Similarity contrast space multi-omics integration method, system, device, and medium

Abstract

The application provides a similarity contrast space multi-omics integration method, system, device and medium, and belongs to the field of bioinformatics. The method preprocesses spatial multi-omics data and constructs a spatial adjacency graph, extracts each modality embedding through a double-branch graph encoder; designs intra-modality contrast learning, aggregates the node embedding and its neighborhood as positive samples, selects negative samples based on feature similarity, and enhances the capture of local spatial structure; designs cross-modality contrast learning, takes the same position cross-modality embedding as a positive sample and different positions as negative samples, and realizes semantic alignment; adaptively learns the modality weight through attention fusion and reconstructs the original data, and obtains joint representation by jointly optimizing three types of losses. The application can avoid misjudgment of biologically similar cells by random negative sampling, overcome the defect of fixed fusion, and be used for tissue spatial heterogeneity analysis and disease mechanism research.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to methods, systems, devices and media for integrating multi-omics in similarity comparison space. Background Technology

[0002] In recent years, the rapid development of sequencing technologies such as spatial transcriptomics, spatial epigenomics, and spatial proteomics has enabled researchers to simultaneously acquire multi-molecular-level characteristic profiles of gene expression, chromatin accessibility, and protein abundance from the same tissue slice, while preserving the in-situ spatial information of cells. Representative technologies such as DBiT-seq, MISAR-seq, and Spatial-ATAC-RNA-seq have been widely applied in fields such as embryonic development, neuroscience, and tumor immunology. With the increasing scale of spatial multi-omics data, effectively integrating heterogeneous data from different omics levels has become a core challenge for fully exploring its biological value.

[0003] However, existing methods still face several bottlenecks: insufficient modeling of local spatial dependencies makes it difficult to accurately capture the local continuity and microenvironment signals of spatially adjacent points; weak cross-modal semantic alignment capabilities, with significant differences in feature distribution and noise levels among different modalities, making it difficult for simple splicing or shallow alignment to effectively utilize cross-modal complementary information; the negative sampling strategy in contrastive learning lacks biological rationale, as random sampling may mistakenly treat biologically similar cells as negative samples, weakening the discriminative power of the representation; and the lack of adaptability in modality fusion, as fixed fusion strategies cannot dynamically adjust the contribution weights of each modality according to spatial location.

[0004] Existing methods still have limitations in addressing the aforementioned issues. While single-cell multi-omics integration tools (such as Seurat v4, MOFA+, and MultiVI) can achieve cross-modal alignment, they do not utilize spatial coordinate information and struggle to capture spatial dependencies. Single-modal spatial analysis methods (such as SpaGCN and SpaceFlow) can model spatial structures but lack cross-modal fusion mechanisms. Among methods specifically designed for spatial multi-omics, spaMultiVAE introduces spatial constraints but has excessively high computational costs; COSMOS's spatial modeling is limited to linear relationships; and while SpatialGlue and SpaMI are based on graph neural networks, their contrastive learning strategies still have shortcomings—intra-modal representations rely on random negative sampling, easily misclassifying biologically similar cells, and cross-modal alignment lacks systematic utilization of spatial proximity. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device, and medium for integrating multi-omics data in a similarity comparison space. By combining cross-modal and similarity-based intra-modal comparison learning strategies, a unified framework integrating spatial structure and cross-modal semantic alignment is constructed to achieve effective integration of heterogeneous multi-omics data.

[0006] To achieve the above objectives, this invention provides a multi-omics integration method for similarity comparison space, comprising the following steps: Step S1: Preprocess the input two modal data respectively to obtain the preprocessed two modal feature matrices. Construct a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point. Use a dual-branch graph encoder to perform graph convolution encoding on the two modal feature matrices respectively to obtain the two modal low-dimensional embedding matrices. Step S2: Based on the two-modal low-dimensional embedding matrix, for each modality, each spatial point is embedded and its neighborhood is aggregated as a positive sample pair. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The local spatial structure of each modality is enhanced through intra-modal contrastive learning to obtain the two-modal enhanced embedding matrix. Step S3: Based on the two-modal low-dimensional embedding matrix, cross-modal embeddings at the same spatial location are taken as cross-modal positive sample pairs, and cross-modal embeddings at different spatial locations are taken as cross-modal negative sample pairs. Cross-modal contrastive learning is used to achieve semantic alignment of the two modalities and obtain the cross-modal aligned two-modal embedding matrix. Step S4: Input the aligned two-modal embedding matrix into the attention fusion module, adaptively learn the modal weights to generate a joint representation; construct a decoder to reconstruct the joint representation to obtain the reconstructed two-modal feature matrix; jointly optimize the intra-modal contrast loss, cross-modal contrast loss and reconstruction loss to obtain the optimized joint representation.

