A picture generation method based on mutation data, a generation system and a cancer metastasis prediction method

Abstract

The application belongs to the technical field of picture processing, and particularly relates to a picture generation method based on mutation data, a generation system and a cancer metastasis prediction method. Step one: constructing a pathway image by using the functional similarity of pathways on a gene interaction network; step two: constructing a patient characteristic image; the prediction method further comprises: training a prediction model by using the constructed patient characteristic image, and predicting the metastasis of mutation data by using the trained prediction model. The application aims to solve the problems of lacking high-performance generation of patient characteristic pictures based on single nucleotide variation and the technical problems of cancer metastasis prediction and determination.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of picture processing, and particularly relates to a picture generation method based on mutation data, a generation system and a cancer metastasis prediction method. BACKGROUND

[0002] Clinicians and researchers mainly use imaging pictures, mRNA expression profiles, microRNA expression profiles and DNA methylation to predict and determine cancer metastasis. Compared with imaging pictures, mRNA expression profiles, microRNA expression profiles and DNA methylation, somatic mutation profiles contain the earliest warning information of cancer progression. However, so far, there is still a lack of high-performance cancer metastasis prediction and determination methods based on mutations, and the existing cancer metastasis prediction and determination methods based on mutations have low accuracy and cannot provide doctors with fast, accurate and reliable assistance. Therefore, the application provides a cancer metastasis prediction method based on a convolutional neural network. Clinicians can use the newly designed prediction method to determine the cancer metastasis condition based on pictures. SUMMARY

[0003] The application provides a picture generation method based on mutation data, a generation system and a cancer metastasis prediction method, to solve the problem of a lack of high-performance generation of patient feature pictures based on single nucleotide variations, and the technical problem of cancer metastasis prediction and determination.

[0004] The application is implemented through the following technical solutions:

[0005] A picture generation method based on mutation data, the generation method comprising the following steps,

[0006] Step one: constructing a pathway image by using the functional similarity of pathways on a gene interaction network;

[0007] Step two: generating a patient feature image based on the pathway image of step one.

[0008] Further, the step one specifically comprises the following steps,

[0009] Step one one: expanding biological pathways on HumanNet;

[0010] Step one two: calculating the similarity of biological pathways on HumanNet;

[0011] Step one three: constructing a pathway image by using the similarity matrix;

[0012] Step one four: compressing the pathway image;

[0013] Step one five: normalizing the pathway image.

[0014] Further, the step one is specifically to propagate the influence of the genes inside the pathway along the HumanNet network by using formula (1);

[0015]

[0016] wherein, A is the weighted adjacency matrix of HumanNet, the weight of the HumanNet network is the LLS of the gene pairs; D is a diagonal matrix, the elements on the diagonal are the sum of the rows or columns of the matrix A; a represents the restart probability, a∈[0.5, 1); Ro is the pathway-gene binary matrix, the row represents the pathway, and the column is the gene in the HumanNet network; P is the pathway set; P = {p1, …, p i ,…,p N}, N is the number of pathways; p i represents the i-th pathway; g j represents the j-th gene in the HumanNet network;

[0017] If there is also a gene g i in the pathway p j , then is 1; otherwise, is 0;

[0018] When t = 0, R t = R0; R t+1 is the pathway-gene matrix at time t+1;

[0019] Iterate formula (1) until ||R t+1 -R t || F ≤ 1×10 -6 ; the converged R t+1 is denoted as R;

[0020] When the iteration is finished, expand each pathway; extract the row vector r i corresponding to the pathway p i from R, arrange the element values corresponding to the genes outside the pathway in r i in descending order, and regard the pathway genes corresponding to the first δ elements as the genes expanded by the pathway.

[0021] Further, the step two is specifically to expand the pathway by using formula (2);

[0022]

[0023]

[0024]

[0025]

[0026] where Ps is the N x N pathway similarity matrix; N is the number of pathways; Ps i，j represents the similarity score of pathway p i and p j ; G(p i ) and G(p j ) represent the gene set of pathway p i and p j , respectively. m represents the number of genes in pathway p i ; n represents the number of genes in pathway p j ; p * represents p i or p j . represents the functional relationship score of gene g t and pathway gene set G(p * ); LLS min and LLS max are the minimum and maximum LLS values in HumanNet network, respectively; LLS'(g t ,g k ) is the normalized LLS(g t ,g k ); Fss(g t ,g k ) represents the functional similarity score of gene g t and g k .

[0027] Further, the step one three is specifically converting the pathway similarity matrix into a pathway distance matrix by using formula (6);

[0028]

[0029] inputting the pathway distance matrix into t-SNE to obtain the two-dimensional coordinates of each pathway; constructing a pathway image by using the pathway coordinates; assuming that the two-dimensional coordinates of pathway p i are (x i ,y i ); x i ∈[x down ,x up ], y i ∈[y down ,y up ]; x down and x up represent the integer lower and upper boundaries of the pathway coordinates on the X axis, respectively; y down and y up represent the integer lower and upper boundaries of the pathway coordinates on the Y axis, respectively. The paths are distributed in a rectangle Rect composed of points A(x down ,y down ), B(x up ,y down ), C(x up ,y up ) and D(x down ,y up ); the rectangle Rect is divided into a rectangle composed of small squares with side length β; the paths are distributed in Rect and also distributed on the small squares; according to the number of paths on the small squares, the small squares are divided into three categories, namely, a square without a path, a square with one path and a square with not less than two paths; a plurality of paths gathered in a square are merged into a new path; the gene set of the new path is the union set of the original path gene set; each small square in the rectangle Rect is taken as a pixel point, a small square containing a path is called a path pixel point, and a small square not containing a path is called a non-path pixel point, to obtain a path image.

