CN110189795B - Sub-space learning-based detection method for subgroup-specific driving genes - Google Patents

Abstract

The invention belongs to the technical field of cancer variant gene detection, and discloses a subspace learning-based subgroup specificity driving gene detection method, wherein variant data of variant genes in a cancer sample are used as input objects, low-dimensional vector output of each gene is obtained through a subspace learning algorithm, and the coordinate values of the low-dimensional vectors of each gene in different dimensions of subspaces can reflect the subgroup specificity of the genes; and analyzing the low-dimensional vector of the gene through outlier detection, and identifying the gene corresponding to the outlier vector in the subspace as a subgroup-specific driving gene. The invention can effectively infer the subgroup affiliation of the driving gene when the subgroup affiliation of the driving gene is unknown; aiming at the problem of multi-subgroup coexistence driving gene missed detection, the method strengthens the outlier significance of the driving genes in the characterization prediction and improves the detection performance of the subgroup specific driving genes by a subgroup specific outlier judgment method of subspace multimodal distribution.

Description

A method for subpopulation-specific driver gene detection based on subspace learning

technical field

The invention belongs to the technical field of computational detection of cancer mutation genes, and in particular relates to a subgroup-specific driver gene detection method based on subspace learning.

Background technique

Currently, the closest prior art: Existing methods mainly detect driver genes by gene variation frequency or frequency significance, but can only identify shared driver genes in cancer samples. In heterogeneous cancers, some subgroup samples account for a small proportion of all samples, resulting in low mutation frequency of the corresponding subgroup-specific driver genes, making it difficult to be effectively detected. Although domestic and foreign academic circles have achieved certain results in the detection of common driver genes, the detection of subgroup-specific driver genes is still in its infancy. In recent years, the problem of tumor heterogeneity has received extensive attention in the analysis of cancer driver genes. Due to the existence of different subgroup samples in heterogeneous cancers, the subgroup-specific driver genes only show a high mutation frequency in the corresponding subgroup samples, while the mutation frequency in the entire sample is relatively low, so it is difficult to be determined based on the mutation frequency or frequency significance. effective detection method. The existing subgroup-specific driver gene detection methods mainly detect the common driver genes in various samples through known subgroup categories or sample categories obtained by double clustering.

For heterogeneous cancers, the current research is mainly aimed at the ideal situation where the subgroup information is known, and the corresponding subgroup-specific driver genes are detected by respectively detecting the common driver genes in samples of each subgroup. OncodriveCLUST is a commonly used common driver gene detection method. This method clusters genes according to the gene variation positions of different samples, and selects genes with significant deviations in the clustering results as driver genes. Although this method has been shown to be effective in the ideal case where the class of the sample is known, it is no longer applicable when the class of the subpopulation of cancer samples is unknown.

In the absence of subgroup category information, the biclustering algorithm can cluster the cancer gene variation data and obtain the clustering of each sample. Based on the above clues, existing research has begun to develop a detection algorithm based on double clustering to predict the subgroups and subgroup-specific driver genes of heterogeneous cancer samples. However, the existing biclustering algorithm only performs isolated analysis on each cluster of cancer samples. Because the mutation frequency of multi-subgroup coexistence driver genes is unevenly distributed among different subgroups, the biclustering method is not effective in some subgroup samples. There were missed detections in the multi-subgroup coexistence driver genes.

In summary, the problems in the prior art are:

(1) The subgroup affiliation of the driver genes is unknown. Due to the unknown subgroup affiliation of heterogeneous cancer samples, it is difficult to detect subgroup-specific driver genes.

(2) Missing detection of multi-subgroup coexistence driver genes. For multi-subgroup coexistence driver genes that mutate in multiple subgroups at the same time, due to the uneven distribution of variation frequency among different subgroups, the existing methods even pass each subgroup. The group results are detected by intersection, which cannot avoid the missing detection of such genes in some subgroups.

The improvement to solve the above technical problems lies in: for the automatic analysis of in vitro cancer sample data, the subgroup affiliation of the driver gene is unknown, and the existing subspace learning method does not have subgroup-specific guidance, and it is impossible to analyze the driver gene. Subgroup-specific guidance positioning; at the same time, for the problem of missed detection of multi-subgroup coexistence driver genes, since the existing methods mainly make separate judgments for each subgroup, the variation frequency of multi-subgroup coexistence driver genes is not distributed among the subgroups. It loses the significance of outliers, which makes it difficult to be effectively detected by existing methods. The technical solution of the present application improves the validity of the existing analysis model, and when the method is applied to the automatic screening and analysis system of cancer genes, the validity of the detection results can be improved.