[0007] This invention also provides a similarity comparison space multi-omics integration system, applied to the above-mentioned similarity comparison space multi-omics integration method, including: The data preprocessing and graph coding module is used to preprocess the input two-modal data to obtain the preprocessed two-modal feature matrices. It constructs a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point, and uses a dual-branch graph encoder to perform graph convolution coding on the two-modal feature matrices to obtain the two-modal low-dimensional embedding matrix. The intramodal contrastive learning module is used to embed each spatial point of each modality and aggregate it with its neighborhood as positive sample pairs based on the two-modal low-dimensional embedding matrix. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The intramodal contrastive learning enhances the local spatial structure of each modality and obtains the two-modal enhanced embedding matrix. The cross-modal contrastive learning module is used to align the semantics of the two modalities based on the two-modal low-dimensional embedding matrix. It takes cross-modal embeddings at the same spatial location as cross-modal positive sample pairs and cross-modal embeddings at different spatial locations as cross-modal negative sample pairs. Through cross-modal contrastive learning, the two-modal semantic alignment is achieved, and the cross-modal aligned two-modal embedding matrix is ​​obtained. The attention fusion and joint optimization module is used to input the aligned two-modal embedding matrix into the attention fusion module to adaptively learn the modal weights to generate a joint representation. The decoder is constructed to reconstruct the joint representation to obtain the reconstructed two-modal feature matrix. The intra-modal contrast loss, cross-modal contrast loss and reconstruction loss are jointly optimized to obtain the optimized joint representation.

[0008] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described similarity comparison space multi-omics integration method.

[0009] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described similarity comparison space multi-omics integration method.

[0010] Therefore, the present invention employs the above-mentioned method, system, device, and medium for integrating multiple omics in similarity comparison space, and the beneficial technical effects are as follows: (1) The present invention constructs positive sample pairs by embedding each spatial point and its neighborhood aggregation representation through the intramodal contrast learning module. Based on the cosine similarity of the original features, nodes below the preset threshold are selected as negative sample pairs, which avoids the problem of randomly negative sampling mistaking biologically similar cells as negative samples and improves the discrimination ability of the representation.

[0011] (2) The present invention constructs positive sample pairs by constructing the embeddings of the same spatial location in different modalities and negative sample pairs by constructing the embeddings of different spatial locations through the cross-modal contrast learning module. This effectively realizes cross-modal semantic alignment, solves the problem of inconsistent feature distribution and significant differences in noise level between different modalities, and preserves the unique information of each modality.

[0012] (3) The present invention adaptively learns the modal weights and generates joint representations through the attention fusion module, so that the model can dynamically allocate modal contribution weights according to the features of each spatial location, thus overcoming the lack of adaptability of the fixed fusion strategy.

[0013] (4) By jointly optimizing intramodal contrast loss, cross-modal contrast loss and reconstruction loss, the present invention provides a high-quality joint representation for spatial domain recognition and organizational structure analysis. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the multi-omics integration method for similarity comparison space of the present invention. Figure 2This is a comparison diagram of the spatial domain recognition results of the present invention and the comparison method on a simulated dataset, wherein... Figure 2 In the figure, 'a' represents a comparison of the spatial domain recognition results of the present invention and the comparison method on a simulated dataset. Figure 2 In the figure, b represents a comparison of the UMAP visualization results of the joint embedding of the present invention and the comparison method onto a simulated dataset; Figure 3 This is a comparison image showing the results of spatial domain recognition of the present invention and the comparative method on the E15.5 mouse embryo brain dataset. Figure 3 In the figure, 'a' is a manually annotated reference image of the actual anatomical region from the E15.5 mouse embryo brain dataset. Figure 3 In the figure, b represents a bar chart showing the quantitative evaluation index of spatial domain identification of the present invention and the comparative method on the E15.5 mouse embryonic brain dataset. Figure 3 In the diagram, 'c' represents the visualization results of spatial domain recognition on the E15.5 mouse embryonic brain dataset and the UMAP dimensionality reduction visualization comparison chart of the present invention and the comparison method. Figure 3 In the diagram, 'd' represents a comparison between the original expression profile of a representative marker gene and its enhanced expression profile as described in this invention. Figure 3 In the figure, 'e' represents a heatmap of the expression of marker genes in different spatial domains on the E15.5 mouse embryo brain dataset. Figure 4 This is a comparison image showing the spatial domain recognition results of the present invention and the comparison method on the P22 mouse coronal slice dataset. Figure 4 In the figure, 'a' is a manually annotated reference image of the actual anatomical region in the P22 mouse coronal section dataset. Figure 4 In the figure, b represents a bar chart showing the quantitative evaluation index of spatial domain recognition of the present invention and the comparative method on the P22 mouse coronal slice dataset. Figure 4 In the diagram, 'c' represents the visualization results of spatial domain recognition on the P22 mouse coronal slice dataset and the comparison diagram of UMAP dimensionality reduction using the present invention and the comparison method. Figure 4 In the diagram, 'd' represents a comparison between the original expression of a representative marker gene and its enhanced expression profile as described in this invention. Detailed Implementation

[0015] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0017] Example 1 like Figure 1 As shown, this embodiment provides a multi-omics integration method for similarity comparison space, including the following steps: Step S1: Preprocess the input two modal data respectively to obtain the preprocessed two modal feature matrices. Construct a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point. Use a dual-branch graph encoder to perform graph convolution encoding on the two modal feature matrices respectively to obtain the two modal low-dimensional embedding matrices.

