Cancer prognosis prediction method and system based on multi-omics fusion

Abstract

The invention discloses a cancer prognosis prediction method and system based on multi-omics fusion. The method comprises the following steps: acquiring multiple groups of omics data; preprocessing the plurality of groups of omics data to obtain a plurality of groups of first omics data features; inputting the plurality of groups of first omics data features into a multi-head attention-based graph convolutional network feature extraction model to obtain a plurality of groups of second omics features; inputting the plurality of groups of second group of characteristics into an attention mechanism fusion module based on biological priori knowledge guidance to obtain fusion characteristics; and inputting the fusion features into a prognostic scoring model to obtain prognostic scores. In addition, potential gene targets related to cancer prognosis are obtained by adopting an omics data feature importance evaluation method. The prognosis prediction conclusion obtained by the method is high in accuracy and high in interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain cancer prognosis prediction, and more particularly to a cancer prognosis prediction method and system based on multi-omics fusion. BACKGROUND

[0002] In recent years, the development of bioinformatics has brought new ideas to the prognosis prediction of cancer. People can process multiple characteristics related to cancer using bioinformatics technology, and the results can provide a basis for doctors' diagnosis.

[0003] Traditional cancer survival analysis methods, such as Cox proportional hazards model, usually rely on manual feature engineering and feature pre-screening, and are difficult to actively learn the complex nonlinear relationship between different omics data, resulting in limited accuracy of prognosis prediction. At the same time, although many deep learning methods have been applied to cancer survival analysis in recent years, the analysis conclusion cannot be based on evidence from a biological perspective, and the results lack interpretability.

[0004] The prior art discloses a colorectal cancer prognosis method based on multi-omics and clinical test data, comprising S1, collecting omics data from different colorectal cancer patients and the survival status of the patients two years after surgical resection of colorectal cancer from different data sources, databases or experiments, and collecting clinical data related to the patients, and preprocessing the omics data and clinical data; S2, constructing a patient omics similarity network through the preprocessed omics data and clinical data and the survival status; S3, encoding the node topology structure information of the patient and the clinical data information of the patient; S4, adding the node topology structure information of the patient and the clinical data information of the patient to the encoding of the graph attention network; S5, predicting the survival status of the patient two years after surgical resection of colorectal cancer through a prediction model; S6, optimizing the prediction model using binary cross-entropy loss function. The omics data collected by the method is relatively single, so the analysis result is easily affected by the single omics data, and the accuracy is low. At the same time, the method cannot establish a connection between the prognosis result and a specific omics data, so the result lacks interpretability. SUMMARY

[0005] The present application provides a cancer prognosis prediction method and system based on multi-omics fusion to overcome the defects of low accuracy of prognosis prediction and lack of interpretability of prognosis prediction conclusion in the prior art. The prognosis prediction conclusion obtained by the method has high accuracy and strong interpretability.

[0006] A cancer prognosis prediction method based on multi-omics fusion comprises: S1: obtaining multiple sets of omics data; S2: preprocessing the multiple sets of omics data respectively to obtain multiple sets of first omics data features; S3: inputting the first set of omics data features into a graph convolution network feature extraction model respectively to obtain a plurality of sets of second omics features; S4: inputting the plurality of sets of second omics features into a prognosis scoring model based on a biological prior knowledge guided attention mechanism to obtain a prognosis score.

[0007] Further, the omics data is preprocessed, including: S201: denoising the omics data to obtain denoised omics data; S202: using F-value and principal component analysis algorithm to extract features from the denoised omics data to obtain feature-extracted omics data; S203: performing data scaling on the feature-extracted omics data to obtain first omics data features.

[0008] Further, the omics data includes mRNA data, miRNA data, and methylation data.

[0009] Further, the graph convolution network feature extraction model includes a first graph convolution network module, a second graph convolution network module, and a third graph convolution network module. The output end of the first graph convolution network module is connected to the input end of the second graph convolution network module, and the output end of the second graph convolution network module is connected to the input end of the third graph convolution network module.

