Single-cell proteomic prediction model, prediction method, prediction device and computer readable storage medium

Abstract

The present application relates to the field of protein multi-omics sequencing, and particularly relates to a single-cell proteomics prediction model based on sequence information and deep learning, a prediction method, a prediction device and a computer readable storage medium. The prediction model comprises a gene coding module, a protein coding module, a cell coding module, a cross-omics translation module and a loss module. The model has large-scale human gene and protein sequence information, saving the time of manually collecting relevant information from various databases, and being more convenient for user operation. Whether in the internal omics translation of the data set or between different cell types, batches and sequencing technologies, the present application can achieve higher accuracy. At the same time, the present application supports zero-shot translation lacking multi-omics data set reference.

Description

Technical Field

[0001] This invention relates to the field of protein multi-omics sequencing, and more particularly to a single-cell proteomics prediction model, prediction method, prediction device, and computer-readable storage medium based on sequence information and deep learning. Background Technology

[0002] Currently, because multi-omics sequencing remains an expensive and complex process, many experiments to date have been conducted in relatively few cells. Furthermore, some proteins are difficult to capture due to a lack of suitable antibodies. Proteomics datasets remain insufficient compared to widely available scRNA-seq datasets, particularly for disease samples and hard-to-obtain tissues.

[0003] To address these challenges, a series of deep learning models based on variational autoencoders have been proposed in recent years, such as moETM, totalVI, scVAEIT, and scButterfly. However, existing deep learning models rely solely on paired multi-omics data, ignoring prior biological knowledge, thus exhibiting limitations such as significant performance variations across different datasets. Furthermore, these methods are limited to predicting proteins contained in a reference dataset and cannot be extrapolated to proteins lacking multi-omics datasets or matching antibodies.

[0004] Therefore, the existing technology still needs further improvement and enhancement. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a single-cell proteomics prediction model, prediction method, prediction device, and computer-readable storage medium based on sequence information and deep learning. The aim is to develop a new prediction model that fuses prior sequence knowledge of genes and proteins through two omics-specific encoders, enabling protein expression profile prediction even without learnable samples, while reducing interference from different dataset sources.

[0006] The above-mentioned objective of this invention is achieved through the following technical solution: a single-cell proteomics prediction model based on sequence information and deep learning, comprising:

[0007] A gene encoding module containing human gene sequences collected from the NCBI database, wherein the gene encoding module uses a Dna2vec pre-trained model to form the embedding of each gene;

[0008] The protein-coding module contains all human protein embeddings encoded by a pre-trained language model.

[0009] The cell-coding module is used to convert the raw transcriptome profile into a standardized transcriptome profile and encode the transcriptome profile from a cellular perspective to generate cell embeddings.

[0010] The cross-omics translation module takes the gene embedding, protein embedding, cell embedding, and the normalized map as input, and uses multi-head cross attention to translate the transcriptome map into a proteome map to obtain predicted protein values.

[0011] The loss module is used to process the raw proteomics map to generate a normalized map, and optimizes the single-cell proteomics prediction model through a loss function and an early stopping mechanism.

[0012] The following are preferred technical solutions of the present invention, but are not intended to limit the technical solutions provided by the present invention. The purpose and beneficial effects of the present invention can be better achieved and realized through the following preferred technical solutions.

[0013] As a preferred technical solution, the single-cell proteomics prediction model, wherein forming the embedding of each gene using the Dna2vec pre-trained model includes:

[0014] The k-mers are converted into vectors of length 100 based on the Dna2vec pre-trained dictionary;

[0015] The vector for each segment is obtained by averaging all the vectors mentioned above.

[0016] The vectors of each segment are concatenated in sequence to form the embedding of each gene.

[0017] As a preferred technical solution, the single-cell proteomics prediction model, wherein the pre-trained language model is a protein language model based on a transpose; the pre-trained language model is trained on a large-scale corpus containing 2.1 million proteins and 393 billion amino acids.