[0018] Step S11: Preprocess the input data for both modalities. Use Scanpy to perform quality control on the transcriptome data, remove low-expression genes, perform library size normalization and log1p transformation, screen for highly variable genes, and then perform dimensionality reduction through principal component analysis; other modalities are processed using the corresponding standardization procedures.

[0019] Step S12, for those containing A spatial point, with coordinates as For each spatial omics dataset, construct a spatial neighborhood graph. ,in It is the set of real numbers and the set of nodes. For each spatial point, the edge set Connect adjacent points in space. It is the first A spatial point, a node With nodes Euclidean distance between Defined as: ; in, and They are nodes x and y coordinates and They are nodes The x and y coordinates; choose Establish a connection by using the nearest point as a neighbor. Values ​​that are too small can lead to insufficient graph connectivity and an inability to capture the global spatial structure, while values ​​that are too large can introduce noisy connections and blur local spatial boundaries. The optimal value depends on the point density and organization of different datasets. Set the number of adjacent nodes to 3 to 6; obtain the initial adjacency matrix. Then, it is symmetricized to ensure that the spatial neighborhood graph is undirected, thus obtaining the adjacency matrix. : ; in, For symbolic functions, This is a transpose.

[0020] Step S13: Assign an independent graph encoder to each omics modality, based on the adjacency matrix. A dual-branch graph encoder is used to perform graph convolution encoding on the two modal feature matrices after preprocessing in step S11. Each graph encoder consists of two layers of graph convolution, which are represented as follows: ; in, For the first The node feature matrix of the layer, For the first The node feature matrix of the layer, , It is an adjacency matrix. It is the identity matrix. for The angle matrix, For the first Layer trainable weight matrix, It is the ReLU activation function; set up and These represent the two modal feature matrices and the low-dimensional node embedding matrices output by the two encoders, respectively. and They are respectively: , ; in, (·) represents the graphical encoder function for mode 1. (·) is the graph encoder function for mode 2.

[0021] Step S2: Based on the two-modal low-dimensional embedding matrix, for each modality, each spatial point is embedded and its neighborhood is aggregated as a positive sample pair. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The local spatial structure of each modality is enhanced through intra-modal contrastive learning to obtain the two-modal enhanced embedding matrix.

[0022] Step S21: Based on the low-dimensional node embedding matrix obtained in step S13, for each node... Computing local context representation : ; in, Spatial adjacency graph Middle node The set of neighboring nodes, For the number of neighbors, For nodes In modality The following embedding will and Construct positive sample pairs. For nodes In modality Embedded below.

[0023] Step S22: Construct negative samples based on feature similarity. To address the problem of traditional random scrambling easily mistaking similar cells as negative samples, a negative sampling strategy based on feature similarity is proposed. This strategy guides sampling through feature space similarity to enhance discriminative power. Given the original feature matrices for each modality... , The feature dimension is used, and then the cosine similarity is calculated. : ; in, It is a node In modality The eigenvectors below, It is a node In modality The eigenvectors below, It is the L2 norm of the vector. and For nodes Choose those with a cosine similarity lower than 1. The nodes are used as candidate negative samples, and the candidate negative sample index set is... for: ; in, It is a similarity threshold, in this embodiment = 0.4, If the candidate set is not empty, randomly sample a node. Use it as a negative sample; if it is empty, select the node with the lowest similarity. Based on this, row-level random permutation decouples the feature vector from the identity of its source sample by shuffling the data rows of each cell / point, thus constructing a damaged feature matrix while preserving spatial adjacency relationships and the overall data distribution. Damaged embeddings obtained through the encoder Damaged nodes in the diagram Local context representation for: ; in, To damage the nodes in the diagram In modality The following embedding will and Construct negative sample pairs.

[0024] Step S23: Based on the positive sample pairs in step S21 and the negative sample pairs in step S22, enhance the ability to capture the local spatial structure of each modality through intra-modal contrastive learning. To achieve this, construct modal... Intramodal contrast loss : ; in, For modality The discriminator function.

