Plant single cell gene expression prediction method, system, equipment and medium

Abstract

The invention relates to the technical field of crossing of bioinformatics, artificial intelligence and plant genomics, and discloses a plant single cell gene expression prediction method, system, device and medium. A plant single cell gene expression prediction model realizes dynamic feature fusion of a DNA sequence and chromatin accessibility signals through a gated cross attention mechanism; the problem that a traditional single-mode model cannot model regulation and control dynamic association is effectively solved, and the result interpretability is enhanced; a hybrid expert system and a load balancing design are adopted to significantly improve the recognition capability of the model for rare cell types, and a decoupling prediction head design supports efficient transfer learning; a DNA long sequence processing mechanism and nucleosome scale feature coding ensure cross-species compatibility; an end-to-end automatic process and a dynamic parameter optimization framework greatly improve the practicability; a'prediction-verification 'closed-loop support system can be constructed for molecular breeding, and high-precision and interpretable prediction of plant single-cell gene expression is realized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bioinformatics, artificial intelligence and plant genomics, and in particular to a method, system, device and medium for predicting gene expression in single-cell plants. Background Technology

[0002] Single-cell gene expression prediction is a crucial technique for studying plant development and stress response mechanisms. Traditional methods based on bulk RNA sequencing (bulk RNA-seq) or simplified models struggle to reveal intercellular heterogeneity. With technological advancements, single-cell multi-omics technologies (such as 10x Genomics Multiome) can simultaneously acquire transcriptome (scRNA-seq) and chromatin open state (scATAC-seq) data from individual cells, providing new technical support for cell type-specific regulatory studies. However, current deep learning models still face the following key challenges in predicting single-cell gene expression in plants:

[0003] 1. Insufficient collaborative analysis of multimodal data

[0004] Current mainstream tools (such as scBasset and DeepSEA) typically process DNA sequences or batch ATAC-seq data individually, failing to effectively integrate the synergistic effects of these two different data types in gene expression regulation. Studies have shown that open chromatin regions require binding to specific promoter sequences to activate gene expression, and single-modality models cannot accurately simulate this dynamic regulatory relationship. For example, tests in rice show that models based solely on DNA sequences have an average accuracy of 0.61, while models based solely on ATAC data have an accuracy of 0.63, both far from meeting the requirements of practical applications.

[0005] 2. Poor cell type adaptability

[0006] Plant cell types are highly specific (e.g., rice contains 54 cell types), and the sample distribution is extremely uneven (rare cell types, such as tillering bud cells, account for less than 0.5%). Existing models use a uniform shared fully connected layer for prediction, resulting in a prediction accuracy of less than 50% for rare cell types. Especially when predicting across organs (e.g., from roots to leaves), model performance deteriorates significantly, making it difficult to meet the requirements of whole-organ studies.

[0007] 3. Poor cross-species migration performance

[0008] Models developed for mammals show a significant decline in effectiveness when applied to plants. Tests indicate that transferring models trained directly on human data to rice results in a substantial decrease in prediction accuracy due to differences in genomic GC content (average 45% in plants vs. 40% in mammals) and variations in regulatory elements. Existing frameworks lack adaptability to plant genome structures and struggle to accommodate differences in promoter length and other aspects between monocot and dicot plants.

[0009] 4. Insufficient explainability and practicality

[0010] Currently, the identification of plant-specific regulatory elements still relies on manual annotation, and existing black-box models cannot provide interpretable motif patterns (characteristic sequence patterns that recur in biological sequences such as DNA, RNA, or proteins). Furthermore, fixing input parameters (such as forcing a 2kb TSS window) introduces boundary truncation errors in genomes like maize. At the engineering application level, the lack of an integrated toolchain from Seurat objects to prediction results forces researchers to manually integrate multiple software programs, which can easily introduce systematic errors.

[0011] 5. Poor integration between breeding and application

[0012] Existing methods can only provide gene expression probability predictions and have failed to establish an effective link with molecular breeding. Taking the screening of OsGW7, a key gene for rice panicle grain number, as an example, traditional models cannot accurately assess the specific impact of promoter mutations on 54 cell types, resulting in a low success rate in field validation. Summary of the Invention

[0013] This invention provides a method, system, device, and medium for predicting gene expression in plant single cells, in order to overcome the shortcomings of existing technologies.

[0014] The present invention provides a method for constructing a plant single-cell gene expression prediction model, comprising:

[0015] RNA expression data and chromatin accessibility data of multiple single cells of the target plant species were obtained. The RNA expression data and chromatin accessibility data of each single cell were annotated with cell type.

[0016] Based on RNA expression data and chromatin accessibility data of multiple single cells of the target plant species, a hybrid expert model architecture designed with multimodal-cell type decoupling is used to train the model to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thus obtaining a plant single-cell gene expression prediction model.

[0017] According to the present invention, a method for constructing a plant single-cell gene expression prediction model is provided, wherein the method utilizes a hybrid expert model architecture designed with multimodal-cell type decoupling, based on RNA expression data and chromatin accessibility data of multiple single cells of a target plant species, to train the model to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thereby obtaining a plant single-cell gene expression prediction model, comprising:

[0018] Data decoupling and reconstruction were performed on RNA expression data and chromatin accessibility data of multiple single cells of the target plant species to obtain reconstructed RNA expression data and reconstructed chromatin accessibility data.

[0019] Genomic features were extracted from the reconstructed RNA expression data and reconstructed chromatin accessibility data to obtain DNA sequence features and chromatin accessibility signal (ATAC signal) features.

[0020] DNA sequence features and chromatin accessibility signal features are respectively encoded. For single cells belonging to rare cell types, data augmentation is performed using a synthetic minority class oversampling strategy.

[0021] For the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features, a learnable feature weighting mechanism and a gated cross-attention mechanism are combined to automatically adjust the contribution ratio of the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features according to the chromatin openness intensity, thereby achieving cross-modal dynamic feature fusion.

[0022] According to a method for constructing a plant single-cell gene expression prediction model provided by the present invention, the step of decoupling and reconstructing RNA expression data and chromatin accessibility data of multiple single cells of a target plant species to obtain reconstructed RNA expression data and reconstructed chromatin accessibility data includes:

[0023] RNA expression data and chromatin accessibility data from multiple single cells of the target plant species were decoupled and reconstructed to obtain scRNA-seq expression matrix, scATAC-seq open signal matrix, and cell metadata. Among these, a chromosome-sensing segmentation strategy was used to ensure that homologous genes were not distributed across the training / test sets.

