Method and system for constructing metabolite and phenotype relation model and storage medium

Abstract

The invention provides a method and system for constructing a metabolite and phenotype relation model and a storage medium, and the method comprises the steps: obtaining original cell morphological characteristic data, carrying out the preprocessing and batch effect correction of the original cell morphological characteristic data, and obtaining corrected cell morphological characteristic data; carrying out dimension reduction processing and clustering analysis on the corrected cell morphological characteristic data to obtain a plurality of cell phenotype clusters; constructing a reference space, projecting all cell data in the corrected cell morphological characteristic data to the reference space to obtain a projection result, and calculating a lever score of each cell based on the projection result; according to the cell phenotype clusters and the lever scores, constructing an incidence relation map between the metabolites and the cell phenotype clusters; and constructing an interpretable metabolite and phenotype prediction model based on the cell phenotype cluster, the corrected cell morphological characteristic data and the incidence relation map. According to the method, the specific influence of the metabolite on the cells can be accurately analyzed, and the reliable and predictable metabolite and phenotype relation model can be constructed.

Description

Technical Field

[0001] This application relates to the field of computer-aided bioinformatics technology, and in particular to a method, system, and storage medium for constructing a model of the relationship between metabolites and phenotypes. Background Technology

[0002] With the rapid development of high-content screening technologies, researchers are able to acquire a wealth of morphological data on cells after metabolite treatment using automated imaging systems. This high-dimensional data contains crucial information about the effects of metabolites on cell phenotype, serving as an important foundation for understanding the biological functions of metabolites and advancing drug discovery and toxicity assessment.

[0003] In existing technologies, to extract the association between metabolites and cell phenotypes from the massive amounts of data mentioned above, phenotypic analysis is usually performed using cell morphology data, such as analyzing cell morphological characteristics through simple clustering methods. However, existing technologies have significant shortcomings in practical applications. For example, due to the high dimensionality of cell morphology data, the technical variation caused by experimental batch effects, and the lack of quantitative indicators and systematic association analysis methods for cells deviating from normal states, it is difficult to accurately and efficiently extract the specific effects of metabolites on cells from the data. It is also impossible to construct a metabolite-phenotype relationship model that is both controllable and predictable, thus limiting the application value of high-content screening technology in related fields. Summary of the Invention

[0004] To address the technical challenge of accurately analyzing the specific effects of metabolites on cells from massive amounts of high-dimensional cell morphology data exhibiting batch effects and constructing reliable and predictable models of the relationship between metabolites and phenotypes, embodiments of this application provide a method, system, and storage medium for constructing such models.

[0005] In a first aspect, this embodiment provides a method for constructing a model of the relationship between metabolites and phenotypes, the method comprising: Obtain raw cell morphology feature data, and perform preprocessing and batch effect correction on the raw cell morphology feature data to obtain corrected cell morphology feature data; The corrected cell morphology feature data is subjected to dimensionality reduction processing to obtain a low-dimensional cell morphology feature representation, and cluster analysis is performed on the low-dimensional cell morphology feature representation to identify multiple cell phenotype clusters; Based on the control group data in the corrected cell morphology feature data, a reference space representing the normal cell state is constructed. All cell data in the corrected cell morphology feature data are projected onto the reference space to obtain the projection result. Based on the projection result, the leverage score of each cell is calculated. A correlation map between metabolites and cell phenotype clusters is constructed based on the cell phenotype clusters and the leverage scores; A supervised classification model is trained based on the cell phenotype clusters and the corrected cell morphology feature data. An interpretable metabolite and phenotype prediction model is then constructed based on the association map and the supervised classification model.

[0006] In some embodiments, the preprocessing and batch correction of the original cell morphology feature data to obtain corrected cell morphology feature data includes: The original cell morphology feature data is filtered to remove features with variance below a preset threshold to obtain filtered cell morphology feature data. The filtered cell morphology data are subjected to in-plate normalization to eliminate in-plate position effects in order to obtain in-plate cell morphology data. The cell morphology feature data within the plate are globally standardized to ensure that each feature has zero mean and unit variance, thus obtaining global cell morphology feature data. The global cell morphology feature data is subjected to inverse normal transformation to make the distribution of each feature normal to obtain normal cell morphology feature data; The empirical Bayesian correction algorithm is applied to perform batch effect correction on the normal cell morphology feature data to obtain the corrected cell morphology feature data.