[0025] The total intra-modal contrast loss is obtained by weighting the intra-modal contrast losses of the two modalities. : ; in, For the intra-modal contrast loss of mode 1, For the intra-modal contrast loss of mode 2, =1.0 is the loss weight coefficient for mode 1. =1.0 is the loss weight coefficient for mode 2.

[0026] Based on the minimized total intra-modal contrast loss, the trainable parameters of the bi-branch graph encoder are updated via backpropagation. and Input the updated bi-branch graph encoder to obtain the modality 1 augmented embedding matrix and the modality 2 augmented embedding matrix.

[0027] Step S3: Based on the two-modal low-dimensional embedding matrix, cross-modal embeddings at the same spatial location are taken as cross-modal positive sample pairs, and cross-modal embeddings at different spatial locations are taken as cross-modal negative sample pairs. Cross-modal contrastive learning is used to achieve semantic alignment of the two modalities and obtain the cross-modal aligned two-modal embedding matrix.

[0028] Step S31: Based on the low-dimensional node embedding matrix obtained in S13, construct cross-modal positive sample pairs by embedding the same spatial location in modality 1 and modality 2. Cross-modal positive sample pairs cosine similarity for: ; in, For nodes Low-dimensional embedding in mode 1, For nodes Low-dimensional embedding in mode 2.

[0029] Step S32: In cross-modal contrastive learning, a batch sampling strategy is used to construct cross-modal negative samples. Cross-modal embeddings from different spatial locations are constructed as cross-modal negative sample pairs, with modality 1 node as the starting point. The set of negative sample indices in mode 2 when anchoring. for: ; With mode 2 nodes The set of negative sample indices in mode 1 when anchoring for: .

[0030] Step S33: Based on cross-modal positive sample pairs and cross-modal negative sample pairs, semantic alignment between the two modalities is achieved through cross-modal contrastive learning. To achieve this, a cross-modal contrastive loss from modality 1 to modality 2 is constructed. for: ; Cross-modal contrast loss from mode 2 to mode 1 for: ; in, For mode 1 node With mode 2 nodes Cross-modal negative sample pair similarity between For mode 2 nodes With mode 1 node Cross-modal negative sample pair similarity; total cross-modal contrastive loss for: ; in, As a weightable coefficient, in this embodiment =0.5.

[0031] Based on the minimized overall cross-modal contrastive loss, the trainable parameters of the bi-branch graph encoder are updated via backpropagation; and Input the updated bi-branch graph encoder to obtain the cross-modal aligned mode 1 embedding matrix. and mode 2 embedding matrix .

[0032] Step S4: Input the two-modal enhanced embedding matrix and the aligned two-modal embedding matrix into the attention fusion module, adaptively learn the modal weights to generate a joint representation; construct a decoder to reconstruct the joint representation to obtain the reconstructed two-modal feature matrix; jointly optimize the intra-modal contrast loss, cross-modal contrast loss and reconstruction loss to obtain the optimized joint representation.

[0033] Step S41: After completing cross-modal contrastive learning, embed the cross-modal aligned Mode 1 embedding matrix obtained in step S3. Mode 2 The input attention fusion module stacks the cross-modal aligned modality 1 embedding matrix and modality 2 embedding matrix along the new dimension to form... ,in For stacking operations, attention weights are calculated using an attention mechanism. And a weighted sum is obtained to obtain a joint statement : ; ; in, The Softmax activation function is used. The hyperbolic tangent activation function is used. and For learnable parameters, For embedded dimensions, and This is the attention weight vector, used to dynamically adjust the contribution weights of each modality. This is for element-wise multiplication.

[0034] Step S42: To better preserve the representation patterns of each modality, an independent decoder is designed for each modality, and the joint representation is... Reconstructing back to the original feature space. Each decoder employs a graph convolutional network architecture symmetric to the encoder; the modality-specific decoder is defined as: = ( ); = ( ); in, and The decoders for mode 1 and mode 2 are respectively... l The output features of the layer and The decoders for mode 1 and mode 2 are respectively... The trainable weight matrix of the layer, and The decoders for mode 1 and mode 2 are respectively... The trainable bias of the layer is used to obtain the reconstructed modality 1 feature matrix and modality 2 feature matrix.

[0035] Step S43: The reconstruction loss uses mean squared error to minimize the difference between the original features and the reconstructed features, so as to ensure that the joint representation retains the key biological information of each modality. = ; in, To reconstruct the loss, and They are nodes The original feature vectors in mode 1 and mode 2, and They are nodes Reconstructing feature vectors in mode 1 and mode 2, and In this embodiment, the weighting coefficients for the mode-specific reconstruction loss are... =10, =15.

[0036] By jointly optimizing the total intra-modal contrast loss, total cross-modal contrast loss, and reconstruction loss, the total loss is obtained. for: ; To verify the effectiveness of this method, this embodiment systematically evaluates its performance from aspects such as spatial domain identification, gene expression enhancement experiments, and heatmap analysis.

