Method and device for constructing a structured molecular landscape map, electronic device and medium
By constructing a structured molecular landscape map, an interpretable model is generated using multimodal data. This solves the problem of molecular interactions in single-cell multimodal data, enabling accurate identification of cell types and states and extraction of key molecules, thus improving the understanding of the molecular mechanisms of biological systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FOSHAN UNIVERSITY
- Filing Date
- 2024-02-19
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for integrating single-cell multimodal data struggle to reveal intermolecular interactions, leading to insufficient understanding of cell types and states. Furthermore, existing models lack interpretability, hindering in-depth research into molecular mechanisms in biological systems.
A structured molecular landscape mapping method was adopted. After collecting multimodal data, the data was cleaned and normalized, and then the data was reduced in dimensionality and transformed into images. An interpretable convolutional neural network model was used to generate a structured molecular landscape map. Combined with attribution algorithms and contour maps, key biomic molecules were extracted.
It has enabled the revelation of molecular distribution relationships in cells and the analysis of the contribution of omics molecules. The prediction results are interpretable and support downstream bioinformatics analysis.
Smart Images
Figure CN118016155B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of single-cell omics technology, and in particular to a method for constructing a structured molecular landscape map, a device for constructing a structured molecular landscape map, an electronic device, and a computer-readable medium. Background Technology
[0002] The development of single-cell technologies has generated an increasing amount of single-cell data, providing excellent opportunities for classifying and characterizing the rich cell types and their spatiotemporal organization within biological systems. In particular, the proliferation of multimodal single-cell technologies, encompassing a variety of features including genome sequences, DNA methylation profiles, chromatin accessibility, lineage origin, and even single-cell immunophenotypes, represents a new frontier in cellular biology and biomedical research. For example, multi-omics ATAC (chromatin accessibility expression) + gene expression simultaneously measures cellular gene expression and DNA accessibility. Each approach has enormous potential to overcome the inherent limitations of single-cell scRNA-seq and can be used to explore more comprehensive and in-depth perspectives on cellular states and interactions. Although multimodal genomic measurements are inherently complex due to inter-molecular interactions, multi-omics data integration has become a reality, inspiring new computational methods for information fusion between different high-throughput data types. To date, numerous single-cell multi-omics integration methods have been proposed, along with predictive models for various downstream tasks such as cell clustering or label regression. For instance, UnitedNet, as a representative state-of-the-art multimodal integration method, directly quantifies the relationships between gene expression and other modalities through a unified encoder-decoder framework.
[0003] In biology, intramolecular and intermolecular interactions play a crucial role in various cellular processes and provide a unique molecular basis for distinguishing cell types and states. Clearly, the typical cellular landscape remains largely untapped, making it difficult to discover these interactions within genomic systems. Typically, information about a single cell is represented by a single vector (such as a column in an expression data matrix), with information about molecular interactions hidden within data without direct indication. This hinders a comprehensive understanding of genomic characteristics and interactions and overlooks much relevant information that could aid downstream data analysis. Therefore, in practice, when studying the dynamic changes of complex phenotypes (such as cell lineages or disease-related traits), a single-modality cellular landscape is insufficient to identify phenotype-related feature sets (such as different types of molecular biomarkers) and explain their dynamic functional transitions during phenotypic changes (such as the decisive role of biomarkers in cell state transitions). In other words, these techniques are primarily used to reveal cell distribution within the traditional cellular landscape, including cell subtype identification and pseudotrajectory analysis, but they fall short in extracting feature distributions from the molecular landscape, hindering the understanding of molecular mechanisms at the biological system level. Therefore, there is an urgent need to understand the molecular landscape based on single-cell multimodal datasets. However, integrating multiple tasks into a unified AI recognition framework, xAI, to obtain organized molecular landscapes remains a computational challenge. Furthermore, existing multimodal machine learning models are black boxes, suffering from insufficient interpretability in their design. Data processors cannot see the reasons behind the model's predictions, often leading to skepticism towards the decisions made by deep neural network models. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a method for constructing a structured molecular landscape map, a corresponding apparatus for constructing a structured molecular landscape map, an electronic device, and a computer-readable medium to overcome or at least partially solve the above problems.
[0005] This invention discloses a method for constructing structured molecular landscape maps, the method comprising:
[0006] Collect multimodal data of the organism to be tested; the multimodal data of the organism to be tested includes single-cell omics data and phenomics data;
[0007] The multimodal data of the organism to be tested is cleaned and normalized to obtain standardized multimodal data of the organism to be tested.
[0008] The standardized biological multimodal data to be tested is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape atlas of the biological multimodal data to be tested. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset to be tested. Each point in the basic molecular structure map represents a molecule, the relative position of the points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map indicates the expression value of the molecule in the cell.
[0009] The simulated image dataset to be tested is input into the structured molecular landscape map to construct the model;
[0010] The structured molecular landscape mapping model outputs a structured molecular landscape map of the multimodal data of the organism under test.
[0011] Optionally, the structured molecular landscape mapping model includes an attribution module, and the step of the structured molecular landscape mapping model outputting a structured molecular landscape map of the tested biological multimodal data includes:
[0012] The attribution module uses a gradient-based attribution algorithm to calculate the attribution value of the 3D molecular pseudo-image of each molecule and generate an attribution value map.