[0024] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, the step of extracting genomic features from reconstructed RNA expression data and reconstructed chromatin accessibility data to obtain DNA sequence features and chromatin accessibility signal features includes:

[0025] For the reconstructed RNA expression data, based on the GTF genome annotation file, the transcription start site (TSS) of each gene is located, and X kb intervals upstream and downstream of the transcription start site are dynamically extracted (X is an optimizable parameter, for example, 1, with an initial value of 2000bp) to obtain DNA sequence features.

[0026] For the reconstructed chromatin accessibility data, the data is segmented according to the same genomic coordinates, and the continuous signal is discretized into feature vectors of 147bp nucleosome units using an adaptive binning algorithm to obtain chromatin accessibility signal features.

[0027] According to the present invention, a method for constructing a plant single-cell gene expression prediction model includes feature encoding of DNA sequence features and chromatin accessibility signal features, respectively, comprising:

[0028] Four-dimensional one-hot encoding is used to encode DNA sequence features;

[0029] The chromatin accessibility signal features were normalized using log(1+x).

[0030] According to the present invention, a method for constructing a plant single-cell gene expression prediction model, the multimodal-cell type decoupled hybrid expert model architecture includes a dual-path feature encoder, a cross-modal dynamic fusion module, a load-balanced expert system, and a cell type-specific prediction head.

[0031] The dual-path feature encoder includes a DNA coding branch and a chromatin accessibility coding branch. The DNA coding branch integrates the HyenaDNA long sequence processing mechanism, captures k-mer hierarchical regulatory patterns through convolutional gating and global recursive operators, supports input sequences up to 5kb, and outputs a 512-dimensional feature vector that covers the semantic information of regulatory elements such as promoters and enhancers. The chromatin accessibility coding branch adopts a sparse positional coding Transformer architecture, focuses on open chromatin patterns at the nucleosome scale (147bp) through local self-attention, and outputs a 256-dimensional feature vector that encodes the synergistic effect of spatial openness and long-range interactions.

[0032] The cross-modal dynamic fusion module is designed with a gated cross-attention module. It uses DNA sequence features as the query and chromatin accessibility signal features as the key-value pair for cross-modal interaction. The fusion weight α is generated from learnable parameters and satisfies the formula:

[0033] α=σ(W[DNA embed ATAC embed ]),

[0034] In the formula, DNA embed Indicates DNA sequence characteristics, ATACembed σ represents the chromatin accessibility signal feature, which is the Sigmoid function. This function automatically reduces the weight of DNA sequence features in regions with high chromatin accessibility (α→0.3), thus strengthening the contribution of open chromatin to gene expression regulation. W represents the weight matrix, a parameter that needs to be optimized during model training. It is a linear transformation matrix responsible for mapping the spliced ​​high-dimensional features to a scalar value. Its dimensions are (1, D_dna + D_atac).

[0035] The load-balanced expert system is deployed with a Dynamic Hybrid Expert Layer (DynamicMoE), which contains multiple dedicated prediction experts equal to the number of cell types. Through a Top-k sparse routing mechanism, only the k most relevant experts are activated for each sample, and an expert load-balancing loss is introduced.

[0036] L balance =λ·std(expert_utilization),

[0037] In the formula, L balabce This represents the load balancing loss value, where λ represents the loss weight hyperparameter, and std(expert_utilization) represents the value.

[0038] This means that the standard deviation function of the rate vector is used by experts to force rare cell type experts to obtain sufficient training samples and improve the recall rate of rare types.

[0039] The cell type-specific prediction head is specifically designed to build an independent binary classifier for each cell type, containing a 128-dimensional hidden layer and a Sigmoid output unit. This decoupled design supports local fine-tuning in transfer learning scenarios—when new cell type data is added, only the new prediction head needs to be trained while keeping the backbone network frozen, thus improving the model iteration efficiency by 3.7 times.

[0040] According to the present invention, a method for constructing a plant single-cell gene expression prediction model is provided. The plant single-cell gene expression prediction model adopts a five-stage cascaded optimization training process, wherein the five stages include multi-task loss design, progressive parameter optimization, dynamic learning rate scheduling, early stopping and model selection, and modal smile experiment.

[0041] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, in the multi-task loss design, the main loss function adopts cell type weighted FocalLoss to alleviate the sample imbalance problem, and the auxiliary loss includes MoE load balancing regularization term (λ=0.01) and chromatin accessibility feature-DNA sequence feature consistency constraint (cosine similarity loss).

[0042] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, the progressive parameter optimization includes:

[0043] Phase 1: Fix the model architecture and optimize the data parameters (sequence length l∈[500,5000]bp, ATAC bin size∈[100,500]bp);

[0044] Phase 2: Freeze data parameters and search model hyperparameters (learning rate ∈ [1e-5, 1e-3], number of MoE experts ∈ [8, 64]);

[0045] Phase 3: Jointly fine-tune all parameters and evaluate 200+ combinations in parallel using the Optuna framework.

[0046] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, the dynamic learning rate scheduling is as follows: a cosine annealing strategy is adopted, the initial learning rate is 3e-4, and it decays by 30% every 10 training rounds. The validation sensitivity monitoring callback is integrated—when the area under the validation ROC curve (AUROC) stagnates, the learning rate is automatically halved.

[0047] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, the early stopping and model selection are specifically as follows: an early stopping mechanism is implemented based on the area under the mean precision-recall curve (AUPRC) of the validation set, and the best-performing checkpoint is saved to the local directory. Each checkpoint contains the complete model state and independent parameters of all prediction heads.

[0048] According to the method for constructing a plant single-cell gene expression prediction model provided by the present invention, the modal ablation experiment specifically includes: a single-modal control experiment built into the training process, which can independently evaluate the prediction performance of prediction models based on DNA sequences or based on ATAC signal features.

[0049] The trained model can be used in downstream application analysis systems to form a closed loop of "prediction-interpretation-verification", with functions including: visualization of cell expression, interpretation of regulatory elements, and analysis of genetic variations.

[0050] This invention also provides a method for predicting gene expression in plant single cells, comprising:

[0051] Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested;

[0052] Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above can predict the cell type of single cells of the plant species to be tested and the gene expression level of the plant species to be tested.

