Diffuse midline glioma classification method, device and program product

Abstract

The invention relates to the field of intelligent medical treatment, in particular to a diffuse midline glioma classification method and device and a program product. The method includes: acquiring molecular data of a to-be-detected sample; performing feature extraction on the molecular data to obtain DNA methylation features; classification is carried out based on the DNA methylation characteristics to obtain a classification result that the to-be-detected sample is in a bridge brain shape or a thalamus shape, the bridge brain shape is a tumor type of which the tumor position is located in the bridge brain, and the thalamus shape is a tumor type of which the tumor position is located outside the bridge brain. Targeted therapy can be carried out on different gliomas, and good clinical value is achieved.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical treatment, and specifically to a classification method, device, program product and computer-readable storage medium for diffuse midline glioma. Background Art

[0002] H3 K27-altered gliomas, especially diffuse midline gliomas (DMGs), are a highly aggressive type of central nervous system malignancy that primarily affects children and adolescents. These tumors are characterized by a specific change in lysine 27 (K27) in histone H3, the most common form of which is the K27M mutation, which replaces lysine (K) with methionine (M). This mutation usually occurs in the H3F3A gene encoding the H3.3 histone variant, but can also occur in the HIST1H3B / C gene encoding H3.1. These tumors are mainly located in the midline structures of the central nervous system, such as the brainstem, thalamus, and spinal cord, which are vital for their function, so tumor growth often leads to serious clinical consequences. H3 K27-altered gliomas have a very poor prognosis, with a median survival time of less than 1 year and a 2-year survival rate of less than 10%. The highly invasive nature and recurrence of the tumor are the main reasons. Although imaging techniques such as MRI are helpful for diagnosis, it can be very difficult to obtain sufficient tissue samples for pathological and molecular testing due to the specific location of the tumor. In addition, gliomas with H3K27 alterations may resemble other types of gliomas in imaging, increasing the risk of misdiagnosis. Summary of the invention

[0003] In view of the above problems, the present invention provides a classification method for diffuse midline glioma, which specifically includes: obtaining molecular data of a sample to be tested; performing feature extraction on the molecular data to obtain a DNA methylation feature; and classifying the sample to be tested based on the DNA methylation feature to obtain a classification result of whether the sample to be tested is pontine-like or thalamic-like, wherein the pontine-like is a tumor type in which the tumor is located in the pons, and the thalamic-like is a tumor type in which the tumor is located outside the pons.

[0004] The classification is performed by a trained classifier to obtain a classification result.

[0005] Optionally, the classifier adopts one or more of the following: convolutional neural network, Transformer, residual network, hole convolutional neural network, decision tree, extreme learning machine, support vector machine, XGBoost, AdaBoost, random forest, naive Bayes.

[0006] Optionally, the DNA methylation features are classified by a trained Naive Bayes to obtain a classification result, and the training process of the trained Naive Bayes is: obtaining a DNA methylation feature data set; inputting the DNA methylation feature data set into the Naive Bayes for training to obtain a trained Naive Bayes, wherein the Naive Bayes conducts DNA methylation feature learning through a kernel function to construct a methylation spectrum.

[0007] The kernel function The calculation formula is: Among them, K represents the kernel function, h>0, h represents the smoothing parameter, n represents the number of samples, x represents the feature vector of all samples, and i represents an element in x.

[0008] The features also include TP53 mutation features, and classification is performed based on the DNA methylation spectrum features and TP53 mutation features to obtain pontine-like or thalamic-like classification results.

[0009] Optionally, the feature also includes an NF1 mutation feature, and classification is performed based on the DNA methylation spectrum feature and the NF1 mutation feature to obtain a pontine-like or thalamic-like classification result.

[0010] Optionally, the features also include TP53 mutation features and NF1 mutation features, and classification is performed based on the DNA methylation spectrum features, TP53 mutation features, and NF1 mutation features to obtain pontine-like or thalamic-like classification results.

[0011] The TP53 mutation features and DNA methylation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0012] Optionally, the DNA methylation spectrum features and NF1 mutation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0013] Optionally, the DNA methylation spectrum features, NF1 mutation features, and TP53 mutation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same categories to obtain the probability results, and the classification results are obtained based on the probability results.

[0014] Optionally, the probability result is replaced by a weighted average calculation of the same category.