[0013] Calculate the contour plot and attribution gradient direction field of the attribution value map;
[0014] The structured molecular landscape map of the biological multimodal data under test is generated using the contour plot and attribution gradient direction field of the attribution value map, as well as the basic molecular structure map.
[0015] Optionally, the structured molecular landscape mapping model further includes a prediction module, and the method further includes:
[0016] The prediction module predicts the simulated image dataset to be tested and outputs the cell type and cell state of the biological multimodal data to be tested.
[0017] Optionally, the structured molecular landscape mapping model further includes a key omics molecule extraction module, and the method further includes:
[0018] The key biomics molecule extraction module extracts a preset number of molecules with the highest to lowest attribution values from the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and outputs the key biomics molecules.
[0019] Optionally, the generation of the structured molecular landscape mapping model includes:
[0020] Collect biological multimodal data; the biological multimodal data includes single-cell omics data and phenomics data;
[0021] Multimodal cell fusion clustering is performed on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and the biological multimodal data is labeled to obtain biological multimodal data with cell type labels and cell state labels;
[0022] The biological multimodal data with cell type labels and cell state labels are cleaned and normalized to obtain standardized biological multimodal data with cell type labels and cell state labels.
[0023] The standardized biological multimodal data with cell type and cell state labels is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set.
[0024] The structured molecular landscape map construction model is obtained by training a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set.
[0025] Optionally, the labels of the simulated image sample set also include attribution value labels, contour map labels of the attribution value map, attribution gradient direction field labels, and key biomics molecular labels. The simulated image sample set is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework to obtain the structured molecular landscape map construction model, including:
[0026] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape map construction model is trained based on the cell type label and the cell state label.
[0027] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the attribution module of the structured molecular landscape map construction model is obtained by training based on the attribution value label, the contour map label of the attribution value map, and the attribution gradient direction field label.
[0028] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the key biomics molecule extraction module of the structured molecular landscape map construction model is obtained by training based on the key biomics molecule labels.
[0029] Optionally, the interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape mapping model is trained based on the cell type label and the cell state label, including:
[0030] The simulated image sample set is divided into a simulated image training set and a simulated image test set. The interpretable convolutional neural network model is used to predict the simulated image training set and output the predicted cell type and cell state.
[0031] The predicted cell type and cell state are used to calculate the loss with the cell type label and cell state label, and the parameters are adjusted using Bayesian optimization techniques based on the loss.
[0032] After iteratively determining the optimal parameters, the simulated image test set is used for testing to obtain the prediction module of the structured molecular landscape map construction model.
[0033] This invention also discloses a structured molecular landscape mapping device, the device comprising:
[0034] The test data acquisition module is used to acquire multimodal data of the test organism; the multimodal data of the test organism includes single-cell omics data and phenomics data;
[0035] The test data standardization module is used to perform data cleaning and normalization on the test biological multimodal data to obtain standardized test biological multimodal data.
[0036] The image-to-test data module is used to perform dimensionality reduction processing on the standardized biological multimodal data and convert it into an image, obtaining the basic molecular structure map of the structured molecular landscape atlas of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset. Each point in the basic molecular structure map represents a molecule, the relative distance between points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map represents the expression value of the molecule in the cell.
[0037] The input module is used to input the simulated image dataset to be tested into the structured molecular landscape map to construct the model;
[0038] The output module is used to output the structured molecular landscape map of the biological multimodal data under test from the structured molecular landscape map construction model.
[0039] Optionally, the structured molecular landscape mapping model includes an attribution module, and the output module includes:
[0040] The attribution value map generation submodule is used by the attribution module to calculate the attribution value of the 3D molecular pseudo-image of each molecule using a gradient-based attribution algorithm, and generate the attribution value map.
[0041] A contour plot and orientation field calculation submodule is used to calculate the contour plot and attribution gradient orientation field of the attribution value map.
[0042] The structured molecular landscape map generation submodule is used to generate a structured molecular landscape map of the biological multimodal data under test by using the contour plot and attribution gradient direction field of the attribution value map and the basic molecular structure map.
[0043] Optionally, the structured molecular landscape mapping model further includes a prediction module, and the device further includes:
[0044] The cell type and state output module is used by the prediction module to predict the simulated image dataset to be tested and output the cell type and cell state of the biological multimodal data to be tested.
[0045] Optionally, the structured molecular landscape mapping model further includes a key biomics molecule extraction module, and the device further includes:
[0046] A bioomics molecular output module is constructed, which is used by the key bioomics molecule extraction module to extract a preset number of molecules with the highest to lowest attribution values as key bioomics molecules based on the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and output the key bioomics molecules.
[0047] Optionally, the device further includes:
[0048] The sample data acquisition module is used to acquire biological multimodal data, which includes single-cell omics data and phenomics data.
[0049] The sample label generation module is used to perform multimodal cell fusion clustering on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and to label the biological multimodal data to obtain biological multimodal data with cell type labels and cell state labels.
[0050] The sample data standardization module is used to perform data cleaning and normalization on the biological multimodal data with cell type labels and cell state labels to obtain standardized biological multimodal data with cell type labels and cell state labels.
[0051] The pseudo-image sample set generation module is used to perform dimensionality reduction processing on the standardized biological multimodal data with cell type labels and cell state labels and convert it into an image to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set.