[0053] This invention also provides a plant single-cell gene expression prediction system, comprising:

[0054] The data acquisition module is used to acquire RNA expression data and chromatin accessibility data of single cells of the plant species to be tested.

[0055] The prediction module is used to: predict the cell type and gene expression level of the single cells of the plant species under test based on the RNA expression data and chromatin accessibility data of the single cells of the plant species under test, using the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above.

[0056] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the method for constructing a plant single-cell gene expression prediction model and the method for predicting plant single-cell gene expression as described above.

[0057] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a plant single-cell gene expression prediction model or the method for predicting plant single-cell gene expression as described above.

[0058] The present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute any of the above-described methods for constructing a plant single-cell gene expression prediction model and for predicting plant single-cell gene expression.

[0059] The present invention provides a method, system, device, and medium for predicting gene expression in plant single cells, which has at least the following beneficial effects:

[0060] 1. Breakthrough improvements in multimodal collaborative analysis

[0061] Dynamic feature fusion mechanism: The plant single-cell gene expression prediction model automatically learns the contribution weights of DNA sequence and chromatin accessibility signals through gated cross attention, effectively overcoming the problem that traditional single-modal models cannot model dynamic correlations.

[0062] Enhanced interpretability: The plant single-cell gene expression prediction model synchronously outputs the DNA sequence motif and the characteristic association of open chromatin regions, which can provide reliable biological evidence for the analysis of regulatory mechanisms.

[0063] 2. Systematic optimization of cell type adaptation

[0064] Hybrid Expert System (MoE) and Load Balancing: Plant single-cell gene expression prediction models for rare cell types can significantly improve the robustness of the model to imbalanced data by using sparse routing and load balancing loss in dynamic MoE layers.

[0065] Decoupled prediction head design: The plant single-cell gene expression prediction model supports transfer learning across organs or new cell types. Only specific prediction heads need to be fine-tuned, avoiding repeated training of the entire model and effectively improving model training efficiency and practicality.

[0066] 3. Innovative designs with cross-species compatibility

[0067] Genome structure adaptation: HyenaDNA long sequence processing mechanism and dynamic TSS window truncation strategy in plant single-cell gene expression prediction model to adapt to the differences in promoter length and GC content of different plant species.

[0068] Nucleosome-scale feature encoding: Plant single-cell gene expression prediction models can preserve plant-specific nucleosome arrangement patterns by discretizing chromatin open signals in 147bp units.

[0069] 4. Improved practicality in engineering implementation

[0070] End-to-end process integration: The entire chain of automation, from inputting multi-omics data into the plant single-cell gene expression prediction model to outputting prediction results, can reduce errors caused by human intervention.

[0071] Dynamic parameter optimization framework: The plant single-cell gene expression prediction model supports automated search and adaptation of key parameters such as sequence length and bin size, which can improve the accuracy of the plant single-cell gene expression prediction model.

[0072] 5. Closed-loop support for breeding applications

[0073] Cell type-specific effect assessment: It can use plant single-cell gene expression prediction models to analyze the effects of promoter mutations on specific cell types, providing a basis for target screening in molecular breeding.

[0074] Visualization and Decision-Making Toolchain: By integrating the visualization of regulatory networks and the analysis of genetic variations, it forms a "prediction-verification" closed loop, improving the interpretability of plant single-cell gene expression prediction. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0076] Figure 1 This is one of the flowcharts for a plant single-cell gene expression prediction method provided by the present invention.

[0077] Figure 2 This is the second flowchart illustrating a method for predicting gene expression in single-cell plants provided by the present invention.

[0078] Figure 3 This is a schematic diagram illustrating the predictive performance of a plant single-cell gene expression prediction model in rice panicles, measured using AUROC.

[0079] Figure 4 This is a schematic diagram of the structure of a plant single-cell gene expression prediction system provided by the present invention.

[0080] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0082] Figure 1 and Figure 2 This is a flowchart illustrating a plant single-cell gene expression prediction method provided by the present invention. The plant single-cell gene expression prediction method provided by the present invention comprises two parts: the construction and application of a plant single-cell gene expression prediction model. The executing entity of the plant single-cell gene expression prediction method provided by the present invention can be any suitable terminal-side device or network-side device, such as a plant single-cell gene expression prediction device.

[0083] See Figure 1 and Figure 2 The present invention provides a method for predicting gene expression in single-cell plants, which may include:

[0084] S110. Obtain RNA expression data and chromatin accessibility data for multiple single cells of the target plant species. Each single cell's RNA expression data and chromatin accessibility data is annotated with cell type information. In this embodiment, rice is used as the target plant species. RNA expression data and chromatin accessibility data for over 110,000 cells from its eight main organs (root, stem, young leaf, flag leaf, shoot tip, tiller bud, young panicle, and seed) are obtained to construct a plant single-cell gene expression prediction model.

[0085] S120. Based on RNA expression data and chromatin accessibility data of multiple single cells of the target plant species, a hybrid expert model architecture designed with multimodal-cell type decoupling is used to train the model to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thereby obtaining a plant single-cell gene expression prediction model.

[0086] S130. Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested.

[0087] S140. Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the cell type of single cells and the gene expression level of the plant species to be tested are predicted using a plant single-cell gene expression prediction model.

[0088] In step S110, RNA expression data and chromatin accessibility data for multiple single cells of the target plant species can be obtained from commercial databases.

[0089] RNA expression data and chromatin accessibility data of multiple single cells of the target plant species can also be obtained through a data acquisition device.

[0090] In this embodiment, the rice dual-omics data are self-tested data, which have now been uploaded to NCBI: GSE232863. The Arabidopsis dual-omics data are collected from NCBI: GSE235510. The remaining single-cell data were collected from the website: https: / / biobigdata.nju.edu.cn / scplantdb / home.

[0091] In one embodiment, the data acquisition device may include: a single-cell isolation module configured to process target plant tissue to isolate multiple single cells; a molecular detection unit including an RNA sequencing submodule and a chromatin accessibility detection submodule, used to acquire RNA expression data and chromatin accessibility data of the single cells, respectively; a data preprocessing unit including a memory and a data filtering component, configured to perform quality control, standardization, and format conversion on the RNA expression data and chromatin accessibility data to generate an input dataset that can be used for model training; and a data transmission interface configured to transmit the preprocessed data to the model training module.