[0015] Optionally, during the training process, the Naive Bayes of the TP53 mutation feature or the NF1 mutation feature performs feature learning through statistical learning of a specific distribution to obtain a Naive Bayes model of the TP53 mutation feature or the NF1 mutation feature.

[0016] The feature extraction also includes clinical feature extraction, acquiring clinical data, performing clinical feature extraction on the clinical data to obtain clinical features, and classifying based on the DNA methyl spectrum features and clinical features to obtain pontine-like or thalamic-like classification results.

[0017] Optionally, the clinical characteristics include one or more of the following: Ki-67 index, age.

[0018] The DNA methyl spectrum features and clinical features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0019] Optionally, the DNA methyl spectrum features and Ki-67 index are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0020] Optionally, the DNA methyl spectrum features and age are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0021] Optionally, the DNA methyl spectrum features, Ki-67 index, and age are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0022] Optionally, during the training process, the Naive Bayes of clinical characteristics performs feature learning through Gaussian distribution to obtain a Naive Bayes model of clinical characteristics.

[0023] The object of the present invention is to provide a computer program product, which includes a computer program or instructions, and the computer program or instructions are executed by a processor to implement the above-mentioned classification method of diffuse midline glioma.

[0024] The object of the present invention is to provide a computer device, which includes a memory, a processor and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-mentioned classification method of diffuse midline glioma.

[0025] An object of the present invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-mentioned classification method for diffuse midline glioma.

[0026] Advantages of the present invention: 1. In response to the poor prognosis and survival of gliomas with H3 K27 changes, a detailed study of gliomas with H3 K27 changes was conducted. H3 K27 changes alone are not sufficient to cause tumorigenesis, and tumor transformation requires at least one partner change. Animal models have shown that H3 K27-changed tumors with different partner changes show different growth rates, diffusion and invasion phenotypes. Therefore, a subtype classification of gliomas with H3 K27 changes is proposed, namely, pontine-like and thalamic-like tumors located inside and outside the pons. This classification type shows different prognoses. Targeted treatment is given for different subtypes to improve the cure effect of tumors.

[0027] 2. The present invention provides a classification basis for the two types, namely, DNA methylation characteristics, trains a classifier according to the methylation data and performs classification based on the classifier, provides a simple and convenient classification scheme with different concepts, and helps to assist doctors in disease diagnosis.

[0028] 3. The classification basis also includes mutation characteristics (TP53 mutation and NF1 mutation) and clinical characteristics (Ki-67 index, age), and classification is performed through multiple dimensional features to improve classification accuracy.

[0029] 4. According to different features, each feature is trained to obtain the corresponding classifier, and each classifier is classified in parallel to obtain the classification probability of each classifier, and then the classification probabilities are integrated to obtain the final classification result. This method can make classification judgments based on multi-dimensional features and finally integrate the results, including single-dimensional features and multi-dimensional relationships, thereby improving the reliability and accuracy of the classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 A schematic flow chart of a classification method for diffuse midline glioma provided in an embodiment of the present invention; Figure 2 A schematic diagram of a classification system for diffuse midline gliomas provided in an embodiment of the present invention; Figure 3A schematic diagram of a diffuse midline glioma classification device provided by an embodiment of the present invention; Figure 4 Data processing process for subtype classification provided in the embodiment of the present invention; (B) Clinical information, data availability and somatic mutation of H3-DSGs included in this study. C / T / L represent the cervical, thoracic and lumbar segments of the human spinal cord, respectively. Summary of diffuse midline glioma driver genes identified by MutSigCV and reported key mutations. (C) cSNF network showing two multi-omics subtypes of H3-DSGs. TP53 and NF1 mutation status are indicated by edge color. Edge width indicates the strength of evidence supporting patient-patient similarity from multi-omics data. Lymph node size is positively correlated with the age of the patient at diagnosis (7 to 61 years old). Node color indicates the H3-DSG subtype defined by the cSNF method. (D) Mutation frequency of key driver genes in different diseases. Brain-M-Pons and brain-M-Medulla are from Chen et al. (PMID: 32555164). (E) Diagnosis age of patients in different disease groups. (F) Kaplan-Meier survival curves of tumor patients in different groups;

[0032] Figure 5 (A) Proportion of cycling cells in pontine-like and thalamic-like H3-DSGs provided by the present invention. (B) Immunohistochemistry of Ki-67 (left) and Ki-67 index in pontine-like and thalamic-like H3-DSGs.