[0010] Further, the prognosis scoring model based on the biological prior knowledge guided attention mechanism includes a first single-modal multi-head attention module, a second single-modal multi-head attention module, a third single-modal multi-head attention module, a cross-modal cross-attention fusioner, and a multi-omics effect enhancement module. The second omics features corresponding to the mRNA data are input into the input end of the first single-modal multi-head attention module, the second omics features corresponding to the miRNA data are input into the input end of the second single-modal multi-head attention module, and the second omics features corresponding to the methylation data are input into the input end of the third single-modal multi-head attention module; the output end of the first single-modal multi-head attention module, the output end of the second single-modal multi-head attention module, and the output end of the third single-modal multi-head attention module are connected to the input end of the cross-modal cross-attention fusioner, the output end of the cross-modal cross-attention fusioner, the output end of the first single-modal multi-head attention module, and the output end of the third single-modal multi-head attention module are connected to the input end of the multi-omics effect enhancement module, and the output end of the multi-omics effect enhancement module outputs the prognosis score.

[0011] Further, the multi-omics effect enhancement module comprises a first biasing module, a second biasing module, a third biasing module, a first mapping module, a second mapping module, a third mapping module, a multi-modal fusion module, and a linear weight biasing module. The output end of the first single-modal multi-head attention module is connected with the input end of the first biasing module and the input end of the first mapping module, the output end of the cross-modal cross attention fusioner is connected with the input end of the second biasing module and the input end of the second mapping module, the output end of the third single-modal multi-head attention module is connected with the input end of the third biasing module and the input end of the third mapping module, the output end of the first biasing module, the output end of the second biasing module, the output end of the third biasing module, the output end of the first mapping module, the output end of the second mapping module, and the output end of the third mapping module are connected with the input end of the multi-modal fusion module, the output end of the multi-modal fusion module is connected with the input end of the linear weight biasing module, and the output end of the linear weight biasing module outputs a prognosis score.

[0012] An omics data feature importance evaluation method comprises: S01: Obtain multi-set omics data; the multi-set omics data contains at least one feature to be evaluated; S02: Process the multi-set omics data using the cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a first score; S03: Set the value of the feature to be evaluated in the multi-set omics data to zero to obtain zeroed multi-set omics data; S04: Process the zeroed multi-set omics data using the cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a second score; S05: Calculate the importance score of the feature to be evaluated according to the first score and the second score.

[0013] Further, in step S05, the calculation formula of the importance score of the feature to be evaluated according to the first score and the second score is as follows:

[0014] represents the feature dimension, represents the first score, represents the second score, represents the importance score, represents the first score, Take the absolute value.

[0015] A cancer prognosis prediction system based on multi-omics fusion comprises: A first omics data acquisition module: acquire multi-set omics data; The preprocessing module: a plurality of sets of the omics data are respectively preprocessed to obtain a plurality of sets of first omics data characteristics; The feature extraction module: a plurality of sets of the first omics data characteristics are respectively input into a graph convolution network feature extraction model to obtain a plurality of sets of second omics characteristics; The scoring module: a plurality of sets of the second omics characteristics are input into a prognosis scoring model based on a biological prior knowledge guided attention mechanism to obtain a prognosis score.

[0016] The omics data characteristic importance evaluation system comprises: The second omics data acquisition module: a plurality of sets of omics data are acquired; the plurality of sets of omics data contain at least one to-be-evaluated characteristic; The first scoring module: the plurality of sets of omics data are processed by using the one cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a first score; The zero setting module: the values of the to-be-evaluated characteristics in the plurality of sets of omics data are set to zero to obtain a plurality of sets of omics data after zero setting; The second scoring module: the plurality of sets of omics data after zero setting are processed by using the one cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a second score; The score calculation module: the importance score of the to-be-evaluated characteristic is calculated according to the first score and the second score.

