A method, system and storage medium for integrating single-cell and spatial proteomic data based on mRNA sequence guidance

Abstract

The present application belongs to the field of bioinformatics and computational biology, and specifically relates to a single-cell and spatial proteomics data integration method based on mRNA sequence guidance, comprising two cores: regulatory embedding extraction and batch effect correction. The method framework firstly encodes the mRNA sequence corresponding to the protein into an embedding vector, and then iteratively reconstructs the cross-cell protein expression profile by using the embedding through a fully connected neural network. The weights learned in this process are used as regulatory representations, which capture the intrinsic cell-specific functional patterns. To correct the batch effect across datasets, SANPRO uses a triplet loss function to pull similar cells identified by mutual nearest neighbor (MNN) pairing, thereby effectively aligning the distributions under different experimental conditions. Comprehensive evaluation results show that SANPRO is superior to existing methods in capturing cell heterogeneity.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, specifically relating to a method, system, and storage medium for integrating single-cell and spatial proteomics data guided by mRNA sequences. Background Technology

[0002] In recent years, breakthroughs in high-sensitivity mass spectrometry platforms and single-cell sequencing technologies have significantly enhanced our ability to characterize the proteome at single-cell resolution. Single-cell proteomics analysis reveals cellular heterogeneity and molecular regulatory mechanisms in complex tissues (such as the tumor microenvironment), providing valuable molecular biomarkers for disease diagnosis and targeted therapy. For example, mass cytometry (CyTOF) and antibody-based sequencing technologies (such as CITE-seq and REAP-seq) have become important tools for elucidating cellular identity and functional status.

[0003] However, the analysis of single-cell proteomics data still faces significant computational challenges. First, due to technical noise, dropout, and limitations in instrument sensitivity, raw data often contain a large amount of non-biological variation. Second, significant batch effects exist between data from different experimental batches, different sequencing platforms, and even different species, leading to inconsistent distributions of the same cell type in different datasets, which severely hinders effective data integration and comparative analysis.

[0004] Existing computational methods often directly borrow from workflows used in single-cell RNA sequencing (scRNA-seq), such as K-nearest neighbor (KNN)-based interpolation algorithms or statistical correction methods like ComBat. However, proteomics data differ fundamentally from transcriptomics data in terms of data structure, feature sparsity, and noise distribution. For example, KNN-based interpolation methods are prone to introducing artifacts into proteomics data, especially in the presence of strong batch effects; while methods like ComBat assume shared biological components across batches, an assumption often violated in highly heterogeneous single-cell proteomics data. Furthermore, existing methods typically treat proteins independently, neglecting the potential regulatory relationships between proteins and their corresponding mRNA sequences. In reality, protein abundance is regulated by translation efficiency, mRNA stability, and sequence motifs; integrating mRNA sequence information holds promise for providing more biologically meaningful prior signals for protein characterization.

[0005] Therefore, there is an urgent need for a novel computational framework that can fully utilize mRNA sequence regulatory information while effectively addressing noise, batch effects, and cross-species differences in single-cell and spatial proteomics. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for integrating single-cell and spatial proteomics data based on mRNA sequence guidance. It aims to enhance the characterization ability of proteomics data by mining regulatory signals in mRNA sequences through deep learning models, and to achieve robust cross-batch, cross-species and spatial data integration.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for integrating single-cell and spatial proteomics data guided by mRNA sequences, comprising the following steps:

[0009] Acquire single-cell or spatial proteomics data, including protein abundance measurements of multiple cells and the corresponding mRNA nucleotide sequences of the proteins;

[0010] Construct a sequence-guided representation learning model, which includes a hierarchical convolutional neural network encoder and a fully connected prediction layer;

[0011] The mRNA nucleotide sequence is input into the encoder to extract sequence features, and the protein abundance of each cell is predicted through the fully connected prediction layer.

[0012] During training, the model parameters are iteratively optimized by minimizing the mean squared error loss between the predicted protein abundance and the actual protein abundance.

[0013] After training, the shared weight matrix in the fully connected prediction layer is extracted, and the shared weight matrix is ​​used as a high-dimensional embedding vector for each cell. The high-dimensional embedding vector encapsulates the regulatory relationship between mRNA sequence motifs and protein expression.

[0014] Based on the high-dimensional embedding vector, a cross-batch mutual nearest neighbor graph is constructed to identify the mutual nearest neighbor relationships between cells from different experimental batches or different species, and to establish topological connections between cells.

[0015] The high-dimensional embedding vector is reconstructed using a symmetric autoencoder architecture and jointly optimized using a triplet loss function to correct batch effects while preserving biological signals, and outputs a corrected cell embedding vector.

[0016] Further design optimization includes constructing a sequence-guided representation learning model, specifically comprising:

[0017] One-hot encoding is performed on a fixed-length mRNA nucleotide sequence, which is then mapped into a multidimensional binary input matrix;

[0018] The input matrix is ​​processed by a first one-dimensional convolutional layer containing multiple convolutional kernels, followed by batch normalization, activation function processing, and max pooling operations to obtain a first feature map.

[0019] The first feature map is input into a convolutional tower consisting of multiple stacked convolutional blocks, each using a convolutional kernel of fixed width, and the number of convolutional kernels gradually increases between layers to capture multi-scale contextual patterns, thus obtaining a second feature map.

[0020] The feature dimension of the second feature map is reduced by pointwise convolution, and the flattened feature vector is obtained. It is then projected into a low-dimensional latent space through a fully connected bottleneck layer containing dropout and batch normalization to obtain the initial sequence embedding.

[0021] Further design optimization includes, specifically, constructing a cross-batch nearest neighbor graph based on the high-dimensional embedding vector:

[0022] Obtain the high-dimensional embedding vectors of all cells from the first and second batches respectively;

[0023] For any cell in the first batch, calculate its distance to all cells in the second batch in the latent space, and determine the K nearest neighbor cells in the second batch;

[0024] Meanwhile, for any cell in the second batch, calculate its distance to all cells in the first batch in the latent space, and determine the K nearest neighbor cells in the first batch.