[0037] 1. Data set and evaluation metrics.

[0038] The datasets used in this embodiment include a simulated dataset and real spatial multi-omics datasets from the Spatial-ATAC-RNA-seq and MISAR-seq platforms, covering gene expression matrices, spatial location information, and histological images. A summary of the datasets is shown in Table 1. Simulated dataset: The simulated dataset was generated according to a predetermined scheme, containing four spatial structure factors and one background region, with a total of five cell populations. Transcriptomic and proteomic data were simulated using zero-inflated negative binomial and negative binomial distributions, respectively, and spatial coordinates were assigned to construct an adjacency graph. Real dataset: From the MISAR-seq platform, it contains spatial transcriptomic and chromatin accessibility data of the E15.5 mouse embryonic brain, with manually labeled spatial distributions of seven anatomical regions. A second real dataset: From the Spatial-ATAC-RNA-seq platform, it contains spatial transcriptomic and chromatin accessibility data of coronal sections of the P22 mouse brain, covering all layers of the cortex, the caudate nucleus, the lateral ventricles, and the nucleus accumbens, among other major brain regions.

[0039] To quantitatively evaluate the clustering performance of spatial domain identification, this embodiment employs three widely used clustering evaluation metrics: Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Homogeneity Score. ARI measures the consistency between predicted clustering results and true labels, correcting for random conditions; its value ranges from [0, 1], with higher values ​​indicating a closer match to the true clustering. NMI quantifies the amount of shared information between the two clustering results, also ranging from [0, 1], with higher values ​​indicating stronger clustering consistency. The Homogeneity Score assesses whether each predicted cluster contains only samples from the same true category; its value ranges from [0, 1], with values ​​closer to 1 indicating higher cluster purity. Combining these three metrics, ARI and NMI evaluate overall clustering consistency, while the Homogeneity Score evaluates intra-cluster purity. These three metrics complement each other, collectively reflecting the accuracy and reliability of the clustering results.

[0040] Table 1. Dataset Overview

[0041] 2. Comparative analysis of experimental results.

[0042] (1) Spatial domain recognition experiment.

[0043] a. Simulated dataset.

[0044] SpaCCS is a similarity-based spatial multi-omics integration method proposed in this invention. Its performance was first evaluated on a simulated spatial multi-omics dataset containing four spatial structure factors and a background region, representing five cell populations. Spatial domain identification results ( Figure 2 As shown in a), SpaCCS accurately recovered all four spatial factors and the background region, with clear boundaries and complete structure. However, SpaMI and SpatialGlue lacked precise boundary delineation, SpaceFlow exhibited blurriness at factor boundaries, and SpaMultiVAE failed to fully recover all factors. UMAP visualization ( Figure 2 b) further demonstrates that the joint embedding of SpaCCS has the clearest cluster separation and the most compact cluster structure.

[0045] b.E15.5 mouse embryonic brain.

[0046] SpaCCS was systematically compared with five state-of-the-art methods (SpaMI, SpatialGlue, SpaceFlow, COSMOS, and SpaMultiVAE) on the E15.5 mouse embryonic brain dataset, which includes annotations of seven real anatomical regions. Quantitative assessment ( Figure 3As shown in b), SpaCCS achieves the best performance in all three metrics—ARI, NMI, and Homogeneity—significantly outperforming all compared methods. Spatial domain identification results ( Figure 3 As shown in c), SpaCCS successfully identified all seven anatomical regions with clear boundaries between them, consistent with manually labeled references. Figure 3 The results in (a) show a high degree of agreement. In contrast, while SpaMI and SpatialGlue can recover the main brain region structures, they exhibit blurred boundaries in some regions; COSMOS shows confusion between several adjacent regions; SpaceFlow produces fragmented clustering results in both RNA and ATAC modalities, failing to accurately reflect the true anatomical divisions; and SpaMultiVAE shows relatively blurred cluster boundaries. UMAP dimensionality reduction visualization ( Figure 3 c) further demonstrates that the joint embeddings learned by SpaCCS achieve the clearest cluster separation in low-dimensional space, with well-defined boundaries between brain regions and the most consistent correspondence with the real labels. These results fully validate the effectiveness and superiority of SpaCCS in the analysis of complex tissue spatial structures.

[0047] c. P22 mouse coronal section.

[0048] SpaCCS was systematically compared with four state-of-the-art methods (SpaMI, SpatialGlue, SpaceFlow, and COSMOS) on the P22 mouse coronal slice dataset. This dataset, based on the Allen Brain Atlas anatomical reference, manually annotated nine major brain regions. Figure 4 a). Quantitative assessment ( Figure 4 As shown in b), SpaCCS achieved the highest scores on all three metrics—ARI, NMI, and Homogeneity—significantly outperforming all the comparison methods. Spatial domain identification results ( Figure 4 Figure c) shows that SpaCCS successfully identified all major brain region structures with clear boundaries between regions, closely matching the manually annotated reference maps. These results validate the superior performance of SpaCCS in resolving fine substructures of brain regions.

