A TP53 gene prediction method and system

Abstract

The present invention discloses a TP53 gene prediction method and system, belonging to the field of gene prediction technology. The present invention predicts the TP53 gene based on the pathological HE sections of diffuse large B-cell lymphoma patients, and at the same time, based on the backbone network of the MobileNet model, introduces a cell feature subnetwork, an interstitial feature subnetwork and a feature fusion subnetwork, constructs a MODTP53Net network model, and identifies cell features, interstitial features and full image block features of the whole slide imaging of the diffuse large B-cell lymphoma HE section to obtain a feature vector, and predicts the negative and positive test results of the TP53 gene through a Transformer network model classifier. The present invention has high prediction accuracy and can quickly detect the TP53 gene, avoiding the high medical cost caused by taking gene testing for all patients, and also makes up for the dilemma that both NGS sequencing and FISH testing must be done, saving cost and time.

Description

Technical Field

[0001] The present invention belongs to the technical field of gene prediction, and in particular relates to a TP53 gene prediction method. Background Art

[0002] Lymphoma can be roughly divided into two categories: Hodgkin's Lymphoma (HL) and Non-Hodgkin's Lymphoma (NHL). The incidence of NHL is much higher than that of HL, which is 5 times that of HL. Among them, diffuse large B cell lymphoma (DLBCL) is the most common and highly heterogeneous invasive lymphoma among non-Hodgkin's lymphomas, accounting for about 30% to 40% of non-Hodgkin's lymphomas, and the incidence is increasing year by year. The gold standard for the diagnosis of DLBCL is histopathological examination.

[0003] TP53 is a tumor suppressor gene located on the short arm of human chromosome 17. It encodes and expresses Tp53 protein, also known as P53, and is the most common mutated gene in human malignant tumors. Its common mutations include missense mutations caused by point mutations, as well as insertion and deletion mutations of gene fragments. At present, the technologies used in clinical laboratories to detect the TP53 gene mainly include gene sequencing and fluorescence in situ hybridization (FISH). Among them, gene sequencing technology focuses more on missense mutations or insertion and deletion mutations of small fragments. Since the TP53 encoding gene is long and the mutation site is uncertain, the traditional first-generation sequencing Sanger method may miss detection. Therefore, the second-generation sequencing (NGS) technology is more used to ensure accuracy. FISH technology focuses on insertion and deletion mutations of large fragments. The two detection technologies each have their own advantages and limitations. Therefore, in order to ensure that the TP53 gene of DLBCL patients is negative, both tests are often needed and both are negative before a negative genetic diagnosis result can be made. However, as long as one of the detection technologies gives a positive result, it can be determined that the patient's TP53 gene is positive.

[0004] Studies have shown that TP53 mutations occur in approximately 22% of DLBCL, and are closely related to the occurrence of DLBCL, disease progression, and worse prognosis. They are an important basis for A53 subtyping in DLBCL molecular classification. However, both NGS sequencing and FISH detection technology are expensive, time-consuming, and have high technical barriers (high requirements for detection instruments and report analysts). At present, many laboratories do not have the conditions for TP53 gene detection, which results in some TP53-positive patients not being able to receive timely, effective, accurate diagnosis and prognosis judgment. Summary of the invention

[0005] The present invention aims to solve the technical problems existing in the prior art and provides a TP53 gene prediction method. The TP53 gene is predicted based on the pathological HE sections of diffuse large B-cell lymphoma patients. At the same time, based on the backbone network of the MobileNet model, a cell feature subnetwork, an interstitial feature subnetwork and a feature fusion subnetwork are introduced to construct a MODTP53Net network model. The cell features, interstitial features and full image block features of the whole slide imaging of the diffuse large B-cell lymphoma HE sections are identified to obtain feature vectors, and the negative and positive detection results of the TP53 gene are predicted by the Transformer network model classifier. The present invention has extremely high sensitivity and specificity to assist doctors in quickly screening out patients who must undergo TP53 gene testing from patients pathologically diagnosed as DLBCL, avoiding the high medical costs caused by blindly taking gene testing for all patients. On the other hand, it can also make up for the dilemma that both NGS sequencing and FISH testing must be done, further saving costs and time.