[0025] If cell A in the first batch and cell B in the second batch are each other's K nearest neighbors, then an edge is established between cell A and cell B to construct an adjacency matrix that reflects the structural correspondence across the dataset.

[0026] Further design optimization, specifically the joint optimization using a symmetric autoencoder architecture combined with a triplet loss function, includes:

[0027] The concatenated cross-batch high-dimensional embedding vectors are input into a multi-layer encoder to generate latent representations.

[0028] The input features are reconstructed from the latent representation using a structurally symmetric decoder, and the mean squared error loss of the reconstruction is calculated.

[0029] Based on the mutual nearest neighbor graph, a triplet sample is constructed, wherein the triplet sample includes an anchor cell, a positive sample cell from different batches or species, and a negative sample cell.

[0030] Calculate the distance between the anchor cell and the positive sample cell in the latent space, and the distance between the anchor cell and the negative sample cell in the latent space, and calculate the triplet loss according to the preset boundary parameters;

[0031] The triplet loss is designed to forcibly bring the positive sample cells and anchor cells closer together, and to push the negative sample cells and anchor cells further apart.

[0032] The overall objective function is obtained by weighted summing of the reconstruction mean squared error loss and the triplet loss, and the model parameters are updated by backpropagation algorithm.

[0033] In a further optimized design, when constructing triplet samples, the selection of positive samples depends on the mutual nearest neighbor relationship. That is, for each anchor cell, the cell with the highest similarity among the cross-batch or cross-species cells that are mutual nearest neighbors is selected as the positive sample; the selection of negative samples is based on the distance in the latent space, and samples are taken from cells that are not connected to the anchor cell or are far away as negative samples.

[0034] Further design optimization involves using the output corrected cell embedding vector for cross-species proteomics data integration, specifically including:

[0035] Obtain proteomics datasets from different species that share biological homology;

[0036] By using a uniformly trained sequence coding module, mRNA sequences from different species are mapped to the same embedding space to eliminate sequence differences between species;

[0037] Biologically similar cells across species are identified through the mutual nearest neighbor matching, and the cell distribution across species is aligned using the triplet loss function.

[0038] The final output is an integrated embedding vector that can simultaneously reflect evolutionary conservation among species and technological differences among individuals.

[0039] Further design optimization involves using the output corrected cell embedding vector for spatial proteomics data analysis, specifically including:

[0040] When processing spatial proteomics data, the spatial coordinate information of cells is preserved, and the corrected cell embedding vector is combined with spatial neighborhood information.

[0041] Cells are grouped in the embedding space using a clustering algorithm, and the clustering results are mapped back to the original spatial coordinates to visualize the spatial distribution patterns of cell types in the tissue microenvironment.

[0042] Further optimization of the design includes the following specific parameter configurations for the hierarchical convolutional neural network encoder:

[0043] The first one-dimensional convolutional layer contains 288 convolutional kernels, with a kernel width of 17.

[0044] The convolution tower contains six convolution blocks, with the number of convolution kernels in each block being 288, 323, 363, 407, 456 and 512 respectively, and the width of the convolution kernels in all blocks being 5.

[0045] The fully connected bottleneck layer projects features into a 256-dimensional latent space with a dropout rate of 0.2.

[0046] Secondly, the present invention provides a single-cell and spatial proteomics data integration system based on mRNA sequence guidance, comprising:

[0047] The data acquisition module is used to acquire single-cell or spatial proteomics data, including protein abundance measurements of multiple cells and the mRNA nucleotide sequences corresponding to the proteins.

[0048] The sequence encoding and prediction module is used to construct a sequence-guided representation learning model, which includes a hierarchical convolutional neural network encoder and a fully connected prediction layer. The mRNA nucleotide sequence is input into the encoder to extract sequence features, and the protein abundance of each cell is predicted through the fully connected prediction layer. During training, the model parameters are iteratively optimized by minimizing the mean squared error loss between the predicted protein abundance and the actual protein abundance.

[0049] The embedding extraction module is used to extract the shared weight matrix in the fully connected prediction layer after training, and use the shared weight matrix as a high-dimensional embedding vector for each cell. The high-dimensional embedding vector encapsulates the regulatory relationship between mRNA sequence motifs and protein expression.

[0050] The graph construction module is used to construct a cross-batch mutual nearest neighbor graph based on the high-dimensional embedding vector, identify the mutual nearest neighbor relationships between cells from different experimental batches or different species, and establish topological connections between cells.

[0051] The batch correction module is used to reconstruct the latent space of the high-dimensional embedding vector using a symmetric autoencoder architecture and perform joint optimization with a triplet loss function to correct the batch effect while preserving biological signals and output the corrected cell embedding vector.

[0052] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. This invention innovatively utilizes mRNA sequences to extract regulatory features through a convolutional neural network and extracts the shared weights of the prediction layer as cell embeddings. This design enables cell characterization to not only rely on protein abundance, which may be affected by noise, but also incorporate stable prior information on sequence regulation, significantly improving clustering resolution in sparse and noisy data.

[0055] 2. By constructing a mutual nearest neighbor (MNN) graph and combining it with a triplet loss function, this invention can force biologically similar cells to come closer in the latent space, even if they come from different batches or species, while pushing away dissimilar cells. This metric learning method preserves local cell topology better than traditional statistical correction.

[0056] 3. This invention is not only applicable to batch correction of single-cell proteomics data, but also effectively handles comparative proteomics analysis across species (such as human-monkey), as well as spatial proteomics data analysis that preserves spatial information, demonstrating strong versatility and scalability. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the SANPRO framework, a method for integrating single-cell and spatial proteomics data guided by mRNA sequences.

[0058] Figure 2 Adjust the box plot of the RAND index for each method;

[0059] Figure 3 Box plots of normalized mutual information for each method;

[0060] Figure 4 Box plots showing cell type affinity for each method;

[0061] Figure 5 Box plots for the Fox-Marlow index for each method;

[0062] Figure 6 UMAP visualization (batch-colored) after integrating PBMC1 and PBMC2 datasets using different methods;

[0063] Figure 7 Visualize UMAP from the same dataset, with cells colored according to their actual cell type;

[0064] Figure 8 For the cell embeddings generated by each method, the ARI, NMI, CA, FMI, ASW_batch and ILS indices were used to quantitatively evaluate each method.

