Single-cell gene expression profile reconstruction method and system based on virtual cell basic model of hybrid expert architecture
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-19
Smart Images
Figure CN122245421A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a method and system for reconstructing single-cell gene expression profiles based on a virtual cell-based model with a hybrid expert architecture. Background Technology
[0002] The development of single-cell sequencing technology has generated massive amounts of transcriptome data, providing a foundation for constructing large-scale virtual cell models. However, existing mainstream models (such as scGPT, Geneformer, scFoundation, scBERT, etc.) still have significant limitations in their architectural design and application: First, excessively long input sequences lead to low computational efficiency. Existing models typically treat genes as "tokens" for modeling. Models like scGPT and scBERT usually input non-zero expressed genes or highly variable genes within cells, with sequence lengths often reaching several thousand (1000~3000+); while models like scFoundation, which attempt whole-genome input, have sequence lengths approaching 20,000. The computational complexity of the self-attention mechanism in the Transformer architecture increases quadratically with sequence length (…). This results in extremely slow model training, huge GPU memory consumption, and difficulty in performing efficient inference on conventional hardware.
[0003] Second, information truncation and loss of numerical precision. To alleviate computational pressure, some models employ lossy compression strategies. For example, Geneformer uses ranking-based encoding, retaining only the top 2048 genes with the highest expression levels and directly discarding their specific expression values. This processing not only loses low-expressed genes (which often contain key regulatory information) but also fails to accurately reconstruct the quantitative changes in gene expression, limiting the model's performance in refined analyses.
[0004] Third, they have poor adaptability to generative tasks. Existing models are mostly discriminative or autoregressive generative models, whose latent space is usually a variable-length, sparse, high-dimensional vector sequence. This irregular feature space makes it difficult to directly combine with generative architectures such as diffusion models and flow matching to construct continuous cell state evolution trajectories, thus limiting their application in generative tasks such as gene and drug perturbation prediction and cell fate simulation.
[0005] Fourth, data semantic ambiguity and batch effects. Existing models often directly use raw genome annotations (such as GENCODE), resulting in many-to-one semantic ambiguity of gene IDs; and when integrating across datasets, there is a lack of effective adversarial mechanisms to completely eliminate non-biological technical noise (batch effects) from different experimental platforms, leading to insufficient robustness of the analysis results.
[0006] Therefore, there is an urgent need for a single-cell basic model architecture that can accept full gene input and retain complete information, achieve extreme sequence compression and computational efficiency through mechanism innovation, and is naturally adapted to generative tasks to solve the above problems. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for reconstructing single-cell gene expression profiles based on a virtual cell model with a hybrid expert architecture, achieving maximum inference efficiency and full information processing.
[0008] The objective of this invention can be achieved through the following technical solutions: A method for reconstructing single-cell gene expression profiles based on a virtual cell-based model using a hybrid expert architecture includes the following steps: The process involves acquiring single-cell gene expression profile data to be reconstructed, inputting it into a pre-constructed virtual cell basic model, and outputting the reconstruction result. The virtual cell basic model includes a data normalization module, an encoder module, a feature space optimization module, and a decoder module connected sequentially. The steps for outputting the reconstruction result include: The data standardization module is used to standardize the single-cell gene expression profile data to obtain the input Token sequence; In the latent space, the encoder module is used to extract and compress features from the input token sequence to obtain a high-density token sequence. At the same time, the feature space optimization module is used to geometrically shape the latent space and remove non-biological batch effects. The high-density token sequence is decoded using a decoder module to perform zero-value probability prediction and numerical regression prediction, resulting in a denoised gene expression profile, which is used as the reconstruction result.
[0009] Furthermore, the step of obtaining the input token sequence includes: The single-cell gene expression profile data is mapped to the HGNC standard gene set to be mapped to a unified dimension, thus obtaining single-cell gene expression profile data with a unified dimension. Based on the single-cell gene expression profile data of the unified dimension, Log1p normalization is performed to obtain normalized single-cell gene expression profile data. The normalized single-cell gene expression profile data is multiplied by a scaling factor to discretize it, generating an input token sequence.
[0010] Furthermore, the encoder module includes multiple cascaded unified mixing blocks and a convolutional adapter at the end. The multiple cascaded unified mixing blocks are used to extract the full set of gene features, and the convolutional adapter is used to compress the full set of gene features to obtain the high-density token sequence.