[0018] As a preferred technical solution, the single-cell proteomics prediction model, wherein the method for calculating the attention score of multi-head cross-attention includes:

[0019]

[0020] head i =Attention(Q) i ,K i ,V)

[0021] MultiHead(Q',K',V)=Concat(head1,head2,...,head h ),

[0022] Where K is the gene-cell combination matrix; Q is the protein-cell combination matrix; and V is the normalized transcriptome profile.

[0023] The output of the multi-head attention mechanism is mapped to a vector using a two-layer linear neural network. With cell i in n p The expression levels of each protein correspond to the expression levels of the proteins.

[0024] As a preferred technical solution, in the single-cell proteomics prediction model, the loss function is:

[0025]

[0026] Among them, y ij This represents the normalized expression value of protein j in cell i. This represents the predicted value of protein j in cell i.

[0027] As a preferred technical solution, the single-cell proteomics prediction model contains 58,766 human gene sequences.

[0028] The protein coding module includes 20,412 human protein sequence information collected from the UniProt database.

[0029] As a preferred technical solution, the single-cell proteomics prediction model, wherein the conversion of the original transcriptome profile into a standardized transcriptome profile includes:

[0030] For the original counting matrix of the transcriptomic profile, the cell-coding module removes genes expressed in fewer than 10 cells and scales the total count per cell to 10,000;

[0031] Each count of each cell is incremented by 1, and a logarithmic transformation is performed to filter out the top 5000 highly variable genes, thus forming a standardized transcriptome profile.

[0032] Secondly, a single-cell proteomics prediction method based on sequence information and deep learning, comprising:

[0033] The single-cell protein data to be predicted is input into the trained prediction model to obtain the predicted protein value; the prediction model is the single-cell proteomics prediction model described above.

[0034] Thirdly, a computer-readable storage medium stores one or more programs that can be executed by one or more processors to implement the steps of the single-cell proteomics prediction method as described above.

[0035] Fourthly, a terminal includes: a processor and a memory; the memory stores a computer-readable program executable by the processor; the processor executes the computer-readable program to implement the steps of the single-cell proteomics prediction method as described above.

[0036] Beneficial effects: Compared with existing technologies, the single-cell proteomics prediction model provided by this invention comes with large-scale human gene and protein sequence information, saving time spent manually collecting relevant information from various databases and making it easier for users to operate. Whether performing omics translation within a dataset or across different cell types, batches, and sequencing technologies, this invention achieves higher accuracy. Furthermore, this invention supports zero-sample translation when multi-omics datasets are lacking. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the single-cell proteomics prediction model architecture provided by the present invention;

[0038] Figure 2 Box plots show the performance evaluation results of proteomics translation within the dataset; the box plots display the MSE, MAE, and correlation between the predicted protein values ​​and the actual protein values ​​obtained by each method. The lower and upper edges of the rectangles correspond to the 1st and 3rd quartiles, respectively, and the center line is the median. Each point represents a dataset.

[0039] Figure 3 The prediction model provided by this invention is compared with the translation results of other methods on 64 cell types; where dots represent cell types. Dashed lines indicate the equivalent performance between the two methods.

[0040] Figure 4 For cross-batch proteomics translation performance evaluation; in the cross-batch processing scenario, s1d1 to s4d9 are from the BMMC dataset GSE194122, and batch1 and batch2 are from the PBMC dataset GSE164378; each batch is used as the test set in turn, and the other cells constitute the training set; (A) The bar chart shows the average correlation between the predicted expression value and the actual expression value of all proteins in each batch; (B) Box plot of the translation results of each method in all batches;

[0041] Figure 5 To employ the prediction model provided by this invention, new protein expression profiles are inferred using zero-shot learning; wherein, (A) is a scatter plot of the expression of the four proteins excluded from the training set, with each point representing a cell and the color corresponding to the protein expression level; (B) is the effect of the training set ratio on the zero-shot translation performance of ProTrans; the gray solid line connects the mean values ​​of the indicators under different ratios.

[0042] Figure 6 This is a structural schematic diagram of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment...