[0033] Figure 6 The workflow of the SCG classifier provided by the embodiment of the present invention; Figure 7 Key features and cutoff values ​​of the SCG classifier provided for the embodiments of the present invention; (A) cpg used to determine DNA methylation clusters. (B) Procedure for subgrouping tumors using DNA methylation features. (C) Distribution of the values ​​of tumor DNA methylation features in two subgroups. (D) Age distribution of tumors in the two subgroups. (E) Distribution of Ki67 index of tumors in the two subgroups. (F) Distribution of TP53 mutations and NF1 mutations in tumors divided into two subgroups. (G) Schematic diagram of the SCG classifier finally constructed. (H) Performance of the model in leave-one-out validation;

[0034] Figure 8 The results of the 15 independent spinal cord H3 K27 subgroups using the SCG classifier provided for the embodiments of the present invention changed the results of gliomas; Fig. 9 The FANCI and CD3e immunohistochemical indicators of the tumors in the validation cohort provided in the embodiment of the present invention; Fig.10Immunohistochemistry (IHC) provided for the examples of the present invention showed differences in thalamic-like and pontine-like CD14. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0037] Figure 1 The schematic diagram of the classification method of diffuse midline glioma provided in the embodiment of the present invention specifically includes: S101: Acquire molecular data of the sample to be tested; In a specific embodiment, in order to clarify the subtype of H3 K27M mutant spinal cord glioma, the present invention uses 36 cases with DNA methylation, WES and RNA sequencing data to perform multi-omics clustering and subtype identification ( Figure 4 B). There were 22 males and 14 females; the age ranged from 7 to 61 years, with a median age of 33 years. The tumors were extensively resected from the first cervical segment (C1) to the first lumbar segment (L1). Two of them had multifocal tumors. Two subtypes of spinal cord gliomas with H3 K27M mutation were found to have different prognoses, age distributions, and mutation spectra, which simulated the mutational characteristics of gliomas in different anatomical locations in the brain.

[0038] In a specific embodiment, to determine the key somatic mutations, the present invention performed sav2 and MutsigCV analysis on 36 patients in the discovery cohort. Interestingly, 2 cases were found to exhibit H3K27I mutations in the cohort of the present invention, and all other cases exhibited H3K27M mutations.

[0039] In one specific embodiment, a network-based multi-omics integration method, named constrained similarity network fusion, was developed to identify subclusters in the cohort. cSNF first constructed a patient similarity network from each omics data, fused these multi-layer networks to represent the comprehensive associations between patients, and finally clustered the patients using a spectral-based approach. Using cSNF, 36 H3-DSGs were stratified into two subtypes with different clinical and genetic characteristics ( Figure 4 C). Compared with subtype 2, patients with subtype 1 had significantly shorter overall survival (P = 4.1e-04) and were diagnosed at a younger age (P = 4.5e-05). The median age of patients with type 1 was 18 years and the median overall survival was 11.9 months, while the median age of patients with type 2 was 37.5 years and the median survival was 27.6 months. Notably, the present invention found that TP53 mutations were enriched in subtype 1 (13 / 16 vs. 3 / 20, P = 1.25e-4), while NF1 mutations were enriched in subtype 2 (1 / 16 vs. 13 / 20, P = 4.31e-4, Figure 4 B and Figure 4 C).

[0040] cSNF method: First, these features are selected from different omics data to construct PPSNs using the weighted technique in the SNF (Similarity Network Fusion) algorithm, which first calculates the distance between two patients and then determines the weight of the edge between them by scaling the exponential similarity kernel. Then, according to the SNF algorithm definition, the normalized weight matrix (P) and local affinity matrix (A) of each PPSN are constructed. SNF further uses iterative fusion and normalized embedding techniques to integrate multi-omics data. In this iterative fusion process, a small modification was made, as shown in the following formula:

[0041] Among them, v represents one of the different omics data, such as s represents somatic cells, c represents CNV, m represents methylation, r represents RNA-seq, A represents the local affinity matrix, P represents the normalized weight matrix, lab() is a function of the data type from the different omics data, lab(s) and lab(c) are 'Genomics', lab(m) is 'Epigenomics', lab(r) is 'Transcriptomics', N represents the unique number of data types, T represents transposition, k represents the omics data label, k=1, Z(k) represents s; k=2, Z(k) represents c; k=3, Z(k) represents m; k=4, Z(k) represents r.

[0042] The modified approach is to acknowledge that information belonging to the same data type should be considered parallel and cannot be integrated at the same label level.