[0011] Furthermore, each unified hybrid block includes a bidirectional state-space model or a Transformer, and a cascaded hybrid expert layer containing shared and routing experts. For the multiple cascaded unified hybrid blocks, the bidirectional state-space model and the Transformer are alternately configured in different unified hybrid blocks. The bidirectional state-space model and the Transformer serve as feature extractors to extract intermediate token sequences, and the hybrid expert layer serves as a feedforward network to perform feature feedforward transformation.
[0012] Furthermore, the hybrid expert layer includes: Shared Expert: All tokens in the intermediate token sequence are always in the active path to capture common features; Routing experts: A router is configured to dynamically select Top-K routing experts from the intermediate token sequence to perform calculations and obtain cell type-specific characteristics; Weighted fusion layer: The general features and cell type-specific features are weighted and fused to obtain the output features of each hybrid expert layer.
[0013] Furthermore, the convolutional adapter includes a channel projection layer, a feature extraction layer, and an adaptive pooling layer connected in sequence, and the execution process of the convolutional adapter includes: The full set of gene features are input into the channel projection layer for dimensionality reduction to obtain the dimensionality-reduced full set of gene features. The reduced full gene features are input into a feature extraction layer to extract local feature context, wherein the feature extraction layer is a one-dimensional convolutional layer. The local feature context is input into the adaptive pooling layer for adaptive average pooling processing to form a high-density token sequence.
[0014] Furthermore, the feature space optimization module includes a latent space geometry shaping unit and an adversarial debater, wherein the latent space geometry shaping unit performs the following steps: Additive angular margin loss is used as a constraint to map high-density token sequences onto the surface of a hypersphere, and angular margins are forced between cells of the same class and the class center during the training process of the virtual cell basic model. Based on the aforementioned angular margin, similar cells are forced to highly aggregate in the potential space, while distancing dissimilar cells are pushed away, thereby achieving smoothing and shaping of the potential space. The adversarial debater unit includes a domain discriminator and a gradient inversion layer, and specifically performs the following steps: The domain discriminator receives the high-density token sequence and identifies the sequencing batch from which it originates; The gradient inversion layer inverts the gradient of the domain discriminator and passes it to the encoder module during backpropagation. Through the aforementioned adversarial game, the encoder module is forced to remove batch-specific information that is irrelevant to biology from the high-density token sequence it generates.
[0015] Furthermore, the decoder module includes a context interaction layer, a feature projection layer, and a fence decoding layer. The fence decoding layer includes a shared MoE feature extraction layer and subsequent zero-value and numerical branches for splitting. The specific execution steps of the decoder module include: The context interaction layer is used to perform global information exchange between tokens in the high-density token sequence; The feature projection layer is used to extend the dimension of the high-density token sequence to the original dimension; Based on the high-density token sequence after dimensional expansion, the shared MoE feature extraction layer is used for processing to obtain a high-dimensional latent representation; Based on the high-dimensional latent representation, zero-value branching and numerical branching are used to perform zero-value probability prediction and numerical regression prediction, respectively, to obtain the non-zero expression probability and potential expression intensity value of each gene. The non-zero expression probability and potential expression intensity value of each gene are then multiplied to obtain the denoised gene expression profile.
[0016] Furthermore, the virtual cell basic model also includes an output module for outputting the high-density token sequence and the denoised gene expression profile.
[0017] The present invention also provides a single-cell gene expression profile reconstruction system based on a virtual cell basic model with a hybrid expert architecture, including a memory, a processor, and a program stored in the memory, characterized in that the processor executes the program to implement the single-cell gene expression profile reconstruction method based on a virtual cell basic model with a hybrid expert architecture as described above.
[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses an encoder module to extract and compress features. Regardless of the length of the input data, it compresses it into a high-density token sequence of fixed length, forming a high-density information bottleneck. At the same time, the feature space optimization module is used to geometrically shape the latent space and remove non-biological batch effects, providing a high-quality latent space for the feature extraction process. Finally, the decoding achieves the ultimate inference efficiency and full information processing without the need for high-variable gene screening or truncation, thus preserving complete biological information.
[0019] (2) This invention breaks the quadratic complexity curse of existing Transformers by using a Bi-SSM architecture with high-density information bottlenecks and unified hybrid blocks, and achieves small memory and fast inference for large models.
[0020] (3) The latent spatial geometric shaping unit of the present invention compresses features onto a hypersphere by adding an additive angular margin loss, thereby forcing the same cell group to form small cone-shaped clusters on the spherical surface, achieving a compact intra-class and isolated inter-class representation. In addition, the angular margin directly introduces geometric constraints into the loss, which can still form obvious separation bands even for sparse samples, reducing the risk of overfitting.