[0007] In some embodiments, the dimensionality reduction processing of the corrected cell morphology data to obtain a low-dimensional cell morphology feature representation includes: Principal component analysis was performed on the corrected cell morphology data to extract the first N principal components as a preliminary dimensionality-reduced representation of cell morphology features. The preliminary dimensionality-reduced cell morphology feature representation is input into a nonlinear dimensionality reduction model. By constructing a similarity graph between cells and performing diffusion mapping, a low-dimensional cell morphology feature representation is obtained.

[0008] In some embodiments, constructing a reference space characterizing the normal cell state based on control group data in the corrected cell morphology feature data includes: Extract the control group data that has not been treated with metabolites from the corrected cell morphology feature data, and perform singular value decomposition on the control group data to obtain the principal component direction matrix that characterizes the core features of normal cell morphology. The core feature dimensions and distribution range of normal cell morphology are determined based on the principal component orientation matrix. A reference space representing the normal cell state is constructed by using the core feature dimensions and distribution range.

[0009] In some embodiments, the clustering analysis of the low-dimensional cell morphological feature representation to identify multiple cell phenotypic clusters includes: Based on the low-dimensional cell morphology features, a cell similarity map is constructed, wherein the cell similarity map is used to represent the adjacency relationship between cells; A community detection algorithm is performed on the cell similarity map to identify multiple cell phenotype clusters.

[0010] In some embodiments, calculating the leverage score for each cell based on the projection results includes: The squared vector of each cell projection vector in the projection result is obtained by squaring the projection vector of each cell. The square norm of each cell projection vector is obtained by summing the elements of the squared vector of each cell projection vector, and the square norm of each cell projection vector is used as the leverage score of the corresponding cell.

[0011] In some embodiments, constructing an association map between metabolites and cell phenotype clusters based on the cell phenotype clusters and the leverage score includes: The leverage scores of all cells treated with the same metabolite are aggregated and calculated based on the leverage score of each cell to obtain the average perturbation intensity of each metabolite. The average perturbation intensity of each metabolite is correlated with its corresponding cell phenotype cluster to form a relation matrix with metabolites as rows and cell phenotype clusters as columns; Normalization is performed on each element in the relation matrix to obtain the normalized perturbation intensity of each metabolite in each cell phenotype cluster; Based on the standardized perturbation intensity, a correlation map between metabolites and cell phenotype clusters is generated.

[0012] In some embodiments, the method further includes: The system receives cell image data processed with unknown metabolites. After the same preprocessing and batch correction, the cell image data processed with unknown metabolites is input into the supervised classification model to predict the cell phenotypic clusters that it will trigger with the highest probability. Based on the association map, the system infers the inference results of the cell image data processed with unknown metabolites.

[0013] Secondly, this embodiment provides a system for constructing a model of the relationship between metabolites and phenotypes. The system includes: a preprocessing module, a dimensionality reduction and clustering module, a data calculation module, and a model construction module; wherein, The preprocessing module is used to acquire raw cell morphology feature data, and to preprocess and perform batch effect correction on the raw cell morphology feature data to obtain corrected cell morphology feature data. The dimensionality reduction clustering module is used to perform dimensionality reduction processing on the corrected cell morphology feature data to obtain a low-dimensional cell morphology feature representation, and to perform cluster analysis on the low-dimensional cell morphology feature representation to identify multiple cell phenotype clusters. The data calculation module is used to construct a reference space representing the normal cell state based on the control group data in the corrected cell morphology feature data, project all cell data in the corrected cell morphology feature data onto the reference space to obtain the projection result, and calculate the leverage score of each cell based on the projection result. The model building module is used to construct a correlation map between metabolites and cell phenotype clusters based on the cell phenotype clusters and the leverage score, train a supervised classification model based on the cell phenotype clusters and the corrected cell morphology feature data, and construct an interpretable metabolite and phenotype prediction model based on the correlation map and the supervised classification model.

[0014] Thirdly, this embodiment provides a computer-readable storage medium having a computer program stored thereon that can run on a processor, wherein when the computer program is executed by the processor, it implements a method for constructing a metabolite-phenotype relationship model as described in the first aspect.