[0092] In the data acquisition device, the single-cell separation module refers to a component that processes target tissue to obtain multiple single cells, including a mechanical separator, an enzymatic digestion system, and a microfluidic chip. The single-cell separation module can be implemented using commercially available equipment or instruments, depending on the intended purpose and function. For example, a mechanical separator could be a MiltenyigentleMACS. TM The dissociator (Miltenyi Biotec, Germany), the enzymatic digestion system (Worthington Plant Protoplast Isolation Kit, Worthington Biochemical, USA), and the microfluidic chip system (Dolomite Bio Nadia) are all suitable options. TM Innovate (Dolomite Bio, UK).

[0093] A molecular detection unit refers to a module used for detecting molecular-level data, including but not limited to RNA sequencing and chromatin accessibility detection devices. The molecular detection unit can be implemented using commercially available instruments or devices, including but not limited to, 10x Genomics Chromium, depending on its intended use and function. TM Controller (supports RNA and ATAC modes), BDRhapsody TM System (Scalable Multi-omics Detection), Parse Biosciences Evercode TM Multi-Omics (supports simultaneous acquisition of RNA and chromatin accessibility information).

[0094] The RNA sequencing submodule refers to the component that performs single-cell RNA sequencing (scRNA-seq) analysis. This RNA sequencing submodule can use commercially available equipment or instruments, such as the Illumina NextSeq. TM / NovaSeq TMSeries of sequencers, such as the BGIMGISEQ-2000 sequencer.

[0095] The chromatin accessibility detection submodule refers to the component that performs ATAC-seq or other chromatin accessibility analyses. This chromatin accessibility detection submodule can use commercially available products, such as 10x Genomics Single Cell ATACSolution, Illumina Tagment DNA Enzyme and Buffer Kit (used in conjunction with a microfluidic platform), Fluidigm C1 TM +ATAC-seq, etc.

[0096] A data preprocessing unit refers to a processing unit that performs quality control, standardization, and format conversion on raw data. Data preprocessing units can use commercially available workstations pre-installed with suitable software, such as Dell Precision. TM Tower workstations, HP Z Series workstations (for batch quality control, standardization, and format conversion), etc.

[0097] A data transmission interface refers to a hardware interface that enables data to be transmitted from a device to an external system.

[0098] In one embodiment, S120 may include:

[0099] A dedicated parser was used to decouple and reconstruct RNA expression data and chromatin accessibility data of multiple single cells of the target plant species, resulting in reconstructed RNA expression data and reconstructed chromatin accessibility data, including scRNA-seq expression matrix (converted into a cell × gene binary tag matrix), scATAC-seq open signal matrix, and cell metadata (including cell type annotation). Among these, a chromosome-sensing segmentation strategy was used to ensure that homologous genes are not distributed across the training / test set, eliminating the risk of data leakage.

[0100] Genome feature extraction was performed on the reconstructed RNA expression data and reconstructed chromatin accessibility data to obtain DNA sequence features and chromatin accessibility signal (ATAC signal) features. Specifically, for the reconstructed RNA expression data, the transcription start site (TSS) of each gene was located based on the GTF genome annotation file, and 1kb intervals upstream and downstream of the transcription start site were dynamically extracted (1 is an optimizable parameter with an initial value of 2000bp) to obtain DNA sequence features. For the reconstructed chromatin accessibility data, the data was segmented according to the same genome coordinates, and the continuous signal was discretized into feature vectors of 147bp nucleosome units using an adaptive binning algorithm to obtain chromatin accessibility signal features.

[0101] DNA sequence features and chromatin accessibility signal features are respectively encoded. Four-dimensional one-hot encoding is used to encode DNA sequence features (A:[1,0,0,0],T:[0,1,0,0],C:[0,0,1,0],G:[0,0,0,1]). Chromatin accessibility signal features are normalized using log(1+x). For single cells belonging to rare cell types (<0.5% proportion), a synthetic minority class oversampling strategy is used for data augmentation to generate synthetic samples in the feature space to balance the data distribution.

[0102] For the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features, a learnable feature weighting mechanism (embedded learnable feature weighting module (DynamicFeatureWeighting)) and a gated cross-attention mechanism are combined to automatically adjust the contribution ratio of the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features according to the chromatin openness intensity, thereby realizing cross-modal dynamic feature fusion. This increases the ATAC feature weight of high open regions (ATAC signal >90 percentile) to 70%-90%, significantly enhancing the characterization ability of regulatory elements.

[0103] In one embodiment, the hybrid expert model architecture of multimodal-cell type decoupled design includes a dual-path feature encoder, a cross-modal dynamic fusion module, a load-balanced expert system, and a cell type-specific prediction head, wherein...

[0104] The dual-path feature encoder includes a DNA coding branch and a chromatin accessibility coding branch. The DNA coding branch integrates the HyenaDNA long sequence processing mechanism, captures k-mer hierarchical regulatory patterns through convolutional gating and global recursive operators, supports input sequences up to 5kb, and outputs a 512-dimensional feature vector that covers the semantic information of regulatory elements such as promoters and enhancers. The chromatin accessibility coding branch adopts a sparse positional coding Transformer architecture, focuses on open chromatin patterns at the nucleosome scale (147bp) through local self-attention, and outputs a 256-dimensional feature vector that encodes the synergistic effect of spatial openness and long-range interactions.

[0105] The cross-modal dynamic fusion module is designed with a gated cross-attention module. It uses DNA sequence features as the query and chromatin accessibility signal features as the key-value pair for cross-modal interaction. The fusion weight α is generated from learnable parameters and satisfies the formula:

[0106] α=σ(W[DNA embed ATAC embed ]),

[0107] In the formula, DNA embed Indicates DNA sequence characteristics, ATAC embed σ represents the chromatin accessibility signal feature, σ represents the Sigmoid function, which automatically reduces the weight of DNA sequence features in regions with high chromatin accessibility signals (α→0.3), thereby strengthening the contribution of open chromatin to the regulation of gene expression, and W represents the weight matrix;

[0108] The load-balanced expert system is deployed with a Dynamic Hybrid Expert Layer (DynamicMoE), which contains multiple dedicated prediction experts equal to the number of cell types. Through a Top-k sparse routing mechanism (k=8 in this embodiment), only the 8 most relevant experts are activated for each sample, and an expert load-balanced loss is introduced:

[0109] L balance =λ·std(expert_utilization),

[0110] In the formula, L balance This represents the load balancing loss value, where λ represents the loss weight hyperparameter, and std(expert_utilization) represents the value.