[0045] This is merely illustrative and is in no way intended to limit the invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0046] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0047] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. Furthermore, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale. For techniques and methods known to those skilled in the art,

[0048] The methods and devices may not be discussed in detail, but where appropriate, the techniques, methods, and devices described should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0049] like Figure 1As shown, this invention provides a single-cell proteomics prediction model, comprising: a gene encoding module 10 containing 58,766 human gene sequences collected from the NCBI (National Center for Biotechnology Information) database. To address the issue of large variations in gene sequence length, this invention trims all gene sequences to 10,000 bp. Short sequences are padded with 'N' to make their corresponding embeddings 0. Dna2vec is a pre-trained model used to generate k-mer representations and has been widely used in genomics research. The gene encoder converts k-mers into vectors of length 100 based on the Dna2vec pre-trained dictionary, and then averages all vectors to obtain the vector for each segment. Finally, the vectors of each segment are concatenated sequentially to form the embedding of each gene.

[0050] The protein-coding module 20 includes 20,412 human protein sequences collected from the UniProt database. Unlike traditional one-hot encoding strategies, the use of pre-trained language models to convert protein sequences into digital representations is becoming increasingly common in proteomics research. ProtT5, a transpose-based protein language model proposed by Elnaggar et al., was trained on a large-scale corpus containing 2.1 million proteins and 393 billion amino acids. This invention embeds all human proteins encoded by ProtT5 into the protein encoder module of ProTrans. Including this prior knowledge enables ProTrans to translate all human proteins, even those not included in the training set.

[0051] The cell encoding module 30 has two functions: first, it transforms the raw transcriptome profile into a standardized transcriptome profile; second, it encodes the transcriptome profile from a cellular perspective to generate cell embeddings. For the original counting matrix of the transcriptome profile, the cell encoder removes genes expressed in fewer than 10 cells and scales the total count for each cell to 10,000. After incrementing each count by 1 and performing a logarithmic transformation, the top 5000 highly variable genes are filtered to form the standardized transcriptome profile. This invention uses the scVI algorithm to generate cell embeddings based on the standardized transcriptome profile.

[0052] The cross-omics translation module 40 includes four inputs: gene embeddings from the gene encoder, protein embeddings from the protein encoder, cell embeddings from the cell encoder, and a normalized atlas. This module extracts only the genes and proteins contained in the normalized atlas. Gene and protein embeddings are identical across different cells. To capture cellular heterogeneity, this module combines them with cell-specific cell embeddings. The gene-cell combination matrix is ​​K, the protein-cell combination matrix is ​​Q, and the normalized transcriptome atlas is V. The attention score is calculated as follows:

[0053]

[0054] head i =Attention(Q) i ,K i V)

[0055] MultiHead(Q',K',V)=Concat(head1,head2,...,head h )

[0056] The output of the multi-head attention mechanism is mapped to a vector using a two-layer linear neural network. This is related to n in cell i p The expression levels of each protein correspond to the expression levels of the proteins.

[0057] Loss module 50 processes the raw proteomics map to generate a normalized map and calculates the loss to optimize model parameters. For the raw proteomics count matrix, this invention scales the total number of each cell to 10,000, increments it by 1, performs a logarithmic transformation on each count, and obtains a normalized proteomics map. ij This represents the normalized expression value of protein j in cell i. This represents the predicted value of protein j in cell i. The loss function is:

[0058]

[0059] The trained predictive model can translate a matching proteomics map based on a newly sequenced transcriptomics map. The cell encoder, when processing a new transcriptomics map, filters for genes consistent with those used in training. Inputs to the cross-omics translation module include training gene embeddings, protein embeddings, new cell embeddings, and a normalized transcriptomics map. By default, the predictive model translates the expression profiles of all proteins present in the training set. In the zero-shot translation task, the protein encoder generates embeddings for new proteins.