[0015] By employing the above method, this application obtains raw cell morphology feature data, preprocesses and corrects for batch effects to obtain corrected cell morphology feature data. Dimensionality reduction is performed on the corrected cell morphology feature data to obtain a low-dimensional cell morphology feature representation, and cluster analysis is performed on this low-dimensional representation to identify multiple cell phenotype clusters. A reference space representing the normal cell state is constructed based on control group data from the corrected cell morphology feature data. All cell data from the corrected cell morphology feature data are projected onto the reference space to obtain the projection result, and a leverage score is calculated for each cell based on the projection result. A correlation map between metabolites and cell phenotype clusters is constructed based on the cell phenotype clusters and leverage scores. A supervised classification model is trained based on the cell phenotype clusters and the corrected cell morphology feature data. An interpretable metabolite-phenotype prediction model is constructed based on the correlation map and the supervised classification model. This addresses the technical problem of accurately analyzing the specific effects of metabolites on cells and constructing a reliable and predictable metabolite-phenotype relationship model from massive amounts of high-dimensional cell morphology data with batch effects. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for constructing a model of the relationship between metabolites and phenotypes provided in this application.

[0017] Figure 2This is a flowchart of a method provided in this application for preprocessing and batch correcting raw morphological feature data to obtain corrected cell morphological feature data.

[0018] Figure 3 This is a flowchart of the method for calculating the leverage score of each cell based on the projection results provided in this application.

[0019] Figure 4 This is a schematic diagram of the overall process for constructing a model of the relationship between metabolites and phenotypes provided in this application.

[0020] Figure 5 This is a schematic diagram of the system connection for constructing a model of the relationship between metabolites and phenotypes provided in this application. Detailed Implementation

[0021] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but is consistent with the broadest scope claimed in this application.

[0022] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0023] The purpose of this application is to achieve automated analysis from cell imaging data to metabolite and phenotypic profiles. By data preprocessing, batch effect elimination, dimensionality reduction clustering, leverage score calculation, and classifier construction, it solves the problems of residual batch effects, inaccurate clustering, and difficulty in quantifying metabolite effects in the prior art. Figure 1 This is a flowchart illustrating a method for constructing a model of the relationship between metabolites and phenotypes, as provided in this application. Figure 1 As shown, a method for constructing a model of the relationship between metabolites and phenotypes includes the following steps: Step S100: Obtain the original cell morphology feature data, and perform preprocessing and batch effect correction on the original cell morphology feature data to obtain the corrected cell morphology feature data.

[0024] This application describes the process from the processing end. The raw cell morphology data comes from an automated imaging system used in high-content screening experiments. Multichannel fluorescence imaging was performed on cells treated with metabolites or controls, and image analysis software, such as CellProfiler, was used to segment cells and extract features from each image, yielding multidimensional features including morphology, texture, intensity, and substructure. These features are organized in a matrix, with each row representing a cell and each column representing a morphological feature, along with metadata such as metabolite identification, experimental plate, and batch. Obtaining the raw cell morphology data in this way provides a structured, high-dimensional input data foundation for subsequent preprocessing and model construction, ensuring the integrity of biological information and automating the analysis process.

[0025] In the overall process of constructing a model of the relationship between metabolites and phenotypes, data preprocessing and batch effect correction are key prerequisites for ensuring the accuracy and reproducibility of subsequent analyses. Figure 2 This is a flowchart illustrating the method provided in this application for preprocessing and batch correcting raw morphological feature data to obtain corrected cell morphological feature data. (Example:) Figure 2 As shown, the steps for preprocessing and batch correction of the original cell morphology data to obtain corrected cell morphology data include: Step S101: Perform feature filtering on the original data morphology feature data to eliminate features with variance below a preset threshold to obtain filtered cell morphology feature data.

[0026] Step S102: Perform intraplate normalization on the filtered cell morphology feature data to eliminate intraplate position effects and obtain intraplate cell morphology feature data.

[0027] Step S103: Perform global standardization on the cell morphology feature data within the plate to make each feature have zero mean and unit variance to obtain global cell morphology feature data.