[0052] The model training module is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set, so as to obtain the structured molecular landscape map construction model.
[0053] Optionally, the labels of the simulated image sample set further include attribution value labels, contour map labels of the attribution value map, attribution gradient direction field labels, and key biological omics molecular labels. The model training module includes:
[0054] The prediction module training submodule is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the prediction module of the structured molecular landscape map construction model according to the cell type label and the cell state label.
[0055] The attribution module training submodule is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the attribution module of the structured molecular landscape map construction model based on the attribution value labels, the contour map labels of the attribution value map, and the attribution gradient direction field labels.
[0056] The training submodule for the key omics molecule extraction module is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the key omics molecule extraction module of the structured molecular landscape map construction model according to the key omics molecule labels.
[0057] Optionally, the prediction module training submodule includes:
[0058] The cell type and state prediction unit is used to divide the simulated image sample set into a simulated image training set and a simulated image test set, and to use the interpretable convolutional neural network model to predict the simulated image training set, and output the predicted cell type and cell state.
[0059] The parameter optimization unit is used to calculate the loss using the predicted cell type and cell state, along with the cell type label and cell state label, and to adjust the parameters using Bayesian optimization techniques based on the loss.
[0060] The testing unit is used to iteratively determine the optimal parameters and then test them using the simulated image test set to obtain the prediction module of the structured molecular landscape map construction model.
[0061] The present invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;
[0062] The memory is used to store computer programs;
[0063] When the processor executes the program stored in the memory, it implements the structured molecular landscape map construction method as described in this invention.
[0064] The present invention also discloses one or more computer-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the structured molecular landscape mapping method as described in the present invention.
[0065] This invention has the following advantages:
[0066] The structured molecular landscape mapping method of this invention involves collecting multimodal data of the target organism, including single-cell omics data and phenomics data. The multimodal data is then cleaned and normalized to obtain standardized multimodal data. This standardized data is then dimensionality-reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape mapping. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image for each molecule, resulting in a pseudo-image dataset. Each point in the basic molecular structure map represents a molecule, the relative distance between points indicates the positional relationship between molecules, and the color of a point in the basic molecular structure map represents the expression value of the molecule in the cell. The pseudo-image dataset is then input into a structured molecular landscape mapping model, which outputs the structured molecular landscape mapping of the target organism's multimodal data. This invention transforms biological multimodal data from vector data into a pseudo-image dataset containing sample information and molecular relationships. This allows a pre-trained structured molecular landscape map construction model to automatically process the pseudo-image dataset and output a structured molecular landscape map of the biological multimodal data. The structured molecular landscape map reveals the correlations in molecular distribution within cells and reflects the contribution of omics molecules, making the prediction results interpretable and laying the foundation for downstream analysis of bioinformatics. Attached Figure Description
[0067] Figure 1This is a flowchart of the steps in a method for constructing a structured molecular landscape map according to an embodiment of the present invention;
[0068] Figure 2 This is the basic molecular structure diagram of the structured molecular landscape map of the biological multimodal data provided in the embodiments of the present invention;
[0069] Figure 3 This is a 3D molecular pseudo-image dataset of biological multimodal data to be tested provided in the embodiments of the present invention;
[0070] Figure 4 This is a structured molecular landscape map of the biological multimodal data to be tested provided in the embodiments of the present invention;
[0071] Figure 5 These are key biomic molecules in the multimodal biological data to be tested provided in the embodiments of the present invention;
[0072] Figure 6 This invention provides a key biomolecular regulatory network constructed using key biomics molecules derived from multimodal biological data of the organism under test.
[0073] Figure 7 This is a structural block diagram of a structured molecular landscape mapping device provided in an embodiment of the present invention;
[0074] Figure 8 This is a block diagram of an electronic device provided in an embodiment of the present invention;
[0075] Figure 9 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation
[0076] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0077] Reference Figure 1 The diagram illustrates a flowchart of a method for constructing a structured molecular landscape map according to an embodiment of the present invention, which may specifically include the following steps:
[0078] Step 101: Collect multimodal data of the organism to be tested; the multimodal data of the organism to be tested includes single-cell omics data and phenomics data;
[0079] This invention constructs an interpretable structured molecular landscape atlas model. This model allows for the determination of the joint definition of cellular states and the distribution of multi-omics molecules within quantified graphic tissues, i.e., a structured molecular landscape atlas. In this invention, the data used to construct the structured molecular landscape atlas can be biological multimodal data. Therefore, the first step is to collect multimodal data of the organism to be tested. This multimodal data can include single-cell omics data and phenomics data. Single-cell omics data and phenomics data can be data from different levels such as genomics, transcriptomics, proteomics, and metabolomics. For example, single-cell omics data can include scRNA-seq gene expression data, methylation expression data, scATAC-seq chromatin accessibility expression data, etc.; phenomics data can include RNA-seq gene expression data, methylation expression data, SNP copy number variation data, etc.
[0080] Among these, the number of molecules in each modality of the biological multimodal data to be tested remains consistent, and there is no limit to the number of samples in the biological multimodal data to be tested. In addition, the basic source of the biological multimodal data to be tested and the training sample data for training the structured molecular landscape map model can be consistent. For example, if the training sample data is liver cancer multi-omics training data, the biological multimodal data to be tested can also be liver cancer multi-omics test data.