[0065] Figure 9 Box plot summaries of ARI scores for different methods on cross-species datasets;

[0066] Figure 10 Box plot summaries of NMI scores for different methods on cross-species datasets;

[0067] Figure 11-13 Integrate UMAP visualizations of cell embeddings for H1N1 and IFR datasets generated by different methods;

[0068] Figure 14 Box plots showing ARI scores for different methods in spatial proteomics data;

[0069] Figure 15 Box plots showing CA scores for different methods in spatial proteomics data;

[0070] Figure 16 Visualization results of cell clusters identified using the Louvain clustering algorithm for cell embeddings generated on spatial proteomics data by different methods;

[0071] Figure 17 UMAP visualization results of raw protein expression data;

[0072] Figure 18 UMAP visualization results of SANPRO on the clinical proteomics dataset;

[0073] Figure 19 A bar chart of 1-metric scores for five methods—SANPRO, scPROTEIN, Harmony, Scanorama, and Combat—on a clinical proteomics dataset is generated, with batch labels used as the true labels for evaluation.

[0074] Figure 20 Distribution of cells from healthy donors (green) and CTCL patients (red) in various clusters on a clinical proteomics dataset;

[0075] Figure 21 The proportion of healthy donor cells and CTCL patient cells in cluster 0 on a clinical proteomics dataset;

[0076] Figure 22 Volcano plot of differentially expressed proteins between healthy donor cells and CTCL patient cells in Cluster 0 of the clinical proteomics dataset;

[0077] Figure 23 The first few results of GeneOntology (GO) biological processes for enrichment of upregulated proteins in CTCL cells in Cluster 0 of the Clinical Proteomics dataset;

[0078] Figure 24 The proportion of healthy donor cells and CTCL patient cells in Cluster 1 of the clinical proteomics dataset;

[0079] Figure 25 A volcano plot of differentially expressed proteins between healthy donor cells and CTCL patient cells in Cluster 1 of a clinical proteomics dataset;

[0080] Figure 26 The first few results of GO biological processes for enriching upregulated proteins in CTCL cells in Cluster 1 of the clinical proteomics dataset;

[0081] Figure 27 The proportion of healthy donor cells and CTCL patient cells in Cluster 2 of the clinical proteomics dataset;

[0082] Figure 28 Volcano plot of differentially expressed proteins between healthy donor cells and CTCL patient cells in Cluster 2 of the clinical proteomics dataset;

[0083] Figure 29 The first few results of GO biological processes for enriching upregulated proteins in CTCL cells in Cluster 2 of the Clinical Proteomics dataset;

[0084] Figure 30 The UMAP visualization results of cell embeddings were integrated into the PBMC1_CITE (Batch 1) and PBMC2_CITE (Batch 2) datasets generated after applying the various methods in conjunction with Harmony, with cells colored according to batch origin;

[0085] Figure 31 Integrate UMAP visualizations of cell embeddings for the same dataset, with cells colored according to their actual cell type;

[0086] Figure 32 The ARI and NMI indices were used to quantitatively evaluate the cell embeddings generated by each method.

[0087] Figure 33 Clustering performance of all methods on a simulated dataset generated by scDesign3;

[0088] Figure 34 Clustering performance of all methods after integrating health and disease datasets;

[0089] Figure 35SANPRO parameters in different neighborhoods Clustering performance under different values;

[0090] Figure 36 Box plots of CA scores for different methods on cross-species datasets;

[0091] Figure 37 Box plots of FMI scores for different methods on cross-species datasets;

[0092] Figure 38-41 UMAP visualization of cell embeddings for H1N1 and IFR datasets generated by MAGIC, Seurat, AutoClass, and Autoencoder, with cells colored by batch origin and real cell type, respectively;

[0093] Figure 42 UMAP visualizations of cell embeddings were integrated into the H1N1 and IFR datasets generated by combining various methods with Harmony, with cells colored by batch origin;

[0094] Figure 43 Integrate UMAP visualizations of cell embeddings for the same dataset, with cells colored according to their actual cell type;

[0095] Figure 44 The ARI and NMI indices were used to quantitatively evaluate the cell embeddings generated by each method.

[0096] Figure 45 Cell embedding maps from the IFH and IFR datasets generated for Combat are visualized using UMAP, with cells colored according to their batch origin and true cell type.

[0097] Figure 46 Cell embedding maps from the IFH and IFR datasets generated for Harmony are visualized using UMAP, with cells colored according to their batch origin and true cell type.

[0098] Figure 47 Cell embedding maps from the IFH and IFR datasets generated for SANPRO are visualized using UMAP, with cells colored according to their batch origin and true cell type.

[0099] Figure 48 Box plots of NMI and FMI scores on the MIBI-TOF dataset;

[0100] Figure 49Visualization results of cell clusters identified by various methods using the Louvain clustering algorithm on the TNBC_MIBI-TOF_2 dataset;

[0101] Figure 50 Visualization results of cell clusters identified by various methods using the Louvain clustering algorithm on the TNBC_MIBI-TOF_3 dataset; Detailed Implementation

[0102] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0103] This embodiment provides a method for integrating single-cell and spatial proteomics data guided by mRNA sequences, which mainly includes two stages: the first stage is characterization learning based on mRNA sequences, and the second stage is batch correction based on mutual nearest neighbor and triplet loss.

[0104] Specifically, it includes the following:

[0105] 1. Sequence-guided protein embedding extraction

[0106] To integrate the mRNA sequence information behind proteomic data, we constructed a deep learning framework to convert raw mRNA sequences into protein-level representations rich in biological information. Specifically, for nucleotide sequences of a fixed length L = 1344, we first employ one-hot encoding to map each base (A, U, G, C) into a 4-dimensional binary vector. Formally, let the sequence... ,in The encoded input matrix Defined as:

[0107]

[0108] in This represents the j-th nucleotide base. It is the input matrix after one-hot encoding, with a matrix dimension of L×4; the i-th row of the matrix represents the i-th base of the mRNA sequence, and the j-th column corresponds to the 4 base channels.

