A method for acute myeloid leukemia subtyping based on prototype contrastive learning

Abstract

The present invention discloses a method for subtyping acute myeloid leukemia based on prototype contrast learning, comprising: extracting and fusing features of AML patient gene expression data, gene mutation data and clinical information, using the fused features as samples; clustering the samples using a clustering method to obtain an initial prototype; constructing a prototype contrast learning loss function to perform prototype contrast learning; dynamically updating the prototype, reallocating the samples to the nearest prototype, and iteratively optimizing; and adaptively determining a reasonable number of AML subtypes based on clustering quality indicators. The present invention comprehensively considers different types of biomarkers, provides a more comprehensive and accurate classification of AML subtypes, effectively identifies subtle differences between different subtypes, and improves classification accuracy. By adaptively determining the number of AML subtypes, the limitation of the fixed number of subtypes in traditional methods is avoided, so that the model automatically adjusts according to the actual structure of the data, thereby improving the flexibility and robustness of subtype classification.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer subtype classification, and in particular to an acute myeloid leukemia subtype classification method based on prototype contrast learning. Background Art

[0002] Acute myeloid leukemia (AML) is a hematologic malignancy that originates in the bone marrow and exhibits significant variability in clinical presentation, treatment response, and prognosis. AML subtyping is not only crucial for understanding the biological mechanisms of AML but also provides a key basis for personalized clinical treatment and prognosis prediction. AML is highly heterogeneous, with distinct subtypes exhibiting distinct molecular features, chromosomal abnormalities, genetic mutations, and immune phenotypes, making its diagnosis and treatment extremely challenging. Therefore, accurate subtyping is crucial for improving patient survival and developing more precise treatment plans. In recent years, with the rapid advancement of molecular biology and high-throughput omics technologies, researchers have been able to identify multiple molecular subtypes of AML. These subtypes are not only closely associated with clinical presentation but may also determine a patient's response to and resistance to various treatments. For example, some AML subtypes exhibit specific mutations in genes such as FLT3, NPM1, and IDH1 / 2, which have important clinical prognostic value. Furthermore, the study of AML subtypes has helped identify new therapeutic targets, laying the foundation for the development of targeted therapies.

[0003] With the rapid development of multi-omics technologies, the multimodal fusion of gene mutation, gene expression, and clinical data has provided a new research direction for AML subtyping. By integrating this multi-level information, the molecular mechanisms of AML can be more comprehensively revealed, improving the accuracy and clinical application value of subtyping classification. The fusion of multi-omics data not only reduces the limitations of single-omics data but also reveals potential correlations between different levels of data, improving the accuracy and reliability of AML subtyping. With the continuous advancement of computational methods, especially the development of machine learning and deep learning technologies, the multimodal fusion of gene mutation, gene expression, and clinical data will play an increasingly important role in AML subtyping.

[0004] Traditional methods for subtyping acute myeloid leukemia (AML) rely primarily on clinical manifestations, cytomorphology, immunophenotyping, and karyotyping. While these methods provide a basis for AML subtyping to a certain extent, they still have several limitations. First, cytomorphology and immunophenotyping rely on the physician's experience and are subject to significant subjectivity, resulting in inconsistent subtyping results between different laboratories and physicians. Second, while karyotyping can reveal certain chromosomal abnormalities, its resolution is low, preventing it from capturing subtle gene mutations and molecular variations, making it difficult to fully reflect the molecular heterogeneity of AML. Existing AML subtyping methods are mostly based on single-omics data and lack comprehensive analysis of diverse omics data. The complex relationships between multi-omics information, such as gene mutations, gene expression, and clinical data, have not been fully explored and utilized, resulting in significant limitations in the accuracy of existing subtyping methods. Summary of the Invention

[0005] The purpose of the present invention is to provide an acute myeloid leukemia subtyping method based on prototype comparative learning in order to address the deficiencies of the existing technology.