[0081] Step 102: Perform data cleaning and normalization on the multimodal data of the organism to be tested to obtain standardized multimodal data of the organism to be tested;
[0082] The collected biological multimodal data can be preprocessed, such as by data cleaning and normalization, to eliminate the adverse effects of outlier data and to standardize the data, eliminating the influence of different dimensions between features, making the data comparable, improving the interpretability of the data, accelerating the convergence speed of the program, and improving the training efficiency and performance of the model.
[0083] Specifically, since molecules expressing zero have no biological significance, the first step is to remove molecules with all zero expressions from the biological multimodal data to be tested. This can be done using Seurat software by setting thresholds or filtering conditions to identify and remove molecular data with all zero expressions, thus eliminating biologically meaningless noise data. Next, the cleaned biological multimodal data can be normalized to obtain standardized biological multimodal data to be tested. Standardized biological multimodal data can be represented as follows: The matrix represents different modes, a, b, c represent the number of molecules corresponding to different modes, and n represents the number of samples to be tested.
[0084] Step 103: The standardized biological multimodal data to be tested is subjected to dimensionality reduction processing and converted into an image to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data to be tested. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining the pseudo-image dataset to be tested. Each point in the basic molecular structure map represents a molecule, the relative position of the points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map represents the expression value of the molecule in the cell.
[0085] After preprocessing the multimodal data of the target organism, further image transformation can be performed on the standardized multimodal data to embed the correlations of omics molecules. Specifically, the standardized multimodal data of the target organism can be dimensionality reduced to two dimensions. Dimensionality reduction methods include PCA, TNSE, KPCA, and UMAP, etc. Then, a two-dimensional graph is constructed from the two-dimensional vectors to obtain the basic molecular structure map (SML). Figure 2 The basic molecular structure map is a two-dimensional molecular clustering map after dimensionality reduction of multimodal data. Each point in the map represents a multimodal molecule, and the relative positions indicate the positional relationships between molecules. Each point in the map represents the position of each molecule in the subsequently constructed molecular landscape map. Next, the expression value of each sample in each modality is matched to the two-dimensional molecular structure map to obtain a 3D molecular pseudo-image for each molecule. Color represents the expression value, resulting in a dataset of pseudo-images to be tested, composed of the 3D molecular pseudo-images of all molecules. Figure 3 This completes the transformation of vector samples into image samples to adapt to the CNN framework and embed the relevant distribution relationships of omics molecules.
[0086] Step 104: Input the simulated image dataset to be tested into the structured molecular landscape map to construct the model;
[0087] Step 105: The structured molecular landscape mapping model outputs a structured molecular landscape map of the multimodal data of the organism to be tested.
[0088] After transforming and processing the biological multimodal data to obtain the simulated image dataset, it can be used as input to a pre-trained structured molecular landscape mapping model. This model then further processes the simulated image data to output a structured molecular landscape map of the biological multimodal data. Figure 4 Structured molecular landscape mapping reveals the correlations in molecular distribution within cells and can reflect the contribution of omics molecules, making the prediction results interpretable and laying the foundation for downstream analysis of bioinformatics.
[0089] In one embodiment of the present invention, the structured molecular landscape mapping model includes an attribution module, and the step of the structured molecular landscape mapping model outputting a structured molecular landscape map of the tested biological multimodal data includes:
[0090] The attribution module uses a gradient-based attribution algorithm to calculate the attribution value of the 3D molecular pseudo-image of each molecule and generate an attribution value map.
[0091] Calculate the contour plot and attribution gradient direction field of the attribution value map;
[0092] The structured molecular landscape map of the biological multimodal data under test is generated using the contour plot and attribution gradient direction field of the attribution value map, as well as the basic molecular structure map.
[0093] The structured molecular landscape mapping (SML) model of this invention can include an attribution module. After inputting the simulated image dataset to be tested into the pre-trained SML model, the attribution module can use a gradient-based attribution algorithm (Grad-CAM) to calculate the attribution value of the 3D molecular simulated image for each molecule. Therefore, each sample's simulated image can obtain a contribution value of each molecule to the prediction, i.e., the attribution value. All molecular attribution values are integrated to generate an attribution value map of the simulated image dataset to be tested. Next, the contour plot and attribution gradient direction field of the attribution value map are calculated. Finally, the basic molecular structure map of the biological multimodal data to be tested, the contour plot of the attribution value map, and the attribution gradient direction field are used to form the structured molecular landscape map (SML) of the biological multimodal data to be tested.
[0094] In one embodiment of the present invention, the structured molecular landscape mapping model further includes a prediction module, and the method further includes:
[0095] The prediction module predicts the simulated image dataset to be tested and outputs the cell type and cell state of the biological multimodal data to be tested.
[0096] The structured molecular landscape mapping model of this invention can also identify cell types and cell states in the multimodal data of the organism to be tested. Specifically, the structured molecular landscape mapping model also includes a prediction module, which can accurately identify cell types and cell states in the multimodal data of the organism to be tested. For example, for single-cell omics data, it can predict that the single cell is a T cell, B cell, epithelial cell, etc.; for phenomics data, it can predict that the cell state of the multimodal data of the organism to be tested is disease, normal, or stage 1, stage 2, stage 3, etc.