[0049] (2) Gene expression enhancement experiment.

[0050] To further verify that the joint representation learned by SpaCCS has good generalization ability, this embodiment conducted gene expression enhancement experiments on multiple datasets. By comparing the original gene expression with the expression maps enhanced by SpaCCS, the model's performance in denoising and spatial expression recovery was evaluated.

[0051] a.E15.5 mouse embryonic brain.

[0052] To further validate the biological effectiveness of the joint representations learned by SpaCCS, gene expression enhancement experiments were conducted on the E15.5 mouse embryonic brain dataset. Representative marker genes Myh8, Col2a1, Pax2, and Tiam2 were selected, and their original expression profiles were compared with those enhanced by SpaCCS. Figure 3 (d) The results show that the enhanced expression map effectively reduced spatial noise, restored spatially continuous expression patterns, and made region boundaries clearer, revealing fine spatial expression structures that were not very obvious in the original data. Specifically, Myh8 was highly expressed in the muscle region, Col2a1 was enriched in the cartilage region, Pax2 showed a stable expression signal within the mesenchymal region cluster, and Tiam2 also showed a consistent expression pattern within its corresponding spatial domain cluster. These distribution patterns are highly consistent with previously reported anatomical markers, strongly supporting the biological relevance of the spatial domains identified by SpaCCS. These results demonstrate that SpaCCS, through cross-modal contrastive learning and attention fusion mechanisms, can effectively remove spatial noise, restore biologically relevant spatial expression patterns, and the enhanced gene expression features have good biological interpretability.

[0053] b. Coronal section of P22 mouse.

[0054] The gene expression enhancement effect was further validated on the P22 mouse coronal slice dataset. Marker genes enriched in regions such as Dach1, Scn4b, Gsn, Nnat, and Prkar2b were selected, and their original expression and enhanced expression profiles were compared. Figure 4 (d) The results showed that the enhanced expression map effectively reduced spatial noise, restored spatially continuous expression patterns, and made region boundaries clearer, revealing fine spatial expression structures that were not very obvious in the original data. Specifically, Dach1 and Scn4b showed a clear layered distribution in the cortical region, Gsn was specifically enriched in the caudate nucleus, Mog was highly expressed in the white matter region, and Nnat and Prkar2b showed obvious spatial enrichment characteristics in the lateral ventricle and nucleus accumbens, respectively. The spatial distribution of each gene was highly consistent with brain region markers reported in Allen Brain Atlas, validating the biological authenticity of the identified spatial domains at the molecular level. This provides reliable molecular evidence for the fine analysis of complex brain region substructures.

[0055] To further verify the molecular consistency of the SpaCCS recognition spatial domain, a heatmap of marker gene expression was plotted on the E15.5 mouse embryonic brain dataset. Figure 3(e). The results showed that the expression patterns of the marker genes Myh8, Col2a1, Pax2, and Tiam2 were highly consistent within each spatial cluster, while significant expression differences were observed between different clusters. Specifically, Myh8 was highly expressed and evenly distributed in the muscle-related region cluster, Col2a1 was specifically enriched in the cartilage region cluster, Pax2 showed a stable expression signal in the mesenchymal region cluster, and Tiam2 also showed a consistent expression pattern within its corresponding spatial domain cluster. The expression consistency of each gene within its own cluster was significantly higher than in other regions. The expression profiles of different spatial clusters differed significantly, and each cluster had unique molecular characteristics, indicating that the spatial domains exhibited good intra-regional molecular homogeneity and inter-regional molecular heterogeneity.

[0056] Example 2 A multi-omics ensemble system for similarity comparison space, including: The data preprocessing and graph coding module is used to preprocess the input two-modal data to obtain the preprocessed two-modal feature matrices. It constructs a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point, and uses a dual-branch graph encoder to perform graph convolution coding on the two-modal feature matrices to obtain the two-modal low-dimensional embedding matrix. The intramodal contrastive learning module is used to embed each spatial point of each modality and aggregate it with its neighborhood as positive sample pairs based on the two-modal low-dimensional embedding matrix. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The intramodal contrastive learning enhances the local spatial structure of each modality and obtains the two-modal enhanced embedding matrix. The cross-modal contrastive learning module is used to align the semantics of the two modalities based on the two-modal low-dimensional embedding matrix. It takes cross-modal embeddings at the same spatial location as cross-modal positive sample pairs and cross-modal embeddings at different spatial locations as cross-modal negative sample pairs. Through cross-modal contrastive learning, the two-modal semantic alignment is achieved, and the cross-modal aligned two-modal embedding matrix is ​​obtained. The attention fusion and joint optimization module is used to input the aligned two-modal embedding matrix into the attention fusion module to adaptively learn the modal weights to generate a joint representation. The decoder is constructed to reconstruct the joint representation to obtain the reconstructed two-modal feature matrix. The intra-modal contrast loss, cross-modal contrast loss and reconstruction loss are jointly optimized to obtain the optimized joint representation.

