A method for spatiotemporal modeling of cell development by fusing single-cell and spatial transcriptomes
Patent Information
- Application Number
- CN202611154866.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-10-16
AI Technical Summary
[0021]向量场构建于纯基因表达虚拟空间，无法与组织真实三维物理坐标耦合，不能模拟细胞空间迁移、器官形态重塑；
[0060] Compared with existing scGPT static pre-trained models and Dynamo single-cell vector field dynamics models, this invention has five core technological advantages, all of which have quantifiable and objective results:
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computational biology, spatiotemporal multi-omics artificial intelligence modeling technology, and in particular to a spatiotemporal modeling method for cell development that integrates single-cell and spatial transcriptomics. Background Technology
[0002] As life sciences enter the era of spatiotemporal multi-omics, single-cell transcriptome sequencing (scRNA-seq) and spatial transcriptomics (ST) have become core pillars for elucidating the mechanisms of organismal development, aging, and complex diseases. However, how to effectively integrate these two types of heterogeneous data to construct a spatiotemporal dynamic evolutionary world model that is cross-scale, continuous, and capable of causal inference remains a core bottleneck that urgently needs to be addressed in the field of computational biology.
[0003] Single-cell transcriptome sequencing (scRNA-seq): This method breaks down tissues into individual single cells and detects the gene expression levels of each cell's entire genome, which can accurately distinguish cell types and functional states. However, the tissue breakdown process is irreversible, and the three-dimensional spatial coordinates of cells within organs are completely lost. Only a static snapshot of cell expression at discrete time points can be obtained, which cannot characterize the continuous spatiotemporal evolution of cells.
[0004] Spatial transcriptomics (ST) involves in situ sequencing of tissue sections to simultaneously record gene expression and the three-dimensional spatial geometric coordinates (X / Y / Z) of cells. However, existing sequencing technologies produce data with high sparsity, high noise, and difficulty in aligning discrete sections. They can only obtain static tissue snapshots and cannot model the continuous differentiation, migration, and tissue remodeling processes of cells over time.
[0005] The development and pathological evolution of living organisms are essentially multi-scale, spatiotemporally synchronous, and continuous dynamic processes involving molecules, cells, tissues, and organs. However, existing computational analysis tools suffer from severe technological fragmentation, with mainstream technical approaches falling into two categories, both of which have inherent defects:
[0006] Similar existing technology 1: Single-cell pre-trained large model based on Transformer architecture.
[0007] Typical examples:
[0008] scGPT framework: Cui, H., et al. "scGPT: towards a generative foundation model for single-cell biology." Nature Methods (2024).
[0009] Geneformer model: Theis, CV, et al. "Transfer learning with Geneformer for single-cell chromatin and expression analytics." Nature (2023).
[0010] This scheme uses genes as tokens and cell expression profiles as text, relying on masked gene prediction self-supervised tasks to complete pre-training, and is only used for static cell typing and batch coarse correction.
[0011] Core flaw:
[0012] The model has no spatial coordinate encoding logic at the underlying level and completely ignores the three-dimensional topology of the tissue and the microenvironment of cell neighborhood communication.
[0013] Statistical distribution fitting based solely on discrete samples lacks continuous-time dynamic constraints, making it impossible to extrapolate long-term cell differentiation trajectories, and long-term time-series predictions will exhibit severe mathematical drift.
[0014] It can only explore gene expression correlations, but it does not have the ability to make causal inferences or simulate counterfactual disturbances of multiple genes / drugs.
[0015] Similar Existing Technology 2: Single-cell dynamics model based on RNA velocity and vector field
[0016] Typical examples:
[0017] Dynamo framework: Qiu, X., et al. "Mapping transcriptomic vector fields of single cells." Cell (2022).PDF
[0018] Fundamentals of RNA Velocity: La Manno, G., et al. "RNA velocity of single cells." Nature (2018).
[0019] This scheme calculates RNA velocity using the ratio of cleaved to uncleaved mRNA and constructs a dynamic vector field in a high-dimensional gene expression space, which can predict short-term functional changes in a single cell.
[0020] Core flaw:
[0021] Vector fields are constructed in a pure gene expression virtual space, which cannot be coupled with the real three-dimensional physical coordinates of tissues and cannot simulate cell spatial migration and organ morphological remodeling;
[0022] Lacking a gene-cell-tissue-organ cross-scale transmission mechanism, underlying gene perturbations cannot be mapped to macroscopic tissue structural changes;
[0023] It has poor robustness to high-noise clinical samples and heterogeneous samples from multiple batches, and the accuracy of prediction of multi-gene combination perturbations drops significantly.
