Single-cell data representation learning method, device and medium for maintaining biological structure topology

Abstract

The application provides a single-cell data representation learning method, device and medium for maintaining biological structure topology, and relates to the technical field of cell data representation learning. The method comprises the following steps: obtaining single-cell RNA sequencing data and preprocessing; constructing a unified representation learning framework, converting a set of input expression vectors into a low-dimensional hidden vector through a linear mapping layer, and obtaining a shared representation through encoding, and then inputting the shared representation into a classification branch and a contrast learning branch respectively to obtain a predicted category probability and a normalized embedding vector; determining a training target according to a preset structure induction bias, and jointly optimizing the framework to obtain a trained model; and performing structure maintenance evaluation on a model embedding space to obtain evaluation results of classification performance, global topology maintenance and local structure continuity. The application can adapt to different cell state organization modes according to different structure induction bias types, and can maintain the global lineage topology relationship and the local continuous relationship between cells while taking into account the classification performance.

Description

Technical Field

[0001] This invention relates to the field of cell data representation learning technology, and more particularly to a single-cell data representation learning method, apparatus, device, and medium that preserves the topology of biological structures. Background Technology

[0002] The hematopoietic system, as a classic biological system for studying cell fate determination, differentiation trajectory, and state transitions, provides an important research background for elucidating the continuous evolution of cell states. Cells at different developmental stages undergo continuous state transitions within a lineage, collectively determining the maintenance of hematopoietic homeostasis, disease development, and treatment response. Therefore, accurately describing the continuous developmental relationships between cells is crucial for understanding hematopoietic regulatory mechanisms and disease heterogeneity. With the development of single-cell transcriptomics technology, this continuous organizational pattern has been further revealed, meaning that an ideal cell representation should not only distinguish different cell populations but also preserve the biological organizational structures inherent in the process of cell state evolution.

[0003] The development of single-cell RNA sequencing (scRNA-seq) technology has enabled researchers to resolve cellular heterogeneity at single-cell resolution and has driven the development of computational methods for cell type identification, cluster analysis, and cell state inference. However, current single-cell analysis still mainly relies on predefined discrete cell labels, mapping continuously changing cell states to a limited number of categories through classification or clustering, which makes it difficult to fully preserve the continuous changes and potential lineage relationships during cell state transitions. Especially in diseases such as acute myeloid leukemia (AML) and mixed phenotype acute leukemia (MPAL), abnormal cells often exhibit cross-differentiation stage and cross-lineage characteristics, making the cell state space present a more complex continuity and heterogeneity. Therefore, how to learn cell representations that can preserve the continuous organizational patterns and potential biological structural relationships of cell states has become a major challenge in single-cell representation learning.

[0004] In recent years, representation learning methods have gradually become important tools for analyzing the latent structure of single-cell data. Unlike traditional dimensionality reduction methods, deep representation learning can learn the latent cell state space from high-dimensional representation data through nonlinear mapping and has been widely applied to tasks such as cell type identification, data integration, and state inference. Furthermore, methods such as self-supervised learning, contrastive learning, and single-cell basic models have provided new directions for constructing more expressive cell representations by introducing different forms of structural constraints. However, different optimization objectives may impart different structural inductive biases to the model, thus affecting the cell representation space it learns: while strengthening the separation between categories can improve classification performance, it may simultaneously disrupt the original topological relationships in the continuous evolution of cell states. Currently, whether the representation spaces induced by different training objectives can maintain the continuous relationships and phylogenetic topology of cell states, and whether this influence depends on cell organization patterns, still lacks systematic research.

[0005] With the development of single-cell representation learning methods, evaluating the cell representations learned by the models has become a crucial issue. Existing research mainly uses task metrics such as cell type classification performance, clustering quality, and expression reconstruction error to evaluate model capabilities. However, these metrics primarily reflect the model's ability to complete specific tasks and are insufficient to measure whether the embedding space preserves the potential biological structural relationships between cells. For continuous cellular systems, optimizing classification boundaries and preserving cell state topology are not necessarily consistent. Therefore, relying solely on task performance to evaluate representation quality may not accurately reflect the model's ability to learn biological organizational patterns.

[0006] Therefore, existing single-cell representation learning methods suffer from at least the following technical problems: First, different representation learning training objectives introduce different levels of structural induction bias, while existing methods lack a systematic consideration of the adaptability between structural induction bias and intrinsic cell organization patterns under different biological topological scenarios. This results in the inability to flexibly adjust the optimization objective according to different cell state organization patterns within the same model, making it difficult to obtain the optimal representation space that preserves biological structure in different topological scenarios. Second, the evaluation system of existing models mainly relies on task-oriented indicators such as classification accuracy, lacking a direct measurement and evaluation mechanism for whether the embedding space preserves the global developmental relationships and local continuity relationships between cell states. Furthermore, the inconsistency between classification performance and biological structure preservation ability has not been effectively resolved. Third, the representation space learned by existing methods is often limited by a single training objective. When faced with cell states that simultaneously possess different organizational characteristics such as hierarchical differentiation, cross-lineage mixing, or discrete populations, a single optimization objective cannot take into account biological structural constraints at different scales, resulting in the learned embedding space failing to simultaneously maintain the overall lineage topology and local continuity relationships of cell states. These problems restrict the structure preservation ability and generalization performance of single-cell representation learning methods under different biological scenarios. Summary of the Invention

[0007] This invention provides a single-cell data representation learning method, apparatus, device, and medium for preserving biological structural topology. It aims to address the technical problems in existing single-cell representation learning, such as the lack of a systematic consideration of the adaptation relationship between different cell state organization patterns and structural inductive biases, and the over-reliance on classification performance in evaluation systems, which makes it difficult to effectively maintain the global phylogenetic topology and local continuity of cell states in the embedding space. This invention constructs a unified representation learning framework including a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. It sets corresponding training objectives based on three different levels of structural inductive biases: weak structural bias, instance separation bias, and phylogenetic structure bias. The unified representation learning framework is jointly optimized using a preset phylogenetic distance matrix, enabling the trained single-cell representation learning model to maintain the corresponding biological topological structure according to different cell state organization patterns. Simultaneously, by evaluating the structure preservation of the embedded space of the trained model, the structure preservation ability of the model is comprehensively evaluated from three dimensions: classification performance, global topology preservation ability, and local structural continuity. This provides a basis for the selection and evaluation of representation learning strategies in different biological topological scenarios.

[0008] In a first aspect, the present invention provides a single-cell data representation learning method that preserves the topology of biological structures, the method comprising: Single-cell RNA sequencing data were acquired and preprocessed to obtain a standardized set of cell expression vectors. A unified representation learning framework is constructed, comprising a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. Responding to the input standardized cell expression vector set, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vector to obtain a cell shared representation. This cell shared representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function, and the contrastive learning branch maps the cell shared representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector. The corresponding training objective is determined according to the preset structural induction bias type, and the unified representation learning framework is jointly optimized based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix to obtain the trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree. The embedding space of the trained single-cell representation learning model is evaluated for structure preservation, and the results of classification performance evaluation, global topology preservation evaluation, and local structural continuity evaluation are obtained.

[0009] In one possible design, the single-cell RNA sequencing data is preprocessed to obtain a standardized set of cell expression vectors, including: Missing genes in different samples were filled with zero values. After unifying the gene space, the original expression matrix was logarithmically transformed to base 2 to obtain a unified cell expression matrix. Based on the gene expression variance ranking in the training set samples, the top-ranked hypervariable genes are selected as the model input features. The input features of the model are Z-score standardized, where the mean and variance parameters are calculated only from the training set and applied to the test set to obtain a standardized set of cell expression vectors.

[0010] In one possible design, the unified representation learning framework, in response to the input standardized set of cell expression vectors, obtains normalized embedding vectors in the following ways: The standardized cell expression vector is converted into a low-dimensional hidden vector using the following formula. : (1) in, x This represents the expression vector of highly variable genes. W p and b p The hidden dimension is set to 128 as a learnable parameter; The low-dimensional hidden vector The cell-shared representation is obtained by taking an encoder consisting of stacked Transformer encoders as input, using the following formula. h : (2) Here, Transformer represents a neural network encoder based on a self-attention mechanism, used for feature extraction and context modeling of input sequences; The cell-shared representation h The input is fed into the classification branch, and the predicted class probability is obtained through the Softmax classifier using the following formula: (3) Where Softmax is a normalized exponential function that transforms a real vector into a probability distribution. W c For classification weight parameters, To predict class probabilities; The cell-shared representation h The input is fed into the contrastive learning branch, where the cell-shared representation is mapped to the contrastive space and normalized using the projection head. The normalized embedding vector is obtained using the following formula. z : (4) Where Proj represents the projection head network consisting of fully connected layers. This represents the L2 norm.