The significance of solving the above technical problems: Subgroup-specific driver genes can not only indicate the divergence of cancer samples at the level of pathogenesis, but also effectively reflect the differences in drug resistance of cancer lesions in clinical treatment. The above-mentioned technology can effectively detect subgroup-specific driver genes, not only has the scientific research value of revealing the pathogenesis of heterogeneous cancers, but also has the prospect of clinical application in providing diagnosis and treatment basis for cancer precision medicine.

Contents of the invention

Aiming at the problems existing in the prior art, the present invention provides a subgroup-specific driver gene detection method based on subspace learning.

The present invention is achieved in this way, a subgroup-specific driver gene detection method based on subspace learning, the subgroup-specific driver gene detection method based on subspace learning uses the variation data of mutated genes in cancer samples as input Object, the low-dimensional vector output of each gene is obtained through the subspace learning algorithm, and the coordinate values of the low-dimensional vector of each gene in different dimensions of the subspace can reflect the subgroup specificity of the gene; the low-dimensional vector of the gene is analyzed by outlier detection , identifying genes corresponding to outlier vectors in the subspace as subgroup-specific driver genes.

Further, the method for detecting subgroup-specific driver genes based on subspace learning specifically includes the following steps:

(1) Unification and data alignment of gene names and symbols of cancer samples, obtaining genetic variation data of heterogeneous cancers from public databases of cancer genomes such as TCGA and ICGC; aiming at the phenomenon of homonyms and synonyms of sample gene names and symbols obtained on different platforms , use Entrez Gene ID to unify the gene names and symbols of cancer samples, match the gene variations such as single nucleotide mutations and copy number variations of each sample to the corresponding genes, and obtain the mutated gene data of cancer samples; all cancer sample data Align by gene name to construct a cancer variation data matrix describing the variation distribution of each gene among samples;

(2) The subspace mapping construction of subgroup-specific guidance, using the subspace learning method to extract the distribution law of gene variation among samples from the data matrix, and describe the gene to be tested in the form of subspace vector points based on this law ; In the construction of the subspace mapping, the residual sum of squares of the original data and the inverse mapping reconstruction data is minimized by the stochastic gradient descent method, so that the inverse mapping can approximately reconstruct the original variation data;

Different dimensions of the subspace are used to guide the potential subordinate subgroups of the driver genes. When the subgroup dimension of the subspace is determined, the positional relationship between the subspace vector points and the coordinate axes/planes can represent the single subgroup of the driver genes, respectively. Group/multi-subgroup affiliation; according to the distribution density of subspace vector points in the space, the orthogonal transformation and rotation of each dimension of the subspace is carried out through the space transformation coordinate guidance method;

(3) The subgroup-specific positioning of the driver genes on the coordinate axis/coordinate plane, introducing non-negative constraints on the subspace coordinate values, so that the coordinate value of the subspace vector reflects the degree of subordination of the driver genes to each subgroup dimension; The subspace vector of a gene should belong to all subgroup dimensions at the same time, while multi-subgroup coexistence driver genes should belong to some subgroup dimensions, and single subgroup specific driver genes should only belong to a single subgroup dimension;

(4) Based on the determination of subgroup subordination based on the subspace vector coordinates of the driver gene, according to the subgroup-specific positioning of the driver gene in the result, whether the subspace vector points at different positions in the subspace are located on the coordinate axis, the coordinate plane or the first hexagram The subspace vector point located on the coordinate axis of the subspace can indicate the specific driver gene of a single subgroup, and the subspace vector point located on the coordinate plane can indicate the coexistence driver gene of multiple subgroups. The far subspace vector points represent the common driver genes of all samples;

(5) The spatial orientation multimodal distribution of subgroup-specific driver genes is represented by spatial coordinates. Since the irrelevant genes in the dimensionality reduction results are mainly concentrated near the origin of the subspace coordinates, the vector points of the driver gene subspace are along the coordinate axis/ The coordinate surface presents a multimodal distribution; through the subgroup affiliation relationship inferred from the vector coordinates of the gene subspace, the above dimensionality reduction results are represented by the spatial orientation multimodal distribution of the subgroup-specific driver genes to describe the distribution of the gene subspace vector points in each The degree of outlier in the dimension direction provides a criterion for subsequent detection of subgroup-specific driver genes based on outlier determination;