[0030] Further, the step two specifically comprises the following steps,

[0031] Step two one: according to formula (10), the somatic mutation data and clinical data of the training set are used to calculate the weight of the gene under different phenotypes;

[0032]

[0033] Where, T j,j＝0,1,…,n-1 represents a tumor phenotype; n is the number of tumor phenotypes, n≥2; w is a gene-phenotype weight matrix; w i,j is the weight of gene g i under phenotype T j ; n(g i ) is the number of tumor phenotypes in which the gene g i mutates, is the total number of mutations of gene g j under phenotype T i ; is the maximum total number of mutations of the gene under phenotype T j ;

[0034] Step two two: the weight distribution of the patient's mutant gene on the path under each phenotype is calculated;

[0035] Specifically, it is assumed that G(pa k ) is the mutant gene set of patient pa k , i is the number of mutant genes; w ij is the weight of gene g i under phenotype T jThe weights are as follows; P' is the set of pathways on the pathway image, P' = {p1, ..., p...} l ,…,p N'}, N' is the number of pathways; G(p l ) represents pathway p l gene set, v represents pathway p l Number of genes in the middle; G(pa) k ,p l ) is G(pa k ) and G(p l The intersection of ) t is G(pa) k ,p l Number of genes in ); D j For phenotype T j The weight distribution vector of the mutated gene in patient pa in each pathway; D j =(d j1 ,…,d jl ,…,d jN' );d jl For phenotype T j The mutated gene of patient pa in the p pathway l The weight distribution values ​​on the above satisfy:

[0036]

[0037] in, Indicates if d jl If the set is empty, then the value is 0;

[0038] Steps 2 and 3: Based on the location of the pathway in the pathway image, transform the weight distribution vector under each phenotype into a weight distribution matrix;

[0039] Specifically, phenotype T j The weight distributions below are mapped onto the path image to obtain the weight distribution matrix D; assuming: md j For patient pa mutation data in phenotype T j The weight distribution matrix below; for md j The element in row u and column v satisfies:

[0040]

[0041] Where Im is the pathway image, Im u,v It is the pixel in row u and column v;

[0042] Step 24: Concatenate the weight distribution matrices under each phenotype to construct the patient feature image;

[0043] Specifically, assuming that Md is a feature image of a patient pa, Md=(md 1 ,md 2 ,…,md n ).

[0044] A cancer metastasis prediction method based on mutation data pictures, the prediction method uses the picture generation method described above, comprising training a prediction model using the constructed patient feature image, and using the trained prediction model to predict mutation data metastasis; the prediction model is a convolutional neural network.

[0045] A picture prediction system based on mutation data, the prediction system uses the cancer metastasis prediction method based on mutation data pictures described above, and the prediction system comprises,

[0046] A pathway image construction module, which constructs a pathway image using the functional similarity of pathways on a gene interaction network;

[0047] A patient feature image construction module, which constructs a patient feature image.

[0048] A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the computer program.

[0049] A computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method described above.

[0050] The beneficial effects of the present application are:

[0051] The key of the cancer metastasis prediction method based on the convolutional neural network of the present application is to convert single nucleotide mutation data into patient feature pictures. Unlike the feature pictures constructed by array data in the general field, the patient feature pictures are self-defined, specialized, based on biological pathways, and oriented to the cancer metastasis prediction task. The biological knowledge involved in constructing the patient feature pictures is that although the common mutations of any two patients are rare, the mutations of patients belonging to the same cancer phenotype may be concentrated on similar biological pathways; the effects of mutation genes on different tumor phenotypes are different, and mutation genes under different phenotypes should have different weights. Because the patient feature pictures integrate more biological knowledge, doctors and researchers can more professionally and accurately assess the cancer metastasis of patients. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a method flowchart of the present application.

[0053] Figure 2is a schematic diagram of the number of pathway pixels in a pathway image of different beta configurations of the present application.

[0054] Figure 3 is a schematic diagram of the average accuracy of M-NET at different beta values of the present application.

[0055] Figure 4 is a schematic diagram of the average accuracy of M-NET at different CNN architectures of the present application.

[0056] Figure 5 is the mean of the results of M-NET and other four methods at each evaluation index of the present application, where (A) is the input of RF, SVM, LR, P-NET and M-NET is SNV; (B) is the input of M-NET is SNVs, and the input of P-NET is SNVs and CNVs. DETAILED DESCRIPTION

[0057] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, technologies, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0058] It is to be understood that the terminology "including", "comprising", "consisting" and "consisting essentially of" used in the specification and the appended claims, indicates the presence of the stated features, integers, steps, operations, elements, and / or components but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0059] It is also to be understood that the terminology used in the present specification and the appended claims is for the purpose of describing the particular embodiments only and is not intended to be limiting.

[0060] The following description is presented to enable any person skilled in the art to practice the present application as claimed. Figures 1-5 The technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0062] Implementation Method 1

[0063] This embodiment provides an image generation method based on mutation data. The prediction method includes the following steps.

