sample.

[0199] like Figure 7 The image shows the obtained RNA-seq transcriptome data, where the first row is the sample name, representing different case names; the first column is the gene number. Figure 7 The values ​​in a row for sample name and column for gene number represent the gene expression level.

[0200] Step S5103: Based on the genome annotation file containing the mapping relationship between DNA methylation data and copy number variation data, obtain the copy number variation data corresponding to the DNA methylation data of the training sample from the database containing copy number variation data.

[0201] Step S5103 can be achieved by writing an automatically executed script.

[0202] like Figure 8 The image shows the obtained copy number variation data. The sampleID column is the sample name, representing different case names; the chrom column is the chromosome number; the chromstart column is the start position; the chromend column is the end position; and the value column is the copy number.

[0203] In one possible implementation, step S5103 includes:

[0204] The associated genes are obtained by finding the gene numbers corresponding to methylation sites in the genome annotation file.

[0205] The copy number variation data corresponding to the DNA methylation data sample is selected from the database containing copy number variation data using the gene number.

[0206] Continuing with the previous example, the gene IDs of the transcriptome data corresponding to methylation sites are found in the genome annotation file to obtain the associated genes. For example... Figure 9 As shown, this represents the gene corresponding to the methylation site in the cg00000029 data sequence. Based on the second mapping relationship, copy number variation data corresponding to DNA methylation data samples are screened by gene IDs of the transcriptome data corresponding to the methylation sites, and used as copy number variation data in the training samples. Figure 10 As shown, this represents the genes corresponding to the chromosome start sites included in the copy number variation data. Furthermore, from... Figure 8 The corresponding copy number mutation data was obtained from the database shown.

[0207] At this point, all DNA methylation data, transcriptome data, and copy number variation data contained in the training samples have been obtained, and the data input layer receives DNA methylation data, transcriptome data, and copy number variation data.

[0208] In a specific example, step S530 includes:

[0209] Step S5301: The first attention mechanism module fuses the sample DNA methylation data and transcriptome data to obtain the first fused feature.

[0210] In one possible implementation, step S5301 includes:

[0211] Compute the attention weight of the transcriptomic data A2 on the methylation data A1:

[0212]

[0213] wherein Q represents A1, K and V represent A2, d k is the dimension of K.

[0214] In another possible implementation, step S5301 comprises:

[0215] using an encoder to feature represent the DNA methylation data and the transcriptomic data of the sample, to obtain representation vectors A1 and A2 of the same dimension, respectively;

[0216] Compute the weight w1 of the DNA methylation data of the sample relative to the transcriptomic data and the weight w2 of the transcriptomic data relative to the DNA methylation data, wherein

[0217]

[0218]

[0219] According to the weights, compute the weighted methylation data representation vector and the transcriptomic data representation vector:

[0220] A'1 = w1A1

[0221] A'2 = w2A2;

[0222] Fuse by averaging to obtain the first fused feature A:

[0223]

[0224] Step S5302, the second attention mechanism module fuses the DNA methylation data and the copy number variation data to obtain a second fused feature.

[0225] In one possible implementation, step S5302 comprises:

[0226] Compute the attention weight of the copy number variation data B2 on the methylation data B1:

[0227]

[0228] wherein Q represents B1, K and V represent B2, d k is the dimension of K.

[0229] In another possible implementation, step S5302 comprises:

[0230] The DNA methylation data and the copy number variation data of the sample are represented by using an encoder to obtain feature vectors B1 and B2 with the same dimension, respectively;

[0231] The weight w3 of the DNA methylation data of the sample relative to the copy number variation data and the weight w4 of the copy number variation data relative to the DNA methylation data are calculated, wherein

[0232]

[0233]

[0234] The weighted methylation data feature vector and the transcriptome data feature vector are calculated according to the weights:

[0235] B'1 = w3B1

[0236] B'2 = w4B2;

[0237] The second fusion feature B is obtained by averaging and fusing:

[0238]

[0239] S5303, the feature fusion module fuses the first fusion feature and the second fusion feature to obtain a global feature.

[0240] S5304, the full connection layer reduces the splicing feature to an m-dimensional vector.

[0241] S5305, the classifier maps the m-dimensional vector to obtain probability values of m classifications of the patient with respect to cancer, wherein m is a natural number greater than or equal to 2.

[0242] S5306, supervised training is performed using a cross-entropy loss function, and the training is stopped when a preset threshold is reached.