[0057] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0059] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0060] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0061] Therefore, this invention employs the aforementioned multi-omics integration method, system, device, and medium for similarity comparison space. It avoids the problem of randomly treating biologically similar cells as negative samples by using an intra-modal contrastive learning module, effectively achieves cross-modal semantic alignment by using a cross-modal contrastive learning module, achieves adaptive allocation of modal weights by using an attention fusion module, and obtains high-quality joint representations by jointly optimizing each loss function.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-omics integration method for similarity comparison space, characterized in that, Includes the following steps: Step S1: Preprocess the input two modal data respectively to obtain the preprocessed two modal feature matrices. Construct a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point. Use a dual-branch graph encoder to perform graph convolution encoding on the two modal feature matrices respectively to obtain the two modal low-dimensional embedding matrices. Step S2: Based on the two-modal low-dimensional embedding matrix, for each modality, each spatial point is embedded and its neighborhood is aggregated as a positive sample pair. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The local spatial structure of each modality is enhanced through intra-modal contrastive learning to obtain the two-modal enhanced embedding matrix. Step S3: Based on the two-modal low-dimensional embedding matrix, cross-modal embeddings at the same spatial location are taken as cross-modal positive sample pairs, and cross-modal embeddings at different spatial locations are taken as cross-modal negative sample pairs. Cross-modal contrastive learning is used to achieve semantic alignment of the two modalities and obtain the cross-modal aligned two-modal embedding matrix. Step S4: Input the aligned two-modal embedding matrix into the attention fusion module, adaptively learn the modal weights to generate a joint representation; construct a decoder to reconstruct the joint representation and obtain the reconstructed two-modal feature matrix; The intramodal contrast loss, cross-modal contrast loss, and reconstruction loss are jointly optimized to obtain the optimized joint characterization.

2. The similarity comparison space multi-omics integration method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Preprocess the input two-modal data, including quality control of transcriptome data, library size normalization, log1p transformation, screening of highly variable genes and principal component analysis for dimensionality reduction, and output the preprocessed two-modal feature matrix. Step S12, for those containing A spatial point, with coordinates as For each spatial omics dataset, construct a spatial neighborhood graph. , where the node set For each spatial point, the edge set Connect adjacent points in space. For the first A spatial point; choose Establish connections between the nearest points as neighbors; obtain the initial adjacency matrix. Then, for the initial adjacency matrix Symmetry processing is performed to obtain the adjacency matrix. ; Step S13: Assign an independent graph encoder to each omics modality, based on the adjacency matrix. A dual-branch graph encoder is used to perform graph convolution encoding on the preprocessed two-modality feature matrices respectively; set up and These represent the initial input feature matrices for the two modalities and the low-dimensional node embedding matrices output by the two encoders, respectively. and They are respectively: ， ； in, (·) represents the graphical encoder function for mode 1. (·) is the graph encoder function for mode 2.

3. The multi-omics integration method for similarity comparison space according to claim 2, characterized in that, Step S2 includes the following steps: Step S21: Based on the low-dimensional node embedding matrix, for each node... Computing local context representation : ； in, Spatial adjacency graph Middle node The set of neighboring nodes, For the number of neighbors, For nodes In modality The following embedding will and Construct positive sample pairs. For nodes In modality Embedded below; Step S22: Based on the original feature matrices of each modality Calculate cosine similarity : ； in, For nodes In modality The eigenvectors below, For nodes In modality The eigenvectors below, Let L2 norm be the vector. and , For the transpose; for the node Choose those with a cosine similarity lower than 1. The nodes are used as candidate negative samples, and the candidate negative sample index set is... for: ； Among them, a damage feature matrix is ​​constructed based on candidate negative samples. Damaged embeddings obtained through the encoder Damaged nodes in the diagram Local context representation for: ； in, To damage the nodes in the diagram In modality The following embedding will and Construct negative sample pairs; Step S23: Based on positive and negative sample pairs, enhance the ability to capture the local spatial structure of each modality through intra-modal contrastive learning, and construct modalities. Intramodal contrast loss for: ； in, For modality The discriminator function; The total intra-modal contrast loss is obtained by weighting the intra-modal contrast losses of the two modalities. : ； in, For the intra-modal contrast loss of mode 1, For the intra-modal contrast loss of mode 2, These are the loss weighting coefficients for mode 1. These are the loss weighting coefficients for mode 2; Based on the minimized total intra-modal contrast loss, the trainable parameters of the bi-branch graph encoder are updated via backpropagation. and Input the updated bi-branch graph encoder to obtain the modality 1 augmented embedding matrix and the modality 2 augmented embedding matrix.

4. The multi-omics integration method for similarity comparison space according to claim 3, characterized in that, Step S3 includes the following steps: Step S31: Based on the low-dimensional node embedding matrix, construct cross-modal positive sample pairs by embedding the same spatial location in modality 1 and modality 2. Cross-modal positive sample pairs cosine similarity for: ； in, For nodes Low-dimensional embedding in mode 1, For nodes Low-dimensional embedding in mode 2; Step S32: Construct cross-modal negative sample pairs from cross-modal embeddings at different spatial locations, using the modality 1 node as an example. The set of negative sample indices in mode 2 when anchoring. for: ； With mode 2 nodes The set of negative sample indices in mode 1 when anchoring for: ； Step S33: Based on cross-modal positive sample pairs and cross-modal negative sample pairs, achieve semantic alignment between the two modalities through cross-modal contrastive learning, and construct a cross-modal contrastive loss from modality 1 to modality 2. for: ； Cross-modal contrast loss from mode 2 to mode 1 for: ； in, For mode 1 node With mode 2 nodes Cross-modal negative sample pair similarity between For mode 2 nodes With mode 1 node Cross-modal negative sample pair similarity; Overall cross-modal contrast loss for: ； in, These are the weighting coefficients; Based on the minimized overall cross-modal contrastive loss, the trainable parameters of the bi-branch graph encoder are updated via backpropagation; and Input the updated bi-branch graph encoder to obtain the cross-modal aligned mode 1 embedding matrix. and mode 2 embedding matrix .

5. The multi-omics integration method for similarity comparison space according to claim 4, characterized in that, Step S4 includes the following steps: Step S41: Embed the cross-modal aligned mode 1 embedding matrix obtained in step S3. Mode 2 The input attention fusion module stacks the cross-modal aligned modality 1 embedding matrix and modality 2 embedding matrix along the new dimension to form... Calculate attention weights through attention mechanisms And a weighted sum is obtained to obtain a joint statement : ； ； in, The Softmax activation function is used. The hyperbolic tangent activation function is used. and For learnable parameters, and For attention weight vectors, For element-wise multiplication; Step S42: Construct the decoder and process the joint representation. Refactor; Step S43: Calculate the reconstruction loss, and jointly optimize the reconstruction loss, intra-modal contrast loss, and cross-modal contrast loss to obtain the total loss.

6. The multi-omics integration method for similarity comparison space according to claim 5, characterized in that, In step S42, the decoder reconstruction process is as follows: = ( )； = ( )； in, and The decoders for mode 1 and mode 2 are respectively... The output features of the layer and The decoders for mode 1 and mode 2 are respectively... The trainable weight matrix of the layer, and The decoders for mode 1 and mode 2 are respectively... The trainable bias of the layer is used to obtain the reconstructed modality 1 feature matrix and modality 2 feature matrix.

7. A multi-omics integration system for similarity comparison space, characterized in that, The method for integrating multiple omics in similarity comparison space according to any one of claims 1-6 includes: The data preprocessing and graph coding module is used to preprocess the input two-modal data to obtain the preprocessed two-modal feature matrices. It constructs a k-nearest neighbor spatial adjacency graph based on the spatial coordinates of each spatial point, and uses a dual-branch graph encoder to perform graph convolution coding on the two-modal feature matrices to obtain the two-modal low-dimensional embedding matrix. The intramodal contrastive learning module is used to embed each spatial point of each modality and aggregate it with its neighborhood as positive sample pairs based on the two-modal low-dimensional embedding matrix. Nodes below the threshold are selected as negative sample pairs based on feature similarity. The intramodal contrastive learning enhances the local spatial structure of each modality and obtains the two-modal enhanced embedding matrix. The cross-modal contrastive learning module is used to align the semantics of the two modalities based on the two-modal low-dimensional embedding matrix. It takes cross-modal embeddings at the same spatial location as cross-modal positive sample pairs and cross-modal embeddings at different spatial locations as cross-modal negative sample pairs. Through cross-modal contrastive learning, the two-modal semantic alignment is achieved, and the cross-modal aligned two-modal embedding matrix is ​​obtained. The attention fusion and joint optimization module is used to input the aligned two-modal embedding matrix into the attention fusion module to adaptively learn the modal weights to generate a joint representation. The decoder is constructed to reconstruct the joint representation to obtain the reconstructed two-modal feature matrix. The intra-modal contrast loss, cross-modal contrast loss and reconstruction loss are jointly optimized to obtain the optimized joint representation.

8. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the similarity comparison space multi-omics integration method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the steps of the multi-omics integration method for similarity comparison space as described in any one of claims 1-6.

Images