[0028] Step S104: Perform inverse normal transformation on the global cell morphology feature data to make the distribution of each feature normal to obtain normal cell morphology feature data.

[0029] Step S105: Apply the empirical Bayesian correction algorithm to perform batch effect correction on the normal cell morphology feature data to obtain the corrected cell morphology feature data.

[0030] Through a series of structured and hierarchical data processing steps, technical variations and noise interference are eliminated, thus providing high-quality input data for subsequent dimensionality reduction, clustering, and modeling. The first step is feature filtering. The aforementioned preset threshold is a configurable parameter, typically set based on data distribution characteristics. For example, it can be set as a certain proportion of the median variance of all features, or the inflection point value can be selected by observing the variance distribution histogram. Feature filtering refers to removing features with minimal variation in expression across all cells from the original cell morphology feature matrix to avoid noise interference. Specifically, the variance of each column in the matrix is ​​calculated, i.e., the variance of each morphological feature in the original data matrix is ​​calculated. The calculated variance of each feature is compared with the preset threshold. All feature columns with variances below the preset threshold are removed from the matrix. This process effectively eliminates features with extremely low information content or even those that are merely measurement noise, achieving data dimensionality reduction and reducing computational load and noise interference for subsequent steps.

[0031] Then, in-plate standardization was performed. The aforementioned in-plate position effect refers to the systematic technical bias in high-content screening experiments caused by environmental differences between the edges and centers of the microplate or systematic errors in sample loading, which is unrelated to the cell biological state. Specifically, for each experimental plate, a linear model was fitted using the bscore algorithm. This model represents the value of each feature as a function of the row and column positions within the plate and estimates the offset of the position effect. Subtracting this offset from the original feature values ​​yielded the corrected in-plate cell morphology feature data, thereby significantly reducing non-biological variation introduced by the experimental plate layout and improving data consistency and comparability.

[0032] Next, global standardization is performed. This global standardization refers to applying z-score standardization to the cell morphology feature data after in-plate standardization, ensuring that each feature has zero mean and unit variance across all cells. Specifically, for each feature column, its mean and standard deviation are calculated across the entire dataset. The mean is subtracted from each value of the feature and divided by the standard deviation to obtain the global cell morphology feature data. This eliminates differences in units and numerical ranges between different features, ensuring that each feature has equal weight in subsequent analysis and summarization, and preventing certain large numerical features from dominating model construction.

[0033] Next, an inverse normal transformation is performed. This inverse normal transformation is a data transformation method designed to convert eigenvalues ​​from an arbitrary distribution into those that approximately follow a standard normal distribution. Normal cell morphology data refers to a data matrix where, after the inverse normal transformation, the distribution of each feature is closer to a normal distribution. Specifically, an inverse normal transformation is performed individually on each feature column, converting the feature value into quantiles. These quantiles are then mapped to the quantiles corresponding to the standard normal distribution, resulting in transformed feature values. This makes the data distribution more symmetrical, reduces the impact of skewness and outliers on subsequent statistical models, and improves the robustness of batch effect correction.

[0034] Finally, batch effect correction is performed. The aforementioned empirical Bayesian correction algorithm refers to using methods such as Combat, based on a Bayesian framework, to correct for batch effects in the data. Batch effect correction involves estimating and removing systematic technical variations introduced by different experimental batches. Specifically, normal cell morphology feature data and corresponding batch labels are used as input. The Combat algorithm is applied, which first estimates the mean and variance batch effect parameters for each feature in each batch, then uses empirical Bayesian methods to shrink these parameters, and finally removes the estimated batch effects from the data, outputting corrected cell morphology feature data. This significantly reduces the technical differences between batches, preserves the true biological signals, and provides a high-quality and reproducible data foundation for subsequent dimensionality reduction, clustering, and modeling.

[0035] This hierarchical and systematic data preprocessing and batch effect correction process not only gradually eliminates various technical interferences such as low variance noise, intra-plate positional errors, scale inconsistencies, distribution skewness, and batch variation, but also ensures a comprehensive improvement in data quality through structured operations. Compared to the single or fragmented correction methods in existing technologies, this method significantly enhances data comparability and model repeatability while preserving biological signals. It provides a high-quality, low-noise input foundation for subsequent dimensionality reduction, clustering, and modeling, thereby indirectly improving the overall accuracy and robustness of the metabolite-phenotype relationship model.