[0243] In one specific example, the preset m classifications correspond to the stage category labels of the cancer, which are the labels of the classifications of the cancer, for example, the m classifications include lung cancer, liver cancer, gastric cancer, nasopharyngeal cancer, cervical cancer, etc., and the lung cancer stage category is divided into lung cancer early stage, lung cancer middle stage and lung cancer late stage, which are the category labels of lung cancer. The preset m classifications correspond to the subtype category labels of the cancer, which are the labels of the classifications of the cancer, for example, the m classifications include lung cancer, liver cancer, gastric cancer, nasopharyngeal cancer, cervical cancer, etc., and the lung cancer subtype category is divided into lung cancer type one, lung cancer type two and lung cancer type three, which are the category labels of lung cancer. In this embodiment, the category labels are taken as the corresponding category labels, so as to realize classification based on the labels.

[0244] As Figure 11As shown, a computer system suitable for use in performing the training method and / or the assisting diagnosis method of the cancer assisting diagnosis apparatus provided by the above-described embodiments includes a central processing module (CPU) that can perform various appropriate actions and processes according to programs stored in a read only memory (ROM) or programs loaded from a storage section into a random access memory (RAM). In the RAM, various programs and data required for the operation of the computer system are also stored. The CPU, the ROM, and the RAM are connected to each other by a bus. An input / output (I / O) interface is also connected to the bus.

[0245] Connected to the I / O interface are an input section including a keyboard, a mouse, etc.; an output section including a liquid crystal display (LCD), a speaker, etc.; a storage section including a hard disk, etc.; and a communication section including a network interface card such as a LAN card, a modem, etc. The communication section performs communication processing via a network such as the Internet. A drive is also connected to the I / O interface as necessary. A removable media such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive as necessary, so that a computer program read out therefrom is installed in the storage section as necessary.

[0246] In particular, according to the present embodiments, the processes described above in the flowcharts can be implemented as a computer software program. For example, the present embodiments include a computer program product comprising a computer program tangibly embodied on a computer readable medium, the computer program containing program code for executing the methods illustrated in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section, and / or installed from a removable media.

[0247] The flowcharts and diagrams in the drawings illustrate the architecture, functionality, and operations of possible implementations of systems, methods and computer program products according to the present embodiments. In this regard, each block in the flowcharts or diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or combinations of hardware and software. The modules described in the present embodiments can be implemented by software, or by hardware.

[0248] The computer system suitable for executing the tumor prediction method provided by the above-mentioned embodiments comprises, in addition to the structure of the computer system shown in the above-mentioned Figure 11 the user interaction interface shown in the above-mentioned Figure 4 the user interaction interface shown in the above-mentioned

[0249] As another aspect, the embodiments of the present application also provide a non-volatile computer storage medium, which can be the non-volatile computer storage medium contained in the above-mentioned device in the above-mentioned embodiments, or can exist independently and not be assembled into the terminal.

[0250] The non-volatile computer storage medium stores one or more programs, and when the one or more programs are executed by a device, the device implements the training method or the diagnosis method according to the above-mentioned embodiments of the present application.

[0251] It should be noted that, in the description of the present application, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0252] Obviously, the above-mentioned embodiments of the present application are only examples for clearly illustrating the present application, and are not limitations on the embodiments of the present application. For those skilled in the art, on the basis of the above-mentioned description, other different forms of changes or variations can also be made, and it is impossible to enumerate all the embodiments here. Any changes or variations falling within the scope of the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A cancer-aided diagnostic device, characterized in that, include: The data input layer is used to receive DNA methylation data, transcriptome data, and copy number variation data. The first attention mechanism module is used to fuse DNA methylation data and corresponding transcriptome data to obtain the first fusion feature. The second attention mechanism module is used to fuse DNA methylation data and corresponding copy number variation data to obtain a second fused feature. The feature fusion module is used to fuse the first fused feature and the second fused feature to obtain the global feature; A fully connected layer is used to reduce the dimensionality of the global features to an m-dimensional vector; A classifier is used to map the m-dimensional vector to obtain probability values ​​for m categories of the cancer. Where m is a natural number greater than or equal to 2; The first attention mechanism module fuses DNA methylation data and transcriptome data to obtain a first fused feature, including: The encoder was used to perform feature representation on DNA methylation data and transcriptome data, resulting in representation vectors A1 and A2 of the same dimension; Calculate the weights w1 of DNA methylation data relative to transcriptome data and w2 of transcriptome data relative to DNA methylation data, where ， ； Calculate the weighted methylation data representation vector and transcriptome data representation vector based on the aforementioned weights: ; The first fusion feature A is obtained by averaging and fusing. 。 2. The apparatus according to claim 1, characterized in that, The classifier is a softmax classifier.

3. The apparatus according to claim 1 or 2, characterized in that, The m categories are cancer types and / or corresponding cancer staging categories or subtype categories.