[0043] Finally, after a certain iterative process, the final fused PPSN can calculate the mean of the fused normalized weight matrix. The spectral-based clustering method is further used to find subgroups on the fused clustering nodes, and the number of clusters is estimated by optimizing RatioCut.

[0044] In a specific embodiment, to better characterize the tumor characteristics of H3-DSG subtype 1 and subtype 2, the present invention compared the somatic mutation spectrum, age at diagnosis, and overall survival outcomes of the H3-DSG subgroups with previously reported brain DMG subgroups in an independent cohort consisting of m-pontine and m-medullary brain DMG17. Tumors of the m-pontine DMG subgroup are mainly located in the pons, while tumors of the m-medullary DMG subgroup are mostly located outside the pons. Strikingly, the newly discovered H3-DSG subtype 1 exhibited somatic characteristics similar to those of the m-pontine DMG subgroup, while the somatic characteristics of the H3-DSG subtype 2 were similar to those of the m-medullary brain DMG subgroup, with the only exceptions being SF3B1 and FGFR1 mutations, which were more frequent in m-medullary brain DMGs than in H3-DSG subtype 2 ( Figure 4 The age distribution at diagnosis was similar in the H3-DSG subtype 1 and M-Pons brain DMG groups and was significantly lower than that in the H3-DSG subtype 2 (P=4.5e-05) and M-Pons brain DMG groups (P=6.4e-06), respectively ( Figure 4 E). The survival rates of the H3-DSG 1 group and the m-pontine DMG group were significantly lower than those of the H3-DSG 2 group (P = 8.7e-04) and the m-medullary DMG group (P = 1.4e-04), respectively ( Figure 4 F). These results indicate that H3-DSG subtype 1 and H3-DSG subtype 2 have similar molecular and clinical characteristics to m-pontine and m-medullary DMG, respectively. Based on these findings, the present invention designates H3-DSG subtype 1 as pontine-like and H3-DSG subtype 2 as thalamic-like.

[0045] In a specific embodiment, by studying the differentially expressed genes between pontine-like and thalamic-like H3-DSGs, the present invention finds that the genes significantly upregulated in pontine-like tumors are mainly related to cell division and cell cycle-related functions, such as MEX3A, KIF2C, FANCI, etc. In contrast, the genes upregulated in thalamic-like tumors include immune and inflammation-related activities. Therefore, compared with pontine-like H3-DSGs, thalamic-like tumors have lower tumor purity (P = 0.006), higher lymphocytes (P = 0.0044) and myeloid scores (P = 0.0006). In fact, scRNA-seq data show that pontine-like and thalamic-like H3-DSGs show different cell type compositions and tumor cell states. Malignant cells account for 80% to 90% of pontine-like tumors, and the remaining components are mostly myeloid cells, and T cells are rare. In contrast, in thalamic-like tumors, malignant cells accounted for only about 50%, and the remaining 50% included myeloid cells, T cells, NK cells, neutrophils, endothelial cells, and oligodendrocytes. Consistent with this, immunohistochemistry (IHC) of 28 H3-DSGs showed that CD14-positive myeloid cells were more abundant in thalamic-like H3-DSGs. We also showed by multicolor staining that myeloid and lymphoid cells were more abundant in thalamic-like H3-DSGs. Correspondingly, CD14 (myeloid), CD3e (T cells), MPO (neutrophils), CD44, HLA-A, and HLA-DR (antigen presentation) were all more abundant in thalamic-like H3-DSGs than in pontine-like H3-DSGs.

[0046] In a specific embodiment, the single cell analysis results also showed that the proportion of CD8-positive cytotoxic T cells in T cells was significantly increased in thalamic-like H3-DSG. The results of immunohistochemistry showed that the proportion of CD3e-positive cells and GZMB-positive cells in thalamic-like H3-DSG was also significantly higher than that of the corresponding cells in pons-like H3-DSG, indicating an increase in the proportion of cytotoxic T cells in thalamic-like H3-DSG.

[0047] In a specific embodiment, the present invention analyzes multi-omics and spatially resolved single-cell data from 36 H3-DSGs and finds two different subgroups: pontine-like and thalamic-like. Thalamic-like tumors are characterized by NF1 mutations and show more immune infiltration. They contain a large number of proinflammatory myeloid cells and CD8+ cytotoxic T cells, and the patients are older and have a relatively good prognosis. In contrast, pontine-like tumors show abundant TP53 mutations, high proliferation niches, genomic instability, and the patients are older and have a relatively poor prognosis.