(6) Estimation of the semi-principal axis of the hyperellipsoid at the interface of multimodal distribution outliers, the missing detection of driving genes for the coexistence of multiple subgroups in the traditional strategy of the isolated interface of each subgroup, and the determination of the boundary of the hyperellipsoid interface for the coexistence of multiple subgroups For this purpose, the covariance matrix is calculated for all subspace vectors, and the eigenvalues and eigenvectors of the covariance matrix are obtained through eigenvalue decomposition; among them, the eigenvectors are used for orthogonal transformation and rotation of the subspace, And the subspace subgroup dimension is indicated by the dimension corresponding to the larger eigenvalue in the rotated subspace; the reciprocal of the corresponding eigenvalue square root is the estimated semi-principal length of the hyperellipsoid, which is used to construct the hyperellipsoid interface.

(7) Subgroup-specific driver gene detection for subspace outlier determination. In the subspace, through the hyperellipsoid interface obtained in step 6, outlier discrimination is performed on the multimodal distribution of gene subspace vector points : Among them, the gene corresponding to the subspace vector point in the inner part of the hyperellipsoid is judged as a cancer-independent gene; for the gene corresponding to the subspace vector point in the outer part of the hyperellipsoid, the coordinate axis/ Coordinate plane, predicted as single-subgroup-specific/multi-subgroup-coexisting driver genes, respectively.

Further, the subgroup-specific positioning of the driver gene in the third step on the coordinate axis/coordinate plane introduces a sparse strategy to the gene subspace vector, performs sparse regularization on the subspace vector, and proposes the subgroup specificity of the gene subspace vector. Sexual guidance subspace learning formula;

Among them, x _i is the variation data vector of the i-th gene in different samples, z _i is the subspace vector of the i-th gene, W is the parameter matrix of subspace mapping/inverse mapping, and λ is the subspace vector z _i The sparse regularization parameter for .

In the process of dimensionality reduction of genetic variation data, the sparse subspace learning strategy can ensure that the subspace vector coordinates of genes with subgroup-specific tendency are sparse, and ensure that the subspace positions of corresponding vector points are locked on the coordinate axis/coordinate plane; In the subspace vector coordinates of a driver gene, the dimension of the sparse non-zero value can indicate the subordinate subgroup of the driver gene. The subgroup specificity of the driver genes is inferred by retrieving the non-zero element dimension index of the sparse subspace vector, locating the coordinate axis/coordinate plane where the subspace vector points lie.

Further, the sixth step realizes the boundary judgment of the hyperellipsoid, and carries out the estimation of the semi-axis of the hyperellipsoid of the interface of multimodal distribution outliers through the subspace vector points; The length of the semi-principal axis is estimated; the hyperellipsoid interface can take into account the outlier significance of single-subgroup specific/multi-subgroup coexistence driver genes at the same time. The specific method is: calculate the covariance matrix for all subspace vectors, and pass the eigenvalue The eigenvalues and eigenvectors of the covariance matrix are decomposed to obtain the eigenvalues and eigenvectors of the covariance matrix; among them, the eigenvectors are used to perform orthogonal transformation and rotation on the subspace, and indicate the dimension of the subspace subgroup through the dimension corresponding to the larger eigenvalue in the rotated subspace; corresponding The reciprocal of the square root of the eigenvalue is the estimated semi-principal length of the hyperellipsoid, which is used to construct the interface of the hyperellipsoid.

Another object of the present invention is to provide a cancer mutation gene detection system using the method for detecting subgroup-specific driver genes based on subspace learning.

In summary, the advantages and positive effects of the present invention are: the present invention performs orthogonal transformation and rotation on each dimension of the subspace through the space transformation coordinate guidance method, so that the coordinate axis/coordinate plane can reflect the subgroup specificity of the driving gene in the representation space sex. Since the subspace vector points of genes are divergent and cannot be positioned on the corresponding coordinate axes/coordinate planes of subgroups, sparse representation is used to ensure that the coordinate values of driver gene characterization points are sparse. face for subpopulation-specific mapping.