[0111] This means that the standard deviation function of the rate vector is used by experts to force rare cell type experts to obtain sufficient training samples and improve the recall rate of rare types.

[0112] The cell type-specific prediction head is specifically designed to build an independent binary classifier for each cell type, containing a 128-dimensional hidden layer and a Sigmoid output unit. This decoupled design supports local fine-tuning in transfer learning scenarios—when new cell type data is added, only the new prediction head needs to be trained while keeping the backbone network frozen, thus improving the model iteration efficiency by 3.7 times.

[0113] In one embodiment, the plant single-cell gene expression prediction model employs a five-stage cascaded optimization training process. The five stages include multi-task loss design, progressive parameter optimization, dynamic learning rate scheduling, early stopping and model selection, and modal smile experiments.

[0114] In the multi-task loss design, the main loss function adopts FocalLoss with cell type weighting to alleviate the sample imbalance problem (30% of root cells vs. 0.5% of tillering cells). The auxiliary loss includes MoE load balancing regularization term (λ = 0.01) and chromatin accessibility feature-DNA sequence feature consistency constraint (cosine similarity loss).

[0115] Progressive parameter optimization includes:

[0116] Phase 1: Fix the model architecture and optimize the data parameters (sequence length l∈[500,5000]bp, ATAC bin size∈[100,500]bp);

[0117] Phase 2: Freeze data parameters and search model hyperparameters (learning rate ∈ [1e-5, 1e-3], number of MoE experts ∈ [8, 64]);

[0118] Phase 3: Jointly fine-tune all parameters and evaluate 200+ combinations in parallel using the Optuna framework;

[0119] The dynamic learning rate scheduling is as follows: a cosine annealing strategy is adopted, with an initial learning rate of 3e-4, which decays by 30% every 10 training rounds. A validation sensitivity monitoring callback is integrated—the learning rate is automatically halved when validation AUROC stalls.

[0120] Early stopping and model selection are as follows: an early stopping mechanism is implemented based on the validation set AUPRC, and the best-performing checkpoint is saved to the local directory. Each checkpoint contains the complete model state and independent parameters of 54 prediction heads.

[0121] The modal ablation experiment specifically involved a built-in single-modal control experiment within the training workflow, which independently evaluated the predictive performance of prediction models based on DNA sequences or ATAC signal features. Tests showed that multimodal fusion improved AUROC predictions compared to single-DNA or single-ATAC predictions. (See [link to relevant documentation]). Figure 3 .

[0122] The environmental dependencies and system deployment for constructing the plant single-cell gene expression prediction model in this embodiment are as follows: The constructed deep learning framework is based on a Linux / Unix operating system environment and adopts modular container technology to achieve cross-platform deployment. System dependency management is implemented through a Conda virtual environment, pre-installed with Python 3.9 and the PyTorch 2.0 deep learning framework, and integrated with a bioinformatics toolchain (including single-cell analysis tools such as Seurat and Scanpy). Users can create and install their own environment with one click via a script, including the 23 core dependency libraries defined in the software. For different hardware configurations, the system provides three-level environment configuration files: basic configuration (base.yaml) defines standard CPU running parameters; test configuration (test_config.yaml) enables lightweight model acceleration for verification; and optimization configuration (best_params.yaml) stores the optimal parameter combination determined through hyperparameter search. The system supports distributed training mode, can automatically detect GPU resources and enable mixed-precision computing, and in rice testing, the training time for a scale of 110,000 cells was reduced from 72 hours to 18 hours.

[0123] The trained model can be used in downstream application analysis systems to form a closed loop of "prediction-interpretation-verification," with functions including:

[0124] Cell expression visualization: The UMAP module reduces the dimensionality of the cell-dimensional expression vectors predicted by the model to 2D space and displays them overlaid with the real scRNA-seq map. It supports interactive screening of specific organs (such as young spikelets) and calculates adjusted mutual information (AMI>0.91) to quantify the consistency between prediction and actual distribution.

[0125] Explanation of control elements:

[0126] Motif discovered that cell type-specific regulatory motifs can be extracted based on an integral gradient algorithm.

[0127] Open region effect analysis: Through systematic mutation scanning, the enhancing effect of ATAC signal on motif function was quantified, and chromatin open region intensity-expression regulation efficiency curves were generated;

[0128] Genetic Variation Analysis: The SNP effect prediction module assesses the perturbation of expression caused by single nucleotide polymorphisms (SNPs) in the promoter region. Inputting a mutation sequence, the model outputs the probability of expression changes in 54 cell types.

[0129] The plant single-cell gene expression prediction method provided in this embodiment achieves high-precision gene expression prediction for over 110,000 single cells across eight major rice organs through a hybrid expert model architecture designed with multimodal and cell type decoupling. This method innovatively employs a dynamic feature fusion mechanism, automatically adjusting the contribution ratio of DNA sequence features and chromatin accessibility signals through gated cross-attention (increasing the ATAC feature weight of highly open regions to 70%-90%), significantly enhancing the characterization ability of regulatory elements. For rare cell types (such as tillering shoot cells accounting for <0.5%), the model effectively improves sample imbalance through a load-balanced expert system and a synthetic minority class oversampling strategy. A specially designed HyenaDNA long sequence processing mechanism (supporting 5kb input) and 147bp nucleosome-scale feature encoding ensure the model's compatibility with different plant species. Experiments show that this method significantly improves the AUROC metric compared to traditional single-modal models, and the efficient deployment achieved through modular container technology reduces large-scale training time from 72 hours to 18 hours. In addition, the model's functions of visualizing cell expression, interpreting regulatory elements, and analyzing genetic variations form a complete "prediction-interpretation-verification" closed loop, providing powerful tools and data support for plant molecular breeding.

[0130] The plant single-cell gene expression prediction system provided by the present invention is described below. The plant single-cell gene expression prediction system described below can be referred to in correspondence with the plant single-cell gene expression prediction method described above.

[0131] See Figure 4 The present invention provides a plant single-cell gene expression prediction system, which may include:

[0132] The data acquisition module is used to acquire RNA expression data and chromatin accessibility data of single cells of the plant species to be tested.

[0133] The prediction module is used to: predict the cell type and gene expression level of the single cells of the plant species under test based on the RNA expression data and chromatin accessibility data of the single cells of the plant species under test, using the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above.