[0011] In one possible design, the preset structural inductive bias type is a weak structural bias, and the training objective is to use only the cross-entropy loss function for supervision; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix includes: Based on the predicted category probability and cell real category labels y The cross-entropy loss value is calculated using the following formula: (5) in, N For the sample size, C For the number of lineage categories, i For sample index, c For category indexing, For the first i The sample at the th c Real labels in each category For the first i The sample at the th c Predicted class probabilities for each category L CE This represents the cross-entropy loss value. The parameters of the unified representation learning framework are updated with the goal of minimizing the cross-entropy loss.

[0012] In one possible design, the preset structural inductive bias type is an instance separation bias, and the training objective is a combination of cross-entropy loss and instance-level contrastive loss; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix includes: Calculate the cross-entropy loss based on the predicted class probability and the true cell class label; According to the normalized embedding vector z ， The InfoNCE contrast loss is calculated using the following formula: (6) in,s ij For the first i The sample and the first j Cosine similarity between samples s ik For the first i The sample and the first k Cosine similarity between samples z i and z j The first i The first sample and the first j The normalized embedding vector of each sample. τ Temperature is a parameter used to control the smoothness of the similarity distribution. k For sample index, N The total number of samples, L infoNCE Compare the loss values ​​for InfoNCE; The cross-entropy loss and the InfoNCE contrastive loss are weighted and summed to obtain a joint loss value. The parameters of the unified representation learning framework are updated with minimizing the joint loss value as the optimization objective.

[0013] In one possible design, the preset structural inductive bias type is a phylogenetic structure bias, and the training objective is a combination of cross-entropy loss and phylogenetic guided contrastive loss; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probabilities, the normalized embedding vectors, and the preset phylogenetic distance matrix includes: The distances between cell types are converted into a weight matrix based on the lineage distance matrix, where the weight values ​​are calculated using the following formula: (7) in, d ij The phylogenetic distance matrix is ​​the th i The cell and the first j Lineage distance between cells σ To control the hyperparameters of the weight decay rate, w ij For the first i The cell and the first j Weight values ​​between individual cells e It is a natural constant; Calculate the cross-entropy loss based on the predicted class probability and the true cell class label; According to the normalized embedding vector z The lineage-guided contrast loss is calculated using the weight matrix and the following formula: (8) in, For the first i The sample and the first j The cosine similarity of the embedding vectors of each sample is normalized in the contrast space. N The total number of samples, Lphy log eny Lineage-guided comparison loss value; The cross-entropy loss and the lineage-guided comparison loss are weighted and summed to obtain a joint loss value. The parameters of the unified representation learning framework are updated with minimizing the joint loss value as the optimization objective.

[0014] In one possible design, the method for evaluating the structure preservation of the embedding space of the trained single-cell representation learning model includes: The expression vector of the test cell is obtained and input into the trained single-cell representation learning model to obtain the normalized embedding vector of the test cell and the predicted test class probability. The classification accuracy is calculated using the following formula: (9) in, N For the number of test samples, i For the test sample index, y i For the first i The true class label of each test cell For the first i Predicted category labels for each test cell I The parentheses () are indicator functions; they take the value 1 if the condition inside the parentheses is true, and 0 otherwise. Accuracy For classification accuracy; Based on the normalized embedding vector of the test cells, the cosine distance between each pair of cells in the embedding space is calculated, and the embedding distance matrix is ​​obtained using the following formula: (10) in, z i and z j The first i The test cells and the first j Normalized embedding vectors of each test cell, For the first in the embedded space i The test cells and the first j Cosine distance between test cells; Spearman rank correlation analysis is performed between the embedding distance matrix and the preset phylogenetic distance matrix, and the Spearman rank correlation coefficient is calculated using the following formula: (11) Where Spearman represents the Spearman rank correlation coefficient calculation method. D embed For the embedding distance matrix, D phylogeny The preset phylogenetic distance matrix, , which is the Spearman rank correlation coefficient, used to measure the consistency between the order of intercellular distances in the embedding space and the order of biological lineage relationships; The embedding vector of each test cell is used to determine the embedding space of each test cell. k The nearest neighbor cells are counted, and the average phylogenetic distance is calculated using the following formula: (12) in, N k ( i ) is the first i The test cells with the closest cosine distance in the embedding space k A set of indexes for each neighbor. The index contains the number of elements in the set. d lineage ( i , j ) is the first i The test cells and the first j The preset lineage distance between test cells, where ALD is the average lineage distance consistency; Based on the classification accuracy, the Spearman rank correlation coefficient, and the average phylogenetic distance consistency, the embedding space structure preservation ability of the trained single-cell representation learning model is comprehensively evaluated from three dimensions: classification performance, global topology preservation ability, and local structural continuity.

[0015] In a second aspect, the present invention provides a single-cell data representation learning device that preserves the topology of biological structures, the device comprising: The data preprocessing module is configured to acquire single-cell RNA sequencing data and preprocess the single-cell RNA sequencing data to obtain a standardized cell expression vector set. The framework construction module is configured to construct a unified representation learning framework, which includes a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. Responding to the input standardized cell expression vector set, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vector to obtain a cell shared representation. This cell shared representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function, and the contrastive learning branch maps the cell shared representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector. The model training module is configured to determine the corresponding training objective based on a preset structural induction bias type, and to jointly optimize the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and a preset phylogenetic distance matrix to obtain a trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree. The structure preservation evaluation module is configured to perform structure preservation evaluation on the embedding space of the trained single-cell representation learning model to obtain classification performance evaluation results, global topology preservation evaluation results, and local structural continuity evaluation results.

[0016] Thirdly, embodiments of the present invention provide an electronic device, comprising: at least one processor and a memory; the memory storing computer execution instructions; the at least one processor executing the computer execution instructions stored in the memory, causing the at least one processor to perform the single-cell data representation learning method for maintaining biological structural topology as described in the first aspect and various possible designs of the first aspect.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the single-cell data representation learning method for maintaining biological structural topology as described in the first aspect and various possible designs of the first aspect.

[0018] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the single-cell data representation learning method for maintaining biological structural topology as described in the first aspect and various possible designs of the first aspect.

[0019] The single-cell data representation learning method, apparatus, device, and medium that preserve the topology of biological structures provided by this invention have at least the following beneficial effects: (1) This invention constructs a unified representation learning framework including a linear mapping layer, a Transformer encoder, a classification branch and a contrastive learning branch, to map high-dimensional single-cell transcriptome data to a low-dimensional embedding space. In the optimization process, a preset phylogenetic distance matrix is ​​introduced as a biological structure prior, so that the model can explicitly maintain the global developmental topological relationship and local continuity relationship between cell states during training, thus overcoming the defect of existing methods that destroy the continuous evolution structure of cell states due to excessive reliance on classification loss.

[0020] (2) By setting three different levels of training objectives—weak structure bias, instance separation bias, and lineage structure bias—the present invention enables the unified representation learning framework to flexibly adapt to the corresponding structure induction bias according to different cell state organization patterns. Under the same basic framework, it achieves compatible representation learning for various biological topological scenarios such as hierarchical continuous differentiation, cross-lineage mixing, and discrete cell populations, which can improve the applicability and structure preservation ability of the single-cell representation learning model in different biological scenarios.