[0097] In one embodiment of the present invention, the structured molecular landscape mapping model further includes a key omics molecule extraction module, and the method further includes:
[0098] The key biomics molecule extraction module extracts a preset number of molecules with the highest to lowest attribution values from the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and outputs the key biomics molecules.
[0099] The structured molecular landscape mapping model of this invention can also identify key omics molecules corresponding to the multimodal data of the organism under test. Specifically, the structured molecular landscape mapping model also includes a key omics molecule extraction module. This module can extract key omics molecules with significant attribution values based on the structured molecular landscape map. Specifically, it uses the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism under test. The contour lines in the structured molecular landscape map represent the attribution values, and the arrows represent the gradient direction of the attribution values. Thus, it can trace back to the molecular position in the structured molecular landscape map from the point with high attribution values, extract a preset number of molecules with the highest attribution values as key omics molecules, and output the key omics molecules. Figure 5 .
[0100] After obtaining key biomolecular molecules, structured molecular landscape mapping and analysis of these molecules can be used for downstream bioinformatics analysis. Specifically, regulatory networks of key biomolecules can be constructed using the bioinformatics database STRING, as described above. Figure 6 The bioinformatics database STRING can include commonly used databases in the field of biology such as KEGG, GO, and pathways. These databases contain a wealth of annotation information on the functions and interactions of biomolecules. Key omics molecules can be compared with these databases to identify related pathways and biological processes, and their roles in these pathways and processes can be analyzed, further revealing their functions and significance in biology. Based on the analysis results, key omics molecules can be functionally annotated and classified to better understand their regulatory mechanisms in organisms. This process contributes to a deeper understanding of the functions and mechanisms of action of biomolecules, providing important references for biological research.
[0101] In one embodiment of the present invention, the generation of the structured molecular landscape mapping model includes:
[0102] Collect biological multimodal data; the biological multimodal data includes single-cell omics data and phenomics data;
[0103] Multimodal cell fusion clustering is performed on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and the biological multimodal data is labeled to obtain biological multimodal data with cell type labels and cell state labels;
[0104] The biological multimodal data with cell type labels and cell state labels are cleaned and normalized to obtain standardized biological multimodal data with cell type labels and cell state labels.
[0105] The standardized biological multimodal data with cell type and cell state labels is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set.
[0106] The structured molecular landscape map construction model is obtained by training a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set.
[0107] In one embodiment of the present invention, the labels of the pseudo-image sample set further include attribution value labels, contour map labels of the attribution value map, attribution gradient direction field labels, and key biomics molecular labels. The pseudo-image sample set is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework to obtain the structured molecular landscape map construction model, including:
[0108] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape map construction model is trained based on the cell type label and the cell state label.
[0109] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the attribution module of the structured molecular landscape map construction model is obtained by training based on the attribution value label, the contour map label of the attribution value map, and the attribution gradient direction field label.
[0110] The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the key biomics molecule extraction module of the structured molecular landscape map construction model is obtained by training based on the key biomics molecule labels.
[0111] In one embodiment of the present invention, the interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape mapping model is trained based on the cell type label and the cell state label, including:
[0112] The simulated image sample set is divided into a simulated image training set and a simulated image test set. The interpretable convolutional neural network model is used to predict the simulated image training set and output the predicted cell type and cell state.
[0113] The predicted cell type and cell state are used to calculate the loss with the cell type label and cell state label, and the parameters are adjusted using Bayesian optimization techniques based on the loss.
[0114] After iteratively determining the optimal parameters, the simulated image test set is used for testing to obtain the prediction module of the structured molecular landscape map construction model.
[0115] The structured molecular landscape map construction model of this invention is based on the ResNet50 framework, which is used to build an interpretable convolutional neural network model and train the interpretable convolutional neural network model using biological multimodal data.
[0116] The ResNet50 network architecture contains 49 convolutional layers and one fully connected layer. The ResNet50 network can be divided into seven parts. The first part does not contain residual blocks and mainly performs convolution, regularization, activation function, and max pooling calculations on the input. The second, third, fourth, and fifth parts all contain residual blocks. In the ResNet50 network architecture, each residual block has three convolutional layers, resulting in a total of 1 + 3 × (3 + 4 + 6 + 3) = 49 convolutional layers, plus the final fully connected layer, for a total of 50 layers. The network input is 224 × 224 × 3. After the convolutional calculations in the first five parts, the output is 7 × 7 × 2048. The pooling layer transforms this into a feature vector, and finally, the classifier calculates the class probability from this feature vector. Convolutional Neural Networks (CNNs) have various hyperparameters, such as convolutional layers, filter size, and learning rate. This invention utilizes Bayesian optimization techniques to optimize parameter selection to achieve optimal model performance.
[0117] This invention collects various biological multimodal data. Due to the training of interpretable convolutional neural network models, the biological multimodal data can include single-cell multi-omics data and phenomics data. Among the collected biological multimodal data are unlabeled cells. Multimodal cell fusion clustering can be performed on the biological multimodal data to determine cell types and cell states, and the biological multimodal data can be labeled to obtain biological multimodal data with cell type and cell state labels. Multimodal cell fusion clustering can include UnitedNet or WNN analysis. Furthermore, attribution value labels, contour plot labels of attribution value maps, attribution gradient direction field labels, and key biological omics molecular labels can also be added to the biological multimodal data.