[0064] Step 1: Construct pathway maps by utilizing the functional similarity of pathways on gene interaction networks;

[0065] Furthermore, step one specifically includes the following steps:

[0066] Step 11: Extend biological pathways on HumanNet (biomolecular network);

[0067] Steps 1 and 2: Calculate the similarity of biological pathways on HumanNet;

[0068] Step 13: Construct a pathway image using a similarity matrix;

[0069] Step 14: Compress the pathway image;

[0070] Step 15: Standardize the pathway image.

[0071] Furthermore, each of the steps is specifically described as follows: using formula (1) to propagate the influence of genes within the pathway along the HumanNet network;

[0072]

[0073] in, A is the weighted adjacency matrix of HumanNet, where the weights of the HumanNet network are the LLS (log-likelihood scores) of gene pairs; D is a diagonal matrix, where the elements on the diagonal are the sum of the rows or columns of matrix A; α represents the restart probability, α∈[0.5,1); Ro is a pathway-gene binary matrix, where the rows represent pathways and the columns are genes in the HumanNet network; P is the pathway set; P={p1,…,p i ,…,p N}, where N is the number of pathways; p i Represents the i-th path; g j This represents the j-th gene in the HumanNet network;

[0074] If path p i Gene g also exists in it. j ,but =1; otherwise, =0;

[0075] When t=0, R t =R0;R t+1 This is the pathway-gene matrix at time t+1;

[0076] Iteratively execute formula (1) until ||R t+1 -R t || F ≤1×10 -6 ; the converged R t+1 Let it be R;

[0077] After the iteration is complete, expand each pathway; extract pathway p from R. i The corresponding row vector r i , will r i The element values ​​corresponding to the external genes of the pathway are arranged in descending order, and the pathway genes corresponding to the first δ elements are regarded as genes for pathway expansion.

[0078] Furthermore, steps one and two specifically involve:

[0079]

[0080]

[0081]

[0082]

[0083] Where Ps is an N×N pathway similarity matrix; N is the number of pathways; Ps i,j Indicates pathway p i and p j Similarity score; G(p) i ) and G(p j ) represent pathway p i and p j gene set, m represents pathway p i Number of genes in the middle; n represents pathway p j Number of genes in the middle; p * p i or p j ; Indicates gene g t and pathway gene set G(p * Functional relationship score; LLS min and LLS max These are the minimum and maximum LLS values ​​in the HumanNet network, respectively; LLS'(g t ,gk ) is the normalized LLS(g t , g k ); Fss(g t , g k ) represents the functional similarity score of genes g t and g k .

[0084] Further, the step one three is specifically converting the pathway similarity matrix into a pathway distance matrix by using formula (6);

[0085]

[0086] Inputting the pathway distance matrix into t-SNE to obtain the two-dimensional coordinates of each pathway; constructing a pathway image by using the pathway coordinates; assuming that the two-dimensional coordinates of pathway p i are (x i , y i ); x i ∈[x down , x up ], y i ∈[y down , y up ]; x down and x up represent the integer lower and upper boundaries of the pathway coordinates on the X axis respectively; y down and y up represent the integer lower and upper boundaries of the pathway coordinates on the Y axis respectively; The pathways are distributed in a rectangle Rect composed of points A(x down , y down ), B(x up , y down ), C(x up , y up ) and D(x down , y up ); the rectangle Rect is divided into a rectangle composed of small squares with a side length of β; the pathways are distributed in the Rect and also distributed on numerous small squares; according to the number of pathways on the small squares, the small squares are divided into three categories, i.e., a square without a pathway, a square with one pathway and a square with not less than two pathways; multiple pathways gathered on a square are merged into a new pathway; the gene set of the new pathway is the union set of the original pathway gene sets; each small square in the rectangle Rect is taken as a pixel point, the small square containing a pathway is called a pathway pixel point, and the small square not containing a pathway is called a non-pathway pixel point, so that the pathway image is obtained.

[0087] Step one four: compressing the pathway image.

[0088] First, the rows and columns of the channel image which do not contain channel pixel points are deleted, and then the row and column compression is iteratively performed for several times until the image cannot be compressed. In the iteration process, if the number of rows of the channel image is greater than the number of columns, the 'row compression' is performed; otherwise, the 'column compression' is performed. The row (column) compression mainly refers to moving the movable channel pixel points from the second row (column) to the last row (column) to the previous row (column) layer by layer. Let Im be an m x n channel image. Im i,j is a pixel point in the i-th row and j-th column of Im. The main operation of the compression from the i-th row to the i-1-th row is:

[0089] 1) find the set S of channel pixel points in the i-th row, 1≤j1<j2<…<j u ≤n. The channel pixel points in the set S can be divided into movable channel pixel points and non-movable channel pixel points. The non-movable channel pixel points can be further divided into directly non-movable channel pixel points and indirectly non-movable channel pixel points. The directly non-movable channel pixel point refers to the channel pixel point in the moving direction of the channel pixel point. If Im i,j , Im i-1,j are channel pixel points, then Im i,j is a directly non-movable channel pixel point. The channel pixel points which are not more than K away from the directly non-movable channel pixel point are indirectly non-movable channel pixel points. The channel pixel points which are not more than K away from the indirectly non-movable channel pixel point are also indirectly non-movable channel pixel points. The movable channel pixel points in S are found by using the knowledge of the queue. The specific process is as follows:

[0090] Let Label be a zero vector of u dimensions. Label is used to record the labels of the elements in S. The channel pixel point can move, and the label value is 0; otherwise, the label value is 1. The elements in Label correspond one by one to the elements in S. Q represents a queue. First, find the set S a of directly non-movable channel pixel points in S. If all the channel pixel points in S are movable channel pixel points. If h1<h2<…<h v ≤u. Set the label value corresponding to the elements in S a to 1, i.e. Label([h1, h2, …, h v ]) = 1. Put the elements in S a into the queue Q in turn. The head element is dequeued. Find the set S b of channel pixel points which are not more than K pixel points away from the head element. If 1≤r1<r2<…<r e ≤u. Set the label value corresponding to the elements in S b to 1, i.e. Label([r1, r2, …, r e]) = 1. Put S b The middle element is put into the queue Q. The above dequeue and enqueue operations are performed until the queue is empty. The passable pixel points with label value 0 are moved forward.