[0006] The object of the present invention is achieved through the following technical solutions: a method for subtyping acute myeloid leukemia based on prototype contrast learning, comprising:

[0007] First, we constructed an AML patient dataset and extracted features from the AML patient data. Each set of AML patient data included gene expression data, gene mutation data, and clinical information.

[0008] Secondly, the gene mutation features, gene expression features, and clinical features of each group of AML patient data were fused, and the fused features were used as samples;

[0009] Then, prototype initialization is performed: the samples are roughly clustered using clustering methods to obtain the initial prototype;

[0010] Then, a prototype contrastive learning loss function is constructed to perform prototype contrastive learning;

[0011] Then, the prototype is dynamically updated, the samples are redistributed to the nearest prototype, and the optimization is iteratively performed;

[0012] Finally, adaptive subtype discovery adaptively determines the number of acute myeloid leukemia (AML) subtypes based on clustering quality metrics.

[0013] Furthermore, encoders are constructed and pre-trained for three modalities: gene expression data, gene mutation data, and clinical information, respectively, to extract good representations of the three modalities, including:

[0014] An encoder is constructed for gene expression data, using log2(TPM+1) normalization and screening variant genes as core features. The encoder uses a Transformer encoder network with a self-attention mechanism to capture nonlinear interactions between genes.

[0015] To construct an encoder for gene mutation data, we first construct a binary matrix, focusing on incorporating highly clinically relevant mutations such as FLT3-ITD, CEBPA, and IDH1 / 2. This encoder structure uses a sparse self-attention mechanism.

[0016] An encoder was constructed for clinical data. First, continuous variables were Z-score standardized and categorical variables were one-hot encoded. The encoder structure consisted of a three-layer deep autoencoder, including a linear layer, a batch normalization layer, and an activation function.

[0017] Furthermore, gene mutations, gene expression characteristics and clinical features are integrated, including:

[0018] The extracted features of different modalities are constrained to align features, and a feature alignment loss function is constructed to minimize the semantic difference between the feature representations of different modalities based on the cosine distance:

[0019]

[0020] in, and They are gene mutation and gene expression characteristics, respectively.

[0021] Furthermore, the clustering method is used to roughly cluster the samples to obtain the initial prototype, including:

[0022] K-means clustering method is used to cluster samples Perform rough clustering to obtain initial pseudo labels , calculate the prototype vector of each cluster as the initial prototype :

[0023]

[0024] in, is the sample set of the current category k, is the center vector of category k.

[0025] Furthermore, the prototype contrast loss function for:

[0026]

[0027] Among them, sim(·) is the cosine similarity, is the i-th sample, N is the number of samples; is the temperature parameter that controls the sensitivity of contrast loss; For prototype; is the center vector of category k.

[0028] Furthermore, the dynamic updating of the prototype redistributes the samples to the nearest prototype, and the iterative optimization is specifically as follows:

[0029] First, calculate the similarity between each sample and all prototypes:

[0030]

[0031] Then, the pseudo labels are reassigned to the samples Belong to the nearest prototype :

[0032]

[0033] Then, update the prototype:

[0034]

[0035] in, For the updated prototype, Control smooth updates to avoid over-reliance on the current sample, so that the prototype remains stable during training;

[0036] In addition, in order to ensure that the pseudo labels of the samples remain stable during the iteration process, the pseudo label consistency loss is introduced and the cross entropy loss function is used. :

[0037]

[0038] in, is the soft pseudo label of sample i in category k, N is the number of samples, K is the number of categories, is the predicted probability that sample i belongs to category k, calculated by Softmax:

[0039]

[0040] in, For samples and prototype The cosine similarity between:

[0041]

[0042] Furthermore, the adaptive subtype discovery adaptively determines a reasonable number of acute myeloid leukemia (AML) subtypes based on clustering quality indicators, specifically:

[0043] First, evaluate the current clustering quality and calculate the Silhouette Score to assess the rationality of the classification. Adaptively determine the appropriate number of AML subtypes, K, and adjust the number of clusters. If the prototype quality is poor, merge / split similar prototypes; refine unstable prototypes; recalculate new pseudo-labels and continue training. The details are as follows:

[0044] First, the Silhouette Score is calculated to evaluate the quality of the current clustering. Then, a threshold is set to decide whether to adjust the number of prototypes: if the Silhouette Score is less than 0.5, the clustering quality is considered poor, and as an unstable prototype, it is necessary to increase the prototype and add a new prototype to the data. If the Silhouette Score is greater than 0.8, it is necessary to reduce redundant prototypes and calculate the two closest prototypes and merge them.

[0045] In addition, regularization terms are introduced, including prototype distribution regularization and prototype aggregation regularization;

[0046] Among them, prototype distribution regularization prevents all samples from being assigned to a single category:

[0047]

[0048] in, is the number of samples in category k, and N is the total number of samples. This is an entropy loss that encourages samples to be evenly distributed across different categories.

[0049] Among them, prototype convergence regularization ensures that the prototypes are sufficiently distinguishable from each other:

[0050]

[0051] Finally, the regularization loss for:

[0052] +

[0053] in, and is the weight coefficient of the loss function

[0054] In summary, the total loss function for:

[0055]

[0056] in, is the prototype contrast loss function, , and are learnable parameters.

[0057] The present invention also provides an acute myeloid leukemia subtype classification system based on prototype comparative learning, comprising:

[0058] Feature extraction module, which constructs an AML patient dataset and extracts features from the AML patient data. Each set of AML patient data includes gene expression data, gene mutation data, and clinical information.

[0059] The feature fusion module is used to fuse the gene mutation features, gene expression features, and clinical features of each group of AML patient data and use the fused features as samples;

[0060] Clustering module, used for prototype initialization: clustering method is used to roughly cluster samples to obtain the initial prototype;

[0061] Contrastive learning module, used to construct prototype contrastive learning loss function and perform prototype contrastive learning;

[0062] Optimization module, used to dynamically update the prototype, redistribute samples to the nearest prototype, and iteratively optimize;

[0063] Subtype number determination module for adaptive subtype discovery: Adaptively determine the number of acute myeloid leukemia (AML) subtypes based on clustering quality indicators.

[0064] The present invention also provides an acute myeloid leukemia subtype classification device based on prototype contrast learning, comprising one or more processors for implementing the above-mentioned acute myeloid leukemia subtype classification method based on prototype contrast learning.

[0065] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-mentioned acute myeloid leukemia subtyping method based on prototype comparative learning.

[0066] Compared with the prior art, the beneficial effects of the present invention are: First, multimodal data fusion can make full use of three different modal data types: gene expression data, gene mutation data, and clinical information, to capture the complex biological characteristics of AML from multiple levels. Traditional single modality analysis often cannot fully reveal the heterogeneity of AML, but this method can effectively integrate information from different data sources by constructing a multimodal fusion module, avoiding data redundancy and information loss, thereby improving the accuracy of classification. Secondly, prototype contrast learning can roughly cluster samples and initialize prototypes, and then continuously optimize the quality of prototypes through contrast learning. This process provides reasonable inspiration for the initial prototype with the help of clustering methods, and then dynamically updates the samples through the prototype contrast learning loss function to ensure that each sample can eventually be accurately assigned to the most appropriate subtype. The dynamic update of the prototype and the contrast learning mechanism can gradually optimize the model and reduce the risk of misclassification and overfitting. Third, adaptive subtype discovery adaptively determines the number of AML subtypes through the evaluation of clustering quality indicators. Compared to the traditional method of manually setting a fixed number of subtypes, this adaptive strategy can flexibly adjust based on the actual data distribution, thus avoiding the bias caused by manually setting the number of subtypes and ensuring the accuracy and reliability of subtype classification. Overall, this subtyping method based on prototype contrastive learning can explore the underlying patterns of AML from multi-angle and multimodal data, and through an efficient adaptive optimization strategy, provide more accurate and clinically valuable subtype classification results. This not only provides strong support for personalized treatment of AML, but also lays a solid foundation for subsequent clinical research and precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0068] Figure 1 A schematic flow chart of a method for subtyping acute myeloid leukemia based on prototype comparative learning provided by an embodiment of the present invention;