[0134] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the following steps:

[0135] Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested;

[0136] Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above can predict the cell type of single cells of the plant species to be tested and the gene expression level of the plant species to be tested.

[0137] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present 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.

[0138] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps:

[0139] Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested;

[0140] Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above can predict the cell type of single cells of the plant species to be tested and the gene expression level of the plant species to be tested.

[0141] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0142] Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested;

[0143] Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model described above can predict the cell type of single cells of the plant species to be tested and the gene expression level of the plant species to be tested.

[0144] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0145] The following are specific embodiments utilizing the method and system of the present invention:

[0146] Example 1: Prediction and Regulation of Rice Grain Number Based on Commercial Single-Cell Capture Platform

[0147] In this embodiment, the single-cell data acquisition device is a modified commercial single-cell capture and sequencing platform (e.g., the 10x Genomics Chromium system), equipped with a fluorescence microscopy imaging module and a microfluidic sorting chip.

[0148] Native cell suspensions were prepared from young panicle tissue (especially floret differentiation region) in the early flowering stage of rice. Single-cell RNA expression data and chromatin accessibility data were acquired simultaneously using microfluidic chips and imaging modules, and cell type annotation was automatically completed by combining floral organ-specific fluorescent markers.

[0149] A multimodal-cell type decoupled hybrid expert model was run on a high-performance computing node connected to the platform to predict the expression patterns and regulatory networks of key regulatory genes affecting the number of grains per ear.

[0150] Based on the prediction results, feature importance scores (such as SHAP values) are calculated, and heatmaps showing genes with high potential for high granulation number and visualizations of open chromatin regions are generated.

[0151] The information marked as "high-potential cells" in the report is fed back to the sorting platform control unit in real time, and the microfluidic channel pressure and sorting valves are adjusted to automatically enrich these cells for subsequent in vitro culture or gene editing.

[0152] Example 2: Optimization of Rice Grain Number per Panicle Based on Greenhouse Intelligent Breeding Platform

[0153] In this embodiment, the breeding control system is a greenhouse automation management platform (such as the Priva or Argus system), which can automatically adjust environmental conditions such as light, temperature, humidity, and irrigation frequency.

[0154] The single-cell data acquisition device is a laser capture microscopy system equipped with cell sorting function, used to obtain single-cell RNA expression and chromatin accessibility data from the young panicle development stage tissues of different rice varieties.

[0155] The trained prediction model is run on the server to generate an interpretable analysis report that includes the contribution ranking of genes regulating floret number and highlights key regulatory factors that affect floret differentiation and degeneration.

[0156] The report was input into the greenhouse control platform via a data interface. The system automatically adjusted the photoperiod, nitrogen supply, and temperature in different experimental areas to verify and accelerate the screening of varieties that ranked high in the high grain number potential prediction, such as extending the light duration and optimizing the nitrogen fertilizer application frequency in the early heading stage.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0158] The present invention also includes the following embodiments:

[0159] Implementation Scheme 1. A computer-based method for constructing a plant single-cell gene expression prediction model, comprising:

[0160] RNA expression data and chromatin accessibility data of multiple single cells of the target plant species were acquired using a data acquisition device. Each single cell's RNA expression data and chromatin accessibility data were labeled with cell type.

[0161] On a computer system including a processor and memory, based on RNA expression data and chromatin accessibility data, a hybrid expert model architecture designed with multimodal-cell type decoupling is used to train the model to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thereby obtaining a plant single-cell gene expression prediction model.

[0162] Output a trained plant single-cell gene expression prediction model, which is configured to predict the cell type and gene expression level of a single cell of a plant species on a computer system based on the single-cell RNA expression data and chromatin accessibility data of the single cells of the plant species to be tested.

[0163] Implementation Scheme 2. The method according to Implementation Scheme 1, wherein the data acquisition device comprises:

[0164] The single-cell separation module is configured to process target plant tissue to separate and obtain multiple single cells;

[0165] The molecular detection unit, including an RNA sequencing submodule and a chromatin accessibility detection submodule, is used to acquire single-cell RNA expression data and chromatin accessibility data, respectively; the data preprocessing unit, including a memory and a data filtering component, is configured to perform quality control, standardization, and format conversion on the RNA expression data and chromatin accessibility data to generate an input dataset that can be used for model training.

[0166] The data transmission interface is configured to transmit the preprocessed data to the model training module.

[0167] Implementation Scheme 3. According to the method described in Implementation Scheme 1, the hybrid expert model architecture of multimodal-cell type decoupled design includes a dual-path feature encoder, a cross-modal dynamic fusion module, a load-balanced expert system, and a cell type-specific prediction head, wherein...

[0168] The dual-path feature encoder includes a DNA coding branch and a chromatin accessibility coding branch. The DNA coding branch integrates a long sequence processing mechanism, captures k-mer level regulatory patterns through convolutional gating and global recursive operators, supports input sequences up to 5kb, and outputs a 512-dimensional feature vector that covers the semantic information of regulatory elements. The chromatin accessibility coding branch adopts a sparse positional coding Transformer architecture, focuses on open chromatin patterns at the nucleosome scale through local self-attention, and outputs a 256-dimensional feature vector that encodes the synergistic effect of spatial openness and long-range interaction.

[0169] The cross-modal dynamic fusion module is designed with a gated cross-attention module. It uses DNA sequence features as the query and chromatin accessibility signal features as the key-value pair for cross-modal interaction. The fusion weight α is generated from learnable parameters and satisfies the formula:

[0170] α=σ(W[DNA embed ATAC embed ]),

[0171] In the formula, DNA embed Indicates DNA sequence characteristics, ATAC embed σ represents the chromatin accessibility signal feature, W represents the sigmoid function, and W represents the weight matrix.

[0172] The load-balanced expert system is deployed with a dynamic hybrid expert layer, containing multiple dedicated prediction experts equal to the number of cell types. Through a Top-k sparse routing mechanism, only the k most relevant experts are activated for each sample, and an expert load-balancing loss is introduced.

[0173] L balance =λ·std(expert_utilization),

[0174] In the formula, L balance This represents the load balancing loss value, where λ represents the loss weight hyperparameter, and std(expert_utilization) represents the value.

[0175] The standard deviation function represents the vector of expert utilization rates;

[0176] The cell type-specific prediction head is specifically designed to build an independent binary classifier for each cell type, containing a 128-dimensional hidden layer and a sigmoid output unit. This decoupled design supports local fine-tuning in transfer learning scenarios—when new cell type data is added, only the new prediction head needs to be trained while keeping the backbone network frozen.