[0091] 2) When the passable pixel points are moved forward, if the row does not have passable pixel points, the row is deleted.

[0092] Step 1V: Normalize the passable image.

[0093] Suppose the size of the compressed passable image is M x L. The image is expanded to an H x H image. In the expanded image, the passable image before expansion is located at the center of the image, and the remaining part is filled with non-passable pixel points. H satisfies:

[0094] H = min(2 i , 3 x 2 i-1 ) (7)

[0095] 2 i ≥ max(M, L) (8)

[0096] 3 x 2 i-1 ≥ max(M, L) (9)

[0097] where i is a positive integer.

[0098] Step 2: Generate a patient feature image based on the passable image of step 1;

[0099] Further, the step 2 specifically comprises the following steps,

[0100] Step 21: According to formula (10), calculate the weight of the gene under different phenotypes using the somatic mutation data and clinical data of the training set;

[0101]

[0102] where T j,j＝0,1,…,n-1 represents a tumor phenotype; n is the number of tumor phenotypes, n ≥ 2; w is a gene-phenotype weight matrix; w i,j is the weight of gene g i under phenotype T j ; n(g i ) is the number of tumor phenotypes in which the mutation of gene g i exists, is the total number of mutations of gene g j under phenotype T i ; is the maximum total number of mutations of the gene under phenotype T j ;

[0103] Step 22: Calculate the weight distribution of the patient's mutated genes on the pathway under each phenotype.

[0104] Specifically, it is assumed that G(pa k ) is a set of mutant genes of patient pa k , i is the number of mutant genes; w ij is the weight of gene g i under phenotype T j ; P' is a set of pathways on the pathway image, P' = {p1, …, p l , …, p N'}, N' is the number of pathways; G(p l ) represents a set of genes of pathway p l , v is the number of genes in pathway p l ; G(pa k , p l ) is the intersection of G(pa k ) and G(p l ), t is the number of genes in G(pa k , p l ); D j is a weight distribution vector of the mutant genes of patient pa under phenotype T j on each pathway; D j = (d j1 , …, d jl , …, d jN' ); d jl is the weight distribution value of the mutant genes of patient pa under phenotype T j on pathway p l , and satisfies:

[0105]

[0106] wherein, d jl is equal to 0 if d jl is an empty set;

[0107] Step 2-3: converting the weight distribution vector under each phenotype into a weight distribution matrix according to the position of the pathway on the pathway image;

[0108] Specifically, the weight distribution under phenotype T j is mapped onto the pathway image respectively to obtain a weight distribution matrix D; it is assumed that md j is the weight distribution matrix of the mutant data of patient pa under phenotype T j ; is the element in the u-th row and v-th column of md j , and satisfies:

[0109]

[0110] wherein, Im is the pathway image, Im u,v is the pixel point of the u-th row and v-th column;

[0111] Step two four: splice the weight distribution matrix under each phenotype to form a patient feature image;

[0112] Specifically, assuming Md is the feature image of the patient pa, Md=(md 1 , md 2 , …, md n ).

[0113] Further, the step three is specifically selecting a convolutional neural network (InceptionV2) as the prediction model.

[0114] Specifically, data preparation

[0115] Download HumanNet interaction network from HumanNetV2 website. HumanNet network consists of 17929 genes and 525537 gene pairs. Each gene pair corresponds to a likelihood log score (LLS). Download 4632 biological pathways containing genes from pathway database PathwayCommons and KEGG. Download somatic mutations and clinical data of prostate cancer from biological website cBioPortal. Prostate cancer data comes from projects TCGA-PRAD, Broad / Cornell, MCTP and SU2C / PCF. The number of prostate cancer patients is 1009. Among them, the number of primary cancer patients is 534, and the number of metastatic cancer patients is 475. Detailed information of prostate cancer data is shown in Table 1.

[0116] Table 1 Detailed information of cancer data

[0117]

[0118] Comparison method

[0119] The cancer metastasis prediction method based on convolutional neural network is referred to as M-NET. To verify the performance of M-NET, five other methods are implemented, namely, Random Forest, Support Vector Machine, Logistic Regression, NBSS, and P-NET. Random Forest, Support Vector Machine, and Logistic Regression belong to traditional machine learning algorithms; NBSS and P-NET belong to deep learning algorithms. Random Forest, Support Vector Machine, and Logistic Regression are implemented using RandomForestRegressor, SGDClassifier, and LinearSVC in the sklearn package of Python. The codes of NBSS and P-NET can be downloaded from GitHub.

[0120] Evaluation index

[0121] The evaluation indexes of model performance are accuracy, precision, recall, F1-score, AUC area, and AUPR area.