[0069] Figure 2 Structural diagram of different modal encoders provided by embodiments of the present invention;

[0070] Figure 3 The overall framework diagram of the model provided by the embodiment of the present invention;

[0071] Figure 4 A prototype subtyping flow chart provided for an embodiment of the present invention;

[0072] Figure 5 A hardware structure diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The present invention will be described in detail below with reference to the accompanying drawings. Unless there is any conflict, the features of the following embodiments and implementations may be combined with each other.

[0074] The present invention provides a method for subtyping acute myeloid leukemia based on prototype contrast learning, such as Figure 1 As shown, the following steps are included:

[0075] (1) First, encoders are constructed and pre-trained for the three modalities of gene expression data, gene mutation data, and clinical information to extract good representations of the three modalities, such as Figure 2 As shown, including:

[0076] Building an encoder for gene expression data , log2(TPM+1) normalization was used, and highly variable genes were selected as core features , encoder A Transformer encoder network with a self-attention mechanism is used to capture nonlinear interactions between genes. The encoder consists of a multi-head attention layer, a first-layer normalization layer, a feedforward network, and a second-layer normalization layer, which are connected in sequence. The first-layer normalization input is the residual combination of the original input and the multi-head attention output, and the second-layer normalization input is the residual combination of the first-layer normalization output and the feedforward network output.

[0077] For gene mutation data Building the encoder First, a binary matrix (presence / absence of mutations) is constructed, focusing on incorporating highly clinically relevant mutations such as FLT3-ITD, CEBPA, and IDH1 / 2. The encoder structure mainly adopts a sparse self-attention mechanism; the encoder includes a sparse embedding layer, a first-layer normalization, a self-attention layer, a second-layer normalization, a feedforward network, and a third-layer normalization layer, which are connected in sequence. Among them, the input of the second-layer normalization is the residual combination of the first-layer normalization output and the self-attention output, and the input of the third-layer normalization is the residual combination of the second-layer normalization output and the feedforward network output.

[0078] Regarding clinical data Building the encoder First, Z-score standardization is performed on continuous variables, and one-hot encoding is performed on categorical variables. The structure mainly includes 3 layers of deep autoencoders, including linear layers, batch normalization layers and activation functions; the encoder includes the first linear layer, the first activation layer (such as ReLU activation function), the first batch normalization, the second linear layer, the second activation layer (such as ReLU activation function), the second batch normalization and the third linear layer.

[0079] Considering that existing AML datasets usually contain fewer AML patient samples, pre-training can help obtain more effective feature representations. Therefore, pre-training is performed on the above encoders. The pre-training task is to construct an encoder for each modality data to reconstruct the data. The modality encoder mainly includes multiple linear layers and activation functions. Finally, the features of different omics data are obtained through the pre-trained encoders. 、 and .

[0080] (2) Secondly, a multimodal fusion module is designed and constructed to effectively fuse three modal information, including:

[0081] The multimodal fusion module mainly uses a cross-modal attention mechanism to fuse information from different modalities and enhance inter-modal interaction. Specifically, the multimodal fusion module includes a feature projection layer, multi-head cross-attention, residual normalization, and weighted splicing. To achieve better fusion effects, the extracted features of different modalities are first subjected to feature alignment constraints. A feature alignment loss function is constructed to minimize the feature semantic differences between the feature representations of different modalities based on the cosine distance:

[0082]

[0083] in, and They are gene mutations, gene expression characteristics and clinical characteristics.