[0021] (4) This invention introduces a hybrid expert layer containing shared and routing experts into a unified hybrid block. The shared experts are responsible for capturing basic general features, while the routing experts focus on processing cell type-specific features. The two are connected in parallel and complement each other, which enables the model to significantly reduce computational latency while maintaining a huge parameter capacity (memory) of 1.4B, requiring only the activation of a small number of parameters (sparse computation).
[0022] (5) This invention achieves efficient inference of a model with 1.4 billion parameters on a consumer-grade graphics card through extreme sequence compression and MoE architecture, and outputs cell features with high semantic density and clear topological structure. It solves the problems of large computational redundancy, difficult batch effect removal and chaotic feature space structure of existing models. It is suitable for cell type annotation, clustering, debatching and gene and drug perturbation generation tasks.
[0023] (6) The high-density token sequence generated by this invention can be directly output. This sequence is a low-dimensional, dense vector space with a clear topological structure. It is naturally adapted to the generation model such as diffusion and can be directly used for gene and drug perturbation simulation.
[0024] (7) The MoE architecture introduced in this invention compensates for the capacity loss caused by sequence compression, so that the compressed token sequence contains extremely high biological semantic density. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0027] Example 1 This embodiment provides a method for reconstructing single-cell gene expression profiles based on a virtual cell basic model using a hybrid expert architecture. This method denoises the input single-cell gene expression profile data (also known as single-cell transcriptome sequencing data) using a virtual cell basic model based on a hybrid expert architecture, thereby reconstructing the single-cell gene expression profile and addressing issues of computational efficiency, denoising, and generation adaptability. In this embodiment, a virtual cell basic model with 1.4B parameters is constructed, comprising several modules: a data normalization module, an encoder module, a feature space optimization module, a decoder module, and an output module. Specifically, as follows... Figure 1 As shown, the method includes the following steps: S1. Data Input: Obtain raw data (Raw Counts) from single-cell transcriptome sequencing, including samples from different batches and sequencing depths.
[0028] S2. Data standardization: Mapped to the HGNC standard gene set (19253 genes), Log1p normalized, and discretized based on a vocabulary of size 1024 to generate the input token sequence.
[0029] In this embodiment, the traditional full GENCODE annotation is abandoned. Instead, a standard gene set of 19253 HGNC (HUGO Gene Nomenclature Committee) genes (19253 coding genes) is used to map the acquired single-cell gene expression profile data to the feature set to map to a unified dimension, thus obtaining single-cell gene expression profile data with a unified dimension, eliminating semantic ambiguity and the many-to-one semantic ambiguity of gene IDs. Log1p (CP10k) normalization is used to obtain normalized single-cell gene expression profile data. After normalization, the data is discretized by multiplying by a scaling factor (e.g., 100) and rounding, mapping continuous values to compact token IDs to obtain the input token sequence (full gene input sequence).
[0030] S3. Feature Extraction: Input the sequence into the hybrid architecture encoder, use bidirectional Mamba to process long sequence dependencies, use Flash Attention to calibrate global information, and perform sparse computation through the MoE layer.
[0031] In this step, the encoder module employs a design combining a "Jamba-style" hybrid architecture with a "DeepSeek-style" MoE, supporting efficient inference at the 1.4B parameter level for cellular feature extraction. This encoder module consists of several cascaded Unified Hybrid Blocks and an end-effector ConvAdapter. The cascaded Unified Hybrid Blocks extract full gene features from the input token sequence, while the ConvAdapter compresses the obtained long sequence's full gene features into a fixed-length (e.g., 32) high-density token sequence (latent space token sequence). The cascaded Unified Hybrid Blocks are alternately configured with a Bidirectional State-Space Model (Bi-SSM) or a Transformer (primarily utilizing the Transformer's Self-Attention mechanism) as sequence feature extractors based on layer indices to extract intermediate token sequences. These are then cascaded with a Hybrid Expert Layer (MoE Layer) as a feedforward network for feature feedforward transformation. Therefore, in this embodiment, Bi-SSM is used as a bidirectional Mamba to handle long sequence dependencies, and a Transformer is used as FlashAttention to calibrate global information.
[0032] In this embodiment, eight unified hybrid blocks are configured, with layers 1-3 and 5-7 using Bi-SSM, and layers 4 and 8 using Flash Attention. The MoE Layer contains one shared expert and seven routing experts, with Top-2 activation during inference and an expert dimension of 2048.