[0048] Immunohistochemistry (IHC) of 28 H3-DSGs showed that CD14-positive myeloid cells were more numerous in thalamic-like H3-DSGs, such as Fig.10 shown.

[0049] S102: extracting features from the molecular data to obtain DNA methylation features; In one embodiment, the feature also includes a TP53 mutation feature, and classification is performed based on the DNA methylation spectrum feature and the TP53 mutation feature to obtain a pontine-like or thalamic-like classification result.

[0050] In one embodiment, the feature further includes an NF1 mutation feature, and classification is performed based on the DNA methylation spectrum feature and the NF1 mutation feature to obtain a pontine-like or thalamic-like classification result.

[0051] In one embodiment, the features further include TP53 mutation features and NF1 mutation features, and classification is performed based on the DNA methylation spectrum features, TP53 mutation features, and NF1 mutation features to obtain pontine-like or thalamic-like classification results.

[0052] In one embodiment, the feature extraction also includes clinical feature extraction, acquiring clinical data, performing clinical feature extraction on the clinical data to obtain clinical features, and classifying based on the DNA methyl spectrum features and clinical features to obtain pontine-like or thalamic-like classification results.

[0053] In one embodiment, the clinical characteristics include one or more of the following: Ki-67 index, age.

[0054] S103: Classifying the sample to be tested as pontine-like or thalamic-like based on the DNA methylation feature, wherein the pontine-like is a tumor type in which the tumor is located in the pons, and the thalamic-like is a tumor type in which the tumor is located outside the pons.

[0055] In one embodiment, the classification is performed by a trained classifier to obtain a classification result.

[0056] In one embodiment, the classifier adopts one or more of the following: convolutional neural network, Transformer, residual network, hole convolutional neural network, decision tree, extreme learning machine, support vector machine, XGBoost, AdaBoost, random forest, naive Bayes.

[0057] In one embodiment, the DNA methylation feature is classified by a trained naive Bayes to obtain a classification result, and the training process of the trained naive Bayes is: Obtain DNA methylation feature dataset; The DNA methylation feature data set is input into Naive Bayes for training to obtain a trained Naive Bayes, wherein Naive Bayes conducts DNA methylation feature learning through a kernel function to construct a methylation spectrum.

[0058] In one embodiment, the kernel function The calculation formula is: Where K represents the kernel function, h>0, h represents the smoothing parameter, n represents the number of samples, x represents the eigenvector of all samples, and i represents an element in x.

[0059] In one embodiment, the TP53 mutation features and DNA methylation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0060] In one embodiment, the DNA methylation spectrum features and NF1 mutation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0061] In one embodiment, the DNA methylation spectrum features, NF1 mutation features, and TP53 mutation features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same categories to obtain the probability results, and the classification results are obtained based on the probability results.

[0062] In one embodiment, the probability result is replaced by a weighted average calculation of the same category.

[0063] In one embodiment, the Naive Bayes model of the TP53 mutation feature or the NF1 mutation feature is obtained by performing feature learning through statistical learning of a specific distribution during the training process.

[0064] In one embodiment, the DNA methylation spectrum features and clinical features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0065] In one embodiment, the DNA methyl spectrum features and Ki-67 index are respectively input into their respective trained naive Bayes to calculate the probabilities of their respective categories, and the probability results of the same category are averaged to obtain the probability results, and the classification results are obtained according to the probability results.

[0066] In one embodiment, the DNA methyl spectrum features and age are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0067] In one embodiment, the DNA methylation spectrum features, Ki-67 index, and age are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, and the probability results are averaged for the same category to obtain the probability results, and the classification results are obtained based on the probability results.

[0068] In one embodiment, the Naive Bayes of clinical features performs feature learning through Gaussian distribution during training to obtain a Naive Bayes model of clinical features.

[0069] In a specific embodiment, the present invention finds through single cell analysis that more cells in pontine-like H3-DSG are in the cycle of cell proliferation ( Figure 5 A), and the index of Ki-67, a marker of cell proliferation, was significantly increased in H3-DSG of pontine-like cells ( Figure 5 B), the present invention then found that the copy number variation and heterozygosity loss rate of pontine-like H3-DSG were significantly higher than those of thalamus-like H3-DSG, indicating that its genome was more unstable. And the expression level of FANCI, a protein related to DNA damage repair, was significantly increased in pontine-like H3-DSG.