[0036] Step S200: Dimensionality reduction is performed on the corrected cell morphology feature data to obtain a low-dimensional cell morphology feature representation, and cluster analysis is performed on the low-dimensional cell morphology feature representation to identify multiple cell phenotype clusters.

[0037] After obtaining high-quality corrected cell morphology feature data as described above, the next step is dimensionality reduction and cell phenotype cluster identification. The core objective of this stage is to transform the high-dimensional corrected cell morphology feature data into low-dimensional identifiers that can clearly reveal the intrinsic structure of the cell state, and on this basis, identify biologically significant cell phenotype clusters. The dimensionality reduction process for obtaining low-dimensional cell morphology feature representations from the corrected cell morphology feature data includes the following steps: Step S201: Perform principal component analysis on the corrected cell morphology data to extract the first N principal components as a preliminary dimensionality reduction representation of cell morphology features.

[0038] Step S202: Input the preliminary dimensionality reduction cell morphology feature representation into the nonlinear dimensionality reduction model, and obtain the low-dimensional cell morphology feature representation by constructing a similarity graph between cells and performing diffusion mapping.

[0039] Principal component analysis (PCA) is a linear dimensionality reduction method that aims to project high-dimensional data onto a new orthogonal basis, maximizing the variance of the projected data. The first N principal components refer to the N principal component directions sorted from largest to smallest variance contribution rate, where N is a preset positive integer determined by whether the cumulative variance contribution rate reaches a preset threshold. The preliminary dimensionality reduction cell morphology feature representation refers to the low-dimensional data representation obtained by projecting the original high-dimensional, corrected cell morphology feature data onto the first N principal component directions. Specifically, the corrected cell morphology feature data matrix is ​​first centered, i.e., each feature is subtracted from its mean. Then, the covariance matrix of this centered matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. Next, the eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the first N eigenvalues ​​are selected as the principal component directions. Finally, the original corrected cell morphology feature data is projected onto the directions of these N principal components to obtain a preliminary dimensionality-reduced cell morphology feature representation. This achieves preliminary dimensionality reduction and noise reduction of high-dimensional data, while preserving the main variation information in the data, providing a simplified structure and reduced noise input basis for subsequent nonlinear dimensionality reduction.

[0040] The aforementioned nonlinear dimensionality reduction model refers to a dimensionality reduction method capable of capturing the nonlinear structure in data; in this application, diffusion mapping is preferred. The cell similarity graph refers to a graph structure constructed with cells as nodes and cell similarity as edge weights. The similarity is typically calculated based on the Euclidean or cosine distance between cells in the initial dimensionality-reduced cell morphology feature representation, and transformed into similarity weights using a Gaussian kernel function. Diffusion mapping is a nonlinear dimensionality reduction technique based on a random walk process, revealing the inherent geometric results of the data by simulating the diffusion process on the similarity graph. The low-dimensional cell morphology feature representation refers to a low-dimensional embedded coordinate matrix obtained after nonlinear dimensionality reduction, which better preserves the continuous changes in cell states and branching structures. Each row represents the coordinates of a cell in the low-dimensional space. Specifically, based on the initial dimensionality-reduced cell morphology feature representation, the similarity between each pair of cells is calculated, and a fully connected similarity matrix is ​​constructed. This similarity matrix is ​​then normalized to obtain a transition probability matrix, which defines the probability of performing a random walk on the graph. Next, eigenvalue decomposition is performed on the transition probability matrix, selecting the top K non-trivial eigenvectors and sorting them according to their eigenvalues. Here, K is a preset low-dimensional dimension, which can be set to 2 or 3. Finally, these eigenvectors are combined row-wise to form a low-dimensional representation of cell morphology features. This step not only further reduces the data dimensionality but also effectively reveals the continuous transitions and branching relationships between cell states, overcoming the limitations of linear dimensionality reduction in processing complex biological data. This approach first removes noise and retains the main linear structure through principal component analysis, and then captures complex nonlinear relationships in the data through diffusion mapping, thereby maximizing the preservation of biological continuity and state branching information in cell morphology data while reducing dimensionality. Compared to using linear or nonlinear dimensionality reduction methods alone, this method can more accurately and robustly reveal the potential structure of cell phenotypes, providing a high-quality, low-dimensional, and structure-preserving data representation foundation for subsequent cluster analysis to identify biologically significant cell phenotype clusters, thus indirectly improving the interpretability and predictive accuracy of the metabolite-phenotype relationship model.