[0024] In summary, existing technologies either lack spatial topological information and can only perform static statistical fitting, or they can only construct dynamics in the single-cell expression space and cannot achieve spatiotemporal alignment with real tissues. There is a lack of a complete technical solution that can realize multi-source omics fusion, cross-scale continuous dynamic modeling, causal counterfactual perturbation simulation, and self-evolutionary closed loop of dry and wet experiments. Summary of the Invention
[0025] In view of all the shortcomings of the existing technology, the technical problems to be solved by the present invention are divided into four categories:
[0026] (1) It is impossible to model continuous spatiotemporal dynamic evolution. The mainstream models can only statically represent learning, lack the mechanism of life dynamics, and cannot systematically simulate the cross-scale continuous evolution from molecules to organs.
[0027] (2) Spatial information is disconnected from developmental time topology. Traditional pseudo-time trajectories have no spatial constraints, making it difficult to mathematically align discrete slices and resulting in insufficient alignment accuracy at the gene-cell-tissue level.
[0028] (3) It lacks the ability to make causal inferences and counterfactual simulations, and can only describe expression correlations. It cannot predict the long-term evolution of cells and tissues after gene knockout or drug intervention.
[0029] (4) The model relies on a massive amount of parameters and labeled data, which is not robust to clinical samples with high heterogeneity and high noise. It has weak generalization ability across species and in scenarios without disturbance, and poor black box interpretability.
[0030] The purpose of this invention is to provide a spatiotemporal modeling method for cell development that integrates single-cell and spatial transcriptomics. Through a complete set of technologies, including individual heterogeneity decoupling preprocessing, multi-task customized pre-training, hierarchical cross-scale graph neural networks, continuous dynamic modeling of neural ordinary differential / stochastic differential equations, biological causal constraint counterfactual simulation, and closed-loop iteration of multi-agent dry and wet experiments, this method achieves high-fidelity, interpretable, and interventionist spatiotemporal digital twin simulation of virtual organs, overcoming the shortcomings of existing static fitting, spatiotemporal fragmentation, lack of causal inference, and high computational and data requirements.
[0031] The technical solution provided by this invention is: a method for spatiotemporal modeling of cell development integrating single-cell and spatial transcriptomics, characterized by comprising the following steps:
[0032] Acquire multi-source heterogeneous spatiotemporal multi-omics data, including at least two of the following: single-cell transcriptome data, spatial transcriptome data, epigenomics data, and cell lineage tracing data;
[0033] The multi-source heterogeneous spatiotemporal multi-omics data is subjected to noise robust cleaning, and individual-specific embedding vectors are introduced at the data input end. Through adversarial training or decoupled representation learning, individual heterogeneous noise and common laws of life development are mathematically decoupled to eliminate interference from experimental batches and donor differences.
[0034] The cleaned and decoupled data is input into a multi-task table learning pre-training framework, and the model is pre-trained through a joint loss function to obtain a unified shared representation space; the joint loss function includes at least masked gene expression prediction loss, cross-modal contrastive learning loss, and dynamic continuity constraint loss.
[0035] A bottom-up hierarchical graph neural network structure is constructed, which includes at least a gene regulation network layer, a cell function embedding layer, a tissue topology microenvironment layer, and an organ morphology and anatomy layer. The perturbation of the transcriptional expression of the underlying genes is dynamically transmitted and mapped to the macroscopic tissue morphology remodeling through a cross-scale attention mechanism.
[0036] The pre-trained representation is input into a continuous spatiotemporal evolution model based on neural regular differential equations or neural stochastic differential equations. The rate of change of cell state over time is modeled as a continuous velocity field parameterized by a neural network. Combined with the optimal transmission algorithm as a boundary constraint, the continuous interpolation alignment of spatial geometric coordinates and time axis is achieved to obtain the spatiotemporal continuous trajectory of cell development.
[0037] A perturbation control vector is introduced into the neural constant differential equation or neural stochastic differential equation, and a biological knowledge graph or causal constraint matrix is embedded in the derivation trajectory as a hard boundary condition. Virtual intervention counterfactual simulation of single genes or combination genes is performed, and the predicted results of cell fate shift after perturbation are output.
[0038] As an improvement, the noise robustness cleaning includes: processing high-noise, sparse, and heterogeneous data through a noise robust layer and a weighted loss function; the individual-specific embedding vector and the cell expression matrix are input in parallel, and the decoupling of individual differences and common patterns is achieved through a domain adversarial network or a conditional variational autoencoder.
[0039] As an improvement, the joint loss function is expressed as:
[0040] L_total = L_masked + L_contrastive + L_dynamic
[0041] Among them, L_masked is the masked gene expression prediction loss, which is used to mask 15% to 30% of the gene expression values in the cell expression matrix to learn the co-expression and regulatory network among genes; L_contrastive is the cross-modal contrastive learning loss, which is used to maximize the mutual information between different modalities in the same tissue microenvironment; and L_dynamic is the dynamic continuity constraint loss, which is used to constrain the spatiotemporal continuity of the physical topology of life evolution.
[0042] As an improvement, the hierarchical graph neural network structure is constructed as follows:
[0043] At the gene level, regulatory characteristics are learned through gene association topological networks;
[0044] At the cellular level, expression patterns are aggregated into cellular state vectors;
[0045] At the organizational level, a local microenvironment dynamics graph is constructed using cells as graph nodes and spatial physical coordinates and intercellular communication networks as graph edges.
[0046] At the organ level, evolutionary inferences are made regarding long-term topological remodeling;
[0047] The cross-scale attention mechanism employs a hierarchical perception mechanism, transmitting the dynamic changes in gene expression layer by layer to tissue morphology evolution.
[0048] As an improvement, the continuous spatiotemporal evolution model is as follows:
[0049] dZ(t) / dt = fθ(Z(t), t, u)
[0050] Where Z(t) represents the state trajectory encoding of the cell in the high-dimensional semantic space, t represents the developmental stage or continuous time axis, u represents the input perturbation label vector, and fθ is the continuously changing velocity field parameterized by the neural network; the optimal transmission algorithm is used to process the state alignment between discrete time slices to achieve dynamic fusion of continuous time and three-dimensional spatial geometric coordinates.
[0051] As an improvement, the virtual intervention counterfactual simulation includes single gene knockout, multi-gene combination knockout, gene overexpression, or drug molecular intervention; the causal constraint matrix contains known developmental regulatory pathways, ligand-receptor databases, and physical constraint information, which are used to impose hard boundary constraints on the inference trajectory generated by the neural constant differential equation, ensuring the biological rationality of the counterfactual inference.
[0052] As an improvement, a top-level multi-agent cooperative scheduling step is also included:
[0053] The scientific research intelligent agent architecture automatically performs problem planning, code generation, and algorithm execution within a Docker sandbox.
[0054] Output the target or mechanism hypothesis predicted by counterfactual simulation to the wet experiment verification stage;
[0055] The wet experiment validation results, including both successful and unsuccessful cases, are automatically converted into a negative sample stream and fed back into the model for incremental parameter fine-tuning and iterative optimization, forming a self-evolving closed loop.
[0056] As an improvement, after continuous interpolation and alignment of the spatial geometric coordinates, the prediction error of the cell's spatial geometric position is less than 15 micrometers; when performing gene combination perturbation prediction on an authoritative perturbation dataset, the simultaneous hit rate of two genes reaches 48.57%, and the hit rate of a single gene reaches 94.29%.
[0057] As an improvement, the scale of the multi-source heterogeneous spatiotemporal multi-omics data is in the tens of millions to hundreds of millions of cells; the model parameters are in the range of 50M to 1B, without relying on the stacking of hundreds of billions of parameters, and the dependence on human-annotated data is reduced through a cross-species transfer fine-tuning strategy.
[0058] As an improvement, the method is used to construct a virtual organ spatiotemporal evolution model, which can output the continuous developmental morphology, cell migration trajectory and tissue structure remodeling trend of a 3D digital twin organ.
[0059] Beneficial technical effects of the present invention:
[0060] Compared with existing scGPT static pre-trained models and Dynamo single-cell vector field dynamics models, this invention has five core technological advantages, all of which have quantifiable and objective results:
[0061] This invention achieves mechanism-driven continuous dynamics extrapolation and eliminates the drift defect of static fitting: existing models are based only on statistical fitting of discrete samples, resulting in severe drift in long-term predictions; this invention constructs a continuous cell evolution velocity field through Neural ODE / SDE, simulating the entire process of cell differentiation from a physical dynamics perspective, which greatly improves the stability of long-term organ development extrapolation and has forward-looking prediction capabilities.
[0062] Achieving cross-scale, spatial-temporal dual-dimensional integrated alignment of genes, cells, tissues, and organs: Existing technologies cannot couple the virtual space of gene expression with the real three-dimensional geometric space of tissues; This invention uses a four-layer hierarchical graph network with cross-scale attention to open up the perturbation transmission link from molecules to organs, significantly reducing the spatial positioning error of cells and generating complete 3D digital twin organs.
[0063] It possesses high-confidence causal counterfactual simulation capabilities, reducing wet experimental R&D costs: Relying on the hard constraints of biological knowledge graphs, this invention has high accuracy in predicting perturbations of single genes and multi-gene combinations; it can complete large-scale virtual screening of genes and drugs in a computer, significantly shortening the traditional wet experimental screening cycle and significantly reducing overall R&D costs.
[0064] Lightweight small model architecture, adapted to high-noise and heterogeneous clinical samples, reduces the computational and data threshold: Decouple batch and individual noise through individual-specific embedding, eliminating the need for large models with a large number of parameters; Large-scale cell data modeling can be completed within a moderate parameter range, cross-species transfer fine-tuning reduces the need for labeled data, supports privacy-compliant federated learning clinical deployment, and has stronger model interpretability.
[0065] A multi-agent, fully automated dry and wet experiment closed-loop system with autonomous model evolution capabilities: The automated agents complete the entire process of data analysis, code generation, and experimental design, and the automated task processing capabilities are significantly better than similar general-purpose tools; Successful / failed experimental samples are automatically fed back for training, and the model performance continuously improves with scientific research data, without the need for repeated manual retraining and parameter tuning. Detailed Implementation
[0066] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on these embodiments are within the protection scope of the present invention.
[0067] I. Data Acquisition and Integration of Multi-Source Heterogeneous Spatiotemporal Multi-omics Data
[0068] The first step of this invention is to acquire multi-source heterogeneous spatiotemporal multi-omics data. The data includes at least two or more of the following categories: single-cell transcriptome sequencing (scRNA-seq) data, spatial transcriptome (ST) data, epigenomics data (such as ATAC-seq, methylation data), and cell lineage tracing data. The data scale can reach tens of millions to hundreds of millions of cells.
[0069] In practice, for scRNA-seq data, gene expression matrices generated by the 10× Genomics or Drop-seq platforms are used; for spatial transcriptome data, spatial coordinates and expression matrices generated by the 10× Visium or Slide-seq platforms are used. Data from different platforms are uniformly converted into standardized gene expression counting matrices, and the three-dimensional geometric coordinates (X / Y / Z) of each cell or spatial point are recorded.
[0070] II. Noise-robust cleaning and individual heterogeneity decoupling pretreatment
[0071] After acquiring the data, a noise-robust cleaning step is performed. This includes processing the high-noise, sparse, and batch heterogeneous data using a noise-robust layer (such as outlier filtering based on the absolute deviation of the median) and a weighted loss function (such as likelihood loss for a zero-inflated negative binomial distribution).
[0072] At the data input end, a learnable individual-specific embedding vector is introduced for each sample (donor or batch), and this vector is input into the network in parallel with the cell expression matrix. Adversarial training or decoupled representation learning is performed through a Domain-Adversarial Neural Network (DAN) or Conditional Variational Autoencoder (VAE) to mathematically decouple individual heterogeneous noise (such as age, sex, sequencing batch) from common developmental patterns, thereby eliminating interference from batch and donor differences. This embedding vector is jointly optimized with the main model parameters during training, but can be fixed or removed during inference.
[0073] III. Multi-task table learning pre-training and joint loss function
[0074] The cleaned and decoupled data is input into a multi-task tabular learning pre-training framework (such as one based on Transformer or TabNet architecture), and the model is pre-trained using a joint loss function to obtain a unified shared representation space. This representation space simultaneously encodes gene expression, spatial location, and developmental stage information.
[0075] The joint loss function is specifically expressed as follows:
[0076] L_total = L_masked + L_contrastive + L_dynamic
[0077] in:
[0078] L_masked is the loss for masking gene expression prediction: 15% to 30% of the gene expression values in the cell expression matrix are randomly masked, and the model needs to predict the masked values in order to learn the co-expression and regulatory network among genes.
[0079] L_contrastive is a cross-modal contrastive learning loss: it uses InfoNCE or contrastive predictive coding (CPC) to maximize the mutual information between different modalities (such as gene expression and spatial neighborhood structure) in the same tissue microenvironment.
[0080] L_dynamic is the dynamic continuity constraint loss: based on the representational similarity of adjacent developmental time points or spatially neighboring cells, it constrains the spatiotemporal continuity of the physical topology of life evolution, for example, by using smooth L2 regularization or cycle consistency loss.
[0081] The Adam optimizer was used for pre-training, with the learning rate set to 1e-4 to 1e-3 and the batch size adjusted according to the GPU memory. The training epochs continued until the validation set loss converged.
[0082] IV. Construction of Hierarchical Graph Neural Networks and Cross-Scale Attention Transmission
[0083] A bottom-up, four-level graph neural network structure is constructed as follows:
[0084] Gene regulation network layer: Genes are used as nodes, and gene co-expression correlations or known regulatory relationships (such as those from the STRING database) are used as edges. Graph Convolutional Network (GCN) or Graph Attention Network (GAT) are used to learn the regulatory features at the gene level.
[0085] Cellular function embedding layer: The gene expression patterns of each cell are aggregated into a cell state vector, and gene layer features are mapped to cell-level representations through graph pooling or attention aggregation.
[0086] Tissue topology microenvironment layer: Using cells as graph nodes and spatial physical coordinates (within the Euclidean distance threshold) and intercellular communication networks (ligand-receptor pair expression) as graph edges, a local microenvironment dynamic graph is constructed, and graph attention or message passing networks are used to simulate cell neighborhood interactions;
[0087] Organ morphological anatomy layer: Evolutionary inference of long-term topological remodeling is performed, and graph-level pooling or differential homeomorphic mapping is used to converge tissue layer features into an overall organ morphological representation.
[0088] The cross-scale attention mechanism employs hierarchical attention (such as HaloNet or U-Net-style skip connections) to transmit dynamic changes in gene transcription and expression layer by layer to the morphological evolution of cells, tissues, and even organs, thereby realizing the mapping of low-level perturbations to macroscopic structures.
[0089] V. Continuous Spatiotemporal Evolution Model (Neuro-Frequency Differential Equations / Stochastic Differential Equations)
[0090] The pre-trained representations are input into a continuous spatiotemporal evolution model based on neural ordinary differential equations (Neural ODEs) or neural stochastic differential equations (Neural SDEs). This model models the rate of change of cell states over time as a continuous velocity field parameterized by a neural network, specifically in the form:
[0091] dZ(t) / dt = fθ(Z(t), t, u)
[0092] in:
[0093] Z(t) represents the state trajectory encoding of the cell in a high-dimensional semantic space (including gene expression, spatial coordinates and cell state).
[0094] t represents a developmental stage or a continuous timeline (such as the number of days of embryonic development).
[0095] u represents the input perturbation label vector (set to zero vector when there is no perturbation);
[0096] fθ is a continuously changing velocity field parameterized by a multilayer perceptron or Transformer.
[0097] Simultaneously, the Optimal Transport algorithm is used as a boundary constraint to handle state alignment between discrete time slices (such as samples with different developmental days), achieving dynamic fusion of continuous time axes and three-dimensional spatial geometric coordinates. Specifically, the Sinkhorn algorithm or Wasserstein distance minimization is employed to ensure that the spatial distribution of the predicted trajectory at different time points remains consistent with the distribution of the measured slices.
[0098] This step yields a continuous spatiotemporal trajectory of cell development, including the predicted spatial location and expression state of each cell at any given time point. After continuous interpolation and alignment, the prediction error of the cell's spatial geometric location is controlled to be less than 15 micrometers.
[0099] VI. Causal Constraints, Counterfactual Simulation, and Virtual Intervention
[0100] In the neural constant differential equation or neural stochastic differential equation, a perturbation control vector u (corresponding to gene knockout, overexpression or drug treatment tags) is introduced, and a biological knowledge graph (such as KEGG pathway, GO database) or a causal constraint matrix (containing known developmental regulatory pathways, ligand-receptor database and physical diffusion constraint information) is embedded in the deduction trajectory as a hard boundary condition.