[0070] In a specific embodiment, the different clinical outcomes, biological characteristics, and responses to IR / DNA damage of pontine and thalamic H3-DSGs prompted the present invention to attempt to develop a classification method that accurately identifies pontine and thalamic H3-DSGs subtypes from a limited number of key clinical and molecular features and is easily accessible in clinical practice. To this end, the present invention developed a machine learning model, named SCG classifier. The model uses age, Ki-67 index, TP53 mutation, NF1 mutation, and DNA methylation profile to distinguish pontine and thalamic H3-DSGs ( Figure 6 ). First, the cut-off value of each feature is determined according to the distribution through the naïve Bayes model ( Figure 7 Then, the AUC and AUPRC values ​​of the models using a single feature or different combinations were compared by leave-one-out cross-validation ( Figure 7 The results show that there is a trend that more features can improve the classification accuracy.

[0071] In a specific embodiment, the present invention applies SCG Classifier to assign subtypes in an independent cohort of 15 patients with spinal cord glioma using all available features ( Figure 8 According to the prediction results, the present invention compared the expression of FANCI and CD3e in 15 patients by IHC ( Fig. 9 ). There were more FANCI-positive cells and fewer CD3e-positive cells in pontine-like H3-DSGs.

[0072] In a specific embodiment, the classifier training process is: 1. training data set; The training dataset used consisted of 36 spinal cord glioma samples with H3K27 mutations, which contained clinical information, whole exome, transcriptome, DNA methylation multi-omics data, and grouping labels pons-like (P) and thalamus-like (T) based on multi-omics integration.

[0073] 2. Input features; The features used by the classifier include age, MKI67 immunohistochemistry value, TP53 mutation status (0 / 1), NF1 mutation status (0 / 1), and probe value of DNA methylation.

[0074] In the above features, binary is used to represent mutation status, 0 means no mutation is detected in the sample, and 1 means mutation exists. For methylation probes, 659 key methylation probes screened by the regulatory factor ranking algorithm based on constrained multi-omics integrated regression (CM3R for short) were used. Subsequently, the principal component analysis (PCA) technology was used to reduce the dimensionality of the data of these key probes.

[0075] 3. Naive Bayes classifier; The classifier was implemented using the naivebayes package in R language, and the model performance was evaluated using the Leave-One-Out Cross-Validation (LOOCV) method to ensure the comprehensiveness and reliability of the results.

[0076] The naive Bayes classifier is built based on the Bayesian probability theory. Its core is to use the prior probability information of the feature and derive the posterior probability of the feature through precise mathematical calculations. The formula is as follows: This formula can be further transformed into an expression for the classification task as follows: Naive Bayes learns different features, specifically: (a) Learning of numerical features For numerical features, such as the patient's age and MKI67 immunohistochemistry value, we used Gaussian distribution for in-depth learning. By accurately fitting the Gaussian distribution, the model can capture the distribution characteristics behind these numerical features, thereby constructing an age Bayesian classifier with predictive ability ( ) and MKI67 immunohistochemistry Bayesian classifier ( ). The distribution is defined as follows: in, and represents the class-conditional mean and variance.

[0077] (b) Learning of Discrete Features For discrete features, especially variables with binary classification attributes such as TP53 mutation and NF1 mutation status, the present invention uses a statistical learning method based on a specific distribution to conduct in-depth feature exploration, thereby revealing the possible relationship between these discrete features and the category to which the sample belongs. Subsequently, a Bayesian classifier based on TP53 mutation status was constructed ( ) and a Bayesian classifier based on NF1 mutation status ( ), the formula for this particular distribution is as follows:

[0078] in, .

[0079] (c) Learning of continuous features For continuous features, such as DNA methylation features, we used non-parametric techniques based on kernel functions to conduct deep mining and precise modeling, which can more flexibly capture the complex changes and potential laws in methylation features. Based on these in-depth learning results, we constructed a Bayesian classifier with methylation feature values ​​as the core input ( ), the classifier can make efficient and accurate category judgments based on the methylation spectrum of the sample. The kernel function is defined as follows: Among them, K represents the kernel function, h>0, h represents the smoothing parameter, n represents the number of samples, x represents the feature vector of all samples, and i represents an element in x.