[0060] To evaluate the performance of the prediction model (denoted as ProTrans) provided in this invention on real-world multi-omics data translation tasks, benchmark experiments covering 14 datasets were designed. These datasets encompassed different batches, cell types, tissues, sequencing technologies, and dataset sizes (cell count: 4,115–161,764). The prediction model was compared with state-of-the-art omics translation methods: moETM, scVAEIT, totalVI, sciPENN, Seurat, and scButterfly.

[0061] Example 1

[0062] Proteomics translation within a dataset

[0063] For each dataset, 60% of the cells are randomly selected for training, and the remaining 40% are used for testing. For example... Figure 2 As shown, ProTrans and Seurat achieved lower mean squared error (MSE) and mean absolute error (MAE), with ProTrans consistently outperforming all other methods (mean MSE reduction of 0.04 and mean MAE reduction of 0.02). Compared to scButterfly and Seurat (both 91.9%), ProTrans showed a higher correlation between translated protein expression values ​​and true protein expression values ​​(mean: 92.2%). Simultaneously, ProTrans achieved the smallest interquartile range (IQR: 0.04), demonstrating the highest stability.

[0064] Example 2

[0065] Proteomics translation across cell types and batches

[0066] While paired transcriptomics-proteomics data remain scarce for certain cell types, scRNA-seq data are widely available. One feasible approach is to train the model using a multi-omics dataset and then transfer it to other cell types. Leave-one-out cross-validation was performed using three datasets covering 64 cell types, with each cell type used sequentially as the test set and the remainder for training. In comparative analysis (… Figure 3 In the data, almost all data points are above the diagonal, indicating that ProTrans outperforms state-of-the-art methods and demonstrates its robustness to cross-cell type migration.

[0067] Similarly, the effectiveness of ProTrans in cross-batch proteomics translation was evaluated. The cross-batch scenario involved 14 batches from two datasets. Figure 4 ProTrans, Seurat, and moETM performed best in GSE194122, while Seurat performed moderately in GSE164378. Notably, ProTrans achieved the highest median correlation (94.8%) across all batches, highlighting its superior reliability.

[0068] Example 3

[0069] ProTrans translates unavailable proteins using a zero-shot translation mechanism.

[0070] Current multi-omics translation methods can only infer proteins observed in the training set. However, CI TE-seq only detects cell surface proteins, leaving many more proteins unmeasured and unpredictable due to a lack of matching multi-omics references. Unlike other tools, ProTrans enables the translation of these unavailable proteins through a zero-sample mechanism. By leveraging sequence-derived biological prior information, ProTrans can generalize to any protein of interest, even without experimental measurements.

[0071] Figure 5 Figure A in the figure demonstrates the zero-shot translation performance of ProTrans on four proteins excluded from the training set. CD98 and CD274 exhibit cell cluster-specific expression patterns, a feature accurately reconstructed by ProTrans. Furthermore, the ProTrans translation results reproduced the uniform distribution of expression levels for CD144 and CD324 across all cells, with MSE values ​​of 0.51 and 0.87, respectively. We further quantified the impact of training set size on zero-shot translation performance by systematically varying the proportion of proteins allocated to the training set. Figure 5 (B in the original text). Reducing the proportion of proteins in the training set from 90% to 60% resulted in a slight increase in the mean MSE (+0.4) and MAE (+0.9). These results highlight the robustness of ProTrans, suggesting that extending it to larger multi-omics datasets would further enhance its ability to predict novel proteins.

[0072] Based on the same inventive concept, this invention also provides a single-cell proteomics prediction method based on sequence information and deep learning. This method includes: inputting the acquired single-cell protein data to be predicted into a trained prediction model to obtain predicted protein values. The prediction model used is the one described above; its specific structure has been detailed above and will not be repeated here. The acquisition of the dataset and the training process involved in model training are conventional techniques in this field and are not specifically limited here.

[0073] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the single-cell proteomics prediction method described above.