[0084] (3) Then, the prototype is initialized and the samples are roughly clustered using the clustering method to obtain the initial prototype, including:

[0085] Use K-means clustering method to roughly cluster the samples and obtain the initial pseudo labels , calculate the prototype vector of each cluster as the initial prototype :

[0086]

[0087] in, is the sample set of the current category k, is the center vector of category k, is the fused feature, i.e., the sample.

[0088] (4) Then, construct the prototype contrast loss function and perform prototype contrast learning, such as Figure 3 As shown, including:

[0089] The purpose of constructing the prototype contrast loss function is to make the sample close to its prototype , away from other categories of prototypes:

[0090]

[0091] in, Is the pseudo label category corresponding to sample i The prototype vector of is the center vector of all samples of category k. The former is the prototype of the category corresponding to the sample, and the latter is the center of the global category; sim(·) is the cosine similarity, is a temperature parameter that controls the sensitivity to contrast loss.

[0092] (5) Subsequently, the prototype is dynamically updated and the samples are redistributed to the nearest prototype. The iterative optimization is as follows:

[0093] First, calculate the similarity between each sample and all prototypes:

[0094]

[0095] Then, the pseudo labels are reassigned to the samples Belong to the nearest prototype :

[0096]

[0097] Then, update the prototype:

[0098]

[0099] in, Control smooth updates to avoid over-reliance on the current sample, so that the prototype remains stable during training and prevents model instability or excessive fluctuations in the prototype vector.

[0100] In addition, in order to ensure that the pseudo labels of the samples remain stable during the iteration process and avoid drastic changes, the pseudo label consistency loss is introduced and the cross entropy loss function is used:

[0101]

[0102] in, is the soft pseudo label of sample i in category k, is the predicted probability that sample i belongs to category k, usually calculated by Softmax:

[0103]

[0104] in, For samples and prototype The cosine similarity between:

[0105]

[0106] (6) Finally, adaptive subtype discovery, based on clustering quality indicators, adaptively determines the number of AML subtypes, thereby constructing a prototype library, specifically:

[0107] First, the current clustering quality is evaluated, the Silhouette Score is calculated to assess the rationality of the classification, and the reasonable number of AML subtypes K is adaptively determined, that is, the number of clusters is adjusted. If the prototype quality is poor, it can be merged / split: similar prototypes are merged; unstable prototypes are refined; new pseudo-labels are recalculated, and training continues.

[0108] In addition, regularization terms are introduced to prevent model collapse, such as all samples being assigned to the same category or features being overly concentrated. Among them, prototype distribution regularization prevents all samples from being assigned to a single category:

[0109]

[0110] in, is the number of samples in category k, and N is the total number of samples. This is an entropy loss that encourages samples to be evenly distributed across different categories.

[0111] Among them, prototype convergence regularization ensures that the prototypes are sufficiently distinguishable from each other:

[0112]

[0113] Finally, the regularization loss is:

[0114] +

[0115] in, and is the weight coefficient of the loss function, which is set as a learnable hyperparameter and optimized during model training.

[0116] To sum up, the total loss function is:

[0117]

[0118] in, , and are learnable parameters.

[0119] Finally, the final subtype is obtained after model training and optimization through the loss function, such as Figure 4 As shown in the figure, when there is a new patient sample, the fusion features of the new patient sample are calculated through the trained model, and then the similarity is calculated with the current prototype (the prototype in the prototype library) respectively, and the subtype prototype with the highest similarity is selected as the subtype of the new patient sample.

[0120] The present invention also provides an acute myeloid leukemia subtype classification system based on prototype comparative learning, comprising:

[0121] Feature extraction module, which constructs an AML patient dataset and extracts features from the AML patient data. Each set of AML patient data includes gene expression data, gene mutation data, and clinical information.

[0122] The feature fusion module is used to fuse the gene mutation features, gene expression features, and clinical features of each group of AML patient data and use the fused features as samples;

[0123] Clustering module, used for prototype initialization: clustering method is used to roughly cluster samples to obtain the initial prototype;

[0124] Contrastive learning module, used to construct prototype contrastive learning loss function and perform prototype contrastive learning;

[0125] Optimization module, used to dynamically update the prototype, redistribute samples to the nearest prototype, and iteratively optimize;

[0126] Subtype number determination module for adaptive subtype discovery: Adaptively determine the number of acute myeloid leukemia (AML) subtypes based on clustering quality indicators.

[0127] It should be noted that the system embodiment shown in this embodiment matches the content of the above method embodiment. You can refer to the content of the above method embodiment and will not repeat it here.

[0128] Corresponding to the aforementioned embodiment of a method for subtyping acute myeloid leukemia based on prototype contrast learning, the present invention also provides an embodiment of a device for subtyping acute myeloid leukemia based on prototype contrast learning.

[0129] See also Figure 5 An embodiment of the present invention provides an acute myeloid leukemia subtyping device based on prototype contrast learning, which includes one or more processors for implementing an acute myeloid leukemia subtyping method based on prototype contrast learning in the above embodiment.

[0130] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0131] An embodiment of the acute myeloid leukemia subtyping device based on prototype contrast learning of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 5 As shown in the figure, it is a hardware structure diagram of any device with data processing capability where the acute myeloid leukemia subtype classification device based on prototype contrast learning of the present invention is located. Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0132] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0133] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0134] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for subtyping acute myeloid leukemia based on prototype comparative learning in the above embodiment is implemented.

[0135] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium may also be any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0136] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A method for subtyping acute myeloid leukemia based on prototype contrast learning, characterized in that: include: First, we constructed an AML patient dataset and extracted features from it. Each set of AML patient data included gene expression data, gene mutation data, and clinical information. We also constructed an encoder for the gene expression data, using a Transformer encoder network with a self-attention mechanism. We also constructed an encoder for the gene mutation data, using a sparse self-attention mechanism. Secondly, the gene mutation features, gene expression features, and clinical features of each group of AML patient data were fused, and the fused features were used as samples; Then, prototype initialization is performed: the samples are roughly clustered using clustering methods to obtain the initial prototype; Then, a prototype contrastive learning loss function is constructed to perform prototype contrastive learning; Then, the prototype is dynamically updated, the samples are redistributed to the nearest prototype, and the optimization is iteratively performed; Finally, adaptive subtype discovery adaptively determines the number of acute myeloid leukemia (AML) subtypes based on clustering quality metrics.

2. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 1, characterized in that: Encoders are constructed and pre-trained for three modalities of data: gene expression data, gene mutation data, and clinical information, respectively, to extract good representations of the three modalities, including: An encoder was constructed for gene expression data, using log2(TPM+1) normalization and selecting variant genes as core features to capture nonlinear interactions between genes. To construct an encoder for gene mutation data, we first constructed a binary matrix, focusing on including FLT3-ITD, CEBPA, and IDH1 / 2 mutations with high clinical relevance; An encoder was constructed for clinical data. First, continuous variables were Z-score standardized and categorical variables were one-hot encoded. The encoder structure consisted of a three-layer deep autoencoder, including a linear layer, a batch normalization layer, and an activation function.

3. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 1, characterized in that: Integration of gene mutations, gene expression signatures, and clinical features, including: The extracted features of different modalities are constrained to feature alignment, and a feature alignment loss function is constructed to minimize the feature semantic differences between the feature representations of different modalities based on the cosine distance: ； in, and They are gene mutation and gene expression characteristics, respectively.

4. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 1, characterized in that: The clustering method is used to roughly cluster the samples to obtain the initial prototype, including: K-means clustering method is used to cluster samples Perform rough clustering to obtain initial pseudo labels , calculate the prototype vector of each cluster as the initial prototype : ； in, is the sample set of the current category k, is the center vector of category k.

5. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 1, characterized in that: The prototype contrast loss function for: ； Among them, sim(·) is the cosine similarity, is the i-th sample, N is the number of samples; is the temperature parameter that controls the sensitivity of contrast loss; For prototype; is the center vector of category k.

6. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 3, characterized in that: The dynamic updating of prototypes redistributes samples to the nearest prototypes, and the iterative optimization is specifically as follows: First, calculate the similarity between each sample and all prototypes: ； Then, the pseudo labels are reassigned to the samples Belong to the nearest prototype : ； Then, update the prototype: ； in, For the updated prototype, Control smooth updates to avoid over-reliance on the current sample, so that the prototype remains stable during training; In addition, in order to ensure that the pseudo labels of the samples remain stable during the iteration process, the pseudo label consistency loss is introduced and the cross entropy loss function is used. : ； in, is the soft pseudo label of sample i in category k, N is the number of samples, K is the number of categories, The predicted probability that sample i belongs to category k is calculated by Softmax: ； in, For samples and prototype The cosine similarity between: 。 7. The method for acute myeloid leukemia subtyping based on prototype contrast learning according to claim 6, characterized in that: The adaptive subtype discovery adaptively determines the appropriate number of acute myeloid leukemia (AML) subtypes based on clustering quality indicators, specifically: First, evaluate the current clustering quality and calculate the Silhouette Score to assess the rationality of the classification. Adaptively determine the appropriate number of AML subtypes, K, and adjust the number of clusters. If the prototype quality is poor, merge / split similar prototypes; refine unstable prototypes; recalculate new pseudo-labels and continue training. The details are as follows: First, calculate the Silhouette Score to evaluate the quality of the current clustering. Then, set a threshold to decide whether to adjust the number of prototypes: if the Silhouette Score is less than 0.5, the clustering quality is considered poor, and as an unstable prototype, it is necessary to increase the prototype and add a new prototype to the data. If the Silhouette Score is greater than 0.8, it is necessary to reduce redundant prototypes and calculate the two closest prototypes and merge them. In addition, regularization terms are introduced, including prototype distribution regularization and prototype aggregation regularization; Among them, prototype distribution regularization prevents all samples from being assigned to a single category: ； in, is the number of samples of category k, N is the total number of samples; Among them, prototype convergence regularization ensures that the prototypes are sufficiently distinguishable from each other: ； Finally, the regularization loss for: + ； in, and is the weight coefficient of the loss function; In summary, the total loss function for: ； in, is the prototype contrast loss function, , and are learnable parameters.

8. An acute myeloid leukemia subtyping system based on prototype contrastive learning, characterized in that: include: The feature extraction module constructs an AML patient dataset and extracts features from the AML patient data. Each set of AML patient data includes gene expression data, gene mutation data, and clinical information. An encoder is constructed for the gene expression data, using a Transformer encoder network with a self-attention mechanism. An encoder is constructed for the gene mutation data, using a sparse self-attention mechanism. The feature fusion module is used to fuse the gene mutation features, gene expression features, and clinical features of each group of AML patient data and use the fused features as samples; Clustering module, used for prototype initialization: clustering method is used to roughly cluster samples to obtain the initial prototype; Contrastive learning module, used to construct prototype contrastive learning loss function and perform prototype contrastive learning; Optimization module, used to dynamically update the prototype, redistribute samples to the nearest prototype, and iteratively optimize; Subtype number determination module for adaptive subtype discovery: Adaptively determine the number of acute myeloid leukemia (AML) subtypes based on clustering quality indicators.

9. An acute myeloid leukemia subtype classification device based on prototype contrast learning, characterized in that: The system comprises one or more processors for implementing the acute myeloid leukemia subtyping method based on prototype contrast learning according to any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, it is used to implement the acute myeloid leukemia subtyping method based on prototype contrast learning according to any one of claims 1 to 7.

Images