[0021] (3) This invention comprehensively evaluates the embedding space of the trained single-cell representation learning model in three dimensions: classification accuracy, Spearman rank correlation coefficient and average phylogenetic distance consistency. It comprehensively characterizes the degree of preservation of biological organizational relationships of the model from three dimensions: classification performance, global topology preservation ability and local structural continuity. This overcomes the shortcomings of the existing evaluation system, which only relies on task performance indicators and cannot measure the ability to preserve biological structures in the embedding space. It provides a reliable basis for the selection and evaluation of representation learning strategies. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0023] Figure 1 A conceptual framework diagram of topology-bias compatibility for evaluating the ability to preserve biological structures in single-cell representation learning, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a single-cell data representation learning method that preserves the topology of biological structures, provided as an embodiment of the present invention; Figure 3This is a schematic diagram illustrating evidence of continuous transcriptional states in AML cells provided by an embodiment of the present invention. A represents the distribution of confidence scores for marker-based annotations in AML cells. The wide distribution of low confidence indicates significant overlap in transcriptional expression among predefined cell states. B represents the relationship between the confidence threshold and the proportion of unassigned cells; as the threshold tightening increases, cells with intermediate state characteristics are gradually excluded, revealing a continuous trade-off between the reliability of discrete annotations and the coverage of cell states. Figure 4 The figure shows a comparison of the methods introduced by SVM and MLP+Cross-Entropy as discrete classification targets in the embodiments of the present invention. Among them, A–C compare the performance of different modeling strategies on the AML dataset under continuous differentiation topology conditions, and evaluate them from three aspects: classification accuracy, global lineage topology preservation ability (Spearman ρ), and local lineage continuity preservation ability (average lineage distance, ALD). The deep learning-based models all use the same encoder architecture and are evaluated by leave-one-patient-out cross-validation. D represents the relationship between classification performance and topology preservation ability, and each point represents a modeling strategy. The horizontal axis represents classification accuracy and the vertical axis represents Spearman ρ. The distribution differences of different methods in two-dimensional space indicate that higher classification performance does not necessarily correspond to better preservation of biological lineage structure. Figure 5 This is a schematic diagram of the structural inductive bias reshaping of AML cell state embedding topology provided in an embodiment of the present invention; wherein A–C respectively show the UMAP visualization results of the AML cell embedding space learned by CE-only, CE+InfoNCE and CE+PGCL, and different optimization targets induce different cell state organization patterns; D–F respectively show the spatial distribution of representative hematopoietic marker genes CD34, LYZ and NKG7 in the embedding space, revealing the consistency between the learned representation space and the known cell differentiation program; Figure 6The graph shows the evaluation results of the topology preservation ability under different cell organization patterns provided in the embodiments of the present invention. Among them, A–C evaluates the impact of different representation learning objectives on the cell representation space topology preservation ability on the MPAL dataset, which represents mixed lineage cell organization patterns; D–F is evaluated on the PBMC dataset, which represents discrete cell population organization patterns. Accuracy, Spearman ρ, and Average Lineage Distance (ALD) are used to evaluate the model's classification performance, global topology preservation ability, and local lineage continuity preservation ability, respectively. The comparison results under different topological scenarios show that the effectiveness of the structure inductive bias depends on the matching relationship between the intrinsic organization pattern of the cell state and the structural assumptions introduced by the model. Figure 7 This diagram illustrates the topological dependence of structural induction bias on cell representation organization patterns, as provided in this embodiment of the invention. A–C represent the UMAP visualization results of the MPAL cell embedding space learned by CE-only, CE+InfoNCE, and CE+PGCL, respectively. Under mixed lineage topology conditions, different optimization objectives induce different cell state organization patterns. D–F represent the UMAP visualization results of the PBMC embedding space under discrete cell topology conditions. This comparison shows that the effectiveness of structural constraints depends on the intrinsic organization patterns of cell states. Figure 8 This is a schematic diagram illustrating the applicability of the phylogenetic structure constraint provided in different patient contexts according to embodiments of the present invention; Figure 9 This is a structural diagram of a single-cell data representation learning device that preserves the topology of biological structures, provided as an embodiment of the present invention.

[0024] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0026] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision and disclosure of relevant data and information comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0027] It should be noted that in the embodiments of the present invention, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of the present invention. However, they do not mean that the inventor has used or necessarily used the solution.

[0028] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0029] Cell states typically change gradually along a continuous trajectory during development and disease evolution, rather than strictly corresponding to discrete cell categories. These continuously changing cell states form a cell state topology with hierarchical, branching, or overlapping features. Accurately preserving this topology is crucial for understanding cell fate determination, disease heterogeneity, and functional state transitions. However, existing single-cell representation learning methods primarily focus on optimization for classification tasks, and whether improvements in classification performance reflect the model's ability to preserve cell state topology remains insufficiently assessed. Therefore, this invention provides a single-cell data representation learning method that preserves the topology of biological structures.

[0030] Specifically, different hematopoietic mononuclear cell systems provide representative biological scenarios for studying the relationship between representation learning and the maintenance of biological structures. AML exhibits a continuous cell state with hierarchical differentiation characteristics; MPAL shows a clear cross-lineage continuous tissue pattern; and peripheral blood mononuclear cells (PBMCs) are mainly composed of mature immune cells, presenting a relatively discrete cell population structure. Therefore, these three types of data represent hierarchical continuity, cross-lineage continuity, and discrete tissue patterns, respectively, and can be used to compare the performance of different representation learning targets in different topological contexts.

[0031] This embodiment focuses on comparing three representative representation learning training objectives: CE-only (using only category supervision), CE+InfoNCE (introducing instance-level contrastive constraints), and CE+PGCL (combining hierarchy prior constraints). These represent three different levels of structural inductive bias: category discrimination bias, instance relation bias, and hierarchy structure bias, respectively. To verify the universality of the obtained patterns, the experiment further introduces the traditional machine learning classifier SVM and the MLP based on cross-entropy optimization as performance comparison baselines. Figure 1 This paper outlines the overall analysis framework of this embodiment, focusing on how the structural inductive bias introduced by different representation learning objectives affects the biological organization in the single-cell representation space, and establishes a unified evaluation framework for task performance and biological structure preservation ability.

[0032] To verify the applicability of different structural induction biases to various cellular state organization patterns, this embodiment conducts a systematic evaluation around three representative biological topologies: continuous differentiation, cross-lineage mixing, and discrete cell populations. It compares the differences in classification performance, global lineage topology preservation, local cellular state continuity, and embedded spatial organizational features among different learning strategies. Based on this, further cross-dataset validation, patient-level heterogeneity analysis, and parameter sensitivity and model capacity perturbation experiments are used to verify the robustness and universality of the obtained patterns. By integrating task performance and biological structure preservation metrics, this embodiment aims to reveal the matching relationship between the structural assumptions introduced by the model optimization objective and the actual cellular state organization patterns, and to demonstrate that relying solely on classification performance is insufficient to comprehensively evaluate the ability of single-cell representation learning models to characterize biological structures.

[0033] like Figure 1As shown, this conceptual framework demonstrates how the compatibility between the intrinsic topology of cells and the representation learning objective determines the organization of the learned embedding space. Different biological systems possess different intrinsic topological patterns, including hierarchical developmental topology in AML, cross-lineage overlapping topology in MPAL, and discrete population topology in PBMC. Different representation learning objectives introduce different structural inductive biases, including class-discriminative bias (CE-only), instance-contrast bias (CE+InfoNCE), and lineage-aware structure bias (CE+PGCL). The degree of compatibility between the biological topology and the inductive biases introduced by the model determines the structural organization of the learned embedding space. To characterize this relationship, the framework evaluates from two complementary perspectives: task-oriented evaluation, reflecting the model's predictive performance; and structure-oriented evaluation, reflecting the ability to preserve biological topology. The potential inconsistency between the two suggests that improved task performance does not necessarily imply enhanced biological tissue structure preservation. Overall, this framework provides a conceptual foundation for systematically analyzing how structural inductive biases affect biological structure preservation capabilities under different cellular topological scenarios.

[0034] exist Figure 1 Based on the conceptual framework shown, further integration Figure 2 As shown, the single-cell data representation learning method that preserves the topology of biological structures includes the following steps S10-S40 in its specific implementation.

[0035] S10: Acquire single-cell RNA sequencing data and preprocess the single-cell RNA sequencing data to obtain a standardized set of cell expression vectors.

[0036] In some exemplary embodiments, to systematically evaluate the structure preservation capabilities of different representation learning strategies under different cell state organization patterns, this embodiment selects publicly available single-cell RNA sequencing datasets with different cell state organization characteristics, covering three representative evaluation scenarios: continuous differentiation topology, cross-lineage mixed topology, and discrete cell population topology. Specifically, the AML dataset GSE116256 is used for the continuous differentiation topology scenario, where cell states exhibit hierarchical lineage evolution along the hematopoietic development process; the MPAL dataset GSE139369 is used for the cross-lineage mixed topology scenario, where cell states exhibit the coexistence of different lineage characteristics; and a subset of normal peripheral blood mononuclear cells (PBMCs) selected from the MPAL dataset is used for the discrete cell population topology scenario, where cell types mainly exhibit relatively independent cluster organization patterns. Basic information about the three datasets is summarized in Table 1.

[0037] Table 1 Data Summary

[0038] To ensure the comparability of experimental results across different topological scenarios, a unified data processing workflow was used for all three datasets. The sequencing data included samples from multiple patients to evaluate the stability of the representation learning strategy across individual cell states. Considering potential expression differences among patients, this embodiment employed a leave-one-patient-out (LOPO) cross-validation strategy for model training and evaluation. In each round of experiments, one patient sample was used as the test set, and the remaining patient samples were used for model training.

[0039] All samples were first subjected to a unified gene space matching. Missing genes between different samples were filled with 0 to construct a unified expression space. Subsequently, the original expression matrix was transformed using log2(x+1).

[0040] In each round of the LOPO experiment, the model input features are calculated solely based on the training set. First, the top 2000 highly variable genes (HVGs) are selected as model input features based on gene expression variance ranking. This step preserves gene features with high expression variability and reduces the impact of low-information genes on model training. Subsequently, the input features are Z-score standardized, where the mean and variance parameters are calculated solely from the training set and applied to the test set to avoid information leakage.

[0041] Cell tags were primarily based on the CellType annotation results provided in the original data. For cells that were missing or marked as Unknown, supplementary annotations were performed based on classic hematopoietic system marker genes, including CD34, MPO, CD14, CD3D, MS4A1, and NKG7. Specifically, the average expression level of the marker gene set corresponding to each cell type was first calculated and Z-score normalized to obtain a cell type matching score. Cells with low scores were screened based on confidence levels according to classification margins and marked as Unassigned.

[0042] Subsequently, based on known hematopoietic differentiation relationships, fine-grained cell types were mapped to coarse-grained lineage categories such as Progenitor, Myeloid, Lymphoid, Other, and Unassigned for subsequent structural modeling.

[0043] S20: Construct a unified representation learning framework, which includes a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. Responding to the input standardized set of cell expression vectors, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vectors to obtain a shared cell representation. This shared cell representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function, and the contrastive learning branch maps the shared cell representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector.

[0044] In some exemplary embodiments, to reduce the impact of model architecture differences on the comparison results, this embodiment constructs a unified representation learning framework and compares the differences in representation space induced by different training objectives under the same encoder structure.

[0045] For each cell, the input is a 2000-dimensional highly variable gene expression vector. First, the input features are transformed into a low-dimensional hidden space through a linear mapping layer: (1) in, Represents the input representation vector. W p and b p The hidden dimension is set to 128 as a learnable parameter.

[0046] Then, The input consists of an encoder composed of four stacked Transformer Encoder layers. Each layer contains a multi-head self-attention mechanism and a feedforward network, and uses dropout regularization.

[0047] The encoder output is represented as: (2) Here, Transformer represents a neural network encoder based on a self-attention mechanism, used for feature extraction and context modeling of input sequences.

[0048] Based on shared representations, the model includes two task branches: (3) Where Softmax is a normalized exponential function that transforms a real vector into a probability distribution. Indicates the predicted class probability. W c These are the classification weight parameters. Simultaneously, the contrastive learning branch maps the representations to the contrastive space using the contrastive learning projection head: (4) Where Proj represents the projection head network consisting of fully connected layers. This represents the L2 norm.

[0049] This branch maps the representation to a 64-dimensional contrast space using two fully connected projection heads and performs L2 normalization.

[0050] S30: Determine the corresponding training objective based on the preset structural induction bias type, and jointly optimize the unified representation learning framework based on the training objective, predicted class probability, normalized embedding vector, and preset phylogenetic distance matrix to obtain the trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree.

[0051] In some exemplary embodiments, three training objectives are designed within a unified encoder framework to compare the impact of different structural inductive biases on the representation space.

[0052] Weak structure bias (CE-only) uses only cross-entropy loss to supervise coarse-grained family classification prediction: (5) in, N For the sample size, C For the number of lineage categories, i For sample index, c For category indexing, For the first i The sample at the th c Real labels in each category For the first i The sample at the th c Predicted class probabilities for each category L CE This represents the cross-entropy loss value. This objective only optimizes class boundaries and does not explicitly model continuous topological relationships between cells.

[0053] Instance Separation Bias (CE+InfoNCE) adds standard InfoNCE contrastive loss to the cross-entropy loss: (6) in, s ij For the first i The sample and the first j Cosine similarity between samples s ik For the first i The sample and the first k Cosine similarity between samples zi and z j The first i The first sample and the first j The normalized embedding vector of each sample. τ Temperature is a parameter used to control the smoothness of the similarity distribution. k For sample index, N The total number of samples, L infoNCE This represents the contrastive loss value for InfoNCE. The objective is to optimize the sample representation space through instance-level contrastive constraints.

[0054] To further incorporate biological structural priors into the cell state transition process, this embodiment proposes Phylogeny-Guided Contrastive Learning (PGCL). First, a phylogenetic distance matrix is ​​constructed based on known hematopoietic differentiation levels.

[0055] For any two cell types, the distance is defined based on their topological relationship in the hematopoietic differentiation tree. Cell types in the same branch are assigned a smaller distance, while the distance between cell types across lineages is calculated based on the hierarchical relationship of their lowest common ancestor (LCA) in the differentiation tree

[38] . The resulting distance matrix is ​​normalized and used to define the lineage-aware contrast weights. Based on the lineage distance matrix, the distance between cell types is converted into a weight matrix: (7) in, d ij The phylogenetic distance matrix is ​​the th i The cell and the first j Lineage distance between cells σ To control the hyperparameters of the weight decay rate, w ij For the first i The cell and the first j Weight values ​​between individual cells e is a natural constant. Therefore, the lineage-guided contrast loss is defined as: (8) in, For the first i The sample and the first j The cosine similarity of the embedding vectors of each sample is normalized in the contrast space. N The total number of samples, Lphy log eny Genealogy-guided comparison loss value.

[0056] S40: Evaluate the structure preservation of the embedding space of the trained single-cell representation learning model to obtain classification performance evaluation results, global topology preservation evaluation results, and local structural continuity evaluation results.

[0057] Traditional evaluation of single-cell representation learning typically relies on task-oriented metrics such as classification accuracy. These metrics primarily reflect the model's ability to distinguish discrete cell categories, but they cannot determine whether the learned representation space preserves the potential biological organizational relationships between cell states. Therefore, in some exemplary embodiments, this embodiment establishes a multi-level representation space evaluation system to systematically evaluate the impact of different inductive biases on the organization of single-cell embedding space from three levels: classification performance, global topology preservation, and local structural continuity. Classification performance measures the model's basic discriminative ability; the structure preservation metric evaluates whether the model can recover the biological relationships between cell states.

[0058] First, the model's ability to predict cell types is evaluated using accuracy: (9) in, N Indicates the number of test samples. Indicates real cell label, Indicates the model's predicted label. I(.) This is an indicator function. It takes the value 1 when the predicted label matches the true label, and 0 otherwise. In addition to classification performance, this embodiment further evaluates the ability of the embedding space to preserve the biological structure on two scales: global topological consistency and local structural continuity. Global topological consistency measures whether the model retains the overall developmental relationships between cell states, while local structural continuity evaluates the model's ability to preserve overall developmental topological relationships. To assess the model's ability to preserve these relationships, this embodiment uses the Spearman rank correlation coefficient (Spearman ρ) to measure the consistency between the embedding space structure and the predefined lineage distance matrix.

[0059] First, calculate the cosine distance between every pair of cells in the embedding space: (10) in, z i and z j The first i The test cells and the first j Normalized embedding vectors of each test cell, For the first in the embedded space i The test cells and the first j The cosine distance between the test cells.

[0060] Subsequently, its distance matrix from the predefined phylogenetic tree is calculated. Spearman correlation analysis: (11) Where Spearman represents the Spearman rank correlation coefficient calculation method. D embed For the embedding distance matrix, D phylogeny The preset phylogenetic distance matrix, ρ is the Spearman rank correlation coefficient, used to measure the consistency between the order of intercellular distances in the embedding space and the order of biological lineage relationships.

[0061] Spearman ρ measures the consistency between the ordering of inter-cell distances in the embedding space and the ordering of biological lineage relationships. A higher value indicates that the model is better able to maintain the global developmental topology.

[0062] Furthermore, to evaluate the ability to maintain local lineage continuity, this embodiment proposes the Average Lineage Distance (ALD) index. For each cell, its position in the embedding space is first determined. k Find the nearest neighbor cells and calculate the average distance between these neighbors in the phylogenetic distance space: (12) in, The total number of cells, Represents cells i The closest cosine distance in the embedding space k A set of indexes for each neighbor. This represents the predefined lineage distance between cells. A lower ALD value indicates that local neighboring cells in the embedding space have a closer developmental relationship in the actual lineage structure.

[0063] In summary, this embodiment evaluates the single-cell representation space through three complementary dimensions: classification performance, global topological consistency, and local structural continuity. Accuracy reflects the model's class discriminative ability, while Spearman ρ and ALD characterize the degree to which the embedding space preserves overall developmental organization relationships and local state transition relationships, respectively. This evaluation system can distinguish representations that "improve classification performance but fail to preserve biological structure," thus enabling the analysis of the fit between different inductive biases and cell state topology.

[0064] The feasibility and progressiveness of the method of the present invention will be further illustrated below with specific experimental results.

[0065] To assess whether discrete cell tags can adequately describe the AML cell state space, we first analyzed the distribution of cell state confidence scores based on weakly supervised marker gene annotation. This metric reflects the degree of matching between a single cell and a predefined cell state; lower confidence scores typically correspond to overlapping transcriptional features among multiple states.

[0066] The results showed that most AML cells exhibited low annotation confidence, with only a small percentage of cells having a clear class affiliation.

[0067] like Figure 3 As shown in Figure A, this result indicates that AML cell states do not exhibit a completely separated discrete category structure, but rather form a continuously changing state distribution in the transcriptional space. Particularly in the progenitor-to-mature cell transition region, some cells simultaneously exhibit multiple lineage-related characteristics, suggesting a clear continuity in the cell state transition process.

[0068] To further evaluate the impact of discrete classification boundaries on continuous-state cells, we analyzed the changes in the proportion of unassigned cells under different confidence thresholds. Figure 3 (B) As the threshold increases, the proportion of unassigned cells continues to rise, indicating that strict class boundaries gradually exclude cells in state transition regions. Based on the balance between label reliability and data coverage, we select an appropriate threshold for subsequent representation learning analysis.

[0069] In summary, the AML single-cell data exhibits clear continuous state organizational characteristics, suggesting that representation learning evaluation based on discrete category labels may not fully reflect the potential structural relationships between cell states.

[0070] To explore the relationship between classification performance and the ability to preserve cell state structure, this embodiment compares representation learning strategies under different structural inductive biases on the AML dataset. For three Transformer-based representation learning methods (CE-only, CE+InfoNCE, and CE+PGCL), the encoder structure and training process are kept consistent, and only the optimization objective is adjusted to isolate the impact of different structural inductive biases on the representation space organization. SVM and MLP+Cross-Entropy are also introduced as methods based on discrete classification objectives for comparison. The results are as follows: Figure 4 As shown.

[0071] In terms of classification performance, the methods generally maintained similar levels, with SVM achieving the highest accuracy, while the differences between different depth representation learning methods were limited. This indicates that supervised objectives based on discrete cell labels can all learn effective class discrimination boundaries. However, classification performance mainly reflects the model's ability to distinguish predefined cell categories and cannot directly evaluate whether the learned representation space preserves the potential biological organizational relationships between cell states.

[0072] Further analysis of the structure preservation index revealed significant differences in the representation spaces induced by different training objectives. CE-only and MLP+Cross-Entropy achieved similar Spearman ρ levels, indicating that when relying solely on category supervision, the model can form a certain degree of cellular state organization structure. However, since the optimization objective mainly focuses on the separation between categories and does not explicitly constrain the continuous developmental relationships between cells, its ability to preserve the global phylogenetic topology is limited.

[0073] It is noteworthy that while CE+InfoNCE maintained high classification performance, its Spearman ρ was significantly reduced, and it exhibited a high ALD value. This result indicates that simply introducing instance-level contrast constraints does not guarantee that the representation space conforms to the biological organization patterns of cell states. A possible reason is that InfoNCE optimizes the representation space by maximizing positive sample similarity and enhancing negative sample separation, its implicit assumption favoring the learning of separable structures between samples. However, single-cell transcriptional states typically change gradually along a continuous differentiation process, exhibiting progressive transcriptional transitions between adjacent states. When instance-level separation constraints dominate, cell states with biological continuity but different labels may be further pushed apart, weakening the original lineage continuity structure in the embedding space and leading to a decrease in its consistency with actual developmental topology.

[0074] In contrast, CE+PGCL achieved the highest Spearman ρ and lowest ALD values ​​while maintaining similar classification performance, indicating that introducing phylogenetic information can effectively guide the representation space to preserve the relative developmental relationships between cell states. This result suggests that appropriate biological structural constraints do not simply improve classification performance, but rather alter the organization of the embedding space, making it more consistent with the hierarchical phylogenetic characteristics during AML cell differentiation.

[0075] In summary, while different training objectives can achieve similar classification performance, their ability to preserve cell state topology varies significantly. This result indicates that classification accuracy and biological structure preservation ability are not equivalent indicators, and relying solely on downstream classification tasks is insufficient to fully evaluate the ability of single-cell representation learning methods to characterize the patterns of continuous cell state transitions. To further understand the source of this difference, subsequent analyses will explore how the structural inductive bias introduced by different training objectives shapes the organization of the representation space.

[0076] To further analyze the potential reasons for the differences in representation space caused by different training objectives, this embodiment compares the embedding space organization patterns induced by three loss functions from the perspective of structural inductive bias, such as... Figure 5 As shown in Figures A–C, different training objectives alter the relative organizational relationships of cell states in the low-dimensional representation space by introducing varying degrees of structural constraints.

[0077] like Figure 5 As shown in Figure A, when only cross-entropy loss (CE-only) is used, the optimization objective mainly revolves around cell category discrimination, with the implicit assumption that there are relatively clear category boundaries between different cell types. Therefore, this strategy can form a relatively clear cell population distribution, but for intermediate-state cells in the continuous differentiation process, its spatial organization ability is limited, manifested as a weakening of the transition relationship between different differentiation stages.

[0078] like Figure 5 As shown in Figure B, the representation space is further affected by instance-level similarity optimization after incorporating the InfoNCE contrast constraint. Since standard InfoNCE typically imposes a uniform exclusion constraint on negative samples without considering the potential biological distance between cell states, cell states with developmental associations but different labels may experience excessive segregation in continuous lineage data. This results in enhanced local clustering of cell populations, but a shift in the overall lineage continuity structure.

[0079] In contrast, the CE+PGCL representation space, after incorporating pedigree distance constraints, exhibits an organizational pattern that better reflects the hematopoietic differentiation process. For example... Figure 5 As shown in Figure C, this strategy adjusts the similarity constraint strength based on the lineage relationships between cell states, ensuring high continuity between lineage-close states while maintaining necessary distinctions between distant states. The corresponding embedding space exhibits more pronounced state transition regions and hierarchical branching structures. This result demonstrates that the biological organization of the representation space depends not only on the model's expressive power but also on the structural assumptions contained in the training objective.

[0080] Furthermore, to verify whether known hematopoietic differentiation characteristics are preserved in different embedding spaces, this embodiment analyzed the expression distribution of representative hematopoietic marker genes in the embedding spaces, such as... Figure 5 As shown in Figures D–F. The results showed that the gene expression gradient and the spatial distribution of cell states were consistent. The progenitor-related marker gene CD34 was mainly enriched in the early progenitor cell region and gradually decreased along the differentiation direction; the myeloid marker gene LYZ was mainly distributed in myeloid-related regions; and the NK / lymphoid-related marker gene NKG7 was mainly enriched in the lymphocyte region, while its expression was low in the progenitor and myeloid regions.

[0081] The above results indicate that different structural inductive biases can induce different cell state organization patterns, and biologically consistent structural constraints are more conducive to maintaining continuous differentiation relationships in single-cell data. Therefore, embedding space evaluation should not only focus on class separation ability, but also consider its ability to preserve the potential cell state topology.

[0082] To further verify whether the effect of structural inductive bias depends on the cell state organization mode, this embodiment performs cross-topological comparisons on two datasets with different topological features, MPAL and PBMC. The results are as follows: Figure 6 As shown. Unlike the relatively continuous hierarchical differentiation structure in AML, MPAL has significant cross-lineage expression characteristics and highly heterogeneous cell state distribution, which can be used to assess the consistency between predefined structural priors and real cell state relationships.

[0083] In the MPAL dataset, different structural inductive biases exhibit different response patterns compared to the continuous differentiation system in AML. Unlike phylogenetic constraints in AML, which promote the preservation of continuous state structures, CE+PGCL, while achieving higher classification performance, did not show a corresponding improvement in its structure preservation index. CE-only, however, outperformed the global topology preservation index Spearman ρ. This result indicates that the effectiveness of phylogenetic structural constraints does not depend on the constraint form itself, but rather on whether the biological assumptions it encodes match the cell state organization patterns in the target data. MPAL cell states are often accompanied by aberrant phylogenetic remodeling, cross-phylogenetic marker co-expression, and intermediate transitional states, and their cell state space may deviate from the classic hematopoietic differentiation hierarchy. Therefore, structural priors constructed based on normal hematopoietic phylogenetic relationships may not accurately describe the state transition relationships in MPAL, and may even introduce additional biases, thus limiting the representation space's ability to characterize the true cell state structure.

[0084] It is noteworthy that CE+InfoNCE exhibits lower classification performance in MPAL, but simultaneously achieves a lower ALD value. This phenomenon suggests that a decrease in the local neighborhood index does not necessarily represent an enhanced ability to preserve biological structures. Since ALD is calculated based on nearest-neighbor relationships in the embedding space, smaller neighborhood lineage distances can also result when the representation space is subjected to excessively strong instance-level constraints, leading to local clustering or structural compression. However, this change may reflect a shift in the geometry of the representation space rather than a restoration of the continuity of actual cell states. Combined with its lower Spearman ρ result, it is evident that standard InfoNCE may enhance inter-sample separation in the mixed lineage structure of MPAL, while weakening its ability to preserve potentially biologically continuous state relationships.

[0085] In the discrete cell population system of PBMC, all methods achieved high classification performance, while the structure preservation indices showed different trends. Since PBMC mainly consists of relatively well-defined cell populations and lacks the typical state transition processes found in continuous developmental lineages, structural constraints constructed based on continuous lineage relationships did not show a stable advantage. Specifically, the Spearman ρ of CE+PGCL was lower than that of the traditional methods SVM and MLP+CE, suggesting that lineage structure priors may not provide additional structural information in data lacking obvious continuous topology. This result further illustrates that the effect of structure inductive bias depends on the topological organization pattern of the data itself, rather than being solely determined by introduced biological priors.

[0086] By integrating three different topological scenarios—AML, MPAL, and PBMC—it can be observed that the impact of the representation learning objective on the embedding space exhibits a significant data dependency. When cell states are organized along continuous differentiation trajectories, structural constraints consistent with biological developmental relationships can promote the preservation of global topology and local continuity. However, in cross-lineage mixed or discrete cell population structures, priors that do not match the data topology may translate into additional biases that limit the model's expressive power. Therefore, the evaluation and selection of single-cell representation learning methods should not solely rely on classification performance but should comprehensively consider the consistency between the structural inductive bias introduced by the model and the cell state organization pattern.

[0087] To further observe the effects of different structural inductive biases on the topology of hybrid spectral systems, this embodiment compares the embedding space organization patterns induced by different methods, and the results are as follows: Figure 7 As shown in Figures A–C.

[0088] To further analyze the impact of structural induction bias on the representation spatial organization under different cellular topological environments, this embodiment visualizes and compares cell embeddings from the MPAL and PBMC datasets, such as... Figure 7 As shown. Figure 7As shown in Figures A–C, in MPAL data with mixed lineage characteristics, different optimization objectives induced significantly different state organization patterns. CE-only can form a relatively clear class structure, but the continuous relationship between different cell states is limited. CE+InfoNCE further enhances the separation trend between cells, making some potentially related state relationships more discretized, indicating that instance-level contrastive constraints may strengthen class boundaries while weakening structure preservation. In contrast, CE+PGCL introduces a certain degree of spatial smoothness while maintaining class differentiation, but its embedding organization still does not fully follow the predefined lineage structure, suggesting that the applicability of fixed lineage priors is limited in complex mixed lineage environments.

[0089] like Figure 7 As shown in Figures D–F, in the PBMC dataset, cell states mainly exhibit a discrete distribution pattern of mature cell populations, and different methods can all form relatively stable cell cluster structures. However, different inductive biases still lead to changes in distance relationships in the embedding space. CE-only maintains a relatively stable inter-cluster organization structure, consistent with the relatively clear cell type boundaries in PBMC. After introducing InfoNCE, the separation between cell populations is further enhanced, but the overall spatial organization still exhibits a discrete cluster structure, indicating that instance-level contrastive constraints mainly regulate the relative distance between samples without actively inducing continuous state transitions. In contrast, although PGCL can increase local continuity, its structural constraints fail to further improve the overall structure preservation ability due to the lack of significant continuous lineage processes in PBMC.

[0090] In summary, the results from both MPAL and PBMC indicate that the impact of structural inductive bias on the cell representation space is not determined by the form of the constraint itself, but rather by its consistency with the intrinsic topological structure of the data. Continuous structural constraints can only provide effective information when there are clear developmental relationships between cell states. In lineage-mixed or discrete cell organization environments, mismatched structural assumptions may become additional sources of bias, thus limiting the representation space's ability to characterize real cell state relationships.

[0091] Cross-dataset analysis revealed that the effectiveness of structural inductive bias is topologically dependent. To further determine whether this dependency also exists at the individual patient level, this embodiment further conducted a patient heterogeneity analysis. This embodiment analyzed the change in structural retention ability (Δρ = ρ(PGCL) - ρ(CE-only)) with and without PGCL in different AML patients to assess the applicability of lineage structural constraints in different patient contexts.

[0092] like Figure 8As shown, the patient-level heterogeneity of lineage-guided representation learning responses is altered by the introduction of PGCL at the pedigree topology preservation level compared to the CE method alone. The Pearman correlation coefficient represents the difference between the Spearman correlation coefficients of PGCL and CE alone; a positive value indicates improved consistency between the learned embedding space and the predefined lineage relationships. The heterogeneous distribution of Δρ across different patients suggests that the effect of lineage-based structural constraints may depend on the patient's unique cellular state arrangement characteristics.

[0093] Patient-level response analysis further revealed differentiated response trends to lineage structure constraints among different AML samples. Most patients showed positive responses, suggesting that introducing lineage structure priors may help enhance the consistency between the embedding space and predefined lineage relationships, consistent with observed trends in the overall AML dataset. However, some patients did not show significant structural preservation enhancements, or even showed negative responses, suggesting that the effectiveness of fixed lineage priors may be influenced by patient-specific cell state organization patterns. For these patients with weaker or opposite responses, their cell state distributions may contain more complex aberrations, mixed lineage states, or transcriptional features deviating from classic hematopoietic pathways, making it difficult for predefined lineage relationships to fully describe their true state topology. Therefore, patient-level differential responses further suggest that the effectiveness of structural inductive bias is not universal, but depends on the degree of matching between the structural assumptions encoded by the model and the individual cell state topology.

[0094] To eliminate the influence of model parameter settings and network capacity differences on the above conclusions, this embodiment further conducts parameter sensitivity analysis and model capacity perturbation experiments to verify the stability of the obtained patterns.

[0095] Table 2. Results of sensitivity analysis of the λ parameter

[0096] Table 3 Comparison of model performance under different network architecture configurations

[0097] Table 2 illustrates the impact of the lineage loss weight λ on model performance. The results show that the overall classification performance remains stable under different λ settings (accuracy approximately 0.87), indicating that the class discrimination ability is mainly driven by the supervised classification objective, while changes in the lineage constraint strength do not alter the classification boundary. In contrast, the structure preservation index shows a more significant response to changes in λ. As λ increases, Spearman ρ generally decreases, while ALD gradually increases, suggesting that excessively strong lineage constraints may enhance reliance on predefined lineage relationships, thereby reducing the flexibility of the representation space in expressing the intrinsic state relationships of the data.

[0098] It is worth noting that under different λ settings, the model always maintains the basic pattern of "relatively stable classification performance, while the ability to preserve structure is adjusted by inductive bias", indicating that the phylogenetic constraint mainly affects the topological organization of the embedding space, rather than the ability to distinguish categories.

[0099] To eliminate the influence of differences in model expressive power on the experimental conclusions, Table 3 further evaluates the stability of each training objective under different network capacity configurations. The results show that the relative behavior patterns among different training objectives remain consistent after changing the hidden dimension, the number of Transformer layers, and the number of attention heads. CE-only exhibits moderate structure preservation ability under different network configurations; CE+InfoNCE consistently shows low Spearman correlation and high ALD; while CE+PGCL maintains high Spearman correlation under different model configurations, indicating that its structure preservation advantage does not depend on specific network capacity or architecture settings.

[0100] The combined results of parameter sensitivity and model capacity perturbation reveal that the differences in representation space induced by different training objectives exhibit strong stability. While constraint strength and model configuration can influence the degree of structure preservation, they do not alter the relative interaction patterns between different inductive biases. This result further supports the proposed viewpoint that the organization of the single-cell representation space is not primarily determined by model capacity, but rather by the structural assumptions introduced by the training objectives and their consistency with the intrinsic topology of the data.

[0101] This embodiment systematically evaluates the impact of different structural induction biases on the organization of cell state space during single-cell representation learning and explores the relationship between task performance and biological structure preservation ability. We construct and compare five representation learning strategies within a unified encoder framework, including the traditional classification method SVM, the multilayer perceptron-based cross-entropy optimization method (MLP+CE), the representation learning method using only cross-entropy supervision (CE-only), the method combining instance-level contrastive constraints (CE+InfoNCE), and the method incorporating phylogenetic structure priors (CE+PGCL). First, in the AML system with continuous developmental hierarchy features, we evaluate the impact of different training objectives on cell type classification ability and cell state topology preservation ability. Subsequently, we further utilize the MPAL system with cross-phylogenetic mixing features and the PBMC system with discrete cell population organization patterns to systematically analyze the applicability of structural induction biases in different cell state topological contexts. Furthermore, we conduct patient-level heterogeneity analysis to explore the differences in structural constraint responses among different individuals under the same disease background. Through multi-level data validation, this embodiment reveals the structural factors behind the performance differences of representation learning models and proposes the analytical perspective of topology–bias compatibility.

[0102] This embodiment first reveals that classification performance cannot fully reflect the biological quality of a single-cell representation space. Traditional evaluation systems typically rely on cell type classification accuracy to judge model performance, but our results show that even when different optimization objectives achieve similar classification performance, their embedding spaces may still exhibit significant differences in their ability to preserve cell state relationships. In the AML system, CE+InfoNCE maintains high classification performance, but its Spearman correlation is significantly reduced, indicating that while instance-level discriminative objectives enhance class separation capabilities, they may disrupt the structural relationships between continuous cell states. This phenomenon may stem from the contradiction between the unified negative sample rejection mechanism in the InfoNCE loss function and the continuity of single-cell states: in real biological processes, some cells, although belonging to different categories, still exhibit developmental associations and state transition relationships, while instance-level contrastive learning may force these potentially related cells into different representation regions. Therefore, relying solely on classification performance may mask the model's deviation from the real-world cell state organization patterns.

[0103] Further analysis of the impact of different structural constraints on the representation space reveals that the effect of structural inductive bias depends on its degree of matching with the intrinsic topological structure of the data. In the AML system, cell states change continuously along the hematopoietic differentiation pathway, and lineage relationships can effectively describe the state transitions between cells. Therefore, introducing lineage structural constraints can promote the maintenance of continuous developmental relationships in the representation space. However, in the MPAL system, due to the existence of aberrant lineage remodeling, cross-lineage expression, and mixed states, classical hematopoietic lineage relationships cannot fully cover the real cell state space. Therefore, fixed lineage priors do not show a stable advantage. This indicates that structural priors themselves do not necessarily improve representation learning performance; their effectiveness depends on whether the prior knowledge can accurately reflect the cellular organization patterns in the target data.

[0104] The results in the PBMC system further support the above view. Since PBMCs are mainly composed of mature immune cell populations, their cell states are closer to discrete population distributions, lacking a clear continuous developmental trajectory. Therefore, structural constraints based on lineage relationships are unlikely to provide additional information. This result indicates that structural biases applicable to continuously differentiated systems cannot be directly transferred to all single-cell data scenarios. In other words, there is no single optimal structural assumption applicable to all biological systems in representation learning models; model performance depends on the consistency between the inductive bias encoded by the training objective and the topological features of the data itself.

[0105] Furthermore, patient-level analysis revealed individual heterogeneity in structural learning responses under disease states. Even within the same AML disease context, different patients exhibited significant differences in their responses to lineage structural constraints, suggesting that the disease-related cell state space may be influenced by factors such as abnormal differentiation pathways, the proportion of mixed states, and individual transcriptional background. Therefore, future single-cell representation learning models need to evolve from fixed-structure prior-driven methods to adaptive learning frameworks capable of automatically adjusting structural constraints based on data features.

[0106] From an application perspective, this embodiment emphasizes that the evaluation system for single-cell artificial intelligence models needs to be expanded from single-task performance indicators to the analysis of structure preservation capabilities. In tasks such as cell trajectory inference, disease state transition analysis, and treatment response prediction, whether the representation space can preserve the true cell state relationships may be more important than simple classification boundaries.

[0107] In summary, this invention establishes a systematic evaluation framework spanning three topological patterns: continuous differentiation, mixed lineages, and discrete cell populations, revealing the interaction between structural inductive bias and cell state topology. The results demonstrate that the effectiveness of single-cell representation learning models is not solely determined by task performance, but rather by whether the learned representations can reflect the organizational patterns of real biological states. This method provides new theoretical support for understanding the relationship between task optimization and biological structure learning in single-cell artificial intelligence models and offers guidance for future development of more structure-aware and biologically consistent representation learning methods.

[0108] This invention also provides a single-cell data representation learning device that preserves the topology of biological structures, used to implement the methods described in any of the above embodiments, such as... Figure 9 As shown, the device includes: The data preprocessing module 901 is configured to acquire single-cell RNA sequencing data and preprocess the single-cell RNA sequencing data to obtain a standardized cell expression vector set. The framework construction module 902 is configured to construct a unified representation learning framework, which includes a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. In response to the input standardized cell expression vector set, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vector to obtain a cell shared representation. The cell shared representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function. The contrastive learning branch maps the cell shared representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector. The model training module 903 is configured to determine the corresponding training objective based on a preset structural induction bias type, and to jointly optimize the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and a preset phylogenetic distance matrix to obtain a trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree. The structure preservation evaluation module 904 is configured to perform structure preservation evaluation on the embedding space of the trained single-cell representation learning model to obtain classification performance evaluation results, global topology preservation evaluation results, and local structural continuity evaluation results.

[0109] This invention provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0110] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0111] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0112] The electronic device provided in this embodiment of the invention can be the terminal device described in the above embodiments.

[0113] This invention also provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the technical solution of the single-cell data representation learning method for maintaining biological structural topology as described in the above embodiments.

[0114] This invention also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the single-cell data representation learning method for maintaining the topology of biological structures in the above embodiments.

[0115] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0116] The modules described as separate components may or may not be physically separate. The components shown as modules 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 can be selected to implement the solution of this embodiment according to actual needs.

[0117] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0118] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute certain steps of the methods of the various embodiments of the present invention.

[0119] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0120] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0121] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0122] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0123] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0124] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-cell data representation learning method that preserves the topology of biological structures, characterized in that, The method includes: Single-cell RNA sequencing data were acquired and preprocessed to obtain a standardized set of cell expression vectors. A unified representation learning framework is constructed, comprising a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. Responding to the input standardized cell expression vector set, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vector to obtain a cell shared representation. This cell shared representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function, and the contrastive learning branch maps the cell shared representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector. The corresponding training objective is determined according to the preset structural induction bias type, and the unified representation learning framework is jointly optimized based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix to obtain the trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree. The embedding space of the trained single-cell representation learning model is evaluated for structure preservation, and the results of classification performance evaluation, global topology preservation evaluation, and local structural continuity evaluation are obtained.