[0177] Implementation Scheme 4. According to the method described in Implementation Scheme 1, wherein,

[0178] The plant single-cell gene expression prediction model employs a five-stage cascaded optimization training process. These five stages include multi-task loss design, progressive parameter optimization, dynamic learning rate scheduling, early stopping and model selection, and modal smile experiments.

[0179] In the multi-task loss design, the main loss function adopts cell type weighted Focal Loss, and the auxiliary loss includes MoE load balancing regularization term and chromatin accessibility feature-DNA sequence feature consistency constraint.

[0180] Progressive parameter optimization includes:

[0181] Phase 1: Fix the model architecture and optimize data parameters;

[0182] Phase Two: Freeze data parameters and search for model hyperparameters;

[0183] Phase 3: Jointly fine-tune all parameters and evaluate 200+ combinations in parallel using the Optuna framework;

[0184] The dynamic learning rate scheduling is as follows: a cosine annealing strategy is adopted, with an initial learning rate of 3e-4, which decays by 30% every 10 training rounds. A validation sensitivity monitoring callback is integrated—when the area under the validation ROC curve stagnates, the learning rate is automatically halved.

[0185] Early stopping and model selection are specifically implemented as follows: an early stopping mechanism is implemented based on the area under the average precision-recall curve of the validation set, and the best-performing checkpoint is saved to the local directory. Each checkpoint contains the complete model state and independent parameters of all prediction heads.

[0186] The modal ablation experiment specifically involves a built-in single-modal control experiment in the training process to independently evaluate the predictive performance of prediction models based on DNA sequence features or chromatin accessibility signal features.

[0187] Implementation Scheme 5. The method according to Implementation Scheme 1, wherein the RNA expression data is single-cell transcriptome sequencing (scRNA-seq) data, and the chromatin accessibility data is single-cell ATAC sequencing (scATAC-seq) data.

[0188] Implementation Scheme 6. The method according to Implementation Scheme 1, wherein the method includes generating an interpretability analysis report on the computer system, the report visualizing the cell type of single cells of the plant species under test and the gene expression level of the plant species under test.

[0189] Implementation Scheme 7. A plant single-cell gene expression prediction system, comprising:

[0190] Single-cell transcriptome sequencing device: used to acquire RNA expression data from multiple single cells of the plant species to be tested;

[0191] Single-cell chromatin accessibility sequencing device: used to acquire chromatin accessibility data corresponding to the plurality of single cells;

[0192] Computing device: electrically connected to the sequencing apparatus, the computing device including at least one processor and a memory storing instructions executable on the processor, the processor being configured to:

[0193] Receive the RNA expression data and the chromatin accessibility data, and align them with the cell type annotation of a single cell;

[0194] Based on a hybrid expert model architecture designed with multimodal-cell type decoupling, the model is trained to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thus obtaining a plant single-cell gene expression prediction model.

[0195] Train and optimize a plant single-cell gene expression prediction model so that when given RNA expression data and chromatin accessibility data of single cells of a plant species, it outputs the cell type and gene expression level of the single cells of the plant species to be tested.

[0196] Implementation Scheme 8. The system according to Implementation Scheme 7 further includes generating an interpretability analysis report for the prediction results, preferably providing the prediction results and / or the report to the breeding control system for developing breeding programs and conducting plant development and stress response analysis.

[0197] Implementation Scheme 9. A non-transitory computer-readable storage medium having instructions stored thereon, the instructions, when executed by one or more processors, causing the one or more processors to perform the following operations:

[0198] RNA expression data of multiple single cells of the target species were obtained from a single-cell transcriptome sequencing device, and corresponding chromatin accessibility data were obtained from a single-cell chromatin accessibility sequencing device. Each single cell was labeled with a cell type.

[0199] The RNA expression data and chromatin accessibility data are input into a hybrid expert model architecture designed with multimodal-cell type decoupling. The model is trained to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, and a plant single cell gene expression prediction model is obtained.

[0200] Train and optimize a plant single-cell gene expression prediction model so that when given RNA expression data and chromatin accessibility data of single cells of a plant species, it outputs the cell type and gene expression level of the single cells of the plant species to be tested.

[0201] Implementation Scheme 10. The non-transitory computer-readable storage medium according to Implementation Scheme 9, wherein the processor further performs the following operations:

[0202] Generate an interpretable analysis report on the prediction results. Preferably, the prediction results and / or the report are provided to the breeding control system for developing breeding programs and conducting plant development and stress response analysis.

[0203] 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 method for constructing a plant single-cell gene expression prediction model, characterized in that, include: RNA expression data and chromatin accessibility data of multiple single cells of the target plant species were obtained. The RNA expression data and chromatin accessibility data of each single cell were annotated with cell type. Based on RNA expression data and chromatin accessibility data of multiple single cells of the target plant species, a hybrid expert model architecture designed with multimodal-cell type decoupling is used to train the model to learn the characteristics of RNA expression data and chromatin accessibility data of different types of single cells, thus obtaining a plant single-cell gene expression prediction model.

2. The method for constructing a plant single-cell gene expression prediction model according to claim 1, characterized in that, The method involves using RNA expression data and chromatin accessibility data from multiple single cells of a target plant species, employing a hybrid expert model architecture designed with multimodal-cell type decoupling, to train the model to learn the characteristics of RNA expression data and chromatin accessibility data from different types of single cells, thereby obtaining a plant single-cell gene expression prediction model, including: Data decoupling and reconstruction were performed on RNA expression data and chromatin accessibility data of multiple single cells of the target plant species to obtain reconstructed RNA expression data and reconstructed chromatin accessibility data. Genomic features were extracted from the reconstructed RNA expression data and reconstructed chromatin accessibility data to obtain DNA sequence features and chromatin accessibility signal features; DNA sequence features and chromatin accessibility signal features are respectively encoded. For single cells belonging to rare cell types, data augmentation is performed using a synthetic minority class oversampling strategy. For the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features, a learnable feature weighting mechanism and a gated cross-attention mechanism are combined to automatically adjust the contribution ratio of the feature-encoded DNA sequence features and feature-encoded chromatin accessibility signal features according to the chromatin openness intensity, thereby achieving cross-modal dynamic feature fusion.

3. The method for constructing a plant single-cell gene expression prediction model according to claim 2, characterized in that, The process of decoupling and reconstructing RNA expression data and chromatin accessibility data from multiple single cells of the target plant species to obtain reconstructed RNA expression data and reconstructed chromatin accessibility data includes: RNA expression data and chromatin accessibility data of multiple single cells of the target plant species were decoupled and reconstructed to obtain scRNA-seq expression matrix, scATAC-seq open signal matrix, and cell metadata. Among them, a chromosome-sensing segmentation strategy was used to ensure that homologous genes are not distributed across the training set / test set. The process of extracting genomic features from the reconstructed RNA expression data and reconstructed chromatin accessibility data to obtain DNA sequence features and chromatin accessibility signal features includes: For the reconstructed RNA expression data, based on the GTF genome annotation file, the transcription start site of each gene is located, and the X kb intervals upstream and downstream of the transcription start site are dynamically extracted to obtain DNA sequence features; For the reconstructed chromatin accessibility data, the data is segmented according to the same genomic coordinates, and the continuous signal is discretized into the feature vector of nucleosome unit by an adaptive binning algorithm to obtain the chromatin accessibility signal features. The feature encoding of DNA sequence features and chromatin accessibility signal features respectively includes: Four-dimensional one-hot encoding is used to encode DNA sequence features; The chromatin accessibility signal features were normalized using log(1+x).

4. The method for constructing a plant single-cell gene expression prediction model according to claim 3, characterized in that, The hybrid expert model architecture with multimodal-cell type decoupled design includes a dual-path feature encoder, a cross-modal dynamic fusion module, a load-balanced expert system, and a cell type-specific prediction head. The dual-path feature encoder includes a DNA coding branch and a chromatin accessibility coding branch. The DNA coding branch integrates a long sequence processing mechanism, captures k-mer level regulatory patterns through convolutional gating and global recursive operators, supports input sequences up to 5kb, and outputs a 512-dimensional feature vector that covers the semantic information of regulatory elements. The chromatin accessibility coding branch adopts a sparse positional coding Transformer architecture, focuses on open chromatin patterns at the nucleosome scale through local self-attention, and outputs a 256-dimensional feature vector that encodes the synergistic effect of spatial openness and long-range interaction. The cross-modal dynamic fusion module is designed with a gated cross-attention module. It uses DNA sequence features as the query and chromatin accessibility signal features as the key-value pair for cross-modal interaction. The fusion weight α is generated from learnable parameters and satisfies the formula: α=σ(W[DNA embed ;ATAC embed ]), In the formula, DNA embed Indicates DNA sequence characteristics, ATAC embed σ represents the chromatin accessibility signal feature, W represents the sigmoid function, and W represents the weight matrix. The load-balanced expert system is deployed with a dynamic hybrid expert layer, containing multiple dedicated prediction experts equal to the number of cell types. Through a Top-k sparse routing mechanism, only the k most relevant experts are activated for each sample, and an expert load-balancing loss is introduced. L balance =λ·std(expert_utilization), In the formula, L balance λ represents the load balancing loss value, λ represents the loss weight hyperparameter, and std(expert_utilization) represents the standard deviation function of the expert utilization vector. The cell type-specific prediction head is specifically designed to build an independent binary classifier for each cell type, containing a 128-dimensional hidden layer and a sigmoid output unit. This decoupled design supports local fine-tuning in transfer learning scenarios—when new cell type data is added, only the new prediction head needs to be trained while keeping the backbone network frozen.

5. The method for constructing a plant single-cell gene expression prediction model according to claim 4, characterized in that, The plant single-cell gene expression prediction model employs a five-stage cascaded optimization training process. These five stages include multi-task loss design, progressive parameter optimization, dynamic learning rate scheduling, early stopping and model selection, and modal smile experiments. In the multi-task loss design, the main loss function adopts cell type weighted Focal Loss, and the auxiliary loss includes MoE load balancing regularization term and chromatin accessibility feature-DNA sequence feature consistency constraint. Progressive parameter optimization includes: Phase 1: Fix the model architecture and optimize data parameters; Phase Two: Freeze data parameters and search for model hyperparameters; Phase 3: Jointly fine-tune all parameters and evaluate 200+ combinations in parallel using the Optuna framework; The dynamic learning rate scheduling is as follows: a cosine annealing strategy is adopted, with an initial learning rate of 3e-4, which decays by 30% every 10 training rounds. A validation sensitivity monitoring callback is integrated—when the area under the validation ROC curve stagnates, the learning rate is automatically halved. Early stopping and model selection are specifically implemented as follows: an early stopping mechanism is implemented based on the area under the average precision-recall curve of the validation set, and the best-performing checkpoint is saved to the local directory. Each checkpoint contains the complete model state and independent parameters of all prediction heads. The modal ablation experiment specifically involves a built-in single-modal control experiment in the training process to independently evaluate the predictive performance of prediction models based on DNA sequence features or chromatin accessibility signal features.

6. A method for predicting gene expression in single-cell plants, characterized in that, include Obtain RNA expression data and chromatin accessibility data of single cells of the plant species to be tested; Based on the RNA expression data and chromatin accessibility data of single cells of the plant species to be tested, the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model according to any one of claims 1-5 is used to predict the cell type of single cells of the plant species to be tested and the gene expression level of the plant species to be tested.

7. A plant single-cell gene expression prediction system, characterized in that, include: The data acquisition module is used to acquire RNA expression data and chromatin accessibility data of single cells of the plant species to be tested. The prediction module is used to: predict the cell type and gene expression level of the single cells of the plant species under test based on the RNA expression data and chromatin accessibility data of the single cells of the plant species under test, using the plant single cell gene expression prediction model obtained by the construction method of the plant single cell gene expression prediction model according to any one of claims 1-5.

8. A downstream application analysis system, characterized in that, The plant single-cell gene expression prediction model obtained based on the construction method of the plant single-cell gene expression prediction model according to any one of claims 1-5 is implemented, and its functions include any one of the following or any combination thereof: cell expression visualization, regulatory element interpretation, and genetic variation analysis.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for constructing a plant single-cell gene expression prediction model as described in any one of claims 1 to 5, or the method for predicting plant single-cell gene expression as described in claim 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a plant single-cell gene expression prediction model as described in any one of claims 1 to 5, or the method for predicting plant single-cell gene expression as described in claim 6.

Images