[0006] The present invention provides a TP53 gene prediction method, comprising the following steps:

[0007] Localization of tumor regions in multiple sample images of whole slide imaging of HE sections of diffuse large B-cell lymphoma;

[0008] The tumor area image located in each sample image is divided into multiple image blocks, and some image blocks are randomly selected from the divided image blocks as the training set of the MODTP53Net network model;

[0009] The training set of the MODTP53Net network model is input into the MODTP53Net network model to train the MODTP53Net network model, and the trained MODTP53Net network model is obtained to extract the features of the input image block, and a set of feature vector sequences is obtained for each sample; the feature vector sequence of each sample is input into the trained Transformer network model classifier to obtain the TP53 gene detection result category of each sample; finally, the trained MODTP53Net network model and the Transformer network model classifier are obtained;

[0010] The whole slide imaging images of diffuse large B-cell lymphoma HE sections were collected and cut into multiple image blocks, some image blocks were randomly selected and input into the trained MODTP53Net network model, the features of the input image blocks were extracted and combined into a feature vector sequence, and the feature vector sequence was input into the trained Transformer network model classifier to classify the TP53 gene detection results;

[0011] The backbone network of the MODTP53Net network model adopts the MobileNet model, which also includes a cell feature subnetwork, an interstitial feature subnetwork and a feature fusion subnetwork. Cell labels, interstitial labels and image block labels are used simultaneously in the learning process. The MODTP53Net network model performs supervised training on a single image block and learns to obtain cell features, interstitial features and image block features respectively. The MODTP53Net network model uses the second-to-last global average pooling layer to output an image block feature vector, and the feature vector of the image block fuses cell features and interstitial features.

[0012] Among them, the training set of the Transformer network model classifier is the same as the training set of the MODTP53Net network model. The trained MODTP53Net network model is used to extract features of different image blocks of each sample in turn. A set of feature vector sequences is obtained for each sample to construct the Transformer network model classifier training set.

[0013] Preferably, the loss function of the MODTP53Net network model includes cell loss, interstitial loss and classification loss, and the weights of the cell loss, interstitial loss and classification loss in the loss function are adjusted according to the training effect.

[0014] Preferably, the cell feature subnetwork, interstitial feature subnetwork and feature fusion subnetwork in the MODTP53Net network model are synchronously trained through a shared loss function during the model training process.

[0015] Preferably, each image block includes a cell label, a stroma label and a classification label.

[0016] Preferably, the cell labels refer to all tumor cells in the image block.

[0017] Preferably, the interstitial label is the area after excluding all cells and invalid areas in the image block; the invalid area refers to blood vessels, bubbles and areas of missing film caused by artificial film making.

[0018] Preferably, the classification label refers to the result of TP53 gene detection of the sample corresponding to the image block, including positive and negative.

[0019] Preferably, the classification label of each image block is the same as the classification label of the sample corresponding to the image block. Specifically, a negative classification label indicates that the detection results of both NGS and FISH detection technologies for the TP53 gene of the sample are negative, and a positive classification label indicates that at least one of the detection results of both NGS and FISH detection technologies for the TP53 gene of the sample is positive.

[0020] Preferably, the TP53 gene category prediction process is divided into two steps: sample feature extraction and sample category prediction. The first step is to extract a set of feature vector sequences of the sample by the trained MODTP53Net network model, and the second step is to input this set of feature vector sequences of the sample into the trained Transformer network model for classification.

[0021] Preferably, the Transformer network model classifier uses a cross entropy loss function during training.