[0122] Experimental process and analysis

[0123] Five 5-fold cross-validations are performed to train and test the model. In each 5-fold cross-validation, 1 part of data is used as the test set, and the remaining 4 parts are divided into the training set and the validation set. The ratio of the training set and the validation set is 4:1. The training set and the validation set are used to determine the hyperparameters of the model, such as the number of channels in the convolutional layer, the learning rate, the L2 penalty (also known as weight decay), and the momentum size. The test set is used to evaluate the performance of the model. When evaluating the performance of the model, the training set and the validation set are used to train the model. During the training process, the batch size is set to 50, and the epoch is set to 1000. If the model is not optimized after 200 iterations, the training is ended early. The input of the InceptionV2 architecture is a patient feature image of HxHx2. When β is set to 0.03, H is 96. The output of InceptionV2 is the probability distribution of the primary and metastatic phenotypes. The prediction result of the model is the phenotype with the highest probability. The weight initialization in the model uses xavier. After each convolution operation is completed, batch normalization is performed. The loss function of the model is categorical cross entropy. The optimizer of the model is MomentumOptimizer. The default momentum of the optimizer is 0.9, the initial value of the learning rate is 0.01, and it is decayed by 0.00001 every epoch. The parameters α in the pathway expansion process are 0.5 and 50, respectively.

[0124] The number of pass pixels in the pass image changes with the side length β (as shown in FIG. 3). The number of pass pixels in the pass image affects the performance of the model. Multiple pass images based on different side lengths β are constructed. Based on these pass images, patient feature images are constructed. M-NET is performed on the constructed patient feature images. The average accuracy of M-NET under different side lengths β is shown in FIG. 4. It can be seen from FIG. 4 that when the side length β is 0.03, the average accuracy of M-NET reaches the maximum, which is 0.852. Therefore, the side length β is set to 0.03. Figure 2 Figure 3 Figure 3

[0125] Embodiment Two

[0126] The embodiment provides a cancer metastasis prediction method based on a mutation data picture, the prediction method using the picture generation method in Embodiment One, comprising training a prediction model by using the constructed patient feature image, and performing mutation data metastasis prediction by using the trained prediction model; and the prediction model is a convolutional neural network.

[0127] Specifically, as shown in FIG. 5, when the CNN architecture changes, the performance of the model changes. The performance of M-NET under four CNN architectures, namely VGG19, InceptionV1, InceptionV2 and InceptionV3 [1, is compared. The average accuracy of M-NET under different CNN architectures is shown in FIG. 6. The performance of M-NET (InceptionV1), M-NET (InceptionV2) and M-NET (InceptionV3) is better than that of M-NET (VGG19). M-NET (InceptionV2) and M-NET (InceptionV3) are better than M-NET (InceptionV1). M-NET (InceptionV2) is not significantly worse than M-NET (InceptionV3). Based on these results, the InceptionV2 architecture is selected in the method. M-NET (InceptionV1), M-NET (InceptionV2) and M-NET (InceptionV3) all have better performance, indicating that the CNN architecture used in the present study can be replaced by other high-performance CNN architectures. Figure 2 3 Figure 4

[0128] Figure 4 ​​​​​​​As shown, the performance of M-NET and similar methods is compared. First, M-NET, RF, SVM, P-NET, and LR are performed on single nucleotide polymorphism (SNV) data. The average results of M-NET and the other four methods under different evaluation metrics are shown below. Figure 5 As shown in (A), M-NET significantly outperforms the other four methods, while P-NET does not significantly outperform RF, SVM, and LR (see [A]). Figure 2 The poor performance of classic machine learning algorithms may be due to the heterogeneity of mutation data, making it difficult to extract key information from them. Secondly, M-NET was performed on single nucleotide polymorphism (SNV) data, and NBSS and P-NET were performed on SNV and copy number variation (CNV) data. The average results of M-NET and the other two methods under different evaluation metrics are shown below. Figure 5 As shown in (B). From Figure 5 (B) It can be seen that M-NET significantly outperforms the other two multi-omics methods, while P-NET outperforms NBSS. The reason why M-NET has better performance is that M-NET introduces appropriate prior biological knowledge, solves the problem of mutational heterogeneity, and adopts a high-performance CNN architecture.

[0129] Implementation Method 3

[0130] This embodiment provides an image prediction system based on mutation data. The prediction system uses a cancer metastasis prediction method based on mutation data images as described in Embodiment Two. The prediction system includes...

[0131] The pathway image construction module utilizes the functional similarity of pathways on gene interaction networks to construct pathway images;

[0132] Furthermore, the specific working process of the pathway image construction module is as follows:

[0133] Extending biological pathways on HumanNet (a biomolecular network);

[0134] Computational biological pathway similarity on HumanNet;

[0135] Construct pathway images using similarity matrices;

[0136] Compressed pathway image;

[0137] Normalized pathway image.