[0118] Next, the biological multimodal data with cell type and cell state labels can be cleaned and normalized to obtain standardized biological multimodal data with cell type and cell state labels. Further, the standardized biological multimodal data with cell type and cell state labels is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape atlas of the biological multimodal data. The expression value of each molecule is then matched to the basic molecular structure map to obtain a 3D molecular pseudo-image for each molecule, resulting in a pseudo-image sample set that can be used to train an interpretable convolutional neural network model.
[0119] An interpretable convolutional neural network model is trained using a quasi-image sample set to obtain a structured molecular landscape atlas construction model, which may include:
[0120] An interpretable convolutional neural network model was trained using a pseudo-image sample set. Based on cell type and cell state labels, a prediction module for the structured molecular landscape mapping model was obtained. Specifically, the pseudo-image sample set was divided into a training set and a test set. The interpretable convolutional neural network model was used to predict the cell type and cell state on the training set. The predicted cell type and cell state were then compared with the cell type and cell state labels to calculate the loss. Based on the loss, Bayesian optimization techniques were used to adjust the parameters. After iteratively determining the optimal parameters, the model was tested using the pseudo-image test set. After testing, the prediction module for the structured molecular landscape mapping model was obtained.
[0121] An interpretable convolutional neural network model was trained using a quasi-image sample set. Based on attribution value labels, contour map labels of the attribution value map, and attribution gradient direction field labels, an attribution module for the structured molecular landscape mapping model was obtained. Specifically, the interpretable convolutional neural network model calculated the attribution values of molecules in the quasi-image training set, outputting the calculated attribution values and attribution value maps. The calculated attribution values were compared with the attribution value labels to calculate the loss, and the gradient-based attribution algorithm was optimized based on the loss. The contour map and attribution gradient direction field of the attribution value map were calculated and output. The calculated attribution value map's contour map and attribution gradient direction field were compared with the attribution value map's contour map labels and attribution gradient direction field labels to calculate the loss, and the gradient-based attribution algorithm and related parameters of the contour map and attribution gradient direction field were optimized based on the loss. After iteratively determining the optimal parameters, the model was tested using a quasi-image test set, and the attribution module for the structured molecular landscape mapping model was obtained after testing.
[0122] An interpretable convolutional neural network model was trained using a pseudo-image sample set, and a key biomics molecule extraction module for the structured molecular landscape mapping model was obtained based on key biomics molecule labels. Specifically, the interpretable convolutional neural network model extracted key biomics molecules from the pseudo-image training set, outputting the extracted key biomics molecules. The loss was calculated using the extracted key biomics molecules and their labels, and the parameters were adjusted using Bayesian optimization techniques based on the loss. After iteratively determining the optimal parameters, the model was tested using a pseudo-image test set. After testing, the key biomics molecule extraction module for the structured molecular landscape mapping model was obtained.
[0123] In this embodiment of the invention, multimodal data of the organism to be tested is collected, including single-cell omics data and phenomics data. The multimodal data of the organism to be tested is cleaned and normalized to obtain standardized multimodal data of the organism to be tested. The standardized multimodal data of the organism to be tested is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape atlas of the multimodal data of the organism to be tested. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset of the organism to be tested. Each point in the basic molecular structure map represents a molecule, the relative position of the points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map indicates the expression value of the molecule in the cell. The pseudo-image dataset of the organism to be tested is input into the structured molecular landscape atlas construction model, and the structured molecular landscape atlas construction model outputs the structured molecular landscape atlas of the multimodal data of the organism to be tested. This invention transforms biological multimodal data from vector data into a pseudo-image dataset containing sample information and molecular relationships. This allows a pre-trained structured molecular landscape map construction model to automatically process the pseudo-image dataset and output a structured molecular landscape map of the biological multimodal data. The structured molecular landscape map reveals the correlations in molecular distribution within cells and reflects the contribution of omics molecules, making the prediction results interpretable and laying the foundation for downstream analysis of bioinformatics.
[0124] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.
[0125] Reference Figure 7 The diagram illustrates a structural block diagram of a structured molecular landscape mapping device provided in an embodiment of the present invention, which may specifically include the following modules:
[0126] The test data acquisition module 701 is used to acquire multimodal data of the test organism; the multimodal data of the test organism includes single-cell omics data and phenomics data;
[0127] The test data standardization module 702 is used to perform data cleaning and normalization processing on the test biological multimodal data to obtain standardized test biological multimodal data;
[0128] The image processing module 703 is used to perform dimensionality reduction processing on the standardized biological multimodal data and convert it into an image to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset. Each point in the basic molecular structure map represents a molecule, the relative position of the points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map indicates the expression value of the molecule in the cell.
[0129] Input module 704 is used to input the simulated image dataset to be tested into the structured molecular landscape map construction model;
[0130] Output module 705 is used to output the structured molecular landscape map of the biological multimodal data under test from the structured molecular landscape map construction model.
[0131] Optionally, the structured molecular landscape mapping model includes an attribution module, and the output module includes:
[0132] The attribution value map generation submodule is used by the attribution module to calculate the attribution value of the 3D molecular pseudo-image of each molecule using a gradient-based attribution algorithm, and generate the attribution value map.