[0109] The matrix Processing is performed through a series of convolutional layers to extract local and hierarchical patterns relevant to protein inference. The first layer applies a 1D convolution. ,Include One filter, convolution kernel width Then, batch normalization, GELU activation, and max pooling with a width of 3 are performed. The resulting feature map Given by the following formula:

[0110]

[0111] In order to capture multi-scale contextual patterns, The output is passed through a convolutional tower consisting of multiple stacked convolutional blocks, with the filter size of each layer gradually increasing [288,323,363,407,456,512]. Each layer uses a convolutional kernel with a width of 5, and the final output... The hierarchical sequence features were summarized.

[0112] Next, a point-to-point convolution (1×1) with 256 filters is used to reduce the feature dimension, resulting in... Expanding this output yields a vector. Then, it is projected into a low-dimensional latent space through a fully connected bottleneck layer:

[0113]

[0114] in d represents batch normalization, b is the learnable bias vector of the bottleneck layer, and d is the dimension. It is the learnable weight matrix of the bottleneck fully connected layer, with a dimension of d×1792.

[0115] The learned z-vector is then input into a d-dimensional (d=256) bottleneck layer to capture the low-dimensional embeddings of the protein. These embeddings are then passed through a weight matrix... Dense layers for prediction The model is trained end-to-end using mean squared error (MSE) loss, which is calculated as the difference in protein abundance between predicted and observed values.

[0116]

[0117] Where P represents the total number of proteins in the proteomics dataset. and Let represent the predicted abundance and the actual abundance of the p-th protein, respectively. In the first stage, all learnable parameters are iteratively optimized using the mean squared error (MSE) loss between the predicted and observed protein abundances. In this study, the learned weights in the dense layer... To form a target The 256-dimensional embedding vector of each cell is then used for downstream inference tasks.

[0118] 2. Graph construction based on protein embedding

[0119] To capture topological relationships between cells based on their proteomics features, we constructed a cross-batch cell map using protein-level embeddings. These embeddings, obtained by a sequence-guided feature extractor, encapsulate high-order molecular features and provide a reliable basis for measuring intercellular similarity.

[0120] make and These represent the protein embedding vectors of batch 1 and batch 2, respectively. We employ a mutual nearest neighbor (MNN) strategy to construct an adjacency matrix across the dataset. When cell i in the first batch and cell j in the second batch are each other's k nearest neighbors, an edge is established between them:

[0121]

[0122] This results in cross-batch charts It reflects the structural correspondence between batches and serves as input for downstream processing such as comparative learning.

[0123] 3. Batch effect correction module

[0124] Column-guided embedding extractors generate initial protein representations that capture sequence-based intrinsic features while preserving dataset-specific variation. However, these representations may still contain batch effects due to technical differences between datasets. To address this, we perform cross-batch alignment using the reconstructed embeddings as input. First, we identify nearest neighbor (MNN) pairs between batches to establish correspondences between biologically similar cells; then, a triplet learning approach is used to bring matching cells closer together while pushing dissimilar cells apart. This process refines the embedding vectors into a shared representation space, thereby reducing batch-specific variation while preserving biologically relevant structures.

[0125] Here, we designed a module specifically for correcting batch effects based on an autoencoder architecture. This is done in constructing a batch graph based on mutual nearest neighbor (MNN). Subsequently, instead of directly using the graph structure as the encoder input, we utilized the MNN graph as prior information to guide the generation of metric learning training samples. This module enables SANPRO to extract learned representations that preserve biological signals while suppressing technological variation.

[0126] The autoencoder consists of a multi-layer encoder and a symmetric decoder. Its optimization process combines mean squared error (MSE) loss for feature reconstruction and triplet loss to facilitate cross-batch alignment in the embedding space.

[0127] make This represents the protein embedding vectors from all batches, where N is the number of samples (i.e., and The sum of ( ), where d is the feature dimension. The encoder processes each input through L stacked non-linear layers. Transform into a latent representation :

[0128]

[0129] in Let be the learnable weight matrix of the k-th layer, and σ(⋅) be a non-linear activation function (e.g., ReLU). The decoder employs a symmetric structure to reconstruct the input features from the latent space:

[0130]

[0131] The final output of the decoder is:

[0132]

[0133] The reconstruction target is defined as the mean squared error of all samples:

[0134] Ternary sampling loss: To facilitate meaningful alignment of cells across different batches or species, we employ a ternary loss strategy that operates within the latent embedding space. (Cross-batch MNN graph) This provides a biological basis for constructing ternary samples. For each anchored cell, we select "k" most similar cross-batch neighbors as positive samples based on MNN similarity, and select "k" least similar cells (e.g., cells not connected and far apart in the latent space) as negative samples. This setup ensures semantic consistency between different batches by bringing similar cells closer and pushing dissimilar cells apart. The ternary loss is defined as follows:

[0135]

[0136] in, , , and Let represent the Lth layer embeddings of the anchor sample, positive sample, and negative sample, respectively. The marginal parameter τ controls the minimum separation between positive and negative sample pairs.

[0137] Joint optimization, the overall objective function combines reconstruction loss and triplet loss:

[0138]

[0139] Where α∈[0,1] is a tradeoff hyperparameter used to balance the autoencoder’s ability to faithfully reconstruct the input with its ability to learn batch-invariant representations.

