Evaluation method, system and equipment of cell specific heritable variation disturbance and storage medium

Abstract

The invention discloses an evaluation method, system and device for cell specific genetic variation disturbance and a storage medium. An integrated single cell data set is used for training a variational auto-encoder, and a high-dimensional gene expression profile is mapped into a cell state vector in a low-dimensional hidden space through an encoder of the variational auto-encoder; freezing parameters of the encoder, training a conditional disturbance module by using training data with a mononucleotide variation tag, and generating conditional variation embedding; performing anti-fact prediction for target cell and target mononucleotide variation, and generating a reference prediction expression profile and a disturbance prediction expression profile through a conditional disturbance module and a decoder of a variational auto-encoder based on the original genotype and the overturned genotype of the target cell; and calculating the difference between the two to obtain the disturbance effect score of the single nucleotide variation in the target cell, and quantitatively describing and predicting the specific disturbance mode of the SNV in different cell states.

Description

Technical Field

[0001] This invention belongs to the field of single-cell sequencing and relates to a method, system, device and storage medium for evaluating cell-specific genetic variation perturbations. Background Technology

[0002] With the development of high-throughput sequencing technology, genome-wide association studies (GWAS) have successfully identified tens of thousands of genetic loci associated with complex diseases and traits. However, the vast majority (over 90%) of these associated loci are located in non-coding regions, and their specific biological functions and pathogenic mechanisms remain unclear. Therefore, accurately identifying truly functional variants from a massive amount of genetic association signals and elucidating their mechanisms of action in specific cell types and states is a core challenge currently facing the field of functional genomics.

[0003] Single-cell sequencing technology, especially single-cell RNA sequencing (scRNA-seq), has provided an unprecedented high-resolution perspective for solving this problem. By simultaneously acquiring genotype and transcriptome information of individual cells, researchers are able to explore the regulatory role of gene variation on gene expression within detailed cellular lineages.

[0004] Against this backdrop, existing technologies mainly analyze the functional effects of SNVs through the following methods: Single-cell expression quantitative trait locus analysis: Through statistical testing, the genotype of a specific SNV is associated with the expression levels of one or more genes (usually neighboring genes) in a specific cell type.

[0005] Genotype-based differential expression analysis: Cells are grouped according to the genotype of a specific SNV (e.g., wild type, heterozygous mutation, homozygous mutation), and then the differences in gene expression profiles between different groups are directly compared.

[0006] Traditional machine learning and deep learning prediction models: use cell gene expression data and genotype data to train models (such as linear models, tree models or standard neural networks) to predict a global, non-cell state-specific perturbation score.

[0007] While the aforementioned existing technologies provide some clues for understanding SNV functionality, they suffer from three core, insurmountable drawbacks when processing high-dimensional, sparse, and technically noisy single-cell data: First, the inability to effectively distinguish between "causality" and "correlation" leads to confusion in the results.

[0008] Single-cell populations exhibit high heterogeneity, encompassing diverse cell types and states. A specific SNV may happen to be enriched in a particular cell subpopulation. Existing techniques (such as sc-eQTL and differential expression analysis) cannot isolate the gene expression characteristics of the cell subpopulation itself from the additional perturbations that SNVs may introduce when making comparisons. This often results in analytical results that are "spurious associations" between cell type or state and SNV genotype, rather than a "causal effect" of SNVs on gene expression, thus producing a large number of false positive results and seriously misleading downstream functional validation.

[0009] Second, it ignores the "cell state specificity" of gene variation effects, making the predictions too rough.

[0010] Biological laws dictate that the functional effects of the same gene variant can be drastically different, or even completely opposite, in different cell types or states (e.g., resting vs. active). Most existing technologies provide a one-size-fits-all, averaged assessment of global perturbations, neglecting this crucial cellular context-dependent nature. This not only misses key biological information but also fails to explain why the genetic effects of many diseases exhibit high tissue or cell type specificity.

[0011] Third, it is difficult to integrate multi-source data and is sensitive to noise and sparsity in single-cell data.

[0012] Single-cell data commonly suffer from strong batch effects, sparse expression profiles (with a large number of zero values), and high levels of inherent noise. Traditional statistical models and machine learning methods often lack sufficient power to address these challenges. In particular, when it is necessary to integrate multiple datasets from different experiments and platforms to enhance statistical power, the strong batch effects can completely mask weak biological signals, making it difficult for existing techniques to perform robust analyses on a large scale across datasets. Summary of the Invention

[0013] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, device and storage medium for evaluating cell-specific genetic variation perturbations, accurately inferring the causal effects of SNVs in the context of complex cellular heterogeneity, and quantitatively describing and predicting the specific perturbation patterns of SNVs in different cellular states.

[0014] To achieve the above objectives, the present invention employs the following technical solution: A method for assessing cell-specific genetic variation perturbations includes the following procedures: S1. Obtain single-cell transcriptome data and corresponding genotype data; train a variational autoencoder using the integrated single-cell dataset, and map the high-dimensional gene expression profile into a cell state vector in a low-dimensional latent space through the encoder of the variational autoencoder. S2, freeze the encoder parameters, train the conditional perturbation module using training data with single nucleotide variant labels, the conditional perturbation module takes the cell state vector and the single nucleotide variant embedding vector as input, and generates conditional variant embeddings; S3 performs counterfactual prediction for the target cell and the target single nucleotide variant. While keeping the cell state vector generated by the encoder unchanged, it generates a reference predicted expression profile and a perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the conditional perturbation module and the decoder of the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

[0015] Optionally, the steps for training the variational autoencoder include: By concatenating batch embedding vectors at the decoder input, the encoder is forced to learn cell state vectors that are independent of batches, thus removing the technical batch effect from different experimental sources. We model the counting features of gene expression data using a negative binomial distribution likelihood function, and denoise high-dimensional sparse single-cell data by learning a low-dimensional latent space.

[0016] Optionally, in S1: a variational autoencoder backbone network is trained using a large-scale public single-cell dataset to learn a general representation of cell biological states. In S2: Load the pre-trained variational autoencoder backbone network weights and train modules related to single nucleotide variation perturbations using a task-specific dataset with genotype labels.

[0017] Optionally, the steps for generating conditional variant embeddings include: The cell state vector in the low-dimensional latent space is concatenated with the original embedding vector of the single nucleotide variant. The spliced ​​vector is input into a multilayer perceptron, which outputs a conditionalized variant embedding that represents the manifestation of the single nucleotide variant in a specific cellular state.

[0018] Optionally, the step of performing counterfactual prediction includes generating a perturbation vector: The conditional variant embeddings of multiple single nucleotide variants within the cell are aggregated using the attention mechanism to generate the final perturbation vector; The weights of the attention mechanism are conditioned on the cell state vector, and the contribution of different variants is dynamically calculated. The final perturbation vector is superimposed on the cell state vector to obtain the perturbed cell state vector.

[0019] Optionally, the steps of performing counterfactual prediction include generating potential counterfactual representations: Establish a causal effect inference framework based on counterfactual prediction and simulate a control experiment; While keeping the cell state vector constant, the perturbation module is used to calculate the perturbation of the cell state vector when a single nucleotide variant is transformed into a reference allele and a mutant allele, respectively. The decoder is used to map the cell state vectors with and without perturbations back to the gene expression space, generating two counterfactual gene expression profiles.

[0020] Optionally, the step of calculating the perturbation effect fraction of the single nucleotide variant in the target cell includes: By comparing the reference predicted expression profile and the perturbation predicted expression profile, the absolute difference, cosine distance or logarithmic fold change between the two are calculated to quantify the multidimensional perturbation fraction of single nucleotide variants at the single-cell level. Obtain cell type annotation information, and group and aggregate perturbation effect scores at the single-cell level based on the cell type annotation information; The aggregated statistical values ​​are calculated to obtain the specific perturbation effect score of single nucleotide variants in a specific cell type.

[0021] A system for assessing cell-specific genetic variation perturbations, comprising: The pre-training module is used to acquire single-cell transcriptome data and corresponding genotype data; the integrated single-cell dataset is used to train a variational autoencoder, which maps the high-dimensional gene expression profile into a cell state vector in a low-dimensional latent space through the encoder of the variational autoencoder. The conditional perturbation training module is used to freeze the encoder parameters and train the conditional perturbation module using training data with single nucleotide variant labels. The conditional perturbation module receives cell state vectors and single nucleotide variant embedding vectors as inputs and generates conditional variant embeddings. The perturbation effect score acquisition module is used to perform counterfactual prediction for target cells and target single nucleotide variants. While keeping the cell state vector generated by the encoder unchanged, it generates a reference predicted expression profile and a perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the conditional perturbation module and the decoder of the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

[0022] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for evaluating cell-specific genetic variation perturbations.

[0023] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for evaluating cell-specific genetic variation perturbations.

[0024] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a deep generative model based on variational autoencoders, utilizing conditional perturbation modules and counterfactual prediction mechanisms. It maps high-dimensional, sparse single-cell data into a low-dimensional latent space to capture the intrinsic biological state of cells, while simultaneously achieving effective integration and denoising of multi-source data through batch embedding at the decoder end. This method simulates gene expression changes after single nucleotide variant (SNV) state flips while maintaining cell state constancy, effectively removing the confounding effects of cell state and accurately inferring the causal effect of SNVs on gene expression rather than simple statistical correlation. Furthermore, by modeling perturbation effects as functions of cell state, it can capture the differentiated performance of the same variant in different cell types or states, thus providing a high-resolution, cell-state-specific functional perturbation map. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process for evaluating cell-specific genetic variation perturbations according to the present invention. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0028] The core objective of this invention is to address a key challenge in the field of functional genomics: how to accurately identify biologically functional variants from a vast amount of gene variants (especially single nucleotide variants, SNVs) that are specific to cell type / state.

[0029] like Figure 1 As shown, to achieve this goal, this invention proposes a complete deep learning framework. This framework takes single-cell transcriptome (scRNA-seq) and corresponding genotype data as input, and through an innovative conditional and generative model, quantitatively evaluates the perturbation effect of each SNV in each single cell using counterfactual prediction, ultimately outputting a cell-state-specific, quantifiable perturbation map.

[0030] Phase 1: Model pre-training (general knowledge learning).

[0031] Data integration: Merging large-scale, multi-source public single-cell datasets with the user's own partial training data (if any).

[0032] Unified batch correction: Create a unified batch label and use the decoder-side conditionalization mechanism in the architecture of this invention to correct batch effects on the merged data.

[0033] Backbone Model Training: On the calibrated data, a variational autoencoder (VAE) backbone network is trained. The purpose of this step is to allow the model to learn a general, batch-independent, low-dimensional latent space z that can represent universal cellular biological states. This z-space serves as the "base map" for all our subsequent analyses.

[0034] Phase Two: Model Fine-tuning (Task-Specific Learning).

[0035] Loading the backbone: Load the weights of the pre-trained VAE backbone (mainly including the encoder and decoder) into a more complete SNVPerturbationModel.

[0036] Freeze and train: Typically, the encoder weights are frozen (because it has already learned how to build the “cell map”), and then user-owned data with SNV labels is used to specifically train modules related to SNV perturbations (such as snv_embedding, cond_mlp, attn_mlp, snv_proj).

[0037] Phase 3: Inference and scoring of disturbance effects.

[0038] Counterfactual prediction: For each SNV that needs to be evaluated, perform a "counterfactual" calculation once for each cell: Reference prediction (mu_ref): Predicts the gene expression profile based on the original genotype of the cell.

[0039] Perturbation prediction (mu_alt): While keeping the cell state z unchanged, flip the state of the SNV in its genotype (e.g., change from wild type to mutant type) and predict its gene expression profile again.

[0040] Effect quantification: The multidimensional perturbation score of the SNV in the cell is obtained by comparing the differences between the reference prediction and the perturbation prediction (e.g., absolute difference, cosine distance, log2 fold change, etc.).

[0041] Cell type-specific aggregation: Finally, based on the cell type annotation, the scores at the single-cell level are grouped and aggregated (e.g., the median) to obtain the final perturbation effect score for each SNV in each cell type.

[0042] The method described in this embodiment is designed in the following parts: Encoder and latent space z: z_mu, z_log_var=Encoder(X) Function: To map a high-dimensional, sparse gene expression profile X into a low-dimensional, dense latent variable z that contains the state of cell biology.

[0043] Decoder-side batch calibration: X_reconstructed=Decoder(z,batch_embedding) Function: By concatenating batch_embedding at the decoder input, the encoder is forced to learn a batch-independent z, which is key to achieving data integration.

[0044] Conditional perturbation module: e_cond=cond_mlp(concat(z,e_j)) Function: e_j is the original embedding of SNVj, and z is the cell state. This module computes a conditional SNV embedding e_cond, which means "the specific representation of SNVj in cell state z". This is the basis for achieving cell state specificity.

[0045] Counterfactual perturbation generation: Delta_latent=snv_proj(attention_agg(e_cond,G)) z_perturbed=z+Delta_latent mu=Decoder(z_perturbed,batch_embedding) Function: Based on the cell's original genotype G and conditional embedding e_cond, an attention mechanism is used to aggregate and generate a final perturbation vector Delta_latent, which is then applied to z to generate the counterfactual expression spectrum mean mu. The entire scoring process essentially compares the difference between the two mu generated when G is the original value and the flipped value.

[0046] The above three stages of this embodiment have the following technical points: 1. Achieve causal effect inference and accurately identify functional variations: To address the significant shortcoming of existing technologies that cannot effectively distinguish between the "causality" and "correlation" between gene mutations and cellular states, leading to confusing results, this invention constructs a deep generative model based on counterfactual prediction. Its specific advantages are as follows: Decoupling cell state from genotype effect: The encoder of this invention first compresses high-dimensional, complex single-cell transcriptome data into a low-dimensional latent variable z, which captures the intrinsic biological state of the cell (such as cell type, cell cycle, etc.).

[0047] Simulated “control experiment”: When evaluating the effect of a certain SNV, this invention calculates the perturbation vector Delta_latent of z when the SNV is the reference allele (-1) and the mutant allele (+1) respectively through the perturbation module, and generates two counterfactual expression profiles (mu_ref and mu_alt) by the decoder.

[0048] Inferring causal relationships: By comparing the differences between mu_ref and mu_alt, this invention can effectively isolate expression profile differences caused by different cell states (z), thereby separating the causally oriented perturbation effect directly caused by the SNV. This design fundamentally solves the problem of "causal-correlation" confusion in the prior art, significantly reduces the false positive rate, and makes the identified highly perturbation SNV more likely to be a true functional variant.

[0049] 2. Revealing cell state specificity and providing high-resolution perturbation maps: To address the shortcomings of existing technologies that ignore cell context dependence and provide a "one-size-fits-all" average effect, this invention designs a conditional perturbation module. Its advantages are: Dynamic, context-aware perturbation modeling: In this invention, the perturbation effect vector Delta_latent of an SNV is not a fixed value that is the same for all cells. Instead, it is dynamically calculated and generated by a neural network (cond_mlp and attn_mlp) conditioned on the cell's hidden state z.

[0050] Achieving cell state-specific prediction: Since the perturbation effect is a function of the cell state z, and z precisely reflects the cell type and state, the predicted perturbation effect of the same SNV differs in different cells (i.e., different z vectors). This allows the present invention to generate high-resolution, cell state-specific perturbation maps, accurately depicting in which cell subpopulations a specific SNV has a strong effect and in which subpopulations it is neutral, which highly matches the actual working pattern of complex genetic regulation in organisms.

[0051] 3. The model is robust and scalable, and can integrate multi-source heterogeneous data: To address the challenges posed by existing technologies in handling the inherent noise, sparsity, and batch effects of single-cell data, this invention employs a combined set of technical solutions to ensure the model's high robustness and scalability.

[0052] Denoising and Representation Learning Based on Deep Generative Models: This invention uses a variational autoencoder (VAE) as the backbone architecture. By learning a low-dimensional latent space z, it can effectively denoise high-dimensional, sparse single-cell data and capture its inherent manifold structure, thereby enhancing the model's resistance to technical noise.

[0053] Conditional batch effect correction: This invention introduces batch embedding as a condition at the decoder end, which can explicitly model and remove technical batch effects from different experimental sources. This allows the encoder to learn a batch-invariant biological representation, enabling the integration of large-scale, heterogeneous datasets from multiple sources.

[0054] Knowledge Enhancement Based on Transfer Learning: This invention supports a "pre-training-fine-tuning" paradigm. By pre-training on large, diverse, publicly available single-cell datasets, the model can learn general prior knowledge about cell states, and then be fine-tuned on new, smaller datasets, significantly improving the model's learning efficiency and prediction accuracy when data volume is limited.

[0055] 4. A framework for causal effect inference based on counterfactual prediction: This invention does not employ traditional direct association analysis. Instead, it uses a deep generative model to compare counterfactual gene expression profiles when there is / is no specific gene mutation (or state reversal) while keeping the cell's intrinsic state (latent variable z) constant. This allows the confounding effects of cell state to be removed, and the causal effect directly caused by the mutation to be inferred.

[0056] 5. The mechanism for generating conditional perturbations based on cellular latent states: This invention proposes that the perturbation effect of gene mutations is not static or globally averaged, but depends on the biological context of the cell. Its core technology lies in the fact that the perturbation effect vector (Delta_latent) is dynamically generated through a neural network module conditioned on the cell's latent state z and the mutation identity e_j, thereby achieving cell-state-specific modeling and prediction of the perturbation effect.

[0057] 6. Robust model architecture integrating batch effect correction and transfer learning: This invention designs a unique deep learning architecture, which is as follows: a) Introducing batch information as a condition at the decoder end enables the encoder to learn batch-invariant pure cell representations, effectively integrating heterogeneous data.

[0058] b) It supports pre-training on large, diverse public datasets and then fine-tuning on task-specific data (transfer learning), which significantly enhances the robustness of the model and its generalization ability on small sample data.

[0059] 7. Cell state-dependent attention mechanisms for aggregating multivariable effects: To address the situation where multiple gene variants may coexist within a single cell, this invention employs an attention mechanism. The weights of this mechanism are also conditioned on the cell's latent state z, enabling dynamic and selective aggregation of contributions from different variants to calculate the final combined perturbation effect.

[0060] In summary, this invention overcomes many limitations of existing technologies through its unique counterfactual prediction, conditional perturbation modeling, and robust data integration capabilities, providing an innovative and effective technical solution for accurately and efficiently deciphering the function of non-coding genetic variations at the single-cell level.

[0061] The following explains part of the process: Variational autoencoders: The encoder compresses high-dimensional gene expression into the latent space, while the decoder reconstructs the expression. RNA expression data is count-type data and suffers from excessive dispersion; NBlikelihood can realistically model expression noise.

[0062] Batch Embedding: Different experimental batches can introduce systematic biases. Batch embedding can model these biases in the latent space, preventing the backbone from learning "batch effects" instead of biological signals.

[0063] Pre-trained Backbone: Learns generalized encoding / decoding structures on large-scale public RNA data. This prevents model instability caused by insufficient data during joint training of SNV+RNA.

[0064] Conditional Embedding: The same SNV may have different effects in different cellular contexts. Conditional embedding can make the SNV effect dependent on the cell state.

[0065] Attention Mechanism: A cell may carry multiple SNVs, but not all of them significantly perturb gene expression. Attention can automatically focus on "functional SNVs".

[0066] CounterfactualDecoding: Directly simulates "what would happen if / without this SNV", achieving causal inference instead of relying solely on correlation.

[0067] PerturbationMetrics: Captures both “intensity” and “pattern variation” simultaneously, avoiding the need to look only at mean differences.

[0068] Implementation application scenarios: Implementation Scenario: In skin cancer research, researchers collected single-cell sequencing data from a batch of squamous cell carcinoma patients and examined the single-cell mutations (SNVs). They hoped to identify, from thousands of SNVs, the key mutations that have the greatest impact on the function of tumor cells or immune cells.

[0069] Application of this invention: Researchers integrate their own data with multi-source single-cell data on skin cancer, head and neck cancer, etc., from public databases such as TCGA.

[0070] Using the first stage of this invention, a powerful skin / squamous cell carcinoma state model is obtained by pre-training on the integrated "super dataset".

[0071] Using the second stage of this invention, fine-tuning is performed on one's own data.

[0072] Using Phase 3 of the present invention, cell type-specific scoring is performed on all SNVs.

[0073] Output: A detailed report indicated, for example, that "SNVrs12345 significantly upregulated proliferation-related pathways in cancerous basal cells, but had almost no effect in tumor-infiltrating T cells," while "SNVrs67890 primarily attenuated the cytotoxic function of T cells." These specific, cell-type-specific conclusions provide invaluable clues for subsequent experimental validation and drug target development.

[0074] Regarding causal inference capabilities: Existing technologies mainly perform statistical correlations and cannot distinguish between causation and correlation. This invention, through counterfactual prediction, compares the results after controlling for cell state variables, and is able to infer causal effects.

[0075] Regarding cell specificity: Existing technologies provide global, averaged effect assessments that ignore cell context. This invention, through a conditional perturbation module, predicts different effects for the same SNV in different cell types.

[0076] Regarding data integration and robustness: Existing technologies struggle to handle batch effects and are sensitive to noise and sparsity. This invention, through the cVAE architecture and transfer learning, effectively integrates multi-source data and is robust to noise and sparsity.

[0077] Explanation of terms in this embodiment: Single-cell transcriptomics (scRNA-seq): a technique that can measure the expression levels of thousands of genes at the single-cell level.

[0078] SNV (Single Nucleotide Variation): A change in a single nucleotide in the genomic DNA sequence.

[0079] VAE (Variational Autoencoder): A deep generative model that excels at learning low-dimensional latent representations (latent space) of data and generating new data.

[0080] Latent Space z: A low-dimensional, dense, and meaningful mathematical representation of the original high-dimensional data (such as cell expression profiles) learned by the VAE.

[0081] Counterfactual prediction: Simulating a "what if...?" scenario. In this invention, it means "what would the expression profile be like if this cell did not have this mutation?".

[0082] Batch effect: Systematic bias in data caused by non-biological factors (such as different experimental dates, reagents, or operators).

[0083] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not omitted in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0084] In another embodiment of the present invention, an evaluation system for cell-specific genetic variation perturbations is provided. This evaluation system can be used to implement the above-mentioned evaluation method for cell-specific genetic variation perturbations. Specifically, the evaluation system for cell-specific genetic variation perturbations includes a pre-training module, a conditional perturbation training module, and a perturbation effect score acquisition module.

[0085] The pre-training module is used to acquire single-cell transcriptome data and corresponding genotype data; the integrated single-cell dataset is used to train a variational autoencoder, which maps the high-dimensional gene expression profile into a cell state vector in a low-dimensional latent space.

[0086] The conditional perturbation training module is used to freeze the encoder parameters. The conditional perturbation module is trained using training data with single nucleotide variant labels. The conditional perturbation module receives cell state vectors and single nucleotide variant embedding vectors as inputs and generates conditional variant embeddings.

[0087] The perturbation effect score acquisition module is used to perform counterfactual prediction for target cells and target single nucleotide variants. While keeping the cell state vector generated by the encoder unchanged, it generates a reference predicted expression profile and a perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the conditional perturbation module and the decoder of the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

[0088] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment can be used in the operation of a method for evaluating cell-specific genetic variation perturbations, including: S1, acquiring single-cell transcriptome data and corresponding genotype data; training a variational autoencoder using the integrated single-cell dataset, and mapping the high-dimensional gene expression profile to a cell state vector in a low-dimensional latent space through the encoder of the variational autoencoder; S... 2. Freeze the encoder parameters and train the conditional perturbation module using training data with single nucleotide variant labels. This conditional perturbation module receives the cell state vector and the single nucleotide variant embedding vector as input and generates the conditional variant embedding. S3. Perform counterfactual prediction for the target cell and the target single nucleotide variant. While keeping the cell state vector generated by the encoder unchanged, generate the reference predicted expression profile and the perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the decoder of the conditional perturbation module and the variational autoencoder. Calculate the difference between the reference predicted expression profile and the perturbation predicted expression profile to obtain the perturbation effect score of the single nucleotide variant in the target cell.

[0089] In another embodiment, the present invention also provides a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here may include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0090] One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for evaluating cell-specific genetic variation perturbations in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps: S1, acquire single-cell transcriptome data and corresponding genotype data; train a variational autoencoder using the integrated single-cell dataset, and map the high-dimensional gene expression profile to a cell state vector in a low-dimensional latent space through the encoder of the variational autoencoder; S2, freeze the encoder parameters, and use the training with single nucleotide variant tags. The data training conditional perturbation module receives cell state vectors and single nucleotide variant embedding vectors as inputs to generate conditional variant embeddings. S3, counterfactual prediction is performed on the target cell and the target single nucleotide variant. While keeping the cell state vector generated by the encoder unchanged, reference predicted expression profiles and perturbation predicted expression profiles are generated based on the original genotype and the flipped genotype of the target cell, respectively, through the decoder of the conditional perturbation module and the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

[0091] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0096] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

[0100] It should be understood that the above description is for illustrative purposes and not for limitation. Many embodiments and applications beyond the provided examples will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of this patent should not be determined by reference to the above description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed inventive subject matter.

Claims

1. A method for assessing cell-specific genetic variation perturbations, characterized in that, Includes the following processes: S1. Obtain single-cell transcriptome data and corresponding genotype data; train a variational autoencoder using the integrated single-cell dataset, and map the high-dimensional gene expression profile into a cell state vector in a low-dimensional latent space through the encoder of the variational autoencoder. S2, freeze the encoder parameters, train the conditional perturbation module using training data with single nucleotide variant labels, the conditional perturbation module takes the cell state vector and the single nucleotide variant embedding vector as input, and generates conditional variant embeddings; S3 performs counterfactual prediction for the target cell and the target single nucleotide variant. While keeping the cell state vector generated by the encoder unchanged, it generates a reference predicted expression profile and a perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the conditional perturbation module and the decoder of the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

2. The method for evaluating cell-specific genetic variation perturbations according to claim 1, characterized in that, The steps for training a variational autoencoder include: By concatenating batch embedding vectors at the decoder input, the encoder is forced to learn cell state vectors that are independent of batches, thus removing the technical batch effect from different experimental sources. We model the counting features of gene expression data using a negative binomial distribution likelihood function, and denoise high-dimensional sparse single-cell data by learning a low-dimensional latent space.

3. The method for evaluating cell-specific genetic variation perturbations according to claim 1, characterized in that, In S1: A variational autoencoder backbone network is trained using a large-scale public single-cell dataset to learn general cell biological state representations. In S2: Load the pre-trained variational autoencoder backbone network weights and train modules related to single nucleotide variation perturbations using a task-specific dataset with genotype labels.

4. The method for evaluating cell-specific genetic variation perturbations according to claim 1, characterized in that, The steps for generating conditional variant embeddings include: The cell state vector in the low-dimensional latent space is concatenated with the original embedding vector of the single nucleotide variant. The spliced ​​vector is input into a multilayer perceptron, which outputs a conditionalized variant embedding that represents the manifestation of the single nucleotide variant in a specific cellular state.

5. The method for evaluating cell-specific genetic variation perturbations according to claim 4, characterized in that, The steps for performing counterfactual predictions include generating a perturbation vector: The conditional variant embeddings of multiple single nucleotide variants within the cell are aggregated using the attention mechanism to generate the final perturbation vector; The weights of the attention mechanism are conditioned on the cell state vector, and the contribution of different variants is dynamically calculated. The final perturbation vector is superimposed on the cell state vector to obtain the perturbed cell state vector.

6. The method for evaluating cell-specific genetic variation perturbations according to claim 1, characterized in that, The steps involved in performing counterfactual prediction include generating potential representations of counterfactual facts: Establish a causal effect inference framework based on counterfactual prediction and simulate a control experiment; While keeping the cell state vector constant, the perturbation module is used to calculate the perturbation of the cell state vector when a single nucleotide variant is transformed into a reference allele and a mutant allele, respectively. The decoder is used to map the cell state vectors with and without perturbations back to the gene expression space, generating two counterfactual gene expression profiles.

7. The method for evaluating cell-specific genetic variation perturbations according to claim 1, characterized in that, The steps for calculating the perturbation effect fraction of the single nucleotide variant in the target cell include: By comparing the reference predicted expression profile and the perturbation predicted expression profile, the absolute difference, cosine distance or logarithmic fold change between the two are calculated to quantify the multidimensional perturbation fraction of single nucleotide variants at the single-cell level. Obtain cell type annotation information, and group and aggregate perturbation effect scores at the single-cell level based on the cell type annotation information; The aggregated statistical values ​​are calculated to obtain the specific perturbation effect score of single nucleotide variants in a specific cell type.

8. A system for assessing cell-specific genetic variation perturbations, characterized in that, include: The pre-training module is used to acquire single-cell transcriptome data and corresponding genotype data. A variational autoencoder was trained using the integrated single-cell dataset. The encoder of the variational autoencoder mapped the high-dimensional gene expression profile into a cell state vector in a low-dimensional latent space. The conditional perturbation training module is used to freeze the encoder parameters and train the conditional perturbation module using training data with single nucleotide variant labels. The conditional perturbation module receives cell state vectors and single nucleotide variant embedding vectors as inputs and generates conditional variant embeddings. The perturbation effect score acquisition module is used to perform counterfactual prediction for target cells and target single nucleotide variants. While keeping the cell state vector generated by the encoder unchanged, it generates a reference predicted expression profile and a perturbation predicted expression profile based on the original genotype and the flipped genotype of the target cell, respectively, through the conditional perturbation module and the decoder of the variational autoencoder. The difference between the reference predicted expression profile and the perturbation predicted expression profile is calculated to obtain the perturbation effect score of the single nucleotide variant in the target cell.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for evaluating cell-specific genetic variation perturbations as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for evaluating cell-specific genetic variation perturbations as described in any one of claims 1 to 7.

Images