[0022] The present invention provides a TP53 gene prediction system, comprising a processor, wherein the processor executes the steps of the above-mentioned TP53 gene prediction method.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. The present invention designs a TP53 gene prediction model MODTP53Net based on multi-task decomposition, which decomposes the image block TP53 gene prediction task into multiple learning tasks such as cell feature learning, interstitial feature learning and TP53 gene category prediction learning. The model can simultaneously utilize cell features and interstitial features to improve prediction accuracy, with low cost, fast detection, no technical barriers, and the test results can be obtained by inputting samples.

[0025] 2. Compared with other black box models, the present invention can observe the intermediate results learned by the model, such as cell segmentation effect or interstitial segmentation effect. In the process of class prediction of samples, when the MODTP53Net network model extracts features from different image blocks of the samples, its cell feature subnetwork will output the cell segmentation map of the corresponding image block, and the interstitial feature subnetwork will output the interstitial segmentation map of the corresponding image block.

[0026] 3. The present invention can flexibly set the weight value of the model loss function to explore the contribution of cell characteristics or interstitial characteristics to gene prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The present invention is a flowchart of a method for predicting the TP53 gene based on HE sections of diffuse large B-cell lymphoma according to an embodiment of the present invention.

[0028] Figure 2 A schematic diagram of the structure of a MODTP53Net network model according to an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of the structure of a Transformer network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] According to a specific embodiment of the present invention, the TP53 gene prediction method of the present invention is described in detail below.

[0032] The present invention provides a TP53 gene prediction method, comprising the following steps:

[0033] Localization of tumor regions in multiple sample images of whole slide imaging of HE sections of diffuse large B-cell lymphoma;

[0034] The tumor area image located in each sample image is divided into multiple image blocks, and some image blocks are randomly selected from the divided image blocks as the training set of the MODTP53Net network model;

[0035] The training set of the MODTP53Net network model is input into the MODTP53Net network model to train the MODTP53Net network model, and the trained MODTP53Net network model is obtained to extract the features of the input image block, and a set of feature vector sequences is obtained for each sample; the feature vector sequence of each sample is input into the trained Transformer network model classifier to obtain the TP53 gene detection result category of each sample; finally, the trained MODTP53Net network model and the Transformer network model classifier are obtained;

[0036] The whole slide imaging images of diffuse large B-cell lymphoma HE sections were collected and cut into multiple image blocks, some image blocks were randomly selected and input into the trained MODTP53Net network model, the features of the input image blocks were extracted and combined into a feature vector sequence, and the feature vector sequence was input into the trained Transformer network model classifier to classify the TP53 gene detection results;

[0037] The backbone network of the MODTP53Net network model adopts the MobileNet model, which also includes a cell feature subnetwork, an interstitial feature subnetwork and a feature fusion subnetwork. Cell labels, interstitial labels and image block labels are used simultaneously in the learning process. The MODTP53Net network model performs supervised training on a single image block and learns to obtain cell features, interstitial features and image block features respectively. The MODTP53Net network model uses the second-to-last global average pooling layer to output an image block feature vector, and the feature vector of the image block fuses cell features and interstitial features.

[0038] The feature vector of the image block is extracted by the feature fusion subnetwork, whose input is composed of the feature map information extracted by the previous cell feature subnetwork and the feature map information extracted by the interstitial feature subnetwork. The MODTP53Net network model uses cell features and interstitial features during training, and finally outputs a feature vector in the fusion subnetwork, which represents the global features of the image block; the training set of the Transformer network model classifier is the same as the training set of the MODTP53Net network model. The trained MODTP53Net network model is used to combine the feature vectors extracted from different image blocks of each sample in the training set, and each sample obtains a feature vector sequence. The TP53 gene test result categories include TP53 gene test results that are positive and TP53 gene test results that are negative.

[0039] The backbone network is a convolutional neural network, which is used to extract abstract feature map information of image blocks. The feature map information will be input into the subsequent cell feature sub-network and interstitial feature sub-network respectively. The feature map information extracted by the cell feature sub-network is combined with the feature map information extracted by the interstitial feature sub-network and input into the final feature fusion network to predict the image block category.