[0133] A contour plot and orientation field calculation submodule is used to calculate the contour plot and attribution gradient orientation field of the attribution value map.
[0134] The structured molecular landscape map generation submodule is used to generate a structured molecular landscape map of the biological multimodal data under test by using the contour plot and attribution gradient direction field of the attribution value map and the basic molecular structure map.
[0135] Optionally, the structured molecular landscape mapping model further includes a prediction module, and the device further includes:
[0136] The cell type and state output module is used by the prediction module to predict the simulated image dataset to be tested and output the cell type and cell state of the biological multimodal data to be tested.
[0137] Optionally, the structured molecular landscape mapping model further includes a key biomics molecule extraction module, and the device further includes:
[0138] A bioomics molecular output module is constructed, which is used by the key bioomics molecule extraction module to extract a preset number of molecules with the highest to lowest attribution values as key bioomics molecules based on the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and output the key bioomics molecules.
[0139] Optionally, the device further includes:
[0140] The sample data acquisition module is used to acquire biological multimodal data, which includes single-cell omics data and phenomics data.
[0141] The sample label generation module is used to perform multimodal cell fusion clustering on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and to label the biological multimodal data to obtain biological multimodal data with cell type labels and cell state labels.
[0142] The sample data standardization module is used to perform data cleaning and normalization on the biological multimodal data with cell type labels and cell state labels to obtain standardized biological multimodal data with cell type labels and cell state labels.
[0143] The pseudo-image sample set generation module is used to perform dimensionality reduction processing on the standardized biological multimodal data with cell type labels and cell state labels and convert it into an image to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set.
[0144] The model training module is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set, so as to obtain the structured molecular landscape map construction model.
[0145] Optionally, the labels of the simulated image sample set further include attribution value labels, contour map labels of the attribution value map, attribution gradient direction field labels, and key biological omics molecular labels. The model training module includes:
[0146] The prediction module training submodule is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the prediction module of the structured molecular landscape map construction model according to the cell type label and the cell state label.
[0147] The attribution module training submodule is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the attribution module of the structured molecular landscape map construction model based on the attribution value labels, the contour map labels of the attribution value map, and the attribution gradient direction field labels.
[0148] The training submodule for the key omics molecule extraction module is used to train the interpretable convolutional neural network model using the pseudo-image sample set, and to train the key omics molecule extraction module of the structured molecular landscape map construction model according to the key omics molecule labels.
[0149] Optionally, the prediction module training submodule includes:
[0150] The cell type and state prediction unit is used to divide the simulated image sample set into a simulated image training set and a simulated image test set, and to use the interpretable convolutional neural network model to predict the simulated image training set, and output the predicted cell type and cell state.
[0151] The parameter optimization unit is used to calculate the loss using the predicted cell type and cell state, along with the cell type label and cell state label, and to adjust the parameters using Bayesian optimization techniques based on the loss.
[0152] The testing unit is used to iteratively determine the optimal parameters and then test them using the simulated image test set to obtain the prediction module of the structured molecular landscape map construction model.
[0153] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.
[0154] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 8 As shown, it includes a processor 801, a communication interface 802, a memory 803, and a communication bus 804. The processor 801, communication interface 802, and memory 803 communicate with each other via the communication bus 804.
[0155] Memory 803 is used to store computer programs;
[0156] The processor 801, when executing the program stored in the memory 803, implements the structured molecular landscape map construction method as described in the above embodiments.
[0157] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.
[0158] The communication interface is used for communication between the aforementioned terminal and other devices.
[0159] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0160] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0161] like Figure 9 As shown, in another embodiment of the present invention, a computer-readable storage medium 901 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the structured molecular landscape map construction method described in the above embodiments.
[0162] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the structured molecular landscape map construction method described in the above embodiments.
[0163] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).
[0164] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0165] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0166] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A method of constructing a structured molecular landscape map, characterized by, The method includes: Collect multimodal data of the organism to be tested; the multimodal data of the organism to be tested includes single-cell omics data and phenomics data; The multimodal data of the organism to be tested is cleaned and normalized to obtain standardized multimodal data of the organism to be tested. The standardized biological multimodal data to be tested is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape atlas of the biological multimodal data to be tested. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset to be tested. Each point in the basic molecular structure map represents a molecule, the relative position of the points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map indicates the expression value of the molecule in the cell. The simulated image dataset to be tested is input into the structured molecular landscape map to construct the model; The structured molecular landscape mapping model outputs a structured molecular landscape map of the multimodal data of the organism under test. The structured molecular landscape mapping model includes an attribution module. The steps of the structured molecular landscape mapping model outputting a structured molecular landscape map of the tested biological multimodal data include: The attribution module uses a gradient-based attribution algorithm to calculate the attribution value of the 3D molecular pseudo-image of each molecule and generate an attribution value map. Calculate the contour plot and attribution gradient direction field of the attribution value map; The structured molecular landscape map of the biological multimodal data under test is generated using the contour plot and attribution gradient direction field of the attribution value map, as well as the basic molecular structure map. The structured molecular landscape mapping model also includes a key omics molecule extraction module, and the method further includes: The key biomics molecule extraction module extracts a preset number of molecules with the highest to lowest attribution values from the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and outputs the key biomics molecules. The generation of the structured molecular landscape map construction model includes: Collect biological multimodal data; the biological multimodal data includes single-cell omics data and phenomics data; Multimodal cell fusion clustering is performed on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and the biological multimodal data is labeled to obtain biological multimodal data with cell type labels and cell state labels; The biological multimodal data with cell type labels and cell state labels are cleaned and normalized to obtain standardized biological multimodal data with cell type labels and cell state labels. The standardized biological multimodal data with cell type and cell state labels is dimensionality reduced and converted into images to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set. The structured molecular landscape map construction model is obtained by training a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set.
