A three-dimensional stacked space transcriptome data-based virtual organ simulation method and system

Abstract

The application provides a virtual organ simulation method and system based on three-dimensional stacked spatial transcriptome data, which comprises the following steps: constructing a spatial adjacency network according to the coordinate information of a plurality of spatial transcriptome slice data; using a graph attention autoencoder to perform expression embedding representation learning on each point in the spatial adjacency network; voxelizing the three-dimensional coordinates of each slice, constructing a continuum through Gaussian smoothing, and generating a three-dimensional morphological model; training a latent space diffusion model to generate a latent expression of the spatial transcriptome data; for a given coordinate and condition, sampling an embedding vector from the trained latent space diffusion model, and inputting the embedding vector into the graph attention autoencoder to obtain the gene expression profile of the corresponding point. The advantage of the application lies in that the spatial transcriptome of the whole organ or even the cross-tissue region is uniformly modeled and expressed, which breaks through the small-scale modeling capability of the existing method which is limited to a local region or a single slice.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of bioinformatics and spatial omics, and specifically relates to a virtual organ simulation method and system based on three-dimensional stacked spatial transcriptome data, for constructing tissue organ models with continuous three-dimensional structure. The technology can be widely applied to the generation, completion, super-resolution reconstruction and virtual experiment of spatial transcriptome data, and has the potential to promote disease research, drug development and spatial molecular mechanism analysis. BACKGROUND

[0002] In recent years, spatial transcriptome technology (ST) has developed rapidly and can detect the expression levels of thousands of genes in situ in the spatial coordinate system. Typical technologies such as 10x Genomics Visium, Slide-seq, Stereo-seq and Slide-tags have pushed the spatial resolution from the sub-tissue scale to the sub-cellular and single-nucleus resolution. These technologies have promoted the development of spatial biology, supported researchers to analyze tissue structure, intercellular communication, spatial heterogeneity, and spawned a large number of computational tools for spatial variable gene detection, spatial partition identification, cell atlas construction and three-dimensional slice integration.

[0003] However, current three-dimensional tissue-level modeling based on real ST data still faces major challenges, mainly in the following aspects:

[0004] High-resolution three-dimensional reconstruction capability is limited. Current methods attempt to reconstruct three-dimensional spatial structure from two-dimensional slice data. For example, stitch3D uses a graph matching strategy for slice registration, but does not have generation capability and is difficult to handle missing slices or reconstruct complete organ structure. STAGE generates high-density expression data by interpolating between slices through a position-supervised autoencoder, and has certain expression completion capability, but lacks controllable modeling of tissue structure and scalability for large-scale three-dimensional expression generation.

[0005] There is a lack of generative simulators for 3D tissue modeling. Some studies have explored the simulated generation of spatial transcriptome data. For example, ZINB-WaVE incorporates spatial coordinates as covariates into a zero-inflated negative binomial distribution for modeling; scDesign3 uses a copula model to jointly simulate single-cell and spatial omics data; and SRTsim rearranges gene expression order to preserve spatial structure. While these methods support spatial data simulation to some extent, they generally have several limitations: they rely on strong distribution assumptions (such as Poisson or negative binomial distributions), making it difficult to fit complex real spatial structures; they only support 2D single-slice modeling and lack 3D simulation capabilities; they lack user controllability, making it difficult to support specific tissue structure settings and label control; and they lack scalability in handling large-scale data, making it difficult to apply to modeling tasks at the scale of millions of spatial points or organs.

[0006] There is a lack of controllable generative platforms to support virtual experiments. Most existing simulators are designed for the development and evaluation of computational methods, lacking support for application scenarios, particularly in areas such as virtual tissue generation and structural perturbation simulation. As spatial transcriptomics technology evolves towards "digital organs" or "tissue digital twins," researchers urgently need a simulation platform with capabilities for expression completion, structure generation, conditional control, and interactive operation. Currently, no publicly available tools meet these needs, especially in tasks such as cross-slice 3D integration and tissue-scale expression generation, where a general and scalable generative solution is lacking. Summary of the Invention

[0007] The purpose of this application is to overcome the shortcomings of existing spatial transcriptome data in constructing structurally continuous, fully expressed, and controllably generated three-dimensional organ-level models.

[0008] To achieve the above objectives, this application proposes a virtual organ simulation method based on three-dimensional stacked spatial transcriptome data, comprising:

[0009] Step 1: Construct a spatial adjacency network based on the coordinate information of several continuous or non-continuous spatial transcriptome slice data;

[0010] Step 2: Use a graph attention autoencoder to learn the embedding representation of each point in the spatial adjacency network;

[0011] Step 3: Voxelize the three-dimensional coordinates of each slice, construct a continuum using Gaussian smoothing, and generate a complete three-dimensional morphological model;

[0012] Step 4: Train the potential spatial diffusion model using a three-dimensional morphological model to generate the potential expression of spatial transcriptome data;

[0013] Step 5: For a given coordinate and condition, sample from the trained latent space diffusion model to generate an embedding vector, input the embedding vector into the graph attention autoencoder to obtain the gene expression profile of the corresponding point.

[0014] As an improvement of the above method, the step 2 further comprises:

[0015] For data from multiple slices, a triplet contrast learning mechanism is introduced to set positive and negative samples to train the network to align the embedding representation of cross-slice points.

[0016] As an improvement of the above method, the step 3 further comprises:

[0017] Align and register the slices by manual or automatic means, and label the structural cavities inside the organs.

[0018] As an improvement of the above method, the latent space diffusion model is based on a U-Net architecture, including a forward process and a backward process;

[0019] The forward process forms a latent expression by adding noise step by step, represented as:

[0020] ;

[0021] wherein, represents the latent expression at the t-th step; represents a noise control parameter; represents a standard Gaussian noise;

[0022] The backward process reverses the noise through a denoising network to inversely restore the noise, represented as:

[0023] ;

[0024] wherein, is a control term; is a spatial position encoding;

[0025] The position encoding of the latent space diffusion model is a fractal position encoding, represented as:

[0026] ;

[0027] wherein, is a position encoding; is a time step embedding; is an optional conditional label embedding; and ​Spatial encoding of the xy-plane and z-axis, respectively; the xy-plane is the plane on which the slice is located, and the z-axis is perpendicular to the xy-plane.

[0028] As an improvement of the above method, further comprising:

[0029] The gene expression profile of each point is generated into a three-dimensional virtual space expression image.

[0030] The application also provides a three-dimensional stacked space transcriptome data virtual organ simulation system based on the above method, the system comprises:

[0031] The spatial adjacency network module is used to construct a spatial adjacency network according to the coordinate information of a plurality of continuous or non-continuous spatial transcriptome slice data.

[0032] The graph attention autoencoder module is used to perform expression embedding representation learning on each point in the spatial adjacency network using a graph attention autoencoder.

[0033] The three-dimensional morphological model generation module is used to voxelize the three-dimensional coordinates of each slice, construct a continuum through Gaussian smoothing, and generate a complete three-dimensional morphological model.

[0034] The latent expression generation module is used to train a latent space diffusion model using the three-dimensional morphological model to generate latent expressions of the spatial transcriptome data.

[0035] The gene expression profile generation module is used to generate an embedding vector from the trained latent space diffusion model for a given coordinate and condition, input the embedding vector into the graph attention autoencoder, and obtain the gene expression profile of the corresponding point.

[0036] Compared with the prior art, the application has the following advantages:

[0037] 1. Support high-resolution three-dimensional space continuous modeling, which can be extended to organ scale. The application proposes a modeling framework that can fuse multi-slice spatial information and reconstruct continuous three-dimensional tissue structure. By introducing a spatial adjacency graph and voxel-level contour interpolation mechanism, the spatial context between two-dimensional slices is effectively integrated, and continuous reconstruction of three-dimensional structure is realized. At the same time, the system has good scalability and can adapt to million-level spatial point data, support unified modeling and expression generation of spatial transcriptome of whole organ or even cross-organ region, and break through the small-scale modeling ability of existing methods limited to local region or single slice.

[0038] 2. The controllable and interactive three-dimensional expression generation simulation is realized. Compared with traditional expression simulation methods relying on fixed distribution assumptions (such as Poisson or negative binomial distribution), the present application adopts a distribution-free latent diffusion model, which can more accurately capture the nonlinear changes and complex covariant structures of real space expressions. The designed three-dimensional space denoising network combines fractal position coding and self-attention mechanism, and can flexibly receive structure labels or spatial conditions, so as to realize controllable expression generation in specific regions, structural disturbance and cross-section selection. In addition, the matching three-dimensional visualization module supports user-defined virtual section input, facilitating downstream tasks such as computer experiment simulation and organizational intervention testing.

[0039] 3. Breakthrough the limitations of two-dimensional modeling and static simulation, and provide the current organizational generation platform. The method is mostly limited to the two-dimensional level, lacks three-dimensional modeling capability and expression continuity control, and is difficult to support the systematic modeling and multi-scale expression reconstruction of spatial transcriptome in the 'digital organ' scene. In contrast, the present application constructs a three-dimensional structural framework of the tissue by stacking slices, and realizes the transformation from 'local slice data' to 'complete three-dimensional tissue model' by interpolating, completing and structurally constraining the expression information through the generation model, thereby providing a unified, highly controllable and expandable generative solution for tissue digital twinning, pathological modeling and structural simulation. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The figure shows the overall computational framework schematic diagram of the virtual organ simulation method based on three-dimensional stacked spatial transcriptome data; the framework integrates three-dimensional spatial coordinate information in both expression representation learning and generation modeling stages, can effectively maintain the spatial continuity between slices, and realize three-dimensional gene expression reconstruction and simulation generation of tissue scale;

[0041] Figure 2 The figure shows the technical implementation flowchart of the present application; the complete technical path from multi-slice spatial transcriptome data input, spatial adjacency construction, expression embedding learning, latent space diffusion modeling, decoding generation to three-dimensional expression output is shown;

[0042] Figure 3 The figure shows the system application schematic diagram of example 1; based on 33 spatial transcriptome slice data of capuchin monkey cerebellum, a complete three-dimensional cerebellum expression atlas is constructed; the atlas supports cross-section reconstruction under any viewing angle, and realizes spatial expression visualization of complex brain region structure;

[0043] Figure 4 The figure shows the system architecture diagram of three-dimensional visualization and expression analysis support; the front-end and back-end separation design is adopted, and the function modules such as expression prediction, cross-section generation and atlas rendering are integrated, so as to provide a flexible three-dimensional interactive analysis platform for users;

[0044] Figure 5 The operation flowchart of three-dimensional visualization and expression analysis support is shown; the red box marks the user input area, which supports the user to specify the section parameters (such as section position and normal vector) online, and generates the corresponding virtual section and expression atlas in real time to assist in exploring the spatial heterogeneity and biological laws in the marmoset cerebellum;

[0045] Figure 6 The simulation generation effect diagram of three-dimensional marmoset cerebellum section is shown; the coronal section, sagittal section and virtual section at any angle generated by the system are shown; each column marks the section normal vector n; the right four pictures in the figure are the gene expression heat maps of representative cell types (white matter, granular layer, Purkinje cell layer, molecular layer) in the cortex, which reflect the precise modeling ability of complex spatial patterns. DETAILED DESCRIPTION

[0046] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0047] As shown in Figure 1 and Figure 2 , the present application provides a three-dimensional stacked spatial transcriptome data-based virtual organ simulation method and system, which can be used to generate high-fidelity, structure-continuous, and marker-controllable three-dimensional spatial expression data, and is suitable for expression completion, data enhancement, heterogeneity simulation, virtual experiment and spatial atlas visualization application scenarios. The method comprises:

[0048] Step 1: input a plurality of continuous or non-continuous spatial transcriptome section data, each section containing point spatial coordinates and its corresponding gene expression matrix. First, construct a spatial neighbor network (SNN) according to the two-dimensional or three-dimensional coordinate information of the points in each section to capture the local spatial structure relationship. The graph structure is defined as follows:

[0049]

[0050] wherein, represents the three-dimensional coordinates of the i th point, is the set spatial field radius.

[0051] Step 2: based on the constructed graph structure, a graph attention autoencoder (Graph Attention Autoencoder) is used to perform expression embedding representation learning for each point. The model structure includes an encoder, a decoder and an attention mechanism, which can model the expression co-structure under the spatial neighbor relationship. For data from multiple sections, a triplet contrast learning mechanism is introduced, and positive and negative samples are set to train the network to align the embedding representation of the cross-section points, thereby alleviating the batch effect and realizing the spatial continuity splicing. ​

[0052] Step 3: 3D geometry reconstruction: The 3D coordinates of each slice are voxelized and a continuous body is constructed through Gaussian smoothing. Convert to an intermediate format that supports multi-dimensional spatial data storage for subsequent processing. Align the slices by manual or automated means and label the structural cavities (such as fissures, cavities, etc.) inside the organ. Then apply contour interpolation along the slice axis to generate a complete 3D morphological model for subsequent generation of geometric constraints and visualization references.

[0053] Step 4: Generation modeling based on diffusion model: The invention uses a latent space diffusion model to achieve high-quality generation of latent expression of spatial transcriptomic data. The forward process and inverse process of the diffusion model follow the prior art, and the latent expression is gradually transformed into the original space through a series of steps Add noise to form:

[0054]

[0055] wherein, is a noise control parameter, is a standard Gaussian noise. And through the denoising network The process of reversing the noise:

[0056]

[0057] wherein, is a control term, is a spatial position encoding; ; One of the innovations of the invention is the design of a spatial denoising network specifically for three-dimensional space, which is based on the U-Net architecture and integrates multi-layer self-attention mechanisms to effectively capture spatial context information. To enhance the accurate modeling of three-dimensional spatial expression, a fractional position encoding is introduced, which is specifically represented as:

[0058]

[0059] wherein, is a time step embedding, is an optional conditional label embedding, and are spatial encodings of the xy plane and the z axis, respectively. After processing by a multi-layer perceptron, the composite embedding is injected into each layer of the U-Net, achieving efficient fusion and expression of three-dimensional spatial conditions. The training process uses standard backpropagation strategies and mainstream optimization algorithms (such as Adam), and the model parameters are optimized end-to-end by minimizing the prediction error or reconstruction error.

[0060] Step 5: Decoding and 3D Generation Output: Given coordinates and conditions (optional), sample and generate embeddings from the trained 3D spatial diffusion model. The final generated embedding vector The data is input into a trained decoder to obtain the gene expression profile at the corresponding location. By combining spatial coordinate input, virtual spatial representation data can be simulated and generated at any three-dimensional location, enabling representation interpolation completion or new structure prediction.

[0061] Step 6: 3D Visualization: Users can specify any spatial plane (defined by a point and a normal vector), calculate the intersection of this plane and the 3D model, and generate a virtual spatial representation image on the cross-section. 3D visualization can be used for structural analysis, tissue anatomy, and computer simulation (in silico) experiments.

[0062] Example 1

[0063] Using a virtual organ simulation method based on three-dimensional stacked spatial transcriptome data, a three-dimensional expression generation model was created based on 3D expression generation from 33 coronal slices of the cerebellum of marmosets.

[0064] like Figure 3 As shown, this embodiment uses the cerebellum of a marmoset as the research object, with spatial transcriptome data from 33 coronal sections as input. The section spacing is 250 μm, and the single-point resolution is 25 μm, exhibiting significant spatial anisotropy. This embodiment demonstrates the ability of this invention to achieve three-dimensional expression modeling and virtual section generation under large-scale, high-sparse data conditions.

[0065] (1) Data preparation and spatial alignment

[0066] After standard preprocessing, the raw data yielded a total of 1,919,455 spatial locations and 2,371 hypervariable genes. To establish a three-dimensional coordinate system, the open-source tool 3D Slicer was used to perform coarse registration on 33 slices, initially reconstructing the three-dimensional anatomical structure of the cerebellum and extracting the three-dimensional spatial coordinates of each location.

[0067] (2) Spatial adjacency graph construction

[0068] For each slice, a spatial adjacency graph is constructed based on the 3D coordinates of its points. Points with an Euclidean distance less than a set radius δ = 75 μm are considered adjacent. The adjacency graph is stored as a sparse adjacency matrix for subsequent use by the graph neural network.

[0069] (3) Representation embedding learning and cross-slice alignment

[0070] A graph attention autoencoder is used as the representation embedding model, and its structure includes:

[0071] Encoder: composed of three layers of graph attention networks, each with output dimension of 128, 64, 32;

[0072] Decoder: based on the dual tasks of adjacency reconstruction and representation reconstruction;

[0073] Loss function: contains reconstruction error and triplet contrastive loss, the latter is used to align similar regions across slices;

[0074] Training method: using Adam optimizer, initial learning rate is 1e-3, training batch size is 512, and iteration is 300 rounds.

[0075] (4) Three-dimensional geometric modeling and organizational structure reconstruction

[0076] The aligned three-dimensional coordinates are voxelized (resolution 25 μm³) and spatially smoothed by a Gaussian kernel. Based on the interpolation algorithm, the missing area between slices is reconstructed to form a continuous cerebellum volume model. To adapt to the expression generation task, the model is converted into a structure-constrained grid and stored in the standard intermediate format.

[0077] (5) Diffusion model modeling and expression generation

[0078] The present application adopts a three-dimensional latent space diffusion model to learn the generation distribution of expression data. The key components are as follows:

[0079] Forward process: add Gaussian noise step by step in the embedding space to generate a perturbed trajectory;

[0080] Denoising network: based on three-dimensional U-Net architecture, integrating multi-layer self-attention modules, the specific architecture is shown in the following table;

[0081] Position encoding: introduce fractal position embedding and inject it into each layer of U-Net to strengthen three-dimensional conditional modeling;

[0082] Training method: the optimization goal is L2 reconstruction error, the training round is 800 rounds, the Adam optimizer is used, the initial learning rate is 1e-4, and the cosine decay learning rate is used.

[0083] Model Input Architecture Channel Parameter Three-dimensional U-Net architecture 1×64 NoSkip → Attn → Attn → Attn → Attn → Attn → Attn → NoSkip 64 → 128 → 128 → 256 ~11.81M

[0084] Where NoSkip: represents the convolution layer without skip-connection; Attn: represents the layer using self-attention mechanism (Attention). In other embodiments, Attn can select the appropriate number of layers according to the complexity of the data.

[0085] (6) Virtual slice generation and expression decoding

[0086] After training, latent embeddings are sampled from the diffusion model and input to the decoder to generate corresponding gene expression profiles The present application supports conditional sampling at any three-dimensional spatial point, generating virtual expression data, and automatically rendering into virtual slice images.

[0087] In this embodiment, the system automatically generates 321 virtual coronal slices with an interval of 1 / 10 of the original resolution, achieving three-dimensional expression prediction and reconstruction of more than 20 million spatial points, demonstrating the effectiveness of the method in spatial continuity and large-scale expression completion.

[0088] (7) Three-dimensional visualization and analysis support

[0089] As shown in Figure 4 and Figure 5 To support the visualization and interactive access of the results generated by the present application, this embodiment deploys a three-dimensional visualization module based on a front-end and back-end separation architecture, which is used to display the reconstructed tissue structure and support expression generation and rendering at any spatial location.

[0090] The front-end module: a three-dimensional interactive interface is built with a Web visualization framework, loading a three-dimensional tissue model and supporting user input to customize section parameters (including section center position and normal vector). The interface supports heat map display, gene channel switching, and other functions.

[0091] The back-end module: after receiving the front-end request, the trained diffusion model is called on the server side to generate the expression profile of the corresponding point based on the input spatial section, and preliminary processing is performed (such as standardization, compression, and color mapping).

[0092] Data communication mechanism: the front-end and back-end realize low-latency data exchange through a lightweight asynchronous communication middleware, supporting high-concurrency request response; the middleware plays a coordinating role in request queue management, cache control, and generated data transmission.

[0093] Acceleration and computing resources: expression generation tasks can be run on computing nodes supporting GPU, improving response efficiency through resource scheduling strategies, and adapting to multi-user interactive scenarios.

[0094] After the user selects or draws any three-dimensional section in the front-end interface, spatial point extraction, expression generation, and image rendering are completed in the background, and the results are returned to the front-end in real time, realizing a "what you see is what you get" three-dimensional expression visualization experience. This module provides complete interactive expression generation and tissue exploration capabilities for the present application, and can be widely applied to tissue structure analysis and virtual imaging simulation.

[0095] Embodiment 2

[0096] The application also provides a virtual organ simulation system based on three-dimensional stacked spatial transcriptome data, which is realized based on the above method, and the system comprises:

[0097] A spatial adjacency network module is configured to construct a spatial adjacency network according to coordinate information of a plurality of continuous or discontinuous spatial transcriptome slice data;

[0098] A graph attention autoencoder module is configured to use a graph attention autoencoder to perform expression embedding representation learning on each point in the spatial adjacency network;

[0099] A three-dimensional morphological model generation module is configured to voxelize three-dimensional coordinates of each slice, construct a continuum through Gaussian smoothing, and generate a complete three-dimensional morphological model;

[0100] A latent expression generation module is configured to use the three-dimensional morphological model to train a latent space diffusion model, and generate latent expression of the spatial transcriptome data;

[0101] A gene expression profile generation module is configured to sample an embedding vector from the trained latent space diffusion model for a given coordinate and condition, input the embedding vector into the graph attention autoencoder, and obtain a gene expression profile of the corresponding point.

[0102] The application can also provide a computer device comprising at least one processor, memory, at least one network interface and user interface. The various components in the device are coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between the components. In addition to the data bus, the bus system also includes power supply bus, control bus and state signal bus.

[0103] The user interface can include a display, a keyboard or a clicking device, for example, a mouse, a trackball, a touchpad or a touch screen.

[0104] It can be appreciated that the memory in the embodiments disclosed in the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (Read-Only Memory, ROM), a programmable read-only memory (Programmable ROM, PROM), an erasable programmable read-only memory (Erasable PROM, EPROM), an electrically erasable programmable read-only memory (Electrically EPROM, EEPROM) or a flash memory. The volatile memory can be a random access memory (Random Access Memory, RAM) used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (Static RAM, SRAM), dynamic random access memory (Dynamic RAM, DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (Synchlink DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DRRAM). The memory described herein is intended to include but not limited to these and any other suitable types of memory.

[0105] In some embodiments, the memory stores elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and an application program.

[0106] Among them, the operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application program includes various application programs, such as a media player (Media Player), a browser (Browser), etc., for implementing various application services. The program for implementing the method of the embodiments of the present disclosure can be included in the application program.

[0107] In the above-mentioned embodiments, the processor can also be used to execute the steps of the above-mentioned method by invoking the programs or instructions stored in the memory, in particular, the programs or instructions stored in the application program.

[0108] execute the steps of the above-mentioned method.

[0109] The method can be applied to a processor or implemented by the processor. The processor can be an integrated circuit chip having a signal processing capability. In implementation, the steps of the method can be completed by an integrated logic circuit of hardware in the processor or by an instruction in the form of software. The processor can be a general-purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The methods disclosed above can be implemented or executed by the processor. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed above can be directly embodied as a hardware code executed by the processor or a combination of hardware and software modules in the processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage mediums in the art. The storage medium is located in the storage memory, and the processor reads information in the storage memory and combines the hardware to complete the steps of the method.

[0110] It can be understood that the embodiments described in the present application can be implemented in hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field-Programmable Gate Arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for executing functions described in the present application or a combination thereof.

[0111] For software implementation, the present application can be implemented by executing the functional modules (such as processes, functions, etc.) described in the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0112] The application further provides a nonvolatile storage medium for storing the computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application is described in detail with reference to the embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and they should be covered in the scope of claims of the present application.

Claims

1. A method for simulating a virtual organ based on three-dimensional stacked spatial transcriptome data, comprising: Step 1: constructing a spatial adjacency network according to coordinate information of a plurality of continuous or discontinuous spatial transcriptome slice data; Step 2: using a graph attention autoencoder to perform expression embedding representation learning on each point in the spatial adjacency network; Step 3: voxelizing the three-dimensional coordinates of each slice, constructing a continuum by Gaussian smoothing, and generating a complete three-dimensional morphological model; Step 4: training a latent space diffusion model using the three-dimensional morphological model to generate a latent expression of the spatial transcriptome data; Step 5: for a given coordinate and condition, sampling an embedding vector from the trained latent space diffusion model, inputting the embedding vector into the graph attention autoencoder, and obtaining a gene expression profile of the corresponding point; the latent space diffusion model is based on a U-Net architecture, including a forward process and a backward process; The forward process is by stepwise potential expression Additive noise formation, denoted as: ； wherein, denotes the latent representation of the t-th step; denotes a noise control parameter; denotes a standard Gaussian noise; The backward process passes through a denoising network de-noising, denoted as: ； wherein is a control term; is a spatial position encoding; ; ; the position encoding of the latent space diffusion model is a fractal dimension position encoding, represented as: ； wherein, is a position encoding; is a time step embedding; is an optional conditional label embedding; and are spatial encodings for the xy-plane and z-axis, respectively; the xy-plane is the plane in which the slice lies, and the z-axis is perpendicular to the xy-plane.

2. The virtual organ simulation method based on three-dimensional stacked spatial transcriptome data according to claim 1, wherein, the step 2 further comprises: for data from multiple slices, a triplet contrast learning mechanism is introduced, and positive and negative samples are set to train the network to align the embedding representations of points across slices.

3. The virtual organ simulation method based on three-dimensional stacked spatial transcriptome data according to claim 1, wherein, the step 3 further comprises: aligning and registering the slices by manual or automatic means, and labeling the structural cavities inside the organ.

4. The virtual organ simulation method based on three-dimensional stacked spatial transcriptome data according to claim 1, wherein, further comprising: generating a three-dimensional virtual space expression image from the gene expression profile of each point.

5. A virtual organ simulation system based on three-dimensional stacked spatial transcriptome data, realized based on the method of any one of claims 1-4, characterized in that, the system comprises: a spatial adjacency network construction module for constructing a spatial adjacency network according to coordinate information of a plurality of continuous or discontinuous spatial transcriptome slice data; a graph attention autoencoder training module for using a graph attention autoencoder to perform expression embedding representation learning on each point in the spatial adjacency network; a three-dimensional morphological model generation module for voxelizing the three-dimensional coordinates of each slice, constructing a continuum by Gaussian smoothing, and generating a complete three-dimensional morphological model; a latent expression generation module for training a latent space diffusion model using the three-dimensional morphological model to generate a latent expression of the spatial transcriptome data; and a gene expression profile generation module for generating an embedding vector from the trained latent space diffusion model for a given coordinate and condition, inputting the embedding vector into the graph attention autoencoder, and obtaining a gene expression profile of the corresponding point.

Images