[0138] Furthermore, the extension of biological pathways on HumanNet (biomolecular network) specifically involves using formula (1) to propagate the influence of genes within the pathway along the HumanNet network;

[0139]

[0140] in, A is the weighted adjacency matrix of HumanNet, where the weights of the HumanNet network are the LLS (log-likelihood scores) of gene pairs; D is a diagonal matrix, where the elements on the diagonal are the sum of the rows or columns of matrix A; α represents the restart probability, α∈[0.5,1); Ro is a pathway-gene binary matrix, where the rows represent pathways and the columns are genes in the HumanNet network; P is the pathway set; P={p1,…,p i ,…,p N}, where N is the number of pathways; p i Represents the i-th path; g j This represents the j-th gene in the HumanNet network;

[0141] If path p i Gene g also exists in it. j ,but =1; otherwise, =0;

[0142] When t=0, R t =R0;R t+1 This is the pathway-gene matrix at time t+1;

[0143] Iteratively execute formula (1) until ||R t+1 -R t || F ≤1×10 -6 ; the converged R t+1 Let it be R;

[0144] After the iteration is complete, expand each pathway; extract pathway p from R. i The corresponding row vector r i , will r i The element values ​​corresponding to the external genes of the pathway are arranged in descending order, and the pathway genes corresponding to the first δ elements are regarded as genes for pathway expansion.

[0145] Furthermore, the similarity of the computational biology pathways on HumanNet specifically refers to the following:

[0146]

[0147]

[0148]

[0149]

[0150] Where Ps is an N×N pathway similarity matrix; N is the number of pathways; Ps i,jIndicates pathway p i and p j Similarity score; G(p) i ) and G(p j ) represent pathway p i and p j gene set, m represents pathway p i Number of genes in the middle; n represents pathway p j Number of genes in the middle; p * p i or p j ; Indicates gene g t and pathway gene set G(p * Functional relationship score; LLS min and LLS max These are the minimum and maximum LLS values ​​in the HumanNet network, respectively; LLS'(g t ,g k ) is the normalized LLS(g t ,g k ); Fss(g t ,g k ) represents gene g t and g k Functional similarity score.

[0151] Furthermore, the construction of the path image using the similarity matrix specifically involves converting the path similarity matrix into a path distance matrix using formula (6).

[0152]

[0153] Input the path distance matrix into t-SNE to obtain the two-dimensional coordinates of each path; construct a path image using the path coordinates; assume path p i The two-dimensional coordinates are (x i ,y i );x i ∈[x down ,x up ],y i ∈[y down ,y up ];x down and x up These represent the lower and upper boundaries of the path coordinates on the X-axis, respectively; y down and y up These represent the lower and upper boundaries of the path coordinates on the Y-axis, respectively. The pathway is distributed from point A(x) down ,x down), B(x up , y down ), C(x up , y ip ) and D(x down , y up ) constitute a rectangle Rect; divide the rectangle Rect into small squares with side length β; the paths are distributed in Rect and also in the small squares; according to the number of paths in the small squares, divide the small squares into three categories, i.e. squares without paths, squares with one path and squares with not less than two paths; merge the multiple paths in one square into one new path; the gene set of the new path is the union set of the original path gene sets; take each small square in the rectangle Rect as a pixel point, the small square with paths is called a path pixel point and the small square without paths is called a non-path pixel point, and thus the path image is obtained.

[0154] Further, the compressed path image is specifically,

[0155] First, delete the rows and columns without path pixel points in the path image, and then perform multiple iterations of row and column compression until the image cannot be compressed. In the iteration process, if the number of rows of the path image is greater than the number of columns, perform 'row compression'; otherwise, perform 'column compression'. Row (column) compression mainly refers to moving the movable path pixel points from the second row (column) to the last row (column) to the previous row (column) layer by layer. Let Im be an m x n path image. Im i,j is the pixel point in the i-th row and j-th column of Im. The main operation of compression from the i-th row to the i-1-th row is:

[0156] 1) find the path pixel point set S, 1≤j1<j2<…<j u ≤n. The path pixel points in set S can be divided into movable path pixel points and non-movable path pixel points. The non-movable path pixel points can be further divided into directly non-movable path pixel points and indirectly non-movable path pixel points. The directly non-movable path pixel point refers to the path pixel point in the forward direction of the path pixel point. If Im i,j , Im i-1,j are path pixel points, then Im i,j is a directly non-movable path pixel point. The path pixel points within a distance of K from the directly non-movable path pixel point are indirectly non-movable path pixel points. The path pixel points within a distance of K from the indirectly non-movable path pixel point are also indirectly non-movable path pixel points. Find the movable path pixel points in S using queue knowledge. The specific process is as follows:

[0157] Let Label be a u-dimension zero vector. Label is used to record the label of elements in S. The label value of a passable pixel point that can move forward is 0; otherwise, the label value is 1. The elements in Label correspond to the elements in S one by one. Q represents a queue. First, find a set S a of passable pixel points in S that cannot move forward directly. All passable pixel points in S can move forward. If h1 v ≤u. Set the label value of the element in S a corresponding to [h1, h2, …, h v ] to 1, i.e., Label([h1, h2, …, h a ])=1. Put the elements in S b into the queue Q in turn. Dequeue the head element of the queue. Find a set S e of passable pixel points that are not more than K pixel points away from the head element of the queue. If r1 b ≤u. Set the label value of the element in S e corresponding to [r1, r2, …, r b ] to 1, i.e., Label([r1, r2, …, r i ])=1. Put the elements in S i-1 into the queue Q in turn. Perform the above dequeue and enqueue operations until the queue is empty. Move the passable pixel points with a label value of 0 forward.

[0158] 2) When the passable pixel points move forward, if there is no passable pixel point in the row, delete the row.

[0159] Further, the normalized passable image is specifically,

[0160] Suppose the size of the compressed passable image is M×L. Expand the image into an H×H image. In the expanded image, the passable image before expansion is located at the center of the picture, and the remaining part is filled with non-passable pixel points. H satisfies:

[0161] H=min(2 i ,3×2 i-1 ) (7)

[0162] 2 i ≥max(M,L) (8)

[0163] 3×2 i-1 ≥max(M,L) (9)

[0164] where i is a positive integer.