In the present invention, by estimating the semi-axis of the hyperellipsoid of the interface of outliers, a superellipsoid interface is constructed for single-subgroup-specific/multi-subgroup coexistence driver genes to reflect the different distribution densities of multi-modal distribution in each dimension. Compared with the isolated outlier interface of each subgroup, the hyperellipsoidal outlier boundary can effectively strengthen the outlier significance of multi-subgroup coexisting genes, and realize the synchronization of single-subgroup-specific/multi-subgroup coexistence-driven genes predict.

By establishing a subspace learning algorithm guided by subgroup specificity, the invention can effectively infer the subgroup affiliation of the driver gene when the subgroup affiliation of the driver gene is unknown; aiming at the problem of missed detection of multi-subgroup coexistence driver genes , through the subgroup-specific outlier determination method of subspace multimodal distribution, the outlier significance of the driver genes in the representation prediction is strengthened, thereby improving the detection performance of the subgroup-specific driver genes.

Description of drawings

Fig. 1 is a flow chart of a subgroup-specific driver gene detection method based on subspace learning provided by an embodiment of the present invention.

Fig. 2 is a flow chart of the realization of the method for subgroup-specific driver gene detection based on subspace learning provided by the embodiment of the present invention.

Fig. 3 is a schematic diagram of a dimensionality reduction mapping based on subgroup-specific guidance subspace learning provided by an embodiment of the present invention.

Fig. 4 is a schematic diagram of a subgroup-specific outlier determination method (taking subgroup dimension = 3 as an example) provided by an embodiment of the present invention for subspace multimodal distribution;

In the figure: (A) The missing detection problem of the isolated interface of each subgroup; (B) Subgroup-specific driver gene prediction of the superelliptic interface.

Fig. 5 is a ROC curve diagram of the subgroup-specific driver gene detection performance results provided by the embodiment of the present invention.

Fig. 6 is a schematic diagram of subspace low-dimensional vectors (showing three subgroups) of subgroup-specific driver genes provided by an embodiment of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Aiming at the situation in the prior art that the subgroup category of the heterogeneous cancer sample is unknown, the present invention can solve the problem of inferring the subgroup affiliation of the driver gene. For multi-subgroup co-existing driving genes with uneven variation frequency distribution, the present invention can solve the problem of genetic testing of such genes in the prior art. Aiming at the unknown subgroup affiliation of driver genes, the present invention guides the subgroup specificity through the subspace dimension, thereby effectively inferring the subgroup affiliation of driver genes; The subgroup-specific outlier determination method based on spatial multimodal distribution realizes simultaneous detection of single-subgroup-specific/multi-subgroup coexistence driver genes.

The application principle of the present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the method for detecting subgroup-specific driver genes based on subspace learning provided by the embodiment of the present invention includes the following steps:

S101: Take the mutation data of the mutated gene in the cancer sample as the input object, and obtain the low-dimensional vector output of each gene through the subspace learning algorithm, and the coordinate values of the low-dimensional vector of each gene in different dimensions of the subspace can reflect the subgroup of the gene specificity;

S102: Analyzing the low-dimensional vectors of genes by outlier detection, and identifying the genes corresponding to the outlier vectors in the subspace as subgroup-specific driver genes.

The application principle of the present invention will be further described below in conjunction with the accompanying drawings.

The subgroup-specific driver gene detection method based on subspace learning provided by the embodiment of the present invention specifically includes the following steps:

(1) Gene name symbol unification and data alignment of cancer samples

In order to construct a subgroup-specific driver gene detection algorithm for heterogeneous cancers when the subgroup type is unknown, the gene variation data of heterogeneous cancers were first obtained from public databases of cancer genomes such as TCGA and ICGC. In view of the phenomenon of homonym and synonym of gene names and symbols obtained on different platforms, the gene names and symbols of cancer samples are unified through Entrez Gene ID, and genetic variations such as single nucleotide mutations and copy number variations of each sample are matched to the corresponding Genes to obtain mutation gene data of cancer samples. All cancer sample data were aligned by gene name, and a cancer variation data matrix describing the variation distribution of each gene among samples was constructed.