[0041] After obtaining a low-dimensional cell morphology feature representation that clearly reveals the intrinsic structure and continuous changes of cell states through the above steps, the next step is to identify cell phenotypic clusters. The core objective of this stage is to accurately identify biologically significant, discrete cell phenotypic clusters based on the low-dimensional cell morphology feature representation by constructing an inter-cell similarity graph and executing a community detection algorithm, thus laying the foundation for subsequent construction of metabolite-phenotypic associations. The clustering analysis of the low-dimensional cell morphology feature representation to identify multiple cell phenotypic clusters includes the following steps: Step S203: Construct a cell similarity map based on low-dimensional cell morphology features, wherein the cell similarity map is used to represent the adjacency relationship between cells.

[0042] Step S204: Perform a community detection algorithm on the cell similarity map to identify multiple cell phenotypic clusters.

[0043] The aforementioned inter-cell similarity graph is a graph data structure where nodes represent individual cells and edges represent adjacency relationships between cells. These adjacency relationships are determined based on the relative positions of cells in a low-dimensional morphological feature representation space. Specifically, firstly, based on the low-dimensional morphological feature representation, the Euclidean distance between each pair of cells is calculated to quantify the morphological similarity between cells. Then, the K-nearest neighbor algorithm is applied; for each cell, the M cells with the smallest Euclidean distance are selected as its nearest neighbors, where M is a preset positive integer. Next, an adjacency matrix is ​​constructed based on this nearest neighbor relationship. If two cells are neighbors, the corresponding position in the nearest neighbor matrix is ​​set to 1, or a distance-based weight is assigned, such as distance similarity transformed using a Gaussian kernel function; otherwise, it is set to 0. M can be set through cross-validation or empirically. Finally, the graph structure defined by this adjacency matrix is ​​the inter-cell similarity graph, effectively capturing local adjacency relationships based on morphological features in a cell population and converting continuous low-dimensional embeddings into a discrete graph structure, providing input for subsequent community detection.

[0044] Community detection algorithms are algorithms used to identify sets of nodes with tightly connected internal connections and sparse external connections in a graph structure. This application preferably uses the Leiden algorithm. A cell phenotypic cluster refers to each community identified on a cell similarity graph by the community detection algorithm. Cells within each community have similar morphological features, representing a potential, biologically significant cell state or phenotype. Specifically, the inter-cell similarity graph constructed in step S203 is used as input, where the graph includes its nodes and edges, and the Leiden community detection algorithm is run. This algorithm iteratively searches for community structures by optimizing the modularity of the graph. The process includes locally moving nodes to quickly improve modularity, and community aggregation and network refinement to further optimize the partitioning results. The algorithm includes a resolution parameter, which controls the granularity of the identified communities. The optimal value can be selected through grid search combined with modularity evaluation. The algorithm outputs the community label to which each cell node belongs, dividing the cells into multiple mutually exclusive sets. These sets are the identified cell phenotypic clusters. This graph-based clustering method effectively identifies discrete, biologically consistent cell phenotypic clusters while preserving a low-dimensional representation of continuous cell state changes. It better adapts to complex manifold structures in the data, identifies non-spherical clusters, and exhibits stronger robustness to noise and parameter selection. Furthermore, by constructing a cell similarity map and executing a community detection algorithm, the accuracy and biological interpretability of cell phenotypic cluster identification are ensured. This provides a reliable cell state classification foundation for the subsequent accurate construction of metabolite-phenotype association maps, thereby indirectly improving the overall accuracy and predictive reliability of the final metabolite-phenotype relationship model.

[0045] Step S300: Construct a reference space representing the normal cell state based on the control group data in the corrected cell morphology feature data, project all cell data in the corrected cell morphology feature data onto the reference space to obtain the projection result, and calculate the leverage score of each cell based on the projection result.

[0046] After obtaining high-quality corrected cell morphology data and completing dimensionality reduction and cell phenotypic cluster identification, