Spatial information clustering, integration and deconvolution using spatial transcriptomics of GraphST

Abstract

Systems and methods for processing spatial transcriptomic data and generating insights related to cell and tissue status. The systems and methods include spatial clustering, integration of a plurality of tissue sample data, and integration of scRNA-seq data with spatial transcriptomic data. The systems and methods combine graph self-supervised contrast learning and graph neural networks to perform spatial transcriptomics data processing.

Description

TECHNICAL FIELD

[0001] The present disclosure relates generally to methods and systems for processing spatial transcriptomic data. In particular, the present disclosure provides methods and systems for spatial clustering, spatial transcriptomic data integration, and transferring single-cell RNA sequencing to spatial transcriptomics. BACKGROUND

[0002] The purpose of the background is to present information that can be of background to the present disclosure. The information of the background section is not to be taken as an acknowledgement or any form of suggestion that it constitutes prior art to the present disclosure.

[0003] Human body has a large number of cells, which undergo state transitions such as development, disease, and regeneration during their life cycle. Although derived from the same fertilized egg, cell types and states are significantly influenced by both internal processes and environmental factors.

[0004] During cell proliferation and differentiation, cell structures exhibit complex heterogeneity, resulting in diverse cell types for organ construction. The same cell type can perform different functions due to cell heterogeneity in morphology, function, and gene expression profiles among different tissues. Linking cell gene expression with its spatial distribution is crucial for understanding the emergent properties and pathology of tissues. Current spatial transcriptomics (ST) combines gene profiles and spatial information to provide deeper insights into healthy and diseased tissues. Spatial information also helps to infer intercellular communication.

[0005] Advances in spatial transcriptomics technology enable characterization of gene expression profiles at single-cell resolution while preserving spatial organizational context information. However, traditional transcriptomic data analysis methods often have limitations in accurately resolving tissue gene expression patterns within tissues, systematically analyzing multiple tissue sections, and accurately integrating single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) data. Traditional methods use unsupervised learning, but the clustering performance is often unsatisfactory because the boundaries of the domains identified are usually fragmented and do not match well with pathological annotations. Since true ST partitions are often unavailable, supervised learning cannot be used to improve performance. In addition, most current analysis methods are only applicable to a single tissue section and cannot identify spatial domains from multiple sections jointly. Current technical limitations also hinder the realization of single-cell resolution for ST while maintaining comparable gene coverage to single-cell RNA sequencing (scRNA-seq).

[0006] It would be desirable to overcome or ameliorate at least one of the above-mentioned problems, or to provide at least a useful alternative. SUMMARY

[0007] In one embodiment, the present disclosure provides a transcriptome data processing system comprising a memory and one or more processors, the memory comprising program code executable by the processors to:

[0008] receive a first batch of records of spatial gene expression data for a first tissue sample, wherein the records comprise gene expression data and location data indicative of a location of the gene expression data in the first tissue sample;

[0009] select a first subset of records in the first batch corresponding to spatial gene expression data for a variable gene;

[0010] convert the selected records in the first subset to a graph data structure based on the location data;

[0011] process the graph data structure using a neural network implementing contrastive self-supervised learning to obtain latent gene expression data,

[0012] wherein the latent gene expression data is a reduced dimension representation of the first subset of records; and

[0013] convert the latent gene expression data to reconstructed gene expression data using a decoder neural network,

[0014] wherein the records in the reconstructed gene expression data are associated with the location data in the received spatial gene expression data.

[0015] Some embodiments relate to a transcriptome data processing system comprising a memory and one or more processors, the memory comprising program code executable by the processors to:

[0016] receive a first batch of records of single-cell RNA sequencing data associated with a plurality of cells from a tissue sample;

[0017] receive a second batch of records of spatial gene expression data for the tissue sample (Hs), wherein the records comprise spatial gene expression data and location data indicative of a location of the spatial gene expression data in the tissue sample; s

[0018] process the first batch of records using a self-supervised autoencoder neural network to generate cell representation records (HC);

[0019] project the cell representation records (HC) onto the second batch of records using an untrained mapping matrix (M') to obtain predicted spatial gene expression records (Hs');

[0020] process the predicted spatial gene expression records using a neural network implementing contrastive learning to obtain reconstructed spatial gene expression records (Hs); and

[0021] ​The mapping matrix (M') is trained based on the reconstructed spatial gene expression records (Hs) and the cell characterization records (HC) to generate a trained mapping matrix (M).

[0022] Some embodiments relate to a method of transcriptomic data processing, the method comprising:

[0023] receiving a first batch of records of spatial gene expression data of a first tissue sample, wherein the records comprise gene expression data and position data indicative of a location of the gene expression data in the first tissue sample;

[0024] selecting a first subset of the first batch of records corresponding to spatial gene expression data of a variable gene;

[0025] constructing a graph data structure based on the position data; and

[0026] processing the graph data structure and the selected first subset using a deep learning framework implementing graph self-supervised contrastive learning to obtain latent gene expression data;

[0027] wherein the latent gene expression data is a reduced dimension representation of the first subset of records.

[0028] The method of some embodiments further comprises processing the reconstructed gene expression data using a clustering model to identify a plurality of clusters in the first tissue sample, wherein each cluster contains locations with similar gene expression profiles.

[0029] The method of some embodiments further comprises:

[0030] receiving a second batch of records of spatial gene expression data of a second tissue sample, wherein the second batch of records comprises gene expression data and position data indicative of a location of the gene expression data in the second tissue sample;

[0031] aligning the position data of the first batch of records and the second batch of records to a common coordinate space; and

[0032] selecting a second subset of the second batch of records corresponding to spatial gene expression data of the variable gene,

[0033] wherein the graph data structure is obtained based on position data derived from the first tissue sample and the second tissue sample; and

[0034] the latent gene expression data comprises data derived from the first tissue sample and the second tissue sample.

[0035] The method of some embodiments further comprises converting the latent gene expression data to reconstructed gene expression data using a decoder neural network, wherein the records in the reconstructed gene expression data are associated with locations in the first tissue sample and the second tissue sample.

[0036] The method of some embodiments further comprises processing the reconstructed gene expression data using a clustering model to identify a plurality of clusters in the first tissue sample and the second tissue sample, wherein each cluster contains locations with similar gene expression profiles.

[0037] Some embodiments relate to a transcriptomic data processing method, the method comprising:

[0038] receiving a first batch of records of single-cell RNA sequencing data associated with a plurality of cells from a tissue sample;

[0039] receiving a second batch of records of spatial gene expression data of the tissue sample (Hs), wherein the records comprise spatial gene expression data and location data indicative of a location of the spatial gene expression data in the tissue sample;

[0040] processing the first batch of records using a self-supervised autoencoder neural network to generate cell representation records (HC);

[0041] projecting the cell representation records (HC) onto the second batch of records using an untrained mapping matrix (M’) to obtain predicted spatial gene expression records (Hs’);

[0042] processing the spatial gene expression records using a deep learning framework implementing self-supervised contrastive learning to obtain reconstructed spatial gene expression records (Hs); and

[0043] training the untrained mapping matrix (M’) based on the reconstructed spatial gene expression records (Hs) and the cell representation records (HC) to generate a trained mapping matrix (M).

[0044] The method of some embodiments further comprises projecting the cell representation records (HC) onto the second batch of records using the trained mapping matrix (M) to perform spatial localization of cell types in the tissue sample. BRIEF DESCRIPTION OF DRAWINGS

[0045] Some embodiments of systems and methods for processing transcriptomic data according to the present disclosure will now be described, by way of non-limiting example only, with reference to the accompanying drawings in which:

[0046] Figure 1 and Figure 2 A transcriptomic data processing method is shown.

[0047] Figure 3 An overview of GraphST for spatial clustering, integration, and deconvolution of spatial transcriptomics is shown.

[0048] Figure 4 A system for implementing the method or Figure 1 and Figure 2 framework is shown. Figure 3 The method or Detailed Implementation

[0049] The system and method disclosed herein may also be referred to as GraphST or the GraphST model. GraphST comprises a multifunctional graph contrastive self-supervised learning framework that integrates spatial location information and gene expression profiles for spatial clustering, integration, and transfer in spatial transcriptomics. With its ease of use and rapid computational capabilities, GraphST can identify heterogeneity within spatial transcriptomics data and provide insights into cellular states within each tissue.

[0050] Figure 1 Some methods described herein are illustrated, wherein method 100 is used for transcriptome data processing, including: receiving multiple records (step 102), selecting a subset from these records (step 104), and converting the records in the selected subset into a graph data structure (step 106), which can be processed to generate potential gene expression data (step 108). The potential gene expression data generated in step 108 is a dimensionality-reduced representation of the original subset of records received in step 102. The potential gene expression data is... Figure 3 The data at position 320 of image (C) is obtained by transforming gene expression data using a neural network that implements contrastive self-supervised learning. This potential gene expression data may represent a dimensionality-reduced representation, making gene expression data easier for conventional computer systems to process.

[0051] Step 102 may involve receiving records in any desired manner. Records may come from a single data source or database, or they may be generated by aggregating datasets from multiple sources.

[0052] The record received in step 102 includes spatial gene expression data from the first tissue sample. The record also includes location data indicating the location of the gene expression data within the first tissue sample.

[0053] Transcriptomics data examples

[0054] To perform the spatial segmentation task, following step 102, the first batch of spatial gene expression data records is obtained (e.g., received) through the GraphST model at least partially implemented by method 100. The records may correspond to samples taken from one or more spatial gene expression datasets. In this example, the samples are taken from four datasets. The spatial gene expression data comes from a first tissue sample, and the records include gene expression data and location data indicating the location of that gene expression data within the first tissue sample.

[0055] The datasets used for this purpose can be generated by the Visium platform or any other platform. The publicly available datasets used herein are summarized as follows: (1) The first dataset is the LIBD Human Dorsolateral Prefrontal Cortex (DLPFC) data, containing 12 slides from 3 individuals, with the number of spots per slide ranging from 3,460 to 4,789. (2) The second dataset is a human breast cancer tumor sample, containing 3,798 spots. (3) The third dataset is a mouse breast cancer derived from untreated metastatic tumors, containing 1,950 spots. (4) The fourth dataset consists of mouse forebrain, including 2,695 spots from publicly available 10X Genomics data repository.

[0056] In some embodiments, the data in the first batch of records must be integrated. The integration can include performing one or both of vertical integration and horizontal integration. To perform a multi-sample integration task in an experiment, the GraphST method was performed on two samples from the 10X Visium platform to conduct vertical integration and horizontal integration, respectively.

[0057] For vertical integration, the analysis encompasses mouse breast cancer tissue. For the data being integrated in the vertical integration process, each spatial gene expression data point can be vertically adjacent to one or more other spatial gene expression data points, and these vertically adjacent data points can be integrated. In the experiment, the vertical integration involved using two sets of samples from this tissue, where each set of samples consists of two vertically adjacent slides, with the number of spots per slide ranging from 1,868 to 3,042.

[0058] For horizontal integration, GraphST was used to analyze spatial transcriptomic mouse brain (forebrain and hindbrain) tissue. The mouse forebrain and hindbrain were horizontally separated. Each of the forebrain and hindbrain contained two slides (Slide 1 and Slide 2), with approximately 3,000 spots per slide. All slides were downloaded from the 10X Genomics data repository.

[0059] In step 104, a subset of the records received from step 102 is selected. The subset is selected to correspond to spatial gene expression data of highly variable genes (HVGs). The HVGs are selected using the SCANPY function “sc.pp.highly_variable_genes()”, in which the HVGs are identified by calculating the normalized variance of each gene. First, the data is standardized with the regularized standard deviation (i.e., z-score normalization for each feature). Then, the normalized variance is calculated as the variance of each gene after transformation. The genes are sorted by the normalized variance. Genes with a normalized variance higher than a predetermined threshold variance (e.g., a threshold variance suitable for the current method is preset by experiment), or the top N genes can be designated as highly variable genes.

[0060] Once the variable genes are identified, the selected subset of records in step 104 is converted into a graph data structure in step 106, based on the positional data of the subset of records.

[0061] Step 108 can then be performed using a neural network implementing self-supervised learning. In step 108, the graph data structure is processed using the neural network to obtain latent gene expression data. The latent gene expression data is a reduced dimension representation of the original records of the selected subset.

[0062] The method 200 further described herein can be used for transcriptomic data processing of mixed datasets, such as single-cell RNA (scRNA) sequencing data and spatial gene expression data. The method 200 comprises receiving records of scRNA sequencing data associated with a plurality of cells from a tissue sample (step 202), and records of spatial gene expression data of the tissue sample (H s ). Like step 202, the spatial gene expression data comprises positional data indicative of the location of the spatial gene expression data in the tissue sample.

[0063] Step 206 of the method 200 involves processing the records received in step 204 using a self-supervised autoencoder neural network to generate cell representation records (H c ). An encoder neural network comprises a neural network that converts a high-dimensional dataset into a lower-dimensional dataset while preserving the information embedded within the dataset. These cell representation records (H c ) can then be projected onto the records received in step 202 using an untrained mapping matrix (M')— step 208. This projection obtains predicted spatial gene expression records (H s ). The resulting records H s ' can then be processed using a neural network implementing contrastive learning to obtain reconstructed spatial gene expression records (H s )— step 210. The reconstructed gene expression data is more informationally dense and / or discriminative gene expression data to more accurately and / or efficiently support subsequent analysis tasks such as spatial partitioning.

[0064] Now having the reconstructed records Hs and the cell representation records Hc, the mapping matrix M' can be trained to produce a trained mapping matrix M— step 212. The trained mapping matrix can then be used to map new cell representation records to spatial gene expression records, and vice versa, to map new spatial gene expression records to cell representation records.

[0065] To accomplish the scRNA-seq and ST integration task in experiments, GraphST was tested on three groups of human or mouse datasets. The first dataset received in step 202 was from a mouse prefrontal sample. The spatial data included 2695 loci and 32285 genes. The scRNA-seq data was taken from mouse whole cortex and hippocampus 10x website, which analyzed >110 million cells and 22764 genes from multiple cortical regions and hippocampal structures. The second dataset received in step 204 was from a human breast cancer sample. The scRNA-seq data included a dataset of 46080 cells, covering 5000 gene expression with cell type information. GraphST adopted data obtained from human breast cancer tissue containing 3798 loci and 36601 genes as ST data. The final dataset was from a human brain sample. The ST data of some experiments contained data of 3639 loci and 33538 genes. The single-cell data included snRNA-seq data obtained by analyzing archived post-mortem dorsolateral prefrontal cortex (BA9) tissue through the 10X Genomics Chromium platform. Some experiments involved randomly sampling 78886 cells covering 30062 genes from this dataset.

[0066] The above data sources are merely examples. GraphST provides a flexible technique for processing ST data and scRNA data to identify heterogeneity within spatial transcriptomic data and provide insights into cell states within a tissue, regardless of the data source or the acquisition source / method of the related data.

[0067] Data preprocessing

[0068] For all ST data, the preprocessing includes first filtering out unannotated loci if the dataset contains annotations. Subsequently, the gene expression counts in the records received in step 204 (and 102) can be normalized by library size. For example, the raw gene expression counts can be log-transformed and normalized by library size through the SCANPY package. For steps 104 and 116, this step can also be performed in method 200 for one or both of the ST and scRNA records, selecting the top 3,000 highly variable genes (HVGs) as the input features for the GraphST model - the number of genes and the number of records can be selected according to preferences, applications, or other factors. Similarly, for scRNA-seq data, the raw gene expression counts are first log-transformed and normalized by library size, and then we select the top 3,000 highly variable genes as the cell input features for the model. To accomplish the scRNA-seq and ST integration task, the preprocessed HVGs of scRNA-seq and ST data are further aligned, and the overlapping genes are used as the final input features for the loci and cells.

[0069] Spatial clustering

[0070] Figure 3 The method 100 for spatial clustering of transcriptomic data in the context of the deep learning framework 300 is shown. Based on the spatial transcriptomic dataset provided in step 102, Figure 3 Part (A) of FIG. 1 shows the data preprocessing of the gene expression data 302 associated with the counts and the spatial location data 306 indicating the location of the gene expression data in the tissue sample. The data preprocessing and augmentation includes generating a perturbed graph 310 corresponding to the original graph 308 according to step 106. The perturbed graph 310 is generated by data augmentation with the original graph as input, which is achieved by randomly swapping or shuffling 304 the features between the nodes (i.e., sites) in the original graph.

[0071] An encoder 311 containing a graph neural network (GNN) can be used to analyze and extract features from the original graph 308 and the perturbed graph 310 for latent representation learning and generating the original representation of the latent gene expression data according to step 108.

[0072] Next, the graph contrastive self-supervised learning framework 312 refines the original representation by pairing the local neighborhood representation with the site representation as a positive sample pair and pairing the site representation with the perturbed local neighborhood representation as a negative sample pair.

[0073] Subsequently, the gene expression reconstruction decoder 314 converts the refined representation of the latent gene expression data into reconstructed gene expression data according to step 110 - i.e., the decoder 314 converts the representation Z s back to the original space to reconstruct the gene expression H s The decoder neural network contains a neural network that converts a low-dimensional dataset into a high-dimensional dataset, where the high-dimensional dataset corresponds to the target domain. In this disclosure, the target domain is the gene expression data associated with the location or site. This inverse transformation associates the records in the reconstructed gene expression data with the location data in the received spatial gene expression data.

[0074] The learning process can be guided by loss minimization - e.g., minimizing the self-reconstruction loss and the contrastive loss. The self-reconstruction loss forces the representation H s to fully retain the gene expression features and spatial location information. The contrastive loss leverages the data itself to make the representation H s more discriminative and informative. After the model is trained according to step 110, the output H s is used to cluster the sites into different spatial domains using a non-spatial clustering algorithm such as mclust. Each cluster can be considered as a spatial domain containing sites with similar gene expression profiles and close spatial locations.

[0075] The embodiments may implement alternative deep learning frameworks to perform the methods. The deep learning framework may include neural networks that implement computational logic to process information.

[0076] like Figure 3 As shown in section (B), the reconstructed gene expression data H obtained in section (A) s Spatial clustering is then performed. The existing "mclust" clustering algorithm is used to segment the sites into different spatial domains. "mclust" is an R contribution package based on finite normal mixture modeling to achieve model-based clustering, classification, and density estimation. It provides functions for parameter estimation of normal mixture models with multiple covariance structures through the EM algorithm, as well as functions for simulation from these models. It also includes functions for a comprehensive strategy that combines model-based hierarchical clustering, the EM algorithm for mixture estimation, and the Bayesian Information Criterion (BIC) in clustering, density estimation, and discriminant analysis, and provides additional functionality for displaying and visualizing fitted models and clustering, classification, and density estimation results.

[0077] Vertical and horizontal integration of multiple ST data

[0078] In some embodiments, the GraphST model performs an integrated analysis of multiple ST data sets. Taking ST data obtained from two tissues (steps 102 and 112) as an example, before inputting them into the model, the embodiment can first align the images according to step 114—for example, the positional data of the first and second batches of records received in steps 102 and 112, respectively, can be aligned to a common coordinate space. The algorithm for performing this alignment can be the PASTE algorithm used for aligning hematoxylin-eosin (H&E) stained images. If the two tissue slices come from different regions of the same tissue, such as the anterior and posterior parts of the mouse brain, alignment helps ensure the horizontal continuity of the same regions between the two tissue slices. If the two tissue slices come from the same region, such as slice 1 and slice 2 of the anterior part of the mouse brain, alignment is required to maximize the overlap between the two slices. Using the aligned data (including spatial location and gene expression) as input, GraphST can generate site representations for both types of ST data (…). Figure 3 Image (A) is used for joint spatial clustering. For transformation step 106 and subsequent steps, a subset of aligned second batch records can be selected based on the records representing variable genes.

[0079] Integration of scRNA-seq and ST data

[0080] Some embodiments include a spatial information contrast learning module for integrating ST and scRNA-seq data. Figure 3 Image C), and implement method 200. Following step 206, cell characterization H can be learned from scRNA-seq gene expression using a self-supervised autoencoder. cTo achieve this, in 316, H s and H c The embodiments aim to map scRNA-seq data into a spatial transcriptome by a trainable mapping matrix M (which can refer to an untrained mapping matrix M' and a trained mapping matrix M - note that M can be further trained) that represents the probability of a cell being projected to each site of the spatial data, thereby generating a predicted spatial gene expression record H s '. This process is performed in step 210 and in 318, the predicted spatial gene expression record H s ' is aligned with the reconstructed spatial gene expression H s by a contrastive learning mechanism, where the similarity of positive sample pairs (i.e., site i and its neighborhood) is maximized, while the similarity of negative sample pairs (i.e., site i and its non-neighborhood) is minimized. After the model is trained, the mapping matrix M can be trained in step 212. In some embodiments, the mapping matrix M can then be used in step 214 to project scRNA-seq data into a spatial transcriptome.

[0081] A variable gene or spatially variable gene includes a gene whose expression profile is significantly dependent on spatial location. Such genes contribute significantly to the intercellular variation within a homogeneous cell population - unless context dictates otherwise, the term "highly variable gene" is used interchangeably with "variable gene" herein.

[0082] Location data includes data for identifying sites in a tissue sample that are the source of gene expression data records. Location data can include three-dimensional coordinates or other representations of locations within a tissue sample.

[0083] A graph data structure includes a non-linear data structure consisting of a finite number of nodes or vertices and edges connecting them. The graph data structure of embodiments associates specific locations or sites with nodes and relies on location data to define vertices or edges. The graph data structure of the present embodiment models the locations of gene expression data within a tissue sample.

[0084] Contrastive self-supervised learning includes computational techniques generally aimed at learning comparisons within a domain of records by an objective function exemplified as follows:

[0085]

[0086] where x+ is similar to x, x- is dissimilar to x, and f is an encoder (representation function). The similarity metric in the context of the present disclosure relates to the similarity of gene expression data between records.

[0087] Figure 3 (D) shows downstream analysis tasks. GraphST supports three major tasks, including spatial clustering, multi-sample integration, and ST to scRNA-seq integration.

[0088] Figure 4 A system block diagram for transcriptomic data processing is shown. According to Figure 4 the system includes at least one processor 402, a memory 404 accessible to the processor 402, and a network interface 408 to facilitate communication with a plurality of databases 420 and a user's 460 computer device 470. Program code 406 provided in the memory 404 includes instructions executable by the processor 402 for performing at least a portion of the methods in the embodiments described herein. Notably, although Figure 4 A standalone computer system is described, but any such computer system can be distributed across multiple servers or multiple devices, or certain functionality can be integrated into a single server or device, without departing from the intended scope of the present disclosure.

[0089] The user device 470 can assist in extracting data (e.g., ST or scRNA data) from the databases 420 by the processor 402 of the computing system 400 to initiate the methods 100, 200, or the implementation framework 300. The user device 470 can include a query engine for inputting a query including, for example, a new ST record for mapping to a cell characterization record (e.g., generating a record based on the new ST record, or generating the record and then cross-referencing it with the cell characterization record to identify a record corresponding to the new ST record). The user computing device 470 can be a personal or handheld computing device, such as a smartphone or tablet. The network 430 facilitates communication between devices, such as the user device 470 and the databases 420, and can include one or more communication networks, such as the Internet, a cellular network, etc.

[0090] The system 400 can also access one or more databases 420. Each database 420 can include records such as one or more ST records and / or scRNA records, mapping matrices or neural network models used in the framework 300.

[0091] The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that the prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates.

[0092] Throughout this specification and the claims which follow, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to mean the inclusion of a stated element or step or group of elements or steps but not the exclusion of any other element or step or group of elements or steps.

[0093] The scope of the disclosure encompasses all changes, substitutions, variations, modifications, and alterations of the illustrative embodiments described or illustrated herein that a person having ordinary skill in the art would comprehend. The scope of the disclosure is not limited to the illustrative embodiments described or illustrated in the specification. Furthermore, although the disclosure describes and illustrates particular embodiments with specific components, elements, features, functions, operations or steps, such embodiments can include any one or more combinations or permutations of the components, elements, features, functions, operations, or steps described or illustrated anywhere herein. Although the disclosure describes or illustrates particular embodiments with specific advantages, such embodiments can not exhibit the described, specified advantages, or the embodiments can exhibit other advantages, or modify advantages that are described, specified, or implied in the specification.

Claims

1. A transcriptome data processing system, the system comprising a memory and one or more processors, the memory including program code executable by the processors to: Receive the first batch of records of spatial gene expression data of a first tissue sample, wherein the records include gene expression data and location data indicating the location of the gene expression data in the first tissue sample; Select the first subset of the first batch of records that corresponds to the spatial gene expression data of the variable gene; Based on the location data, the records in the selected first subset are converted into a graph data structure; The graph data structure is processed using a neural network implementing contrastive self-supervised learning to obtain potential gene expression data; and The latent gene expression data is converted into reconstructed gene expression data using a decoder neural network. Wherein, the potential gene expression data is a dimensionality-reduced representation of the first subset of the records; and The records in the reconstructed gene expression data are associated with the location data in the received spatial gene expression data.

2. A transcriptome data processing system, the system comprising a memory and one or more processors, the memory including program code executable by the processors to: Receive the first batch of records of single-cell RNA sequencing data associated with multiple cells from tissue samples; The second batch of records (H) of spatial gene expression data from the tissue samples was received. s The records include spatial gene expression data and location data indicating the location of the spatial gene expression data in the tissue sample; The first batch of records was processed using a self-supervised autoencoder neural network to generate cell characterization records (H). c ); The cell characterization records (H) are generated using an untrained mapping matrix (M'). c Projecting the data onto the second batch of records to obtain predicted spatial gene expression records (H2N) s '); The predicted spatial gene expression record is processed using a neural network that performs contrastive learning to obtain a reconstructed spatial gene expression record (H). s );as well as Based on the reconstructed spatial gene expression record (H s ) and the cell characterization record (H c Train the mapping matrix (M') to generate the trained mapping matrix (M).

3. A method for processing transcriptome data, the method comprising: Receive the first batch of records of spatial gene expression data of a first tissue sample, wherein the records include gene expression data and location data indicating the location of the gene expression data in the first tissue sample; Select the first subset of the first batch of records that corresponds to the spatial gene expression data of the variable gene; Construct a graph data structure based on the location data; as well as The graph data structure and the selected first subset are processed using a deep learning framework that implements graph self-supervised contrastive learning to obtain potential gene expression data; The potential gene expression data is a dimensionality-reduced representation of the first subset of the records.

4. The method according to claim 1, further comprising: The reconstructed gene expression data is processed using a clustering model to identify multiple clusters in the first tissue sample, wherein each cluster includes locations with similar gene expression profiles.

5. The method according to claim 3, further comprising: A second batch of records of spatial gene expression data from a second tissue sample is received, wherein the second batch of records includes gene expression data and location data indicating the location of the gene expression data in the second tissue sample; Align the position data of the first batch of records and the second batch of records to a common coordinate space; as well as Select the second subset from the second batch of records that corresponds to the spatial gene expression data of the variable gene. The graph data structure is obtained based on the location data derived from the first tissue sample and the second tissue sample; and The potential gene expression data includes data derived from the first tissue sample and the second tissue sample.

6. The method according to claim 5, further comprising: The latent gene expression data is converted into reconstructed gene expression data using a decoder neural network. The records in the reconstructed gene expression data are associated with locations in the first tissue sample and the second tissue sample.

7. The method according to claim 6, further comprising: The reconstructed gene expression data is processed using a clustering model to identify multiple clusters in the first and second tissue samples, wherein each cluster contains locations with similar gene expression profiles.

8. A method for processing transcriptome data, the method comprising: Receive the first batch of records of single-cell RNA sequencing data associated with multiple cells from tissue samples; The second batch of records (H) of spatial gene expression data from the tissue samples was received. s The records include spatial gene expression data and location data indicating the location of the spatial gene expression data in the tissue sample; The first batch of records was processed using a self-supervised autoencoder neural network to generate cell characterization records (H). c ); The cell characterization records (H) are generated using an untrained mapping matrix (M'). c Projecting the data onto the second batch of records to obtain predicted spatial gene expression records (H2N) s '); The spatial gene expression record is processed using a deep learning framework that implements graph self-supervised contrastive learning to obtain a reconstructed spatial gene expression record (H). s );as well as Based on the reconstructed spatial gene expression record (H s ) and the cell characterization record (H c The untrained mapping matrix (M') is trained to generate a trained mapping matrix (M).

9. The method of claim 8, wherein the method further comprises: The cell characterization records (H) are generated using a trained mapping matrix (M). c The data is projected onto the second batch of records, thereby performing spatial localization of cell types in the tissue sample.

Images