[0080] 4. Category judgment based on soft voting First, multiple classifiers are used to evaluate the probability of samples belonging to different categories. This process can give full play to the unique advantages of different classifiers in feature capture and pattern recognition. Then, in order to obtain more robust and accurate classification results, the judgment probabilities of all classifiers for the same category are summarized, and the averaging strategy is used to calculate the final category belonging probability. The specific calculation process is as follows:

[0081] Among them, Pro represents probability.

[0082] After obtaining the average value of the probabilities of each category, the present invention adopts a judgment rule based on probability maximization to determine the final category of the sample, that is, selects the category with the highest average probability as the belonging category of the sample. In this way, the present invention can achieve a more accurate and reliable classification decision on the basis of comprehensively considering the opinions of multiple classifiers. The specific judgment rule is defined as follows:

[0083] If Pro(P) > Pro(T), the category of this sample is the P class, that is, the pons-like group; If Pro(P) < Pro(T), the category of this sample is the T class, that is, the thalamus-like group.

[0084] In one embodiment, the molecular data of the sample to be tested is replaced with the tumor cell immune infiltration data of the sample to be tested, the DNA methylation features obtained by the feature extraction are replaced with immune infiltration features, the positive numbers of CD8 and / or CD14 in the immune infiltration data are calculated, and the classification based on the DNA methylation features is replaced with the classification based on the immune infiltration features to obtain the classification result of whether the sample to be tested is pons-like or thalamus-like. The pons-like is a tumor type with the tumor location in the pons, and the thalamus-like is a tumor type with the tumor location outside the pons.

[0085] Optionally, when the positive number of CD8 and / or CD14 in the immune infiltration data is greater than a preset threshold, it is determined as thalamus-like, and when the positive number of CD8 and / or CD14 in the immune infiltration data is less than the preset threshold, it is determined as pons-like.

[0086] In one embodiment, the feature further includes immune infiltration features. The tumor cell immune infiltration data of the sample to be tested is obtained, the positive numbers of CD8 and / or CD14 in the immune infiltration data are calculated, and the classification results of pons-like or thalamus-like are obtained based on the DNA methylation spectrum features and immune infiltration features.

[0087] In one embodiment, the SCG classifier for immune infiltration features is trained for the classification of pons-like and thalamus-like.

[0088] In one embodiment, the SCG classifier for DNA methylation features is trained for the classification of pons-like and thalamus-like; or the SCG classifier for DNA methylation features and mutation features is trained for the classification of pons-like and thalamus-like; or the SCG classifier for DNA methylation features and clinical features is trained for the classification of pons-like and thalamus-like; Or the classification of pontine-like and thalamic-like can be performed by training SCG classifiers of DNA methylation features and immune infiltration features; Or the classification of pontine-like and thalamic-like can be performed by training SCG classifiers of DNA methylation features, mutation features, and clinical features; Or the classification of pontine-like and thalamic-like tumors can be performed by training SCG classifiers of DNA methylation features, mutation features, clinical features, and immune infiltration features.

[0089] The disclosed embodiments of the present invention further provide a computer program product or system, including a computer program, which implements the above-mentioned steps of the classification method for diffuse midline glioma when executed by a processor.

[0090] Figure 2 The schematic diagram of the classification system of diffuse midline glioma provided in the embodiment of the present invention specifically includes: Acquisition unit: acquires the molecular data of the sample to be tested; Feature unit: extracting features from the molecular data to obtain DNA methylation features; Classification unit: Based on the DNA methylation characteristics, the sample to be tested is classified as pontine-like or thalamic-like, wherein the pontine-like is a tumor type in which the tumor is located in the pons, and the thalamic-like is a tumor type in which the tumor is located outside the pons.

[0091] Figure 3 The schematic diagram of the classification device for diffuse midline glioma provided in the embodiment of the present invention specifically includes: A memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, any one of the above-mentioned diffuse midline glioma classification methods is executed.

[0092] The disclosed embodiments of the present invention further provide a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, is any one of the above-mentioned methods for classifying diffuse midline gliomas.

[0093] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of the method relative to the default setting. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units. A person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above-mentioned embodiments may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0094] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned medium storage can be a read-only memory, a disk or an optical disk, etc.

[0095] The above is a detailed introduction to a computer device provided by the present invention. For a person skilled in the art, according to the concept of the embodiments of the present invention, there may be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for classifying diffuse midline gliomas, characterized in that: include: Acquiring molecular data of the sample to be tested; performing feature extraction on the molecular data to obtain DNA methylation features; Based on the DNA methylation characteristics, the classification result of the sample to be tested is obtained as pontine-like or thalamic-like, wherein the pontine-like is a tumor type in which the tumor is located in the pons, and the thalamic-like is a tumor type in which the tumor is located outside the pons.