[0101] The specific implementation method is as follows: at each step of the ODE / SDE solution, the current state is projected onto the causal constraint matrix. If the trajectory deviates from the known biological constraints (such as the intermediate state of a certain differentiation path must express a specific transcription factor), a gradient correction or penalty term is applied to ensure the biological rationality of the counterfactual inference.
[0102] Based on this, virtual counterfactual simulations of single-gene knockout, multi-gene knockout combinations, gene overexpression, or drug molecular intervention are performed. The model outputs predicted cell fate shifts after perturbation, including changes in differentiation direction, migration path shifts, or tissue remodeling trends.
[0103] Validated on authoritative perturbation datasets (such as Perturb-seq), this method achieves a simultaneous hit rate of 48.57% for two genes and a single gene hit rate of 94.29% when predicting gene combination perturbations.
[0104] VII. Top-level multi-agent cooperative scheduling and closed-loop dry and wet experiment (optional and preferred steps)
[0105] As a preferred embodiment of the present invention, a top-level multi-agent cooperative scheduling step is also included:
[0106] Automatic problem planning, Python code generation, and algorithm execution within a Docker sandbox are achieved through a scientific research intelligent agent architecture (such as an LLM-based planner + code generator + executor).
[0107] Output the target or mechanism hypothesis predicted by counterfactual simulation to the wet experimental verification stage (such as CRISPR screening, tissue culture under drug treatment).
[0108] The results of wet experiment validation, including successful and unsuccessful cases, are automatically converted into a negative sample stream (e.g., marking unsuccessful perturbation cases as hard negative samples), and fed back to the pre-trained model for incremental parameter fine-tuning (e.g., using elastic weights for consolidation or online learning) and iterative optimization, forming a self-evolving closed loop of "dry prediction - wet validation - model update".
[0109] VIII. Model Parameter Scale, Transfer Strategy, and Application Output
[0110] The model described in this invention has a total parameter count on the order of 50M to 1B, without relying on the stacking of billions of parameters. It reduces its dependence on human-labeled data through cross-species transfer fine-tuning strategies (such as transferring from mouse models to human data).
[0111] Ultimately, the method is used to construct a spatiotemporal evolution model of a virtual organ, which can output the continuous developmental morphology, cell migration trajectory and tissue structure remodeling trend of a 3D digital twin organ, and present it in a visual form (such as 3D animation or slice overlay), supporting interactive exploration and hypothesis testing.
[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
[0114] The present invention and its embodiments have been described above. This description is not restrictive, and the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and, without departing from the spirit of the invention, design similar structures and embodiments without creative effort, all such designs should fall within the protection scope of the present invention.
Claims
1. A method for spatiotemporal modeling of cell development integrating single-cell and spatial transcriptomics, characterized in that, Includes the following steps: Acquire multi-source heterogeneous spatiotemporal multi-omics data, including at least two of the following: single-cell transcriptome data, spatial transcriptome data, epigenomics data, and cell lineage tracing data; The multi-source heterogeneous spatiotemporal multi-omics data is subjected to noise robust cleaning, and individual-specific embedding vectors are introduced at the data input end. Through adversarial training or decoupled representation learning, individual heterogeneous noise and common laws of life development are mathematically decoupled to eliminate interference from experimental batches and donor differences. The cleaned and decoupled data is input into a multi-task table learning pre-training framework, and the model is pre-trained through a joint loss function to obtain a unified shared representation space; the joint loss function includes at least masked gene expression prediction loss, cross-modal contrastive learning loss, and dynamic continuity constraint loss. A bottom-up hierarchical graph neural network structure is constructed, which includes at least a gene regulation network layer, a cell function embedding layer, a tissue topology microenvironment layer, and an organ morphology and anatomy layer. The perturbation of the transcriptional expression of the underlying genes is dynamically transmitted and mapped to the macroscopic tissue morphology remodeling through a cross-scale attention mechanism. The pre-trained representation is input into a continuous spatiotemporal evolution model based on neural regular differential equations or neural stochastic differential equations. The rate of change of cell state over time is modeled as a continuous velocity field parameterized by a neural network. Combined with the optimal transmission algorithm as a boundary constraint, the continuous interpolation alignment of spatial geometric coordinates and time axis is achieved to obtain the spatiotemporal continuous trajectory of cell development. A perturbation control vector is introduced into the neural constant differential equation or neural stochastic differential equation, and a biological knowledge graph or causal constraint matrix is embedded in the derivation trajectory as a hard boundary condition. Virtual intervention counterfactual simulation of single genes or combination genes is performed, and the predicted results of cell fate shift after perturbation are output.
2. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: The noise robustness cleaning includes: processing high-noise, sparse, and heterogeneous data through a noise robust layer and a weighted loss function; the individual-specific embedding vector and the cell expression matrix are input in parallel, and the decoupling of individual differences and common patterns is achieved through a domain adversarial network or a conditional variational autoencoder.
3. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: The joint loss function is expressed as: L_total = L_masked + L_contrastive + L_dynamic Among them, L_masked is the masked gene expression prediction loss, which is used to mask 15% to 30% of the gene expression values in the cell expression matrix to learn the co-expression and regulatory network among genes; L_contrastive is the cross-modal contrastive learning loss, which is used to maximize the mutual information between different modalities in the same tissue microenvironment; and L_dynamic is the dynamic continuity constraint loss, which is used to constrain the spatiotemporal continuity of the physical topology of life evolution.
4. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: The hierarchical graph neural network structure is constructed as follows: At the gene level, regulatory characteristics are learned through gene association topological networks; At the cellular level, expression patterns are aggregated into cellular state vectors; At the organizational level, a local microenvironment dynamics graph is constructed using cells as graph nodes and spatial physical coordinates and intercellular communication networks as graph edges. At the organ level, evolutionary inferences are made regarding long-term topological remodeling; The cross-scale attention mechanism employs a hierarchical perception mechanism, transmitting the dynamic changes in gene expression layer by layer to tissue morphology evolution.
5. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: The continuous spatiotemporal evolution model is as follows: dZ(t) / dt = fθ(Z(t), t, u) Where Z(t) represents the state trajectory encoding of the cell in the high-dimensional semantic space, t represents the developmental stage or continuous time axis, u represents the input perturbation label vector, and fθ is the continuously changing velocity field parameterized by the neural network; the optimal transmission algorithm is used to process the state alignment between discrete time slices to achieve dynamic fusion of continuous time and three-dimensional spatial geometric coordinates.
6. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: The virtual intervention counterfactual simulation includes single gene knockout, multi-gene knockout, gene overexpression, or drug molecular intervention; the causal constraint matrix contains known developmental regulatory pathways, ligand-receptor databases, and physical constraint information, which are used to impose hard boundary constraints on the inference trajectory generated by the neural constant differential equation, ensuring the biological rationality of the counterfactual inference.
7. The spatiotemporal modeling method for cell development integrating single-cell and spatial transcriptomes according to claim 1, characterized in that: It also includes the top-level multi-agent cooperative scheduling steps: The scientific research intelligent agent architecture automatically performs problem planning, code generation, and algorithm execution within a Docker sandbox. Output the target or mechanism hypothesis predicted by counterfactual simulation to the wet experiment verification stage; The wet experiment validation results, including both successful and unsuccessful cases, are automatically converted into a negative sample stream and fed back into the model for incremental parameter fine-tuning and iterative optimization, forming a self-evolving closed loop.
8. A method for spatiotemporal modeling of cell development integrating single-cell and spatial transcriptomes according to any one of claims 1 to 7, characterized in that: After continuous interpolation and alignment of the spatial geometric coordinates, the prediction error of the cell's spatial geometric position is less than 15 micrometers; when performing gene combination perturbation prediction on authoritative perturbation datasets, the simultaneous hit rate of dual genes reaches 48.57%, and the hit rate of single genes reaches 94.29%.
9. A method for spatiotemporal modeling of cell development integrating single-cell and spatial transcriptomes according to any one of claims 1 to 7, characterized in that: The scale of the multi-source heterogeneous spatiotemporal multi-omics data ranges from tens of millions to hundreds of millions of cells; the model parameters range from 50M to 1B, without relying on the stacking of billions of parameters, and reduce the dependence on human-annotated data through a cross-species transfer fine-tuning strategy.
10. A method for spatiotemporal modeling of cell development integrating single-cell and spatial transcriptomes according to any one of claims 1 to 7, characterized in that: The method is used to construct a virtual organ spatiotemporal evolution model, which can output the continuous developmental morphology, cell migration trajectory and tissue structure remodeling trend of a 3D digital twin organ.