[0165] A patient feature image construction module constructs a patient feature image;

[0166] According to formula (10), the weight of the gene under different phenotypes is calculated by using the somatic mutation data and clinical data of the training set;

[0167]

[0168] wherein, T j,j＝0,1,…,n-1 represents a tumor phenotype; n is the number of tumor phenotypes, n≥2; w is a gene-phenotype weight matrix; w i,j is the weight of gene g i under phenotype T j ; n(g i ) is the number of tumor phenotypes in which gene g i mutates, is the total number of mutations of gene g j under phenotype T i ; and is the maximum total number of mutations of gene under phenotype T j .

[0169] The weight distribution of the mutation gene of the patient under each phenotype on the pathway is calculated.

[0170] Specifically, assuming that G(pa k ) is the mutation gene set of patient pa k , u is the number of mutation genes; w ij is the weight of gene g i under phenotype T j ; P' is a set of pathways on a pathway image, P'={p1,…,p l ,…,p N'}, N' is the number of pathways; G(p l ) represents the gene set of pathway p l , v is the number of genes in pathway p l ; G(pa k ,p l ) is the intersection of G(pa k ) and G(p l ), t is the number of genes in G(pa k ,p l ); D j is the weight distribution vector of the mutation gene of patient pa under phenotype T j on each pathway; D j =(d j1 ,…,d jl ,…,d jN' ); d jl is the weight distribution of the mutation gene of patient pa under phenotype T j on pathway pl The weight distribution value on the pathway image satisfies:

[0171]

[0172] The weight distribution vector under each phenotype is converted into a weight distribution matrix according to the position of the pathway on the pathway image.

[0173] Specifically, the weight distribution under the phenotype T j is mapped onto the pathway image to obtain a weight distribution matrix D. j Assume that md j is the weight distribution matrix of the mutation data of the patient pa under the phenotype T j . is the element in the u-th row and v-th column of md j , and satisfies:

[0174]

[0175] wherein, Im is the pathway image, Im u,v is the pixel point in the u-th row and v-th column.

[0176] The weight distribution matrices under each phenotype are spliced to form a patient feature image.

[0177] Specifically, assume that Md is the feature image of the patient pa, and Md=(md 1 , md 2 , …, md n ).

[0178] Embodiment Four

[0179] The embodiment provides a cancer metastasis prediction system based on mutation data pictures, uses the method as described in Embodiment Two, and comprises a prediction module, wherein the prediction module is used for training a prediction model by using the constructed patient feature image, and is used for predicting the metastasis of mutation data by using the trained prediction model; and the prediction model is a convolutional neural network (InceptionV2).

[0180] Embodiment Five

[0181] The embodiment of the present application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the memory is used for storing software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and the processor are connected through a bus. Specifically, the processor realizes any step in the above embodiment one by running the above computer program stored in the memory.

[0182] It should be appreciated that, in the embodiments of the present application, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0183] The memory can include read-only memory, flash memory, and random access memory, and provide instructions and data to the processor. Part or all of the memory can also include non-volatile random access memory.

[0184] It should be understood that the above integrated modules / units, if implemented in the form of software functional units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiments of the method of the present application can also be completed by means of a computer program instructing related hardware, and the above-mentioned computer program can be stored in a computer readable storage medium, and when the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The above-mentioned computer readable medium can include any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier wave signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the above-mentioned computer readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0185] The above description of disclosed implementations enables one of ordinary skill in the art to make and use the application. Various modifications to these implementations will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other implementations without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the implementations shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0186] It can be clearly understood by a person skilled in the art that, for the convenience and conciseness of description, only the division of the above functional units and modules is taken as an example, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the above device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiments can be integrated in one processing unit, or each unit can exist physically independently, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction, and are not used to limit the protection scope of the application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0187] It should be noted that the method and its details provided by the above embodiments can be combined into the device and equipment provided by the embodiments, and mutual reference will not be described here.