(2) Construction of subspace mapping for subpopulation-specific guidance

In cancer data with tumor heterogeneity, effectively inferring the subordinate subgroups of driver genes is the basic premise to realize subgroup-specific driver gene detection. For the problem of unknown subgroup affiliation, the invention adopts the subspace learning method to extract the distribution law of gene variation among samples from the data matrix, and describes the gene to be tested in the form of subspace vector points according to the law. The above subspace learning method is a data dimensionality reduction algorithm that can map high-dimensional data to low-dimensional space. It constructs a subspace dimensionality reduction map through unsupervised learning for high-dimensional input data, and calculates the corresponding The low-dimensional vector output of . In the construction of the subspace mapping, the residual sum of squares of the original data and the inverse mapping reconstruction data is minimized by the stochastic gradient descent method, so that the inverse mapping can approximately reconstruct the original variation data, thereby ensuring that the subspace dimensionality reduction results contain Distribution pattern information of cancer mutation data (Fig. 3).

In order to make the vector coordinates of the subspace guide the subgroup affiliation of the driver genes, different dimensional directions of the subspace are used to guide the potential subgroups of the driver genes. The specific method is: through the orthogonal transformation and rotation of the subspace, the coordinate axes/coordinate planes of the subspace are located near most vector points, so as to ensure that the directions of different dimensions of the subspace can effectively guide different subgroups. Therefore, when the subgroup dimension of the subspace is determined, the positional relationship between the subspace vector point and the coordinate axis/coordinate plane can represent the single-subgroup/multi-subgroup affiliation of the driver genes, respectively. In order for the subspace mapping to have subgroup-specific guidance, it is necessary to ensure that the driver genes with similar variation distribution patterns in the same sample are distributed on the same dimension coordinate axis/coordinate plane of the subspace. For this, according to the subspace vector points in the space Distribution density, through the space transformation coordinate guidance method, the orthogonal transformation and rotation of each dimension of the subspace is carried out, that is, the orthogonal transformation and rotation of the subspace is carried out so that the coordinate axis/coordinate plane of the subspace is located near the majority of vector points, thereby ensuring that the subspaces are different. Dimensional directions can effectively guide different subgroups to ensure that the subspace vector points are mainly concentrated on the coordinate axes/coordinate planes, thereby achieving subgroup-specific guidance for driver genes.

(3) Subgroup-specific positioning of driver genes on coordinate axes/planes

In order to ensure the subgroup-specific positioning of the above-mentioned subspace coordinate axes/coordinate planes, non-negative constraints are introduced to the subspace coordinate values, and the non-negative constraint problem is solved by the Lagrangian multiplier method, so that the subspace vector The size of the coordinate value reflects the degree of subordination of the driver gene to each subgroup dimension. Among them, the subspace vectors of common driver genes should belong to all subgroup dimensions at the same time, while multi-subgroup coexistence driver genes should belong to some subgroup dimensions, and single subgroup-specific driver genes should only belong to a single subgroup dimension. To avoid confusion with shared driver genes, the subspace vector of a subgroup-specific driver gene should take on a value of zero at some coordinates, while the remaining non-zero coordinates can locate the subgroup dimension to which the gene belongs. For this reason, the present invention introduces a sparse strategy to the gene subspace vector, performs sparse regularization on the subspace vector, and then proposes a subgroup specific guidance subspace learning formula of the gene subspace vector;

Among them, x _i is the variation data vector of the i-th gene in different samples, z _i is the subspace vector of the i-th gene, W is the parameter matrix of subspace mapping/inverse mapping, and λ is the subspace vector z _i The sparse regularization parameter for .

In the process of dimensionality reduction of genetic variation data, the sparse subspace learning strategy can ensure that the subspace vector coordinates of genes with subgroup-specific tendency are sparse, and ensure that the subspace positions of corresponding vector points are locked on the coordinate axis/coordinate plane. In the subspace vector coordinates of a driver gene, the dimension of the sparse non-zero value can indicate the subordinate subgroup of the driver gene. By retrieving the non-zero element dimension index of the sparse subspace vector, the coordinate axis/coordinate plane where the subspace vector point is located can be located, and the subgroup specificity of the driver gene can be inferred.

(4) Subgroup affiliation determination based on the vector coordinates of the driver gene subspace

According to the subgroup-specific localization of the driver genes in the above results, subgroup affiliation judgments were made for the subspace vector points at different positions in the subspace. Since each dimension of the subspace corresponds to different potential subgroups, the subspace vector points located on the subspace coordinate axis can indicate single-subgroup-specific driver genes, and the subspace vector points located on the coordinate plane can indicate multi-subgroup coexistence driver genes , and the subspace vector points located in the first hexagram limit and far away from the origin represent the common driver genes of all samples (Fig. 3). In contrast, the subspace vector points of cancer-unrelated genes are concentrated near the origin of the subspace.

