A single-cell gene perturbation response prediction method and system

Abstract

This invention discloses a method and system for predicting single-cell gene perturbation responses. The method includes the following steps: dividing the original single-cell expression matrix into multiple cell populations according to a preset partitioning strategy, with each cell population containing a predetermined number of cell samples; for each gene in each cell population, constructing its statistical distribution model based on the gene's expression value in the cell population, and extracting the corresponding distribution parameter vector as the distribution-level representation of the gene; constructing a distribution condition mapping model based on the gene distribution parameters of the control group; optimizing the distribution condition mapping model during the training phase using the gene distribution parameters corresponding to the perturbation group as a supervision signal; inputting the control group distribution parameters and corresponding condition information of the sample to be predicted into the trained distribution condition mapping model during the inference phase to obtain the gene distribution parameters under the target perturbation state; and sampling based on the predicted distribution parameters to generate corresponding single-cell expression data.

Description

Technical Field

[0001] This invention relates to the field of single-cell gene perturbation response prediction technology, and in particular to a method and system for single-cell gene perturbation response prediction. Background Technology

[0002] Single-cell RNA sequencing (scRNA-seq) technology can quantitatively detect whole-genome transcriptome expression profiles at single-cell resolution, and has become a core tool for studying cellular states and their responses to external perturbations (including gene editing, cytokine stimulation, and small molecule drug treatment). Gene perturbation response prediction aims to calculate and predict whole-genome expression changes in cells given unobserved perturbation conditions, especially accurately identifying differentially expressed genes (DEGs), which is of great significance for drug target discovery and disease mechanism analysis.

[0003] However, scRNA-seq technology is inherently destructive—cells must be lysed before sequencing can be completed, so the gene expression state of the same cell before and after perturbation cannot be obtained simultaneously. This physical limitation means that truly "paired training data" does not exist, which is a fundamental challenge that all methods in this field must face.

[0004] In existing technologies, methods represented by CPA (Compositional Perturbation Autoencoder) utilize variational autoencoders (VAEs) to encode single-cell expression vectors into a low-dimensional continuous latent space, modeling perturbation effects as linear additions of latent space vectors. Specifically, the encoder maps single-cell whole-genome expression to latent vectors; perturbation effects are represented by learnable perturbation embedding vectors; during prediction, the latent vectors of control group cells are added to the perturbation embeddings, and then the decoder restores the gene expression space. This method supports combined perturbations (summing multiple perturbation effect vectors) and has certain advantages in generalization scenarios with a small number of perturbation samples. However, during training, this method requires "pseudo-pairing" cells with similar states in the control and perturbation groups in the latent space to construct a supervision signal. However, the cells in the control and perturbation groups come from different individuals, and the differences between the two groups of latent vectors not only include perturbation signals but also batch effects and inherent heterogeneity between cells, resulting in systematic noise in the training signal and significantly affecting the prediction accuracy of differentially expressed genes. Furthermore, the Gaussian prior assumption of VAE does not match the high zero expansion of scRNA-seq data, resulting in poor modeling performance for genes with a high proportion of zero values.

[0005] Methods such as scGPT and GEARS introduce large-scale Transformer pre-training into single-cell perturbation prediction. scGPT performs masked gene expression prediction pre-training on massive single-cell data to obtain gene context representations, which are then fine-tuned for the perturbation prediction task. GEARS introduces a gene regulatory network graph structure on top of the Transformer, enhancing the modeling of prior relationships between genes through graph neural networks. Both methods use single-cell expression vectors (or their bucketed discretized versions) as the basic units of model input and output to predict the genome-wide expression mean of a single cell. However, this method uses single-cell modeling granularity, but scRNA-seq data has extremely high sparsity (the zero-value rate is usually over 90%), and random technical noise (dropout) at the single-cell granularity dominates, making it difficult for the model to stably learn perturbation effects from high-noise signals. Furthermore, these methods still rely on cell-level "pseudo-pairing" between the control and perturbation groups to construct training samples, failing to eliminate systematic pairing noise. Furthermore, the regression paradigm that directly predicts the mean gene expression ignores the inherent bimodal distribution and zero expansion characteristics of scRNA-seq data, resulting in limited accuracy in predicting differential expression levels of genes with low abundance and bimodal expression. Summary of the Invention

[0006] The purpose of this invention is to solve the following technical problems:

[0007] (1) Existing methods generally rely on cell-level "pseudo-pairing" between the control group and the perturbation group to construct the supervision signal. However, due to the destructive nature of single-cell sequencing, the expression state of the same cell before and after perturbation cannot be obtained simultaneously, resulting in the constructed paired samples being essentially from different individuals or different cell populations. These pseudo-paired samples are simultaneously mixed with multiple factors such as perturbation effect, batch effect, and inherent cell heterogeneity, causing systematic bias in the supervision signal, thereby limiting the model's ability to accurately model the real perturbation effect.

[0008] (2) Existing methods usually use single-cell expression vectors as the basic modeling unit. However, single-cell RNA sequencing data has high sparsity, high noise and significant zero expansion characteristics. The expression values ​​at the single-cell level are significantly affected by technical noise, resulting in a low signal-to-noise ratio, making it difficult for the model to stably learn the expression change rules caused by perturbation from single-point observation.

[0009] (3) Existing methods mostly use regression prediction based on the mean of expression or low-dimensional latent variables, which fail to effectively characterize the multimodal distribution and zero expansion characteristics of gene expression. They are difficult to accurately describe the expression changes of low-expression genes and genes with bimodal distribution characteristics, thus affecting the accuracy of differentially expressed gene identification.

[0010] To achieve the above objectives, the present invention provides a method for predicting single-cell gene perturbation responses, the method comprising the following steps:

[0011] The original single-cell expression matrix is ​​divided into multiple cell populations according to a preset partitioning strategy. Each cell population contains a predetermined number of cell samples. For each gene in each cell population, a statistical distribution model is constructed based on the expression value of the gene in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene.

[0012] Based on the gene distribution parameters of the control group, a distribution condition mapping model is constructed. The distribution condition mapping model receives the distribution parameters of the control group as input, and combines gene identification information, perturbation condition information and cell type information to perform condition modeling, and outputs the gene distribution parameters under the target perturbation state.

[0013] During the training phase, the gene distribution parameters corresponding to the perturbation group are used as supervision signals to optimize the distribution condition mapping model;

[0014] During the inference phase, the control group distribution parameters and corresponding condition information of the sample to be predicted are input into the trained distribution condition mapping model to obtain the gene distribution parameters under the target perturbation state; based on the predicted distribution parameters, sampling is performed to generate the corresponding single-cell expression data.

[0015] On the other hand, the present invention provides a single-cell gene perturbation response prediction system, the system comprising:

[0016] The distribution modeling module is used to divide the original single-cell expression matrix into multiple cell populations according to a preset partitioning strategy. Each cell population contains a predetermined number of cell samples. For each gene in each cell population, a statistical distribution model is constructed based on the expression value of the gene in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene.

[0017] The distribution transformation module is used to construct a distribution condition mapping model based on the gene distribution parameters of the control group. The distribution condition mapping model receives the distribution parameters of the control group as input, and combines gene identification information, perturbation condition information and cell type information to perform condition modeling, and outputs the gene distribution parameters under the target perturbation state.

[0018] The model training module is used to optimize the distribution condition mapping model during the training phase, using the gene distribution parameters corresponding to the perturbation group as a supervision signal.

[0019] The sampling generation module is used to input the control group distribution parameters and corresponding condition information of the sample to be predicted into the trained distribution condition mapping model during the inference phase to obtain the gene distribution parameters under the target perturbation state; and to generate corresponding single-cell expression data based on the predicted distribution parameters.

[0020] The present invention provides a single-cell gene perturbation response prediction method and system, which reduces single-cell noise interference and eliminates systematic errors caused by pseudo-pairing by parameterizing the gene expression distribution within a cell population and performing mapping prediction under perturbation conditions in the distribution space. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart of a single-cell gene perturbation response prediction method provided in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of the ZI-GMM distribution modeling process;

[0024] Figure 3 This is a schematic diagram of a distributed transformation learning framework;

[0025] Figure 4 This is a schematic diagram of a single-cell gene perturbation response prediction system provided in an embodiment of the present invention. Detailed Implementation

[0026] The present specification will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. The described embodiments are only a part of the embodiments described herein, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without inventive effort are within the scope of protection of this application.

[0027] This invention elevates the modeling object from single-cell expression values ​​to statistical distribution parameters of gene expression within a cell population, establishing a mapping relationship from control state to perturbed state at the distribution level. This avoids systematic noise introduced by cell-level pseudo-pairing and improves the ability to characterize the zero expansion, multimodality, and high sparsity of single-cell data.

[0028] Figure 1 This is a schematic flowchart of a single-cell gene perturbation response prediction method provided in an embodiment of the present invention. Figure 1 As shown, the overall process of this embodiment of the invention includes the following steps:

[0029] Step 101: Cell population segmentation and parameterized representation of gene expression distribution.

[0030] Specifically, in this embodiment of the invention, the original single-cell expression matrix is ​​divided into multiple cell populations according to a preset partitioning strategy, with each cell population containing a predetermined number of cell samples. The partitioning strategy can be random partitioning, partitioning based on cell type, or partitioning based on experimental batch, etc. For each gene in each cell population, a statistical distribution model is constructed based on the expression value of the gene in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene, thereby converting the original high-dimensional sparse expression matrix into a compact distribution parameter representation.

[0031] The distribution model in this embodiment of the invention can adopt a zero-inflated Gaussian mixture model (ZI-GMM), in which each gene is characterized by parameters such as zero-inflated parameter, mixture weight, mean and variance; in other embodiments, a negative binomial distribution, a zero-inflated negative binomial distribution, or other probability distribution models that can characterize sparsity and multimodality can also be used.

[0032] Figure 2 A schematic diagram of the ZI-GMM distribution modeling process. (Example) Figure 2 As shown, for the first cell batch... Each gene, its expression level The probability density of a zero-inflated Gaussian mixture model (ZI-GMM) is defined as:

[0033]

[0034] in A zero-inflation parameter indicates that the gene The probability of a technical zero (dropout) in this batch; The Gaussian mixture component is a hyperparameter that can be flexibly set according to the data characteristics. , , The first The weights, mean, and standard deviation of each Gaussian component satisfy the following conditions: , .when At that time, the distribution of each gene is determined by a parameter vector. (common =7-dimensional) complete description.

[0035] During fitting, the zero-inflation parameter is first estimated directly from the proportion of non-zero values:

[0036]

[0037] Then, for non-zero expression values Fitting with the EM algorithm The components are Gaussian mixed, and the resulting components are sorted in ascending order of mean to eliminate label switching issues. Distribution parameters are directly loaded during training to avoid processing the original sparse matrix.

[0038] Step 102: Construction and prediction of distribution condition mapping model.

[0039] Specifically, this embodiment of the invention constructs a distribution condition mapping model based on the gene distribution parameters of the control group. This model is used to learn the variation law of gene expression distribution from the control state to the perturbed state under given perturbation conditions and cell type conditions. The distribution condition mapping model receives the distribution parameters of the control group as input and combines gene identification information, perturbation condition information, and cell type information to perform condition modeling, outputting the gene distribution parameters under the target perturbed state. The mapping model can adopt an encoder-decoder structure based on an attention mechanism; in other embodiments, multilayer perceptrons, graph neural networks, diffusion models, or other nonlinear mapping models can also be used to realize the transformation relationship between distribution parameters.

[0040] Figure 3 This is a schematic diagram of the distribution transformation learning framework. In one embodiment, the Transformer encoder (3-layer multi-head self-attention + feedforward network) receives gene distribution parameter embedding sequences, captures the co-expression relationships between genes in the control state, and outputs gene representations containing global context. The Transformer decoder (6 layers) uses the encoder output as the query and the sequence formed by concatenating the perturbation conditional embedding and the cell type embedding as the key. Through a cross-attention mechanism, it injects specific perturbation and cell type information into each gene representation, and finally outputs the distribution parameter residuals by four linear heads.

[0041] Multi-space residual prediction strategy. The model outputs the residual change in the most suitable parameter space for each parameter constraint, and then combines this output with the control group parameters to synthesize the prediction result.

[0042]

[0043]

[0044]

[0045]

[0046] in, , , , These are the zero-inflation parameter, mixture weight, mean, and standard deviation of the control group, respectively. , , , ε represents the residual prediction of the model's output; ε in the standard deviation formula is a very small normal number that ensures σ is strictly greater than zero. For the sigmoid function, Cut off extreme values ​​to To ensure numerical stability; the weights are superimposed in log space and then normalized using softmax to automatically satisfy the simplex constraint; the standard deviation uses multiplicative exponential residuals to ensure... And it has scale invariance; This represents element-wise multiplication. The weights of all four output heads are initialized to zero, so that the model's output is equivalent to the control group distribution (identity mapping) in the early stages of training, ensuring stability in early training.

[0047] Step 103: Model training and distribution consistency optimization.

[0048] During the training phase, the gene distribution parameters corresponding to the actual perturbation group are used as supervision signals to optimize the distribution condition mapping model. The optimization objective is used to measure the consistency between the predicted distribution and the actual distribution. Optimization can be performed using negative log-likelihood loss based on probability distribution; in other embodiments, KL divergence, Wasserstein distance, or other distribution distance metrics can also be used as the training objective.

[0049] Training loss. Monte Carlo Zig-GMM negative log-likelihood (MC-NLL) is employed: first, the training loss is calculated from the true perturbation group distribution parameters. Medium sampling Synthetic expression values Then calculate the predicted distribution of these samples. Negative log-likelihood:

[0050]

[0051] Where B is the number of cell populations in each batch, N is the total number of genes, and S is the number of Monte Carlo samples (default S=50); b, g, and s are the cell population index, gene index, and sampling index, respectively. Here are the distribution parameters predicted by the model for gene g in cell population b, where the sample-by-sample log-likelihood is:

[0052]

[0053] Where x is the synthetic expression level sample value obtained from Monte Carlo sampling; The vector of distribution parameters predicted by the model, including the predicted zero-inflation parameters. Mixing weights of each Gaussian component mean and variance ; The threshold for distinguishing between zero and non-zero values ​​(the same as the aforementioned numerical stability cutoff constant). , These represent the weight, mean, and variance of the k-th Gaussian component in the prediction distribution.

[0054] Logarithmic probability of mixed components Techniques ensure numerical stability; the Gaussian logarithm probability is... .

[0055] Where x is the expression level sample value; This represents the predicted mean of the corresponding Gaussian component; The prediction variance for the corresponding Gaussian component; This is the normalization constant term for the Gaussian distribution; The scale normalization term; the last term / This is the squared standardized Mahalanobis distance, used to measure the degree to which a sample deviates from the mean.

[0056] Step 104: Distribution sampling generation and downstream applications.

[0057] During the inference phase, the distribution parameters of the control group and corresponding conditional information of the sample to be predicted are input into the trained distribution condition mapping model to obtain the gene distribution parameters under the target perturbation state. Based on the predicted distribution parameters, sampling is performed to generate corresponding single-cell expression data, which can be further processed by normalization, logarithmic transformation, etc., for differentially expressed gene analysis, perturbation effect assessment, and downstream biological analysis tasks. Through the above steps, this embodiment of the invention realizes the prediction process from "observed control distribution" to "unobserved perturbation distribution," transforming the traditional single-point expression prediction problem into a distribution-level modeling and transformation problem, thereby effectively reducing the influence of single-cell noise and improving the perturbation effect modeling capability.

[0058] In specific experimental validation and performance evaluation, standard tests were conducted on multiple publicly available standard datasets. Among them, the Parse PBMC dataset contains 61 cytokine perturbations, 12 donor sources, and approximately 850,000 PBMC cells, covering 18 cell types, and is mainly used to evaluate the model's generalization ability across donors and under no-perturbation conditions; the Replogle-Nadig dataset contains over 1,000 CRISPR gene knockout perturbations and approximately 2.4 million cells, and is used to evaluate predictive performance in large-scale gene editing scenarios; the Adamson dataset contains 87 CRISPRi perturbations, targeting the unfolded protein response pathway, and approximately 22,000 cells, and is used to validate the model's applicability in specific pathway regulation scenarios.

[0059] The evaluation metrics all revolve around differentially expressed genes (DEGs), including:

[0060] (1) The Pearson correlation coefficient (DEG-PCC@20) of the Top-20 differentially expressed genes is used to measure the consistency between the predicted and the actual log2 fold change;

[0061] (2) Top-20 differentially expressed genes Jaccard overlap rate (DEG-Overlap@20) is used to assess the degree of overlap between the predicted DEG set and the actual DEG set;

[0062] (3) Differential expression mean square error (DEG-MSE) is used to measure the deviation between the predicted expression level and the actual expression level.

[0063] In comparative experiments with methods such as CPA, GEARS, and scGPT, the embodiments of this invention demonstrated superior performance on multiple datasets. Specifically, on the Parse PBMC dataset, DEG-PCC@20 achieved a score of 0.873, higher than scGPT (0.842), GEARS (0.801), and CPA (0.763); DEG-Overlap@20 achieved a score of 0.341, higher than scGPT (0.125), GEARS (0.243), and CPA (0.216); DEG-MSE achieved a score of 0.213, lower than scGPT (0.241), GEARS (0.278), and CPA (0.312). On the Replogle-Nadig dataset, DEG-PCC@20 achieved a score of 0.801, DEG-Overlap@20 achieved a score of 0.362, and DEG-MSE achieved a score of 0.271, all outperforming the baseline methods. On the Adamson dataset, DEG-PCC@20 scores 0.834, DEG-Overlap@20 scores 0.421, and DEG-MSE scores 0.231, all of which outperform the baseline methods.

[0064] Accordingly, embodiments of the present invention also provide a schematic diagram of a single-cell gene perturbation response prediction system. For example... Figure 4 As shown, the system includes:

[0065] The distribution modeling module 41 is used to divide the original single-cell expression matrix into multiple cell populations according to a preset partitioning strategy. Each cell population contains a predetermined number of cell samples. For each gene in each cell population, a statistical distribution model is constructed based on the expression value of the gene in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene.

[0066] The distribution conversion module 42 is used to construct a distribution condition mapping model based on the gene distribution parameters of the control group. The distribution condition mapping model receives the distribution parameters of the control group as input, and combines gene identification information, perturbation condition information and cell type information to perform condition modeling, and outputs the gene distribution parameters under the target perturbation state.

[0067] The model training module 43 is used to optimize the distribution condition mapping model during the training phase by using the gene distribution parameters corresponding to the perturbation group as a supervision signal.

[0068] The sampling generation module 44 is used to input the control group distribution parameters and corresponding condition information of the sample to be predicted into the trained distribution condition mapping model during the inference stage to obtain the gene distribution parameters under the target perturbation state; and to generate corresponding single-cell expression data based on the predicted distribution parameters.

[0069] The embodiments of the present invention are applicable to various application scenarios such as cytokine stimulation response prediction, gene editing perturbation prediction, and drug effect assessment, and can be deployed as a core prediction component in a single-cell analysis system.

[0070] It should be noted that the embodiments of the present invention are not limited to single-cell RNA sequencing data, and can be extended to various data modeling and prediction tasks with high sparsity, multi-peak distribution or complex statistical characteristics.

[0071] At the data type level, embodiments of the present invention can be extended to scenarios including but not limited to the following:

[0072] (1) Spatial transcriptome data analysis, in which cells can be grouped based on spatial location and the distribution characteristics of different spatial regions can be modeled and predicted;

[0073] (2) Single-cell ATAC sequencing data analysis: By performing distribution parameterization modeling on chromatin accessibility signals, the changes in chromatin state under perturbation conditions can be predicted.

[0074] (3) Multi-omics data modeling, including proteomics, epigenome and other high-dimensional biomolecular data, achieves cross-modal prediction and fusion through unified distribution representation.

[0075] At the model implementation level, the distribution modeling method and distribution transformation model in this embodiment of the invention can be replaced or extended. The distribution modeling module can adopt different types of probability distribution models; the distribution transformation model can be implemented using neural networks with different structures or other nonlinear mapping methods, thereby adapting to different data characteristics and application requirements.

[0076] At the task extension level, embodiments of the present invention can support various complex perturbation modeling tasks, including but not limited to multi-perturbation combination prediction, continuous condition modeling, and cross-domain generalization prediction. For multi-perturbation scenarios, joint effect prediction can be achieved by combining and modeling multiple perturbation conditions.

[0077] Furthermore, in terms of feature representation, the gene or feature embedding can be initialized or replaced using a representation generated by a pre-trained model to improve the model's generalization ability across different datasets.

[0078] Therefore, the distributed modeling and transformation method proposed in this invention has good versatility and scalability, and can be used as a general data modeling framework for various complex data analysis and prediction tasks.

[0079] In summary, the embodiments of the present invention achieve stable modeling and high-precision prediction of single-cell perturbation response by constructing a distribution-level representation and distribution mapping mechanism, and have good generalization ability and scalability.

[0080] The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of the multiple embodiments disclosed in this specification. It should be understood that the above description is only within the scope of this specification. Any modifications, equivalent substitutions, improvements, etc., made based on the technical solutions of the multiple embodiments disclosed in this specification should be included within the protection scope of the multiple embodiments disclosed in this specification.

Claims

1. A method for predicting single-cell gene perturbation response, characterized in that, Includes the following steps: The original single-cell expression matrix is ​​divided into multiple cell populations according to a preset partitioning strategy. Each cell population contains a predetermined number of cell samples. For each gene in each cell population, a statistical distribution model is constructed based on the expression value of the gene in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene. Based on the gene distribution parameters of the control group, a distribution condition mapping model is constructed. The distribution condition mapping model receives the distribution parameters of the control group as input, and combines gene identification information, perturbation condition information and cell type information to perform condition modeling, and outputs the gene distribution parameters under the target perturbation state. During the training phase, the gene distribution parameters corresponding to the perturbation group are used as supervision signals to optimize the distribution condition mapping model; During the inference phase, the control group distribution parameters and corresponding condition information of the sample to be predicted are input into the trained distribution condition mapping model to obtain the gene distribution parameters under the target perturbation state; based on the predicted distribution parameters, sampling is performed to generate the corresponding single-cell expression data.

2. The method of claim 1, wherein, The distribution model adopts a zero-inflated Gaussian mixture model, where each gene is characterized by a zero-inflated parameter, mixture weights, mean, and variance parameters; or it adopts a negative binomial distribution, a zero-inflated negative binomial distribution, or a probability distribution model that characterizes sparsity and multimodality.

3. The method of claim 1, wherein, The mapping model adopts an encoder-decoder structure based on an attention mechanism; or it uses a multilayer perceptron, graph neural network, diffusion model or nonlinear mapping model to realize the transformation relationship between distributed parameters.

4. The method of claim 1, wherein, The optimization objective is used to measure the consistency between the predicted distribution and the true distribution; optimization can be performed using negative log-likelihood loss based on probability distribution, or KL divergence, Wasserstein distance, or distribution distance metric can be used as the training objective.

5. The method of claim 1, wherein, The partitioning strategies include: random partitioning, partitioning based on cell type, or partitioning based on experimental batch.

6. A single-cell gene perturbation response prediction system, comprising: include: The distribution modeling module is used to divide the original single-cell expression matrix into multiple cell populations according to a preset partitioning strategy, with each cell population containing a predetermined number of cell samples. For each gene in each cell population, a statistical distribution model is constructed based on the gene's expression value in the cell population, and the corresponding distribution parameter vector is extracted as the distribution-level representation of the gene. The distribution transformation module is used to construct a distribution condition mapping model based on the gene distribution parameters of the control group. The distribution condition mapping model receives the distribution parameters of the control group as input, and combines gene identification information, perturbation condition information and cell type information to perform condition modeling, and outputs the gene distribution parameters under the target perturbation state. The model training module is used to optimize the distribution condition mapping model during the training phase, using the gene distribution parameters corresponding to the perturbation group as a supervision signal. The sampling generation module is used to input the control group distribution parameters and corresponding condition information of the sample to be predicted into the trained distribution condition mapping model during the inference phase to obtain the gene distribution parameters under the target perturbation state; and to generate corresponding single-cell expression data based on the predicted distribution parameters.

7. The system of claim 6, wherein, The distribution model adopts a zero-inflated Gaussian mixture model, where each gene is characterized by a zero-inflated parameter, mixture weights, mean, and variance parameters; or it adopts a negative binomial distribution, a zero-inflated negative binomial distribution, or a probability distribution model that characterizes sparsity and multimodality.

8. The system of claim 6, wherein, The mapping model adopts an encoder-decoder structure based on an attention mechanism; or it uses a multilayer perceptron, graph neural network, diffusion model or nonlinear mapping model to realize the transformation relationship between distributed parameters.

9. The system of claim 6, wherein, The optimization objective is used to measure the consistency between the predicted distribution and the true distribution; optimization can be performed using negative log-likelihood loss based on probability distribution, or KL divergence, Wasserstein distance, or distribution distance metric can be used as the training objective.

10. The system according to claim 6, characterized in that, The partitioning strategies include: random partitioning, partitioning based on cell type, or partitioning based on experimental batch.

Images