2. The single-cell data representation learning method for preserving biological structural topology according to claim 1, characterized in that, The single-cell RNA sequencing data were preprocessed to obtain a standardized set of cell expression vectors, including: Missing genes in different samples were filled with zero values. After unifying the gene space, the original expression matrix was logarithmically transformed to base 2 to obtain a unified cell expression matrix. Based on the gene expression variance ranking in the training set samples, the top-ranked hypervariable genes are selected as the model input features. The input features of the model are Z-score standardized, where the mean and variance parameters are calculated only from the training set and applied to the test set to obtain a standardized set of cell expression vectors.

3. The single-cell data representation learning method for preserving biological structural topology according to claim 2, characterized in that, The unified representation learning framework, in response to the input standardized set of cell expression vectors, obtains normalized embedding vectors in the following ways: The standardized cell expression vector is converted into a low-dimensional hidden vector using the following formula. : (1) in, x This represents the expression vector of highly variable genes. W p and b p The hidden dimension is set to 128 as a learnable parameter; The low-dimensional hidden vector The cell-shared representation is obtained by taking an encoder consisting of stacked Transformer encoders as input, using the following formula. h : (2) Here, Transformer represents a neural network encoder based on a self-attention mechanism, used for feature extraction and context modeling of input sequences; The cell-shared representation h The input is fed into the classification branch, and the predicted class probability is obtained through the Softmax classifier using the following formula: (3) Where Softmax is a normalized exponential function that transforms a real vector into a probability distribution. W c For classification weight parameters, To predict class probabilities; The cell-shared representation h The input is fed into the contrastive learning branch, where the cell-shared representation is mapped to the contrastive space and normalized using the projection head. The normalized embedding vector is obtained using the following formula. z : (4) Where Proj represents the projection head network consisting of fully connected layers. This represents the L2 norm.

4. The single-cell data representation learning method for preserving biological structural topology according to claim 3, characterized in that, When the preset structural induction bias type is a weak structural bias, the training objective is to use only the cross-entropy loss function for supervision; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix includes: Based on the predicted category probability and cell real category labels y The cross-entropy loss value is calculated using the following formula: (5) in, N For the sample size, C For the number of lineage categories, i For sample index, c For category indexing, For the first i The sample at the th c Real labels in each category For the first i The sample at the th c Predicted class probabilities for each category L CE This represents the cross-entropy loss value. The parameters of the unified representation learning framework are updated with the goal of minimizing the cross-entropy loss.

5. The single-cell data representation learning method for preserving biological structural topology according to claim 3, characterized in that, When the preset structural induction bias type is instance separation bias, the training objective is a combination of cross-entropy loss and instance-level contrastive loss; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix includes: Calculate the cross-entropy loss based on the predicted class probability and the true cell class label; According to the normalized embedding vector z ， The InfoNCE contrast loss is calculated using the following formula: (6) in, s ij For the first i The sample and the first j Cosine similarity between samples s ik For the first i The sample and the first k Cosine similarity between samples z i and z j The first i The first sample and the first j The normalized embedding vector of each sample. τ Temperature is a parameter used to control the smoothness of the similarity distribution. k For sample index, N The total number of samples, L infoNCE Compare the loss values ​​for InfoNCE; The cross-entropy loss and the InfoNCE contrastive loss are weighted and summed to obtain a joint loss value. The parameters of the unified representation learning framework are updated with minimizing the joint loss value as the optimization objective.

6. The single-cell data representation learning method for preserving biological structural topology according to claim 1, characterized in that, When the preset structural induction bias type is a phylogenetic structure bias, the training objective is a combination of cross-entropy loss and phylogenetic guided contrastive loss; the joint optimization of the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and the preset phylogenetic distance matrix includes: The distances between cell types are converted into a weight matrix based on the lineage distance matrix, where the weight values ​​are calculated using the following formula: (7) in, d ij The phylogenetic distance matrix is ​​the th i The cell and the first j Lineage distance between cells σ To control the hyperparameters of the weight decay rate, w ij For the first i The cell and the first j Weight values ​​between individual cells e It is a natural constant; Calculate the cross-entropy loss based on the predicted class probability and the true cell class label; According to the normalized embedding vector z The lineage-guided contrast loss is calculated using the weight matrix and the following formula: (8) in, For the first i The sample and the first j The cosine similarity of the embedding vectors of each sample is normalized in the contrast space. N The total number of samples, Lphy log eny Lineage-guided comparison loss value; The cross-entropy loss and the lineage-guided comparison loss are weighted and summed to obtain a joint loss value. The parameters of the unified representation learning framework are updated with minimizing the joint loss value as the optimization objective.

7. The single-cell data representation learning method for preserving the topology of biological structures according to any one of claims 1 to 6, characterized in that, The methods for evaluating the structure preservation of the embedding space of the trained single-cell representation learning model include: The expression vector of the test cell is obtained and input into the trained single-cell representation learning model to obtain the normalized embedding vector of the test cell and the predicted test class probability. The classification accuracy is calculated using the following formula: (9) in, N For the number of test samples, i For the test sample index, y i For the first i The true class label of each test cell For the first i Predicted category labels for each test cell I The parentheses () are indicator functions; they take the value 1 if the condition inside the parentheses is true, and 0 otherwise. Accuracy For classification accuracy; Based on the normalized embedding vector of the test cells, the cosine distance between each pair of cells in the embedding space is calculated, and the embedding distance matrix is ​​obtained using the following formula: (10) in, z i and z j The first i The test cells and the first j Normalized embedding vectors of each test cell, For the first in the embedded space i The test cells and the first j Cosine distance between test cells; Spearman rank correlation analysis is performed between the embedding distance matrix and the preset phylogenetic distance matrix, and the Spearman rank correlation coefficient is calculated using the following formula: (11) Where Spearman represents the Spearman rank correlation coefficient calculation method. D embed For the embedding distance matrix, D phylogeny The preset phylogenetic distance matrix, , which is the Spearman rank correlation coefficient, used to measure the consistency between the order of intercellular distances in the embedding space and the order of biological lineage relationships; The embedding vector of each test cell is used to determine the embedding space of each test cell. k The nearest neighbor cells are counted, and the average phylogenetic distance is calculated using the following formula: (12) in, N k ( i ) is the first i The test cells with the closest cosine distance in the embedding space k A set of indexes for each neighbor. The index contains the number of elements in the set. d lineage ( i , j ) is the first i The test cells and the first j The preset lineage distance between test cells, where ALD is the average lineage distance consistency; Based on the classification accuracy, the Spearman rank correlation coefficient, and the average phylogenetic distance consistency, the embedding space structure preservation ability of the trained single-cell representation learning model is comprehensively evaluated from three dimensions: classification performance, global topology preservation ability, and local structural continuity.

8. A single-cell data representation learning device that preserves the topology of biological structures, characterized in that, The device includes: The data preprocessing module is configured to acquire single-cell RNA sequencing data and preprocess the single-cell RNA sequencing data to obtain a standardized cell expression vector set. The framework construction module is configured to construct a unified representation learning framework, which includes a linear mapping layer, a Transformer encoder, a classification branch, and a contrastive learning branch. Responding to the input standardized cell expression vector set, the unified representation learning framework converts each cell expression vector into a low-dimensional hidden vector through the linear mapping layer. The Transformer encoder encodes the low-dimensional hidden vector to obtain a cell shared representation. This cell shared representation is then input to the classification branch and the contrastive learning branch, respectively. The classification branch obtains the predicted class probability through a classification prediction function, and the contrastive learning branch maps the cell shared representation to a contrastive space and normalizes it through a contrastive learning projection head to obtain a normalized embedding vector. The model training module is configured to determine the corresponding training objective based on a preset structural induction bias type, and to jointly optimize the unified representation learning framework based on the training objective, the predicted class probability, the normalized embedding vector, and a preset phylogenetic distance matrix to obtain a trained single-cell representation learning model; the phylogenetic distance matrix is ​​used to characterize the developmental relationship of different cell types in the hematopoietic differentiation tree. The structure preservation evaluation module is configured to perform structure preservation evaluation on the embedding space of the trained single-cell representation learning model to obtain classification performance evaluation results, global topology preservation evaluation results, and local structural continuity evaluation results.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the single-cell data representation learning method for maintaining the topology of biological structures as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the single-cell data representation learning method for maintaining the topology of biological structures as described in any one of claims 1-7.

Images