[0040] Among them, the training set of the Transformer network model classifier is the same as the training set of the MODTP53Net network model. The trained MODTP53Net network model is used to extract features of different image blocks of each sample in turn. A set of feature vector sequences is obtained for each sample to construct the Transformer network model classifier training set.

[0041] In certain specific embodiments of the present invention, the loss function of the MODTP53Net network model includes cell loss, interstitial loss and classification loss. In the loss function, the weights of cell loss, interstitial loss and classification loss are adjusted according to the training effect. The MODTP53Net network model is only trained for image blocks, and the labels used in the loss function refer to the labels of image blocks. The labels of all image blocks will be used in the Transformer network model.

[0042] In certain specific embodiments of the present invention, the cell feature subnetwork, interstitial feature subnetwork, and feature fusion subnetwork in the MODTP53Net network model are synchronously trained through a shared loss function during the model training process to ensure that the features extracted by each can be effectively fused and jointly optimize the prediction ability of the model. The cell feature subnetwork focuses on extracting feature information related to cells, while the interstitial feature subnetwork focuses on capturing features related to interstitial tissue. The feature fusion subnetwork is responsible for stacking and combining the feature maps extracted by the two subnetworks to generate the final prediction results. Through synchronous training, the entire model can learn a more comprehensive and accurate feature representation, thereby improving the prediction accuracy of the TP53 gene.

[0043] In some specific embodiments of the present invention, each image block includes a cell label, a stroma label, and a classification label.

[0044] In some specific embodiments of the present invention, the cell label refers to all tumor cells in the image block.

[0045] In some specific embodiments of the present invention, the interstitial label is the area after excluding all cells and invalid areas in the image block; the invalid area refers to blood vessels, bubbles and areas of missing film caused by artificial film making.

[0046] In certain specific embodiments of the present invention, the classification label refers to the result of TP53 gene detection of the sample corresponding to the image block, including positive and negative.

[0047] In certain specific embodiments of the present invention, the classification label of each image block is the same as the classification label of the sample corresponding to the image block. Specifically, a negative classification label indicates that the test results of both NGS and FISH detection technologies for the TP53 gene of the sample are negative, and a positive classification label indicates that at least one of the test results of both NGS and FISH detection technologies for the TP53 gene of the sample is positive.

[0048] There are two types of classification labels, one is the label of the image block, which is used in the MODTP53Net network model, and the other is the label of the entire sample used in the Transformer classification network. In the present invention, all image blocks corresponding to the sample use the same label as the sample. If the sample is positive, all corresponding image blocks are also positive.

[0049] The classification labels of the sample data used for training in the present invention, and the classification label of each image block is consistent with the classification of its corresponding sample.

[0050] The present invention is based on the TP53 intelligent prediction system of diffuse large B-cell lymphoma HE sections. The fundamental meaning is that no NGS and FISH tests are performed, and only the digital scanning images of HE sections are used to predict the results of TP53, which can be highly consistent with the actual results obtained by these two detection methods. This can save the possibility of opening more clinical test items and save medical expenses.

[0051] In certain specific embodiments of the present invention, the TP53 gene category prediction process is divided into two steps: sample feature extraction and sample category prediction. In the first step, the trained MODTP53Net network model extracts a set of feature vector sequences of the sample, and in the second step, the set of feature vector sequences of the sample is input into the trained Transformer network model for classification.

[0052] In some specific embodiments of the present invention, the Transformer network model classifier uses a cross entropy loss function during training.

[0053] In some specific embodiments of the present invention, the image blocks are segmented according to a specification of 768×768 pixels;

[0054] In some specific embodiments of the present invention, 30 image blocks are randomly selected for each sample.

[0055] In some specific embodiments of the present invention, the length of the feature vector of each image block is 1280 dimensions.

[0056] In some specific embodiments of the present invention, for each sample, 30 1280-dimensional feature vectors of 30 image blocks are combined into a 30×1280-dimensional feature vector sequence.