[0140] The specific implementation method is as follows:

[0141] The SANPRO framework is implemented using PyTorch 1.13.1 with CUDA acceleration. For the sequence encoding module, we perform one-hot encoding on a fixed input sequence of length L = 1344 bp to generate a 4-channel binary matrix. The convolutional network architecture uses an initial layer with 288 filters (kernel size 17), followed by six convolutional blocks with progressively increasing filter sizes [288, 323, 363, 407, 456, 512] (all based on a kernel size of 5). The entire network uses batch normalization and the GELU activation function throughout, and a bottleneck layer with a dropout rate of 0.2 is used to finally map the features to a d = 256-dimensional protein embedding vector. For graph construction, we use the default k = 10 nearest neighbor (MNN) method. The batch correction autoencoder employs a symmetric structure, consisting of a 3-layer encoder (with hidden dimensions [128, 64, 32]) and a corresponding decoder layer, using ReLU activation and batch normalization. A triplet loss introduces a marginal value tau = 0.01 and assigns a weight α = 0.5 relative to the reconstruction loss. Training utilizes the Adam optimizer with an initial learning rate of 10⁻⁴ cosine decay. We used a batch size of 256 and trained on an NVIDIA 4090 GPU using PyTorch's Automatic Mixed Precision (AMP) to achieve mixed precision acceleration.

[0142] The specific evaluation indicators are as follows:

[0143] To quantitatively evaluate clustering performance, we used four widely used metrics: Adjusted Rand Index (ARI)

[18] , Normalized Mutual Information (NMI)

[19] , Clustering Accuracy (CA)

[20] and Folkes-Malos Index (FMI)

[21] .

[0144] The adjusted RAND Index (ARI) measures the similarity between predicted clusters and true labels, and is adjusted for random factors.

[0145]

[0146] in, This represents the number of samples for predicted cluster i and true class j. , , where n is the total number of samples.

[0147] Normalized mutual information (NMI) is used to quantify the dependency between the predicted label U and the true label V:

[0148]

[0149] Where I(U;V) represents mutual information and H(⋅) represents entropy.

[0150] Clustering accuracy (CA) is defined as the optimal one-to-one mapping between predicted labels and true labels:

[0151]

[0152] in, and These are the true label and predicted label of sample i, respectively. This represents the optimal permutation.

[0153] The Fowlkes-Mallows index (FMI) calculates the geometric mean of precision and recall between two clusters:

[0154]

[0155] In this context, “TP”, “FP”, and “FN” represent the number of true positive examples, false positive examples, and false negative examples, respectively.

[0156] The specific results are as follows:

[0157] 1. SANPRO architecture

[0158] like Figure 1 As shown, SANPRO is a cross-scale deep learning framework designed to enhance the analytical capabilities of single-cell and spatial proteomics data by integrating regulatory signals encoded in mRNA sequences. The framework comprises two consecutive phases: mRNA-guided representation learning and cross-dataset alignment.

[0159] In the first stage, SANPRO uses a hierarchical convolutional neural network to process the complete mRNA sequence. Each nucleotide sequence is encoded once and then processed through stacked convolutional layers to generate a d-dimensional embedding vector to capture the regulatory features derived from the sequence. These embedding vectors are then fed into a fully connected network to predict protein abundance across cells. The resulting weight matrix is ​​shared across all samples, serving as a representation between cells and features, encapsulating the regulatory relationship between mRNA sequence motifs and protein expression.

[0160] In the second stage, SANPRO employs a symmetric autoencoder to denoise the initial embeddings and optimize the latent space. To achieve robust integration across datasets, the system utilizes a coprime nearest neighbor (MNN) matching method to identify anchored cells in different batches. Subsequently, a triple loss objective is employed to optimize the embedding space, minimizing the distance between biosimilar cells under different conditions while maximizing separation from dissimilar cell types. This two-stage architecture enables SANPRO to effectively distinguish between biological variations and technical artifacts, outperforming protein-only benchmark methods in both cell state resolution and batch correction.

[0161] 2. Achieving efficient feature extraction using SANPRO

[0162] To systematically evaluate the quality of the embeddings learned by SANPRO, we conducted large-scale benchmarking on six biologically independent single-cell proteomics datasets, including the CITE-seq and REAP-seq platforms, covering a variety of tissue sources such as human peripheral blood mononuclear cells (PBMCs) and spleen. These datasets encompass a wide range of biological and technical differences, enabling rigorous evaluation of the model's robustness and generalization ability. We evaluated clustering performance based on four commonly used metrics: adjusted RAND index (ARI), normalized mutual information (NMI), clustering accuracy (CA), and Falkes-Maslow index (FMI). Clustering was performed using the Louvain algorithm, and default parameters were applied to the cell embeddings generated by each method, including AutoEncoder, MAGIC, AutoClass, Seurat, and SANPRO.

[0163] like Figure 2 As shown, SANPRO outperforms all competing methods across all four evaluation metrics. Each boxplot summarizes the model's performance across multiple datasets, with each data point corresponding to a separate dataset. Regarding ARI metrics ( Figure 2 SANPRO achieved an average ARI of 61.1%, a 12.2% improvement over the second-ranked Seurat method. Unless otherwise specified, percentage improvements reported in this study refer to absolute differences in metrics. This convention is consistent across all charts and tables for clear comparison. AutoEncoder and MAGIC performed similarly in average ARI values. In the NMI assessment ( Figure 3 SANPRO also showed a significant advantage, improving performance by an average of 4.67% compared to the second-best method. Its superior performance is also reflected in CA (…). Figure 4 ) and FMI ( Figure 5In [the study], SANPRO achieved average improvements of 8.08% and 10.31%, respectively. Notably, compared to other methods, SANPRO exhibits less variability across different datasets, indicating stable and reliable performance under diverse biological conditions. In summary, these results demonstrate that SANPRO can not only learn highly discriminative and biologically meaningful characterizations from single-cell proteomics data but also achieve robust generalization across datasets with varying measurement platforms and tissue origins. These findings establish SANPRO as a state-of-the-art feature extraction method for single-cell proteomics analysis.