[0188] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or in a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0189] In the embodiments provided by the present application, it should be understood that the disclosed device / terminal equipment and method can be implemented in other ways. For example, the above-described device / device embodiments are only schematic, and the division of the above modules or units is only a logical function division, and there can be another division in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0190] The above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent ones; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A picture generation method based on mutation data, characterized by, The picture generation method comprises the following steps, Step one: constructing a pathway image using the functional similarity of pathways on the gene interaction network; Step two: generating a patient feature image based on the pathway image of step one; The step one specifically comprises the following steps, Step one one: extending the biological pathways on HumanNet; Step one two: calculating the similarity of the biological pathways on HumanNet; Step one three: constructing a pathway image using the similarity matrix; Step one four: compressing the pathway image; Step one five: normalizing the pathway image; The step two specifically comprises the following steps, Step two one: calculating the weight of genes under different phenotypes using the somatic mutation data and clinical data of the training set according to formula (10); wherein T j,j＝0,1,…,n-1 represents tumor phenotypes; n is the number of tumor phenotypes, n≥2; w is a gene-phenotype weight matrix; w i,j is the weight of gene g i under phenotype T j ; n(g i ) is the number of tumor phenotypes in which gene g i mutates, is the total number of mutations of gene g i under phenotype T j ; and is the maximum total number of mutations of gene under phenotype T j . Step two two: calculating the weight distribution of the patient mutation genes on the pathway under each phenotype; Specifically, it is assumed that G(pa k ) is a set of mutant genes of patient pa k , u is the number of mutant genes; w ij is the weight of gene g i under phenotype T j ; P' is a set of pathways on the pathway image, P' = {p1,…,p l ,…,p N'}, N' is the number of pathways; G(p l ) represents a set of genes of pathway p l , v is the number of genes in pathway p l ; G(pa k ,p l ) is the intersection of G(pa k ) and G(p l ), t is the number of genes in G(pa k ,p l ); D j is the weight distribution vector of the mutant genes of patient pa under phenotype T j on each pathway; D j = (d j1 ,…,d jl ,…,d jN' ); d jl is the weight distribution value of the mutant genes of patient pa under phenotype T j on pathway p l , satisfying: Step two three: converting the weight distribution vector under each phenotype into a weight distribution matrix according to the position of the pathway on the pathway image; Specifically, the weight distribution of the phenotype T j is mapped onto the pathway image respectively to obtain a weight distribution matrix D; it is assumed that md j is the weight distribution matrix of the mutation data of the patient pa at the phenotype T j . is the element of the u-th row and v-th column in md j , and it satisfies: wherein Im is a pass image, Im u,v is a pixel point in the u-th row and v-th column. Step two four: splicing the weight distribution matrix under each phenotype to form a patient feature image; Specifically, assuming Md is the feature image of patient pa, Md = (md 1 ,md 2 ,…,md n ).

2. The picture generating method of claim 1, wherein, The step one one specifically comprises propagating the influence of the genes inside the pathway along the HumanNet network using formula (1); wherein, A is the weighted adjacency matrix of HumanNet, the weight of HumanNet network is the LLS of gene pairs; D is a diagonal matrix, the elements on the diagonal are the sum of rows or columns of matrix A; a represents the restart probability, a e [0.5, 1); R0 is a pathway-gene binary matrix, the rows represent pathways, and the columns are genes in the HumanNet network; P is a pathway set; P = {p1, …, p i ,…,p N}, N is the number of pathways; p i represents the i-th pathway; g j represents the j-th gene in the HumanNet network; If pathway p i gene g j then 1; otherwise, 0; R = R0when t = 0 t = R0; R t+1 is the pathway-gene matrix at time t + 1; The formula (1) is iteratively executed until ||R t+1 -R t || F ≤1×10 -6 ; the converged R t+1 is denoted as R; When the iteration ends, extend each pathway; extract the pathway p from R i The corresponding row vector r i , r i The element values corresponding to the genes outside the pathway are arranged in descending order, and the first δ elements are considered as the genes of the pathway extension.

3. The picture generating method of claim 1, wherein, The step one two specifically comprises, where Ps is the N x N pathway similarity matrix; N is the number of pathways; Ps i,j represents the similarity score of pathway p i and p j ; G(p i ) and G(p j ) represent the gene set of pathway p i and p j , respectively. m represents the number of genes in pathway p i ; n represents the number of genes in pathway p j ; p * represents p i or p j . represents the functional relationship score of gene g t and pathway gene set G(p * ); LLS min and LLS max are the minimum and maximum LLS values in the HumanNet network, respectively; LLS'(g t , g k ) is the normalized LLS(g t , g k ); Fss(g t , g k ) represents the functional similarity score of gene g t and g k .

4. The picture generating method of claim 1, wherein, The step one three specifically comprises converting the pathway similarity matrix into a pathway distance matrix using formula (6); Input the path distance matrix into t-SNE to obtain the two-dimensional coordinates of each path; construct a path image using the path coordinates; assume path p i The two-dimensional coordinates are (x i ,y i );x i ∈[x down ,x up ],y i ∈[y down ,y up ];x down and x up These represent the lower and upper boundaries of the path coordinates on the X-axis, respectively; y down and y up These represent the lower and upper boundaries of the path coordinates on the Y-axis, respectively. The pathway is distributed from point A(x) down ,y down B(x) up ,y down ), C(x) up ,y up ) and D(x down ,y up Within a rectangle (Rect) formed by β, the rectangle is divided into rectangles composed of small squares with side length β. Pathways are distributed within the rectangle and also on the numerous small squares. Based on the number of pathways on each small square, they are divided into three categories: squares without pathways, squares with one pathway, and squares with at least two pathways. Multiple pathways clustered on a single square are merged into a new pathway. The gene set of the new pathway is the union of the gene sets of the original pathways. Each small square within the rectangle is treated as a pixel; small squares containing pathways are called pathway pixels, and small squares without pathways are called non-pathway pixels, thus obtaining the pathway image.

5. A method for predicting cancer metastasis based on a mutation data picture, characterized by, The prediction method uses the picture generation method of any one of claims 1-4, comprising training a prediction model using the constructed patient feature image, and performing prediction of mutation data transfer using the trained prediction model; the prediction model is a convolutional neural network.

6. A cancer metastasis prediction system based on mutational data pictures, characterized by, The prediction system uses the prediction method of claim 5, and the prediction system comprises, a pathway image construction module, which constructs a pathway image using the functional similarity of pathways on the gene interaction network; a patient feature image construction module, which constructs a patient feature image; a prediction module, which trains a prediction model using the constructed patient feature image, and performs prediction of mutation data transfer using the trained prediction model.

7. A computer device, characterized by The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-4.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-4.

Images