[0057] The present invention provides a TP53 gene prediction system, comprising a processor, wherein the processor executes the steps of the above-mentioned TP53 gene prediction method.

[0058] The above description is only an embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and improvement made within the application scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A TP53 gene prediction method, characterized in that: The steps include: Localization of tumor regions in multiple sample images of whole slide imaging of HE sections of diffuse large B-cell lymphoma; The tumor area image located in each sample image is divided into multiple image blocks, and some image blocks are randomly selected from the divided image blocks as the training set of the MODTP53Net network model; The training set of the MODTP53Net network model is input into the MODTP53Net network model to train the MODTP53Net network model, and the trained MODTP53Net network model is obtained to extract the features of the input image block, and a set of feature vector sequences is obtained for each sample; the feature vector sequence of each sample is input into the trained Transformer network model classifier to obtain the TP53 gene detection result category of each sample; finally, the trained MODTP53Net network model and the Transformer network model classifier are obtained; The whole slide imaging images of diffuse large B-cell lymphoma HE sections were collected and cut into multiple image blocks, some image blocks were randomly selected and input into the trained MODTP53Net network model, the features of the input image blocks were extracted and combined into a feature vector sequence, and the feature vector sequence was input into the trained Transformer network model classifier to classify the TP53 gene detection results; Among them, the backbone network of the MODTP53Net network model adopts the MobileNet model, and also includes a cell feature subnetwork, an interstitial feature subnetwork and a feature fusion subnetwork. Cell labels, interstitial labels and image block labels are used simultaneously in the learning process. The MODTP53Net network model performs supervised training on a single image block, and learns to obtain cell features, interstitial features and image block features respectively; the MODTP53Net network model uses the penultimate global average pooling layer to output an image block feature vector, and the feature vector of the image block fuses cell features and interstitial features; the loss function of the MODTP53Net network model includes cell loss, interstitial loss and classification loss, and the weights of cell loss, interstitial loss and classification loss in the loss function are adjusted according to the training effect; Among them, the training set of the Transformer network model classifier is the same as the training set of the MODTP53Net network model. The trained MODTP53Net network model is used to extract features of different image blocks of each sample in turn. A set of feature vector sequences is obtained for each sample to construct the Transformer network model classifier training set.

2. The TP53 gene prediction method according to claim 1, characterized in that: During the model training process, the cell feature subnetwork, interstitial feature subnetwork, and feature fusion subnetwork in the MODTP53Net network model are trained synchronously through a shared loss function.

3. The TP53 gene prediction method according to claim 1, characterized in that: Each image patch includes a cell label, a stroma label, and a classification label.

4. The TP53 gene prediction method according to claim 3, characterized in that: The cell labels refer to all tumor cells in the image patch.

5. The TP53 gene prediction method according to claim 3, characterized in that: The interstitial label is the area after excluding all cells and invalid areas in the image block; the invalid area refers to blood vessels, bubbles and areas of missing film caused by artificial film making.

6. The TP53 gene prediction method according to claim 3, characterized in that: The classification label refers to the result of the TP53 gene test of the sample corresponding to the image block, including positive and negative.

7. The TP53 gene prediction method according to claim 6, characterized in that: The classification label of each image block is the same as the classification label of the sample corresponding to the image block. Specifically, a negative classification label indicates that the test results of both NGS and FISH detection technologies for the TP53 gene of the sample are negative, and a positive classification label indicates that at least one of the test results of NGS and FISH detection technologies for the TP53 gene of the sample is positive.

8. The TP53 gene prediction method according to claim 1, characterized in that: The TP53 gene category prediction process is divided into two steps: sample feature extraction and sample category prediction. The first step is to extract a set of feature vector sequences of the sample by the trained MODTP53Net network model. The second step is to input this set of feature vector sequences of the sample into the trained Transformer network model for classification.

9. A TP53 gene prediction system, characterized in that: Comprising a processor, wherein the processor executes the steps of the TP53 gene prediction method according to any one of claims 1 to 8.

Images