[0033] In this embodiment, Bi-SSM has linear computational complexity ( This allows the model to easily handle a full gene input of nearly 20,000 characters without exceeding GPU memory. Sparse insertion using Transformer enables global calibration.
[0034] Specifically, the hybrid expert layer in the unified hybrid block adopts a shared-routing architecture, including parallel shared experts and routed experts, and a weighted fusion layer at the end. The hybrid expert layer performs the following steps: Shared Expert: All input tokens in the reasoning process are always in the active path, used to process common housekeeper gene features and background signals to obtain common features; Router expert: Configured with a router, it dynamically selects Top-K experts based on input features to perform calculations, which are used to process cell type-specific features and obtain cell type-specific features; Weighted fusion layer: Weighted fusion of general features and cell type-specific features is performed to obtain the output features of each hybrid expert layer.
[0035] In this embodiment, the hidden layer dimension of the expert is set to a wide dimension (e.g., 2048) to maintain information capacity under high compression ratio. In addition, the introduction of the hybrid expert layer allows the model to activate only a small number of parameters during inference (sparse computation), thereby significantly reducing computational latency while maintaining a huge parameter capacity (memory) of 1.4B.
[0036] The output dimension of the above convolutional adapter is configured as [Batch, 32, 128]. This convolutional adapter includes a channel projection layer, a feature extraction layer, and an adaptive pooling layer connected in sequence. Its specific execution process includes: The full gene features are input into the channel projection layer for dimensionality reduction to the adapter dimension, resulting in dimensionality-reduced full gene features. The reduced full gene features are input into the feature extraction layer (a one-dimensional convolutional layer) to extract local feature context. The local feature context is input into the adaptive pooling layer. By using the AdaptiveAvgPool1d operation, the entire gene sequence input from the front end is compressed into a fixed 32 tokens, forming a high-density information bottleneck. These 32 tokens constitute a latent space with a biological topological structure, ultimately resulting in a high-density token sequence.
[0037] The convolutional adapter forces nearly 20,000-dimensional gene features into a fixed-length (32-to-1) token sequence by utilizing one-dimensional convolution and adaptive average pooling. This design reduces the computational cost of subsequent processing by several orders of magnitude.
[0038] For this step, experimental verification showed that the number of model parameters reached more than 1 billion (1B). Thanks to the compression mechanism of the convolution adapter and the sparse computing characteristics of the hybrid expert layer, this method can be configured to support batch inference on a single GPU with less than 10GB of video memory.
[0039] S4. Feature Space Optimization: The latent space is geometrically shaped and smoothed through the ArcFace (Additive Angular Margin) mechanism, and non-biological batch effects are eliminated by combining the Domain Adversarial Neural Network (DANN) mechanism.
[0040] In this embodiment, a feature space optimization module is used for optimization. This module is configured with latent space geometry shaping units and adversarial de-batching units. The execution steps of the latent space geometry shaping unit include: Configure the ArcFace classification head and use the Additive Angular Margin Loss (ArcFace, additive angle) as a constraint. The Additive Angular Margin Loss is simply adding a fixed angle margin m to the original angle loss. In this embodiment, m can be set from 0.15 to 0.25. Cellular features (high-density token sequences) are mapped onto the surface of a hypersphere, and an angular margin is forcibly introduced between the features of the same type of cells and the class center during training. That is, the feature vectors of the same type of cells are squeezed into a minimal cone angle with the class center as the axis. Experiments show that this strategy improves the silhouette coefficient (ASW) of the feature space from negative to positive (e.g., >0.03), realizing the transformation of the feature space from chaos to order.
[0041] By forcing similar cells to be highly cohesive in the latent space through angular margins, while pushing away dissimilar cells, the latent space is smoothed and shaped, eliminating irregular mutations in the characteristic manifold and making it suitable for continuous manifold generation tasks.
[0042] The adversarial debater unit comprises a domain discriminator and a gradient reversal layer, and its execution steps include: The domain discriminator receives the latent space token sequence generated by the encoder and attempts to identify the sequencing batch from which the data originated; The gradient inversion layer inverts the gradient of the domain discriminator and passes it to the encoder during backpropagation. Through the aforementioned adversarial game, the latent space token sequence generated by the encoder is forced to remove batch-specific information that is irrelevant to biology, thus achieving deep batch removal.