[0074] Based on the same inventive concept, the present invention also provides a terminal, such as... Figure 6As shown, it includes at least one processor 30; a display screen 31; and a memory 32, and may also include a communications interface 33 and a bus 34. The processor 30, display screen 31, memory 32, and communications interface 33 can communicate with each other via the bus 34. The display screen 31 is configured to display a preset user guide interface in the initial setup mode. The communications interface 33 can transmit information. The processor 30 can call logical instructions in the memory 32 to execute the methods described in the above embodiments.

[0075] Furthermore, the logic instructions in the aforementioned memory 32 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0076] The memory 32, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 30 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 32, thereby implementing the methods in the above embodiments. The memory 32 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the terminal, etc. Furthermore, the memory 32 may include high-speed random access memory, and the memory 32 may include non-volatile memory. For example, 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, may be used, or they may be transient storage media.

[0077] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal has been described in detail in the above method, and will not be repeated here.

[0078] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-cell proteomics prediction model based on sequence information and deep learning, characterized in that, include: A gene encoding module containing human gene sequences collected from the NCBI database, wherein the gene encoding module uses a Dna2vec pre-trained model to form the embedding of each gene; The protein-coding module contains all human protein embeddings encoded by a pre-trained language model. The cell-coding module is used to convert the raw transcriptome profile into a standardized transcriptome profile and encode the transcriptome profile from a cellular perspective to generate cell embeddings. The cross-omics translation module takes the gene embedding, protein embedding, cell embedding, and the normalized map as input, and uses multi-head cross attention to translate the transcriptome map into a proteome map to obtain predicted protein values. The loss module is used to process the raw proteomics map to generate a normalized map, and optimizes the single-cell proteomics prediction model through a loss function and an early stopping mechanism.

2. The single-cell proteomics prediction model according to claim 1, characterized in that, The process of forming the embedding of each gene using a Dna2vec pre-trained model includes: The k-mers are converted into vectors of length 100 based on the Dna2vec pre-trained dictionary; The vector for each segment is obtained by averaging all the vectors mentioned above. The vectors of each segment are concatenated in sequence to form the embedding of each gene.

3. The single-cell proteomics prediction model according to claim 1, characterized in that, The pre-trained language model is a protein language model based on the transpose; the pre-trained language model is trained on a large-scale corpus containing 2.1 million proteins and 393 billion amino acids.

4. The single-cell proteomics prediction model according to claim 1, characterized in that, The method for calculating the attention score of multi-head cross-attention includes: head i =Attention(Q i ,K i ,V) MultiHead(Q',K',V)=Concat(head1,head2,...,head h ), Where K is the gene-cell combination matrix; Q is the protein-cell combination matrix; and V is the normalized transcriptome profile. The output of the multi-head attention mechanism is mapped to a vector using a two-layer linear neural network. With cell i in n p The expression levels of each protein correspond to the expression levels of the proteins.

5. The single-cell proteomics prediction model according to claim 1, characterized in that, The loss function is: Among them, y ij This represents the normalized expression value of protein j in cell i. This represents the predicted value of protein j in cell i.

6. The single-cell proteomics prediction model according to claim 1, characterized in that, The number of human gene sequences is 58,766; The protein coding module includes 20,412 human protein sequence information collected from the UniProt database.

7. The single-cell proteomics prediction model according to claim 1, characterized in that, The process of converting the raw transcriptome profile into a normalized transcriptome profile includes: For the original counting matrix of the transcriptomic profile, the cell-coding module removes genes expressed in fewer than 10 cells and scales the total count per cell to 10,000; Each count of each cell is incremented by 1, and a logarithmic transformation is performed to filter out the top 5000 highly variable genes, thus forming a standardized transcriptome profile.

8. A single-cell proteomics prediction method based on sequence information and deep learning, characterized in that, include: The single-cell protein data to be predicted is input into the trained prediction model to obtain the predicted protein value; the prediction model is the single-cell proteomics prediction model according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps of the single-cell proteomics prediction method as described in claim 8.

10. A terminal, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; when the processor executes the computer-readable program, it implements the steps of the single-cell proteomics prediction method as described in claim 8.

Images