2. The method for classifying diffuse midline glioma according to claim 1, characterized in that: The classification is performed by a trained classifier to obtain a classification result; optionally, the classifier adopts one or more of the following: convolutional neural network, Transformer, residual network, hollow convolutional neural network, decision tree, extreme learning machine, support vector machine, XGBoost, AdaBoost, random forest, naive Bayes; optionally, the DNA methylation feature is classified by a trained naive Bayes to obtain a classification result, and the training process of the trained naive Bayes is: obtaining a DNA methylation feature data set; inputting the DNA methylation feature data set into the naive Bayes for training to obtain a trained naive Bayes, wherein the naive Bayes performs DNA methylation feature learning through a kernel function to construct a methylation spectrum.

3. The method for classifying diffuse midline glioma according to claim 2, characterized in that: The kernel function The calculation formula is: Where K represents the kernel function, h>0, h represents the smoothing parameter, n represents the number of samples, x represents the eigenvector of all samples, and i represents an element in x.

4. The method for classifying diffuse midline glioma according to claim 1, characterized in that: The feature also includes a TP53 mutation feature, and classification is performed based on the DNA methylation feature and the TP53 mutation feature to obtain a pontine-like or thalamic-like classification result; optionally, the feature also includes an NF1 mutation feature, and classification is performed based on the DNA methylation feature and the NF1 mutation feature to obtain a pontine-like or thalamic-like classification result; optionally, the feature also includes a TP53 mutation feature and a NF1 mutation feature, and classification is performed based on the DNA methylation feature, the TP53 mutation feature, and the NF1 mutation feature to obtain a pontine-like or thalamic-like classification result.

5. The method for classifying diffuse midline glioma according to claim 4, characterized in that: The TP53 mutation feature and DNA methylation feature are respectively input into their respective trained Naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories, and the classification results are obtained according to the probability results; optionally, the DNA methylation feature and NF1 mutation feature are respectively input into their respective trained Naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories, and the classification results are obtained according to the probability results; optionally, the DNA methylation feature, NF1 mutation feature, and TP53 mutation feature are respectively input into their respective trained Naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories, and the classification results are obtained according to the probability results; optionally, the probability results are replaced by weighted average calculations for the same categories; optionally, the Naive Bayes of the TP53 mutation feature or the NF1 mutation feature is a Naive Bayes model of the TP53 mutation feature or the NF1 mutation feature obtained by feature learning through statistical learning of a specific distribution during the training process.

6. The method for classifying diffuse midline glioma according to claim 1, characterized in that: The feature extraction also includes clinical feature extraction, acquiring clinical data, performing clinical feature extraction on the clinical data to obtain clinical features, and classifying based on the DNA methyl spectrum features and clinical features to obtain pontine-like or thalamic-like classification results; optionally, the clinical features include one or more of the following: Ki-67 index, age.

7. The method for classifying diffuse midline glioma according to claim 6, characterized in that: The DNA methylation features and clinical features are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories, and the classification results are obtained according to the probability results; optionally, the DNA methylation features and Ki-67 index are respectively input into their respective trained naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories, and the classification results are obtained according to the probability results; Optionally, the DNA methylation features and age are respectively input into their respective trained Naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories to obtain the probability results, and the classification results are obtained according to the probability results; Optionally, the DNA methylation features, Ki-67 index, and age are respectively input into their respective trained Naive Bayes to calculate the probabilities of each in different categories, the probability results are averaged for the same categories to obtain the probability results, and the classification results are obtained according to the probability results; Optionally, the Naive Bayes of clinical characteristics performs feature learning through Gaussian distribution during the training process to obtain a Naive Bayes model of clinical characteristics.

8. A computer program product comprising a computer program or instructions, characterized in that: The computer program or instructions are executed by a processor to implement the method for classifying diffuse midline glioma according to any one of claims.

9. A computer device comprising a memory, a processor and a computer program or instruction stored in the memory, characterized in that: The computer program or instructions are executed by a processor to implement the method for classifying diffuse midline glioma according to any one of claims.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: The computer program or instructions are executed by a processor to implement the method for classifying diffuse midline glioma according to any one of claims.

Images