In view of the above characteristics of the subspace, the distance between the subspace vector point and the origin can reflect the driver gene tendency of the corresponding gene to a certain extent, and the location of the subspace vector point in the space coordinate axis/coordinate plane can infer the subspace of the driver gene. group affiliation.

(5) Spatial orientation multimodal distribution representation of subpopulation-specific driver genes

Since the irrelevant genes in the above dimensionality reduction results are mainly concentrated near the origin of the subspace coordinates, while the vector points of the driver gene subspace present a multimodal distribution along the coordinate axis/coordinate plane, the subgroups inferred from the gene subspace vector coordinates Affiliation, the spatial orientation multimodal distribution of subgroup-specific driver genes is expressed for the above-mentioned dimensionality reduction results to describe the outlier degree of gene subspace vector points in each dimension direction, which is used for subsequent determination of subgroup specificity based on outlier points. Sex drive gene testing provides discriminative criteria.

(6) Estimation of the semi-principal axis of the hyperellipsoid on the interface of multimodal outliers

In view of the missing detection problem of multi-subgroup coexistence driver genes in the traditional strategy of isolated interface of each subgroup (Figure 4A), the multi-subgroup coexistence driver genes were discriminated through the hyperellipsoid interface. In contrast, the hyperellipsoid outlier interface can further strengthen the outlier significance of multi-subgroup coexistence driver genes when the subspace distribution is determined (Figure 4B), thereby improving the detection performance of such driver genes .

In order to realize the boundary judgment of the hyperellipsoid, the semi-axis estimation of the boundary of the hyperellipsoid with multimodal distribution outliers is carried out through the subspace vector points. In view of the different distribution densities of the subspace multimodal distribution in different dimensions, the semi-principal length of the hyperellipsoid in each dimension is estimated by using the eigenvalues of the covariance matrix. In view of the different distribution densities of subspace vector points in each subgroup dimension, the hyperellipsoid interface can take into account the outlier significance of single-subgroup-specific/multi-subgroup-coexisting driver genes, thereby improving the probability of subgroup-specific driver genes. Detection performance.

(7) Subgroup-specific driver gene detection for subspace outlier determination

In the subspace, since the subspace vector points farther from the origin of the coordinates have a stronger propensity to drive genes, through the above hyperellipsoid interface, the gene subspace vector points with multimodal distribution in the subspace are outlier point of discrimination. Among them, the subspace vector point gene in the inner part of the hyperellipsoid (coordinate origin side) is judged as a cancer-independent gene, while the subspace vector point gene in the outer part of the hyperellipsoid (the opposite side of the coordinate origin) is determined according to the subspace subspace. Axes/planes indicated by the population dimension, which are predicted as single-subgroup-specific/multi-subgroup-coexisting driver genes, respectively.

The application effects of the present invention will be described in detail below in conjunction with experiments.

Evaluation of detection results of group-specific driver genes in the present invention: Due to the scarcity of known cancer driver genes, the present invention does not use the labeling information of known driver genes in the subspace learning process of cancer genes. In order to evaluate the recognition performance of the above prediction algorithm, the known driver genes collected from databases such as COSMIC and IntOGen were compared with the detected subgroup-specific driver genes, and the true positive rate and false positive rate of the detection results were calculated respectively. Draw the ROC curve to evaluate the performance of the detection results. When applied to breast cancer and non-small cell lung cancer data, the ROC curve results show that the proposed method for subgroup-specific driver gene detection based on subspace learning has better detection performance than the existing OncodriveCLUST method and sparse singular value-based method Decomposed biclustering approach (Fig. 5). The positions of different genes in the subspace and coordinate axes/planes can also reflect the corresponding subgroups of driver genes (Figure 6).

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the protection of the present invention. within range.

Claims (4)

1. The method is characterized in that mutation data of mutation genes in cancer samples are used as input objects, coordinate values of low-dimensional vectors output by all genes are obtained through a subspace learning algorithm, the subgroup specificity of the genes is reflected by the coordinate values of different dimensions, and finally genes corresponding to outlier vectors in subspaces corresponding to outlier points are identified as subgroup-specific driving genes;

the subspace learning algorithm of the subspace learning-based subgroup specific driving gene detection method specifically comprises the following steps of:

(1) The gene name symbols of cancer samples are unified and aligned with data, and genetic variation data of heterogeneous cancers are obtained from TCGA and ICGC cancer genome public databases; aiming at the homonymous and heteronymous phenomena of the Gene name symbols of samples obtained by different platforms, carrying out Gene name symbol unification treatment on cancer samples through Entrez Gene ID, and matching single nucleotide mutation and copy number variation Gene variation of each sample to corresponding genes to obtain Gene variation data of the cancer samples; aligning all cancer sample data according to gene names, and constructing a cancer variation data matrix for describing variation distribution of each gene among samples;

(2) Subspace mapping construction of subgroup specificity guidance, extracting a distribution rule of gene variation among samples from a data matrix by adopting a subspace learning method, and describing genes to be detected in a subspace vector point mode according to the rule; in the construction of subspace mapping, the residual square sum of the original data and the inverse mapping reconstruction data is subjected to minimization solution by a random gradient descent method;

guiding potential subordinate sub-populations of the driving genes by adopting different dimension directions of subspaces, and carrying out orthogonal transformation rotation on each dimension of the subspaces by a space transformation coordinate guiding method according to the distribution density of subspace vector points in the space;

(3) The driving genes are specifically positioned on the subpopulations of the coordinate axes/planes, and non-negative constraint is introduced to the subspace coordinate values, so that the coordinate values of the subspace vectors reflect the degree of the dependence of the driving genes on the dimensions of each subpopulation;

(4) Based on the sub-group subordinate judgment of the space vector coordinates of the driving base factors, judging whether the sub-space vector points at different positions in the sub-space are positioned in coordinate axes, coordinate planes or first diagrams according to the sub-group specific positioning of the driving genes in the result;

(5) Carrying out space coordinate representation on space orientation multimodal distribution of the subgroup specific driving genes so as to describe the outlier degree of the gene subspace vector points in each dimension direction, and providing a discrimination standard for the subsequent subgroup specific driving gene detection based on outlier discrimination;

(6) Aiming at multi-subgroup coexistence driving gene omission in the traditional strategy of the isolated interfaces of all subgroups, the super-ellipsoid semi-principal axis estimation of the multimodal distribution outlier interfaces is adopted to judge the multi-subgroup coexistence driving genes by adopting the super-ellipsoid interface demarcation judgment;

(7) And (3) detecting the subgroup-specific driving genes judged by subspace outliers, and judging outliers of subspace vector point multimodal distribution of the genes in subspaces through the super-ellipsoidal interface obtained in step (6).

2. The subspace learning-based subgroup-specific driving gene detection method as claimed in claim 1, wherein the driving gene (3) introduces a sparse strategy to the gene subspace vector in the subgroup-specific positioning of the coordinate axes/coordinate planes, performs sparse regularization to the subspace vector, and proposes a subgroup-specific index subspace learning formula of the base factor space vector;

wherein x is _i Is the variation data vector of the ith gene in different samples, z _i Is the subspace vector of the ith gene, W is the parameter matrix of subspace mapping/inverse mapping, and lambda is the subspace vector z _i Is used for the sparse regularization parameters of (a).

3. The method for detecting a subgroup specific driving gene based on subspace learning according to claim 1, wherein (6) the determination of the boundary of the super-ellipsoid interface is realized, and the super-ellipsoid semi-principal axis estimation of the multimodal distribution outlier interface is carried out by adopting a method of using covariance matrix eigenvalue to the semi-principal axis length of the super-ellipsoid in each dimension through subspace vector points, and the specific method is as follows: calculating covariance matrixes for all subspace vectors, and decomposing the eigenvalues and eigenvectors of the covariance matrixes through eigenvalues; the feature vector is used for carrying out orthogonal transformation rotation on the subspace, and indicates the subspace subgroup dimension through the dimension corresponding to the larger feature value in the subspace after rotation; the inverse number of the corresponding eigenvalue after the evolution is the estimated length of the semi-principal axis of the super-ellipsoid, and is used for constructing the super-ellipsoid interface.

4. A cancer variant gene detection system applying the subspace learning-based subpopulation-specific driver gene detection method according to any one of claims 1 to 3.

Images