[0164] 3. SANPRO enables robust multi-batch proteomics integration

[0165] To evaluate SANPRO's ability to integrate proteomic data across batches, we applied it to a matched PBMC_CITE dataset generated by two different donors (PBMC1 and PBMC2). These datasets represent two independent batches, sharing some cell types but differing in donors and experimental techniques, thus presenting significant integration challenges. Figure 6 As shown, the UMAP visualization results of cell embeddings generated by different methods reveal significant differences in batch mixing performance. Compared to SANPRO, which effectively achieves cross-batch cell alignment—with only minimal batch-to-batch segregation differences—methods such as AutoEncoder, MAGIC, and AutoClass fail to properly integrate the data, resulting in a significant preservation of batch-specific clustering structures. Although some overlap exists in the original data (see...),... Figure 7 This is especially true for certain high-abundance cell populations, where topological misalignment was still observed between the same cell types from two batches, indicating insufficient correction for batch effects. Figure 8 As shown, SANPRO achieved the highest ARI score of 61.6% in clustering performance, 7.31% higher than the second-ranked method, Seurat (default CCA integration; hereinafter referred to as Seurat). In the NMI metric, SANPRO again led, reaching 75.98%, 5.01% higher than the second-ranked method. CA, FMI, and batch hybrid sensitivity metrics ASW_batch and ILS also showed similar trends, confirming that SANPRO achieves strong biological resolution without sacrificing integration quality.

[0166] To further verify the robustness of the SANPRO design, we tested whether external batch calibration could improve the performance of other competing methods. Each baseline method was paired with the widely used integration tool Harmony for testing. Figure 9As shown, all methods exhibited improved batch mixing performance in UMAP space when combined with Harmony. However, key integration issues remained unresolved even with Harmony. For example, CD14 monocytes failed to fully integrate in the Harmony+MAGIC pipeline, while CD4 naive cells still showed partial segregation in the Harmony+AutoEncoder configuration. In contrast, SANPRO achieved superior integration and clustering performance even without external tools, outperforming the best Harmony-enhanced baseline (Seurat+Harmony) by 6.14% in ARI and 11.6% in NMI. These results collectively demonstrate that SANPRO offers significant advantages in multi-batch integration of single-cell proteomics, providing a unified framework that eliminates batch effects while preserving biologically meaningful cellular diversity.

[0167] To conduct a more rigorous evaluation, we further used the reference-based simulation framework scDesign3

[22] to conduct experiments on a simulated ADT dataset with known real labels. Specifically, using the PBMC ADT dataset as a reference, we considered both cell type and batch effects when simulating the count distribution, and set the remaining parameters to their default values. Figure 10 As shown, SANPRO consistently outperformed other methods across all baseline methods. Specifically, SANPRO achieved a 62.03% improvement in the ARI metric, a 7.26% improvement over the second-best method, and also showed consistent improvements in NMI, CA, and FMI. To further evaluate the robustness of SANPRO under different biological conditions, we conducted additional validation experiments on the AML CITE-seq dataset GSE185381, which contains a sample from one AML patient (8,686 cells from 5 individuals) and a sample from one healthy control (5,445 cells from 4 individuals). To ensure stable and comparable evaluation results, we consolidated the original fine-grained immune subtypes into nine major cell type categories based on their annotation hierarchy and evaluated all methods using the same merge labels. Figure 9 As shown, SANPRO consistently outperforms baseline methods, achieving a 2.61% improvement over the second-best method, Autoclass, in the ARI metric, while also demonstrating similar advantages in NMI, CA, and FMI. These results demonstrate that SANPRO can achieve robust integration under disease control conditions while preserving biological structures at the major cell type level.

[0168] To evaluate SANPRO's sensitivity to the parameter k in MNN graph construction, we performed a parameter analysis, varying k from 5 to 25 (k = 5, 10, 15, 20, 25). Figure 35 As shown, SANPRO exhibits stable performance on the ARI metric, with fluctuations within 1.04%. NMI, CA, and FMI also show similar patterns, with only slight variations at different k values. These results indicate that SANPRO is robust to the choice of k value and does not require fine-tuning of parameters.

[0169] 4. Achieving cross-species proteomics mapping using SANPRO

[0170] To comprehensively evaluate SANPRO's ability to integrate proteomic data at the cross-species level, we benchmarked it using six carefully selected species combinations. These datasets included: human whole blood cells exposed to H1N1 virus (H1N1), human blood cells stimulated with IFNγ (IFH), rhesus monkey blood cells stimulated with IFNγ (IFR), and marmoset blood cells stimulated with IFNγ (IFC). These datasets cover the evolutionary gradient from intra-human variation to cross-primate comparisons, providing a robust testing ground for cross-species integration. We evaluated its performance using four complementary metrics: adjusted RAND index (ARI), normalized mutual information (NMI), classification accuracy (CA), and Falkes-Maslow index (FMI). In addition to comparing SANPRO to four representative baseline methods (AutoEncoder, MAGIC, AutoClass, and Seurat), we also tested its integration performance when combined with commonly used batch correction tools Harmony and ComBat.

[0171] like Figure 9 As shown, in the benchmark tests of six species combinations, SANPRO consistently outperformed all other methods. Its average ARI score reached 41%, 8.26% higher than the second-best method, ComBat. Compared to other baseline methods, SANPRO outperformed by 9.46% on AutoEncoder, 8.99% on MAGIC, 9.41% on AutoClass, and 12.3% on Seurat. Similar superiority was also observed in NMI, CA, and FMI metrics. Figure 10 and Figure 36 , 37 As shown in the figure. These results collectively demonstrate that SANPRO is robust in aligning proteomics data from evolutionarily diverse species while preserving biologically meaningful structures.

[0172] To further explore SANPRO's integration capabilities, we compared its performance with a baseline approach based on Harmony enhancements. Figures 42-44 As shown, although Harmony-based calibration typically improves the clustering quality of baseline methods, SANPRO still demonstrates a significant advantage: it outperforms the Seurat+Harmony combination by 11.8% in ARI metrics, exceeds AutoClass+Harmony by 8.89%, and maintains a consistent lead among all other Harmony-enhanced baseline methods. Interestingly, we found that for some methods, adding Harmony actually leads to performance degradation, possibly due to overcalibration. This indicates that SANPRO's built-in integration mechanism is not only more efficient but also avoids the over-tuning problem often encountered with external batch calibration tools.