4. The apparatus according to claim 1, characterized in that, The second attention mechanism module fuses DNA methylation data and copy number variation data to obtain a second fused feature, including: Calculate the attention weight of copy number variation data B2 on methylation data B1: Where Q represents B1, K and V represent B2, and d k Let K be the dimension.

5. The apparatus according to claim 1, characterized in that, The second attention mechanism module fuses DNA methylation data and copy number variation data to obtain a second fused feature, including: The encoder was used to perform feature representation on DNA methylation data and copy number variation data, resulting in representation vectors B1 and B2 of the same dimension; Calculate the weights w3 and w4 of DNA methylation data relative to copy number variation data, where... ， ； Calculate the weighted methylation data representation vector and copy number variation data representation vector based on the aforementioned weights: ; The second fusion feature B is obtained by averaging and fusion. 。 6. A training method for a cancer auxiliary diagnostic device according to any one of claims 1-5, characterized in that, include: Obtain the training dataset and validation set; Design the loss function; The dataset is input into the device for training until a threshold condition is met; Use a validation set for validation; The step of inputting the dataset into the device for training until a threshold condition is met includes: The data input layer receives DNA methylation data, corresponding transcriptome data, and corresponding copy number variation data of the training samples; The first attention mechanism module fuses DNA methylation data and transcriptome data to obtain a first fused feature; The second attention mechanism module fuses DNA methylation data and copy number variation data to obtain a second fused feature; The feature fusion module fuses the first fused feature and the second fused feature to obtain the global feature; The fully connected layer reduces the dimensionality of the global features to an m-dimensional vector. The classifier maps the m-dimensional vector to obtain the probability values ​​of the patient's cancer in m different categories; and Supervised training is performed using the cross-entropy loss function, and training stops when a preset threshold is reached. The first attention mechanism module fuses DNA methylation data and transcriptome data to obtain a first fused feature, including: The encoder is used to perform feature representation on the DNA methylation data and transcriptome data of the sample, and the representation vectors A1 and A2 with the same dimension are obtained respectively; Calculate the weights w1 of the DNA methylation data relative to the transcriptome data and w2 of the transcriptome data relative to the DNA methylation data, where ， ； Calculate the weighted methylation data representation vector and transcriptome data representation vector based on the aforementioned weights: ; The first fusion feature A is obtained by averaging and fusing. 。 7. The training method according to claim 6, characterized in that, The second attention mechanism module fuses DNA methylation data and copy number variation data to obtain a second fused feature, including: Calculate the attention weight of copy number variation data B2 on methylation data B1: Where Q represents B1, K and V represent B2, and d k Let K be the dimension.

8. The training method according to claim 6, characterized in that, The second attention mechanism module fuses DNA methylation data and copy number variation data to obtain a second fused feature, including: The encoder is used to perform feature representation on the DNA methylation data and copy number variation data of the sample, and the representation vectors B1 and B2 with the same dimension are obtained respectively. Calculate the weights w3 and w4 of the DNA methylation data relative to the copy number variation data for each sample, where w3 is the weight of the copy number variation data relative to the DNA methylation data. ， ； Calculate the weighted methylation data representation vector and copy number variation data representation vector based on the aforementioned weights: ; The second fusion feature B is obtained by averaging and fusion. 。 9. The training method according to claim 6, characterized in that, The data input layer receives DNA methylation data, corresponding transcriptome data, and corresponding copy number variation data from the training samples, including: Obtain DNA methylation data of training samples from a database containing DNA methylation data; Obtain transcriptome data from a database containing transcriptome data, under the same sample name as the DNA methylation data of the training sample; Based on a genome annotation file containing a mapping relationship between DNA methylation data and copy number variation data, copy number variation data corresponding to the DNA methylation data of the training sample is obtained from a database containing copy number variation data.

10. The training method according to claim 9, characterized in that, The genome annotation file, based on the mapping relationship between DNA methylation data and copy number variation data, retrieves copy number variation data corresponding to the DNA methylation data of the training samples from a database containing copy number variation data, including: The associated genes are obtained by finding the gene numbers corresponding to methylation sites in the genome annotation file. The copy number variation data corresponding to the DNA methylation data sample is selected from the database containing copy number variation data using the gene number.

11. The training method according to claim 6, characterized in that, Also includes: Supervised training is performed using the cross-entropy loss function.

12. An electronic device comprising a processor and a memory storing a program, characterized in that, When the processor executes the program, it implements the training method according to any one of claims 6-11.

13. A storage medium storing a program, characterized in that, When the program is executed, it implements the training method according to any one of claims 6-11.

14. An electronic device, characterized in that, include: Cancer auxiliary diagnostic device according to any one of claims 1-5; The user interface allows users to input the patient's DNA methylation data, transcriptome data, and copy number variation data, and displays the probability values ​​of the m categories of the patient's cancer.

Images