2. The method of claim 1, wherein, The structured molecular landscape mapping model also includes a prediction module, and the method further includes: The prediction module predicts the simulated image dataset to be tested and outputs the cell type and cell state of the biological multimodal data to be tested.
3. The method according to claim 1, characterized in that, The labels of the simulated image sample set also include attribution value labels, contour map labels of the attribution value map, attribution gradient direction field labels, and key biomics molecular labels. The simulated image sample set is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework to obtain the structured molecular landscape map construction model, including: The interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape map construction model is trained based on the cell type label and the cell state label. The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the attribution module of the structured molecular landscape map construction model is obtained by training based on the attribution value label, the contour map label of the attribution value map, and the attribution gradient direction field label. The interpretable convolutional neural network model is trained using the pseudo-image sample set, and the key biomics molecule extraction module of the structured molecular landscape map construction model is obtained by training based on the key biomics molecule labels.
4. The method of claim 3, wherein, The interpretable convolutional neural network model is trained using the pseudo-image sample set, and a prediction module for the structured molecular landscape map construction model is trained based on the cell type label and the cell state label, including: The simulated image sample set is divided into a simulated image training set and a simulated image test set. The interpretable convolutional neural network model is used to predict the simulated image training set and output the predicted cell type and cell state. The predicted cell type and cell state are used to calculate the loss with the cell type label and cell state label, and the parameters are adjusted using Bayesian optimization techniques based on the loss. After iteratively determining the optimal parameters, the simulated image test set is used for testing to obtain the prediction module of the structured molecular landscape map construction model.
5. A structured molecular landscape mapping device, characterized by, The device includes: The test data acquisition module is used to acquire multimodal data of the test organism; the multimodal data of the test organism includes single-cell omics data and phenomics data; The test data standardization module is used to perform data cleaning and normalization on the test biological multimodal data to obtain standardized test biological multimodal data. The image-to-test data module is used to perform dimensionality reduction processing on the standardized biological multimodal data and convert it into an image, obtaining the basic molecular structure map of the structured molecular landscape atlas of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, resulting in a pseudo-image dataset. Each point in the basic molecular structure map represents a molecule, the relative distance between points indicates the relative positional relationship between molecules, and the color of the point in the basic molecular structure map represents the expression value of the molecule in the cell. The input module is used to input the simulated image dataset to be tested into the structured molecular landscape map to construct the model; The output module is used to output the structured molecular landscape map of the biological multimodal data to be tested from the structured molecular landscape map construction model. The structured molecular landscape mapping model includes an attribution module, and the output module includes: The attribution value map generation submodule is used by the attribution module to calculate the attribution value of the 3D molecular pseudo-image of each molecule using a gradient-based attribution algorithm, and generate the attribution value map. A contour plot and orientation field calculation submodule is used to calculate the contour plot and attribution gradient orientation field of the attribution value map. The structured molecular landscape map generation submodule is used to generate a structured molecular landscape map of the biological multimodal data under test by using the contour plot and attribution gradient direction field of the attribution value map and the basic molecular structure map. The structured molecular landscape mapping model also includes a key omics molecule extraction module, and the device further includes: A bioomics molecular output module is constructed, which is used by the key bioomics molecule extraction module to extract a preset number of molecules with the highest to lowest attribution values as key bioomics molecules based on the contour plot and attribution gradient direction field in the structured molecular landscape map of the multimodal data of the organism to be tested, and output the key bioomics molecules. The device further includes: The sample data acquisition module is used to acquire biological multimodal data, which includes single-cell omics data and phenomics data. The sample label generation module is used to perform multimodal cell fusion clustering on the biological multimodal data to determine the cell type and cell state of the biological multimodal data, and to label the biological multimodal data to obtain biological multimodal data with cell type labels and cell state labels. The sample data standardization module is used to perform data cleaning and normalization on the biological multimodal data with cell type labels and cell state labels to obtain standardized biological multimodal data with cell type labels and cell state labels. The pseudo-image sample set generation module is used to perform dimensionality reduction processing on the standardized biological multimodal data with cell type labels and cell state labels and convert it into an image to obtain the basic molecular structure map of the structured molecular landscape map of the biological multimodal data. The expression value of each molecule is matched to the basic molecular structure map to obtain a 3D molecular pseudo-image of each molecule, thus obtaining a pseudo-image sample set. The model training module is used to train a pre-built interpretable convolutional neural network model based on the ResNet50 framework using the simulated image sample set, so as to obtain the structured molecular landscape map construction model.
6. An electronic device, comprising: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the structured molecular landscape map construction method as described in any one of claims 1-4.
7. One or more computer-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method for constructing structured molecular landscape maps as described in any one of claims 1-4.