[0173] The qualitative analysis provided by UMAP visualization further supports these quantitative results. For example... Figure 11-13 and Figure 38-41 As shown, SANPRO uniquely integrates biologically equivalent cell types from different species. For example, CD8 T cell and B cell subsets in H1N1 and IFR samples formed highly consistent, topologically coherent clusters, unaffected by species origin. In contrast, embeddings generated by other methods still exhibit significant inter-species segregation. Even after applying Harmony correction, some methods failed to successfully integrate specific cell types—for example, in the results of MAGIC+Harmony and AutoClass+Harmony, CD4 T cells and CD8 T cells remained segregated across different species. Furthermore, consistent integration was observed in the IFH and IFR paired datasets, further demonstrating the generalizability of SANPRO under different species conditions. Figures 45-47 In summary, these results highlight SANPRO's unique advantages in cross-species proteome integration. Its architecture effectively eliminates technical differences between species while preserving subtle biological variations, making it a powerful tool for comparative immunology and evolutionary proteomics.

[0174] 5. Spatial proteomics analysis using SANPRO

[0175] To comprehensively evaluate SANPRO's performance on spatial proteomics data, we compared it with several competing methods using five datasets generated by time-of-flight Doppler ion beam imaging (MIBI-TOF). These datasets were derived from five patients diagnosed with triple-negative breast cancer (TNBC). Each dataset was analyzed independently to ensure objectivity, and clustering performance was evaluated using multiple standard metrics. Figure 14As shown, SANPRO consistently outperforms all baseline methods across all five TNBC spatial proteomics datasets. Its average adjusted RAND index (ARI) is the highest, outperforming the second-place method, Seurat, by 12.9%. While AutoClass and Seurat perform similarly in average ARI, AutoEncoder shows a significant performance drop across all spatial datasets. This performance ranking also remains stable across other evaluation metrics, including normalized mutual information (NMI), clustering accuracy (CA), and the Falks-Mass index (FMI), with SANPRO consistently ranking first. Figure 15 and Figure 48 These results highlight SANPRO's robustness and superior adaptability in dealing with the complexity of spatial proteome data.

[0176] To further validate the effectiveness of SANPRO, we performed spatial visualization analysis on cell type prediction. Taking the TNBC_MIBI-TOF_1 dataset as an example (…), Figure 16 SANPRO accurately recovered the spatial structure of the tissue, highly consistent with the reference cell type annotations. The reference cell type labels reflected the spatial distribution of cells within the tissue microenvironment, derived from experimental protein expression profiles obtained through clustering and marker-based annotation in the original MIBI-TOF study. Notably, both SANPRO and Seurat successfully identified local patterns of non-immune T cells, demonstrating their ability to preserve biologically meaningful spatial information. In contrast, other benchmark methods failed to recover these spatial patterns, particularly in accurately locating non-immune cells. This indicates limitations in resolving spatial heterogeneity. A similar trend was observed when analyzing the datasets TNBC_MIBI-TOF_2 and TNBC_MIBI-TOF_3. Figures 49-50 Overall, these results demonstrate that SANPRO not only achieves state-of-the-art clustering accuracy but also exhibits excellent preservation of biologically meaningful spatial structures. Its ability to integrate proteomic signals with spatial context highlights its potential as a powerful tool for advancing spatial proteomics and tumor microenvironment research.

[0177] 6. Clinical applications of SANPRO in single-cell proteomics

[0178] To explore the applicability of SANPRO in clinically derived proteomics datasets, we applied it to antibody-labeled single-cell proteomic profiles collected from human tissue samples. This dataset, generated using ECCITE-seq technology, contains measurements of 49 surface proteins under two different biological conditions: one from a healthy donor and one from a patient diagnosed with cutaneous T-cell lymphoma (CTCL), each contributing 6500 cells. Figure 17 As shown, the original UMAP projection clearly separates cells according to their donor origin, with cells from healthy donors displayed in green and cells from CTCL patients displayed in red. After integration using SANPRO, the resulting cell representation ( Figure 18 The results show that the two datasets are effectively aligned, indicating that the variation between samples has been successfully suppressed.

[0179] We systematically compared the integration performance of SANPRO with other competing methods. Evaluation metrics included 1-NNMI_batch and 1-ARI_batch (… Figure 19 All results consistently showed that SANPRO performed better, highlighting its ability to suppress technical noise while preserving meaningful biological signals. After integration, we performed unsupervised clustering analysis on the embedded data from SANPRO, identifying ten distinct cellular states at the transcriptional level. The donor source composition of these clusters is summarized in... Figure 20 middle.

[0180] To further understand the disease-specific characteristics, we focused on three representative clusters for subsequent analysis. Figure 2 , 22 And 23). Within each cluster, the ratio of healthy samples to CTCL sample cells is represented by a pie chart ( Figure 24 , 25 and 26), while the corresponding volcano map ( Figure 27 , 28 (and 29) demonstrates the differences in protein expression between the two conditions. Enrichment analysis of CTCL-upregulated proteins in each cluster ( Figure 30 , 31 (and 32) showed that these proteins were significantly associated with immune-related gene ontology (GO) terms. Notably, programmed death 1 (PD1) was identified as significantly upregulated in CTCL-derived cells. Figure 33 This is consistent with previous findings that PD1 overexpression is associated with an immunosuppressive microenvironment and impaired antitumor response during CTCL progression.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.

Claims

1. A method for integrating single-cell and spatial proteomics data guided by mRNA sequences, characterized in that, Includes the following steps: Acquire single-cell or spatial proteomics data, including protein abundance measurements of multiple cells and the corresponding mRNA nucleotide sequences of the proteins; Construct a sequence-guided representation learning model, which includes a hierarchical convolutional neural network encoder and a fully connected prediction layer; The mRNA nucleotide sequence is input into the encoder to extract sequence features, and the protein abundance of each cell is predicted through the fully connected prediction layer. During training, the model parameters are iteratively optimized by minimizing the mean squared error loss between the predicted protein abundance and the actual protein abundance. After training, the shared weight matrix in the fully connected prediction layer is extracted, and the shared weight matrix is ​​used as a high-dimensional embedding vector for each cell. The high-dimensional embedding vector encapsulates the regulatory relationship between mRNA sequence motifs and protein expression. Based on the high-dimensional embedding vector, a cross-batch mutual nearest neighbor graph is constructed to identify the mutual nearest neighbor relationships between cells from different experimental batches or different species, and to establish topological connections between cells. The high-dimensional embedding vector is reconstructed using a symmetric autoencoder architecture and jointly optimized using a triplet loss function to correct batch effects while preserving biological signals, and outputs a corrected cell embedding vector.

2. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 1, characterized in that, The construction of the sequence-guided representation learning model specifically includes: One-hot encoding is performed on a fixed-length mRNA nucleotide sequence, which is then mapped into a multidimensional binary input matrix; The input matrix is ​​processed by a first one-dimensional convolutional layer containing multiple convolutional kernels, followed by batch normalization, activation function processing, and max pooling operations to obtain a first feature map. The first feature map is input into a convolutional tower consisting of multiple stacked convolutional blocks, each using a convolutional kernel of fixed width, and the number of convolutional kernels gradually increases between layers to capture multi-scale contextual patterns, thus obtaining a second feature map. The feature dimension of the second feature map is reduced by pointwise convolution, and the flattened feature vector is obtained. It is then projected into a low-dimensional latent space through a fully connected bottleneck layer containing dropout and batch normalization to obtain the initial sequence embedding.

3. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 2, characterized in that, The construction of cross-batch nearest neighbor graphs based on the high-dimensional embedding vectors specifically includes: Obtain the high-dimensional embedding vectors of all cells from the first and second batches respectively; For any cell in the first batch, calculate its distance to all cells in the second batch in the latent space, and determine the K nearest neighbor cells in the second batch; Meanwhile, for any cell in the second batch, calculate its distance to all cells in the first batch in the latent space, and determine the K nearest neighbor cells in the first batch. If cell A in the first batch and cell B in the second batch are each other's K nearest neighbors, then an edge is established between cell A and cell B to construct an adjacency matrix that reflects the structural correspondence across the dataset.

4. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 3, characterized in that, The joint optimization using a symmetric autoencoder architecture combined with a triplet loss function specifically includes: The concatenated cross-batch high-dimensional embedding vectors are input into a multi-layer encoder to generate latent representations. The input features are reconstructed from the latent representation using a structurally symmetric decoder, and the mean squared error loss of the reconstruction is calculated. Based on the mutual nearest neighbor graph, a triplet sample is constructed, wherein the triplet sample includes an anchor cell, a positive sample cell from different batches or species, and a negative sample cell. Calculate the distance between the anchor cell and the positive sample cell in the latent space, and the distance between the anchor cell and the negative sample cell in the latent space, and calculate the triplet loss according to the preset boundary parameters; The triplet loss is designed to forcibly bring the positive sample cells and anchor cells closer together, and to push the negative sample cells and anchor cells further apart. The overall objective function is obtained by weighted summing of the reconstruction mean squared error loss and the triplet loss, and the model parameters are updated by backpropagation algorithm.

5. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 4, characterized in that: When constructing triplet samples, the selection of positive samples depends on the mutual nearest neighbor relationship. That is, for each anchor cell, the cell with the highest similarity is selected from the cross-batch or cross-species cells that are mutual nearest neighbors as the positive sample; the selection of negative samples is based on the distance in the latent space, and samples are taken from cells that are not connected to the anchor cell or are far away as negative samples.

6. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 5, characterized in that, The output corrected cell embedding vector is used for cross-species proteomics data integration, specifically including: Obtain proteomics datasets from different species that share biological homology; By using a uniformly trained sequence coding module, mRNA sequences from different species are mapped to the same embedding space to eliminate sequence differences between species; Biologically similar cells across species are identified through the mutual nearest neighbor matching, and the cell distribution across species is aligned using the triplet loss function. The final output is an integrated embedding vector that can simultaneously reflect evolutionary conservation among species and technological differences among individuals.

7. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 6, characterized in that, The output corrected cell embedding vector is used for spatial proteomics data analysis, specifically including: When processing spatial proteomics data, the spatial coordinate information of cells is preserved, and the corrected cell embedding vector is combined with spatial neighborhood information. Cells are grouped in the embedding space using a clustering algorithm, and the clustering results are mapped back to the original spatial coordinates to visualize the spatial distribution patterns of cell types in the tissue microenvironment.

8. The method for integrating single-cell and spatial proteomics data based on mRNA sequence guidance as described in claim 7, characterized in that, The specific parameter configuration of the hierarchical convolutional neural network encoder includes: The first one-dimensional convolutional layer contains 288 convolutional kernels, with a kernel width of 17. The convolution tower contains six convolution blocks, with the number of convolution kernels in each block being 288, 323, 363, 407, 456 and 512 respectively, and the width of the convolution kernels in all blocks being 5. The fully connected bottleneck layer projects features into a 256-dimensional latent space with a dropout rate of 0.

2.

9. A single-cell and spatial proteomics data integration system guided by mRNA sequence, characterized in that, The system includes: The data acquisition module is used to acquire single-cell or spatial proteomics data, including protein abundance measurements of multiple cells and the mRNA nucleotide sequences corresponding to the proteins. The sequence encoding and prediction module is used to construct a sequence-guided representation learning model, which includes a hierarchical convolutional neural network encoder and a fully connected prediction layer. The mRNA nucleotide sequence is input into the encoder to extract sequence features, and the protein abundance of each cell is predicted through the fully connected prediction layer. During training, the model parameters are iteratively optimized by minimizing the mean squared error loss between the predicted protein abundance and the actual protein abundance. The embedding extraction module is used to extract the shared weight matrix in the fully connected prediction layer after training, and use the shared weight matrix as a high-dimensional embedding vector for each cell. The high-dimensional embedding vector encapsulates the regulatory relationship between mRNA sequence motifs and protein expression. The graph construction module is used to construct a cross-batch mutual nearest neighbor graph based on the high-dimensional embedding vector, identify the mutual nearest neighbor relationships between cells from different experimental batches or different species, and establish topological connections between cells. The batch correction module is used to reconstruct the latent space of the high-dimensional embedding vector using a symmetric autoencoder architecture and perform joint optimization with a triplet loss function to correct the batch effect while preserving biological signals and output the corrected cell embedding vector.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images