[0017] Compared with the prior art, the present application has the following advantages: The present application uses a prognosis scoring model based on a biological prior knowledge guided attention mechanism to autonomously learn the mutual relationship of three omics characteristics in a high-dimensional nonlinear space, and then deeply fuses the omics characteristics through the biological prior knowledge guided attention mechanism. This method can more effectively capture the unique information of each omics and mine the complex nonlinear relationship between different omics characteristics, thereby significantly improving the fusion effect and further improving the accuracy of the prognosis score.

[0018] The present application obtains a prognosis score as a first score by performing prognosis prediction on the plurality of sets of omics data, sets the values of the to-be-evaluated characteristics in the plurality of sets of omics data to zero, performs prognosis prediction on the plurality of sets of omics data after zero setting to obtain a prognosis score as a second score, and calculates the importance score of the to-be-evaluated characteristic by using the first score and the second score. This method makes the prognosis score more interpretable by calculating the importance score of the to-be-evaluated characteristic. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of a cancer prognosis prediction method based on multi-omics fusion provided for Example 1.

[0020] Figure 2A flowchart for preprocessing omics data provided for Example 1.

[0021] Figure 3 A structural diagram of a graph convolution network feature extraction model provided for Example 1.

[0022] Figure 4 A structural diagram of a prognosis scoring model based on a biology prior knowledge guided attention mechanism provided for Example 1.

[0023] Figure 5 A structural diagram of a multi-omics effect enhancement module provided for Example 1.

[0024] Figure 6 A structural diagram of a multi-omics effect enhancement module provided for Example 1.

[0025] Figure 7 A flowchart of an omics data feature importance evaluation method provided for Example 1.

[0026] Figure 8 A schematic diagram of a cancer prognosis prediction system based on multi-omics fusion provided for Example 1.

[0027] Figure 9 A comparison chart of prognosis prediction results of three cancers provided for Example 1.

[0028] Figure 10 A comparison chart of prognosis prediction results of three cancers under three models provided for Example 1. DETAILED DESCRIPTION

[0029] The accompanying drawings are only for illustrative purposes and should not be construed as limiting the patent; In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product; It is understandable that some well-known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0030] The technical solutions of the present application will be further described below in conjunction with the drawings and embodiments.

[0031] Example 1 As shown in Figure 1 A cancer prognosis prediction method based on multi-omics fusion, comprising: S1: obtaining multi-omics data; S2: preprocessing the multi-omics data respectively to obtain multi-group first omics data features; S3: inputting the multi-group first omics data features into a graph convolution network feature extraction model respectively to obtain multi-group second omics features; S4: inputting the plurality of groups of the second omics features into a prognosis scoring model based on a biology prior knowledge guided attention mechanism to obtain a prognosis score.

[0032] Further, as shown in the figure, the omics data is preprocessed, including: Figure 2 S201: denoising the omics data to obtain denoised omics data; S202: using F-value and principal component analysis algorithm to extract features from the denoised omics data to obtain feature-extracted omics data; S203: scaling the feature-extracted omics data to obtain first omics data features.

[0033] In one specific embodiment, the first omics data features can construct a weighted sample similarity network and an adjacency matrix. The cosine similarity is used to construct a weighted sample similarity network for each omics feature data. Then, by setting the number of correlation between nodes, an adjacency matrix is constructed to represent the weight of the correlation between nodes to distinguish the mutual influence between different adjacent nodes on the current node.

[0034] Further, the omics data includes mRNA data, miRNA data, and methylation data.

[0035] It should be noted that the present application adopts the combination of mRNA data, miRNA data, and methylation data for prognosis prediction, fully considering the biological hierarchical regulation relationship between different omics data. The mRNA data, miRNA data, and methylation data constitute a methylation-transcription-post-transcription regulation pathway at the biological level.

[0036] Further, as shown in the figure, the graph convolution network feature extraction model includes a first graph convolution network module, a second graph convolution network module, and a third graph convolution network module. Figure 3 The output end of the first graph convolution network module is connected to the input end of the second graph convolution network module, and the output end of the second graph convolution network module is connected to the input end of the third graph convolution network module.

[0037] It should be noted that the graph convolution uses the number of degrees of each node and its neighbor nodes to perform information transmission and aggregation with a certain weight, integrates a position coding module, and preserves the topological information of biological sequences. The graph convolution network can use the number of degrees to perform information interaction between the node and its neighbor nodes with a certain weight in each update process, update the node layer by layer, and enhance the connection ability between different samples on the relevant dimension data.

[0038] Further, as shown in the figure, Figure 4 ​​As shown, the prognosis scoring model based on the attention mechanism guided by biological prior knowledge comprises a first single-modal multi-head attention module, a second single-modal multi-head attention module, a third single-modal multi-head attention module, a cross-modal cross-attention fusioner, and a multi-omics effect enhancement module. The second-omics features corresponding to the mRNA data are input into an input end of the first single-modal multi-head attention module, the second-omics features corresponding to the miRNA data are input into an input end of the second single-modal multi-head attention module, and the second-omics features corresponding to the methylation data are input into an input end of the third single-modal multi-head attention module; an output end of the first single-modal multi-head attention module, an output end of the second single-modal multi-head attention module, an output end of the third single-modal multi-head attention module, and an input end of the cross-modal cross-attention fusioner are connected; an output end of the cross-modal cross-attention fusioner, the output end of the first single-modal multi-head attention module, and the output end of the third single-modal multi-head attention module are connected to an input end of the multi-omics effect enhancement module; and an output end of the multi-omics effect enhancement module outputs the prognosis score.

[0039] It should be noted that the single-modal multi-head attention module can effectively process high-dimensional data, the cross-modal cross-attention fusioner can capture important global and local features in the data, improve the expression ability and generalization ability of the model, and effectively analyze and utilize the information in each type of omics data. The prognosis scoring model based on the attention mechanism guided by biological prior knowledge can effectively solve the problems of gradient explosion and gradient disappearance, and improve the stability of the biomedical classification task.

[0040] Further, as shown in Figure 5 、 Figure 6 The multi-omics effect enhancement module comprises a first bias module, a second bias module, a third bias module, a first mapping module, a second mapping module, a third mapping module, a multi-modal fusion module, and a linear weight bias module. The output end of the first single-modal multi-head attention module is connected to an input end of the first bias module and an input end of the first mapping module; the output end of the cross-modal cross-attention fusioner is connected to an input end of the second bias module and an input end of the second mapping module; the output end of the third single-modal multi-head attention module is connected to an input end of the third bias module and an input end of the third mapping module; the output ends of the first bias module, the second bias module, and the third bias module, the output ends of the first mapping module and the second mapping module, and the output end of the third mapping module are connected to an input end of the multi-modal fusion module; the output end of the multi-modal fusion module is connected to an input end of the linear weight bias module; and the output end of the linear weight bias module outputs the prognosis score.

[0041] It should be noted that the multi-omics effect enhancement module realizes multi-modal effect enhancement by using a multi-dimensional weight matrix. The multi-omics effect enhancement module dynamically adjusts the weights of different modalities to adjust the prediction orientation of different tasks and omics data. This dynamic adjustment can be automatically completed through the learning process, so that the model can better enhance the connection between different modalities and achieve the effect of mutual enhancement of omics prediction effect. In the multi-modal learning task, the multi-omics effect enhancement module is mainly used to enhance the fused omics data feature expression and reduce the noise between the three omics data, which shows its advantage in processing complex data sets.

[0042] As shown in Figure 7 , the omics data feature importance evaluation method comprises: S01: obtaining multiple sets of omics data; the multiple sets of omics data contain at least one feature to be evaluated; S02: processing the multiple sets of omics data using the cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a first score; S03: setting the value of the feature to be evaluated in the multiple sets of omics data to zero to obtain the multiple sets of omics data after zero setting; S04: processing the multiple sets of omics data after zero setting using the cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a second score; S05: calculating the importance score of the feature to be evaluated according to the first score and the second score.

[0043] Further, in step S05, the calculation formula of the importance score of the feature to be evaluated according to the first score and the second score is as follows:

[0044] denotes the feature dimension, denotes the first score, denotes the second score, denotes the importance score, denotes the value of the feature to be evaluated, and taking the absolute value.

[0045] As shown in Figure 8 , a cancer prognosis prediction system based on multi-omics fusion comprises: a first omics data acquisition module: acquiring multiple sets of omics data; a preprocessing module: preprocessing the multiple sets of omics data respectively to obtain multiple sets of first omics data features; a feature extraction module: inputting the multiple sets of first omics data features into a graph convolution network feature extraction model respectively to obtain multiple sets of second omics features; The scoring module: inputting the plurality of sets of the second omics features into a prognosis scoring model based on a biology prior knowledge guided attention mechanism to obtain a prognosis score.

[0046] The omics data feature importance evaluation system comprises: The second omics data acquisition module: acquiring a plurality of sets of omics data; the plurality of sets of omics data contain at least one to-be-evaluated feature; The first scoring module: processing the plurality of sets of omics data by using the one cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a first score; The zero setting module: setting the value of the to-be-evaluated feature in the plurality of sets of omics data to zero to obtain the plurality of sets of omics data after zero setting; The second scoring module: processing the plurality of sets of omics data after zero setting by using the one cancer prognosis prediction method based on multi-omics fusion to obtain a prognosis score as a second score; The score calculation module: calculating the importance score of the to-be-evaluated feature according to the first score and the second score.

[0047] In one specific embodiment, the loss function used in training of the graph convolution network feature extraction model and the prognosis scoring model based on the biology prior knowledge guided attention mechanism of the present application is a cox proportional loss. It should be noted that, as Figure 9 , Figure 10 The method of the present application is a pan-cancer method, which not only optimizes the ranking consistency and binary classification target, but also considers the continuity of prognosis survival time and the censored information. It can be widely applied to prognosis prediction of various different cancers and is not limited to a specific cancer. Figure 10 As shown (ours represents the method of the present application), compared with the prior art, the present application has advantages in accuracy and stability.

[0048] The same or similar reference numerals correspond to the same or similar components; The terms used to describe the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation on the present application; Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not a limitation on the embodiments of the present application. Based on the above description, those skilled in the art can make other different forms of changes or modifications. Here, it is not necessary and impossible to exhaust all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A cancer prognostic prediction method based on multi-omics fusion, characterized in that, include: S1: Obtain multiple sets of omics data; S2: Preprocess the multiple sets of omics data to obtain multiple sets of first omics data features; S3: Input multiple sets of the first group of omics data features into the graph convolutional network feature extraction model to obtain multiple sets of the second group of omics features; S4: Input multiple sets of the second set of omics features into a prognostic scoring model based on an attention mechanism guided by biological prior knowledge to obtain a prognostic score.

2. The cancer prognosis prediction method based on multi-omics fusion according to claim 1, characterized in that, Preprocessing of omics data includes: S201: Denoise the omics data to obtain denoised omics data; S202: Using F-value and principal component analysis algorithms, feature extraction is performed on the denoised omics data to obtain feature-extracted omics data; S203: Perform data scaling on the omics data after feature extraction to obtain the first omics data features.

3. The cancer prognosis prediction method based on multi-omics fusion according to claim 1, characterized in that, The omics data includes: mRNA data, miRNA data, and methylation data.

4. The cancer prognosis prediction method based on multi-omics fusion according to claim 1, characterized in that, The graph convolutional network feature extraction model includes: a first graph convolutional network module, a second graph convolutional network module, and a third graph convolutional network module; The output of the first convolutional network module is connected to the input of the second convolutional network module, and the output of the second convolutional network module is connected to the input of the third convolutional network module.

5. The cancer prognosis prediction method based on multi-omics fusion according to claim 3, characterized in that, The prognostic scoring model based on biological prior knowledge-guided attention mechanisms includes: a first unimodal multi-head attention module, a second unimodal multi-head attention module, a third unimodal multi-head attention module, a cross-modal cross-attention fusion device, and a multi-omics effect enhancement module; The second omics feature corresponding to the mRNA data is input to the input end of the first unimodal multi-head attention module, the second omics feature corresponding to the miRNA data is input to the input end of the second unimodal multi-head attention module, and the second omics feature corresponding to the methylation data is input to the input end of the third unimodal multi-head attention module. The output ends of the first unimodal multi-head attention module, the second unimodal multi-head attention module, and the third unimodal multi-head attention module are connected to the input end of the cross-modal cross-attention fusion module. The output ends of the cross-modal cross-attention fusion module, the first unimodal multi-head attention module, and the third unimodal multi-head attention module are connected to the input end of the multi-omics effect enhancement module. The output end of the multi-omics effect enhancement module outputs the prognostic score.

6. The cancer prognosis prediction method based on multi-omics fusion according to claim 5, characterized in that, The multi-omics effect enhancement module includes: a first bias module, a second bias module, a third bias module, a first mapping module, a second mapping module, a third mapping module, a multimodal fusion module, and a linear weight bias module; The output of the first unimodal multi-head attention module is connected to the input of the first bias module and the input of the first mapping module. The output of the cross-modal cross-attention fusion module is connected to the input of the second bias module and the input of the second mapping module. The output of the third unimodal multi-head attention module is connected to the input of the third bias module and the input of the third mapping module. The outputs of the first bias module, the second bias module, the third bias module, the first mapping module, the second mapping module, and the third mapping module are connected to the input of the multimodal fusion module. The output of the multimodal fusion module is connected to the input of the linear weight bias module. The output of the linear weight bias module outputs the prognostic score.

7. A method for assessing the importance of omics data features, characterized in that, include: S01: Acquire multiple sets of omics data; The multi-omics data sets contain at least one feature to be evaluated; S02: The multi-omics data are processed using the cancer prognosis prediction method based on multi-omics fusion as described in any one of claims 1 to 6 to obtain a prognosis score, which is used as the first score; S03: Set the values ​​of the features to be evaluated in multiple sets of omics data to zero to obtain the zeroed sets of omics data; S04: The zeroed multi-omics data are processed using the cancer prognosis prediction method based on multi-omics fusion as described in any one of claims 1 to 6 to obtain a prognosis score, which is used as a second score. S05: Calculate the importance score of the feature to be evaluated based on the first score and the second score.

8. The method for assessing the importance of omics data features according to claim 7, characterized in that, In step S05, based on the first score and the second score, the importance score of the feature to be evaluated is calculated using the following formula: Representing feature dimension, This indicates the first rating. This indicates the second rating. Indicates importance score, Indicates to Take the absolute value.

9. A cancer prognostic prediction system based on multi-omics fusion, applied to the prediction method according to any one of claims 1 to 6, characterized in that, include: First omics data acquisition module: Acquire multiple omics data sets; Preprocessing module: preprocesses multiple sets of the omics data to obtain multiple sets of first omics data features; Feature extraction module: Input multiple sets of first omics data features into the graph convolutional network feature extraction model to obtain multiple sets of second omics features; Scoring module: Input multiple sets of second-group features into a prognostic scoring model based on an attention mechanism guided by biological prior knowledge to obtain a prognostic score.

10. An omics data feature importance assessment system, applied to the assessment method described in any one of claims 7-8, characterized in that, include: The second omics data acquisition module: acquires multiple omics data sets; The multi-omics data sets contain at least one feature to be evaluated; First scoring module: The multi-omics data are processed using the cancer prognosis prediction method based on multi-omics fusion as described in any one of claims 1 to 6 to obtain a prognosis score, which is used as the first score; Zeroing module: Sets the values ​​of the features to be evaluated in multiple sets of omics data to zero, resulting in zeroed sets of omics data; Second scoring module: The multi-omics data after being zeroed are processed using the cancer prognosis prediction method based on multi-omics fusion as described in any one of claims 1 to 6 to obtain a prognosis score, which is used as the second score; Score Calculation Module: Calculates the importance score of the feature to be evaluated based on the first score and the second score.

Images