[0043] In this embodiment, a multilayer perceptron is configured as the domain discriminator, and adversarial weights (e.g., 1.0) are set during training. This mechanism forces the model to ignore differences in sequencing depth (e.g., the difference between 10x v2 and v3) and retain only the common biological semantics.
[0044] In addition, the feature space optimization module also includes a dynamic weighted balancing mechanism unit, which is used to dynamically adjust the weight ratio of the classification loss function term, the domain adversarial loss function term, and the expert load balancing loss function term during training, so as to achieve the optimal balance between geometric compactness (ASW) and cluster consistency (ARI) of the feature space.
[0045] In the early stages of training, the backbone network (multiple cascaded unified mixed blocks) is frozen, allowing only the classification head to adapt to the feature distribution. Subsequently, the backbone network is unfrozen, and batches are removed using DANN. In the final stage, the clustering structure is further solidified in a pure feature space by adjusting the classification weights (e.g., increasing them to 2.0 or 5.0).
[0046] S5. Decoding Output: The hurdle architecture is used to perform zero-value probability prediction and numerical regression prediction on the compressed features.
[0047] This implementation utilizes a decoder module to decode the aforementioned high-density token sequence to denoise the single-cell gene expression profile data, and outputs it through an output module. In addition, the output module is also used to output the aforementioned latent space token sequence as a low-dimensional dense representation (Embedding), which can be directly used for cell type annotation (verified to have state-of-the-art classification accuracy and clustering structure) or as input to downstream generative models.
[0048] The decoder module includes a context interaction layer (a 6-layer Transformer Encoder), a feature projection layer, and a fence decoding layer. The specific execution steps include: The context interaction layer is used to perform global information exchange between tokens in the high-density token sequence; A feature projection layer is used to extend the dimension of the high-density token sequence to the original high-dimensional gene space dimension; Based on the high-density token sequence after dimensional expansion, a shared MoE feature extraction layer is used for processing to obtain a high-dimensional latent representation; Based on high-dimensional latent representation, zero-value branching and numerical branching are used to perform zero-value probability prediction and numerical regression prediction, respectively, to obtain the non-zero expression probability and potential expression intensity value of each gene. The non-zero expression probability and potential expression intensity value of each gene are then multiplied to obtain the denoised gene expression profile (matrix form) to adapt to the zero-expansion characteristics of single-cell data.
[0049] The latent space consisting of 32 tokens output in this embodiment is configured to adapt to the flow matching generation task. This latent space has denoising properties and a clear cell type geometry, and can be used as a conditional input to generate simulated cell states.
[0050] The virtual cell base model in this embodiment is designed to consume only about 8GB of video memory (Bfloat16 precision) when processing the full set of 19253 genes with a batch size of 256. This allows the system to perform high-throughput inference on consumer-grade graphics cards such as the RTX 4090, with efficiency far exceeding existing models such as scFoundation.
[0051] Experiments show that the model in this embodiment achieves state-of-the-art performance of ARI > 0.32 and ASW > 0.1 on a full-coverage validation set containing 1029 batches, and the generated embeddings exhibit a clear and smooth topology structure in UMAP visualization.
[0052] Example 2 This embodiment provides a single-cell gene expression profile reconstruction system based on a virtual cell basic model with a hybrid expert architecture, including a memory, a processor, and a program stored in the memory. The feature is that when the processor executes the program, it implements a single-cell gene expression profile reconstruction method based on a virtual cell basic model with a hybrid expert architecture as described in Embodiment 1.
[0053] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0054] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.
[0055] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0058] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0059] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for reconstructing single-cell gene expression profiles based on a virtual cell-based model using a hybrid expert architecture, characterized in that, Includes the following steps: The process involves acquiring single-cell gene expression profile data to be reconstructed, inputting it into a pre-constructed virtual cell basic model, and outputting the reconstruction result. The virtual cell basic model includes a data normalization module, an encoder module, a feature space optimization module, and a decoder module connected sequentially. The steps for outputting the reconstruction result include: The data standardization module is used to standardize the single-cell gene expression profile data to obtain the input Token sequence; In the latent space, the encoder module is used to extract and compress features from the input token sequence to obtain a high-density token sequence. At the same time, the feature space optimization module is used to geometrically shape the latent space and remove non-biological batch effects. The high-density token sequence is decoded using a decoder module to perform zero-value probability prediction and numerical regression prediction, resulting in a denoised gene expression profile, which is used as the reconstruction result.
2. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 1, is characterized in that... The step of obtaining the input token sequence includes: The single-cell gene expression profile data is mapped to the HGNC standard gene set to be mapped to a unified dimension, thus obtaining single-cell gene expression profile data with a unified dimension. Based on the single-cell gene expression profile data of the unified dimension, Log1p normalization is performed to obtain normalized single-cell gene expression profile data. The normalized single-cell gene expression profile data is multiplied by a scaling factor to discretize it, generating an input token sequence.
3. The method for reconstructing single-cell gene expression profiles based on a virtual cell-based model with a hybrid expert architecture according to claim 1, characterized in that, The encoder module includes multiple cascaded unified mixing blocks and a convolutional adapter at the end. The multiple cascaded unified mixing blocks are used to extract full gene features, and the convolutional adapter is used to compress the full gene features to obtain the high-density token sequence.
4. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 3, is characterized in that... Each unified hybrid block includes a bidirectional state-space model or a Transformer, and a cascaded hybrid expert layer containing shared and routing experts. For the multiple cascaded unified hybrid blocks, the bidirectional state-space model and the Transformer are alternately configured in different unified hybrid blocks. The bidirectional state-space model and the Transformer act as feature extractors to extract intermediate token sequences, and the hybrid expert layer acts as a feedforward network to perform feature feedforward transformation.
5. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 4, is characterized in that... The hybrid expert layer includes: Shared Expert: All tokens in the intermediate token sequence are always in the active path to capture common features; Routing experts: A router is configured to dynamically select Top-K routing experts from the intermediate token sequence to perform calculations and obtain cell type-specific characteristics; Weighted fusion layer: The general features and cell type-specific features are weighted and fused to obtain the output features of each hybrid expert layer.
6. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 3, is characterized in that... The convolutional adapter includes a channel projection layer, a feature extraction layer, and an adaptive pooling layer connected in sequence. The execution process of the convolutional adapter includes: The full set of gene features are input into the channel projection layer for dimensionality reduction to obtain the dimensionality-reduced full set of gene features. The reduced full gene features are input into a feature extraction layer to extract local feature context, wherein the feature extraction layer is a one-dimensional convolutional layer. The local feature context is input into the adaptive pooling layer for adaptive average pooling processing to form a high-density token sequence.
7. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 1, is characterized in that... The feature space optimization module includes a latent space geometry shaping unit and an adversarial debater, wherein the latent space geometry shaping unit performs the following steps: Additive angular margin loss is used as a constraint to map high-density token sequences onto the surface of a hypersphere, and angular margins are forced between cells of the same class and the class center during the training process of the virtual cell basic model. Based on the aforementioned angular margin, similar cells are forced to highly aggregate in the potential space, while distancing dissimilar cells are pushed away, thereby achieving smoothing and shaping of the potential space. The adversarial debater unit includes a domain discriminator and a gradient inversion layer, and specifically performs the following steps: The domain discriminator receives the high-density token sequence and identifies the sequencing batch from which it originates; The gradient inversion layer inverts the gradient of the domain discriminator and passes it to the encoder module during backpropagation. Through the aforementioned adversarial game, the encoder module is forced to remove batch-specific information that is irrelevant to biology from the high-density token sequence it generates.
8. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 1, is characterized in that... The decoder module includes a context interaction layer, a feature projection layer, and a fence decoding layer. The fence decoding layer includes a shared MoE feature extraction layer and subsequent zero-value and numerical branches for splitting. The specific execution steps of the decoder module include: The context interaction layer is used to perform global information exchange between tokens in the high-density token sequence; The feature projection layer is used to extend the dimension of the high-density token sequence to the original dimension; Based on the high-density token sequence after dimensional expansion, the shared MoE feature extraction layer is used for processing to obtain a high-dimensional latent representation; Based on the high-dimensional latent representation, zero-value branching and numerical branching are used to perform zero-value probability prediction and numerical regression prediction, respectively, to obtain the non-zero expression probability and potential expression intensity value of each gene. The non-zero expression probability and potential expression intensity value of each gene are then multiplied to obtain the denoised gene expression profile.
9. The method for reconstructing single-cell gene expression profiles based on a virtual cell fundamental model using a hybrid expert architecture, as described in claim 1, is characterized in that... The virtual cell basic model also includes an output module for outputting the high-density token sequence and the denoised gene expression profile.
10. A single-cell gene expression profile reconstruction system based on a virtual cell-based model with a hybrid expert architecture, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements a single-cell gene expression profile reconstruction method based on a virtual cell basic model with a hybrid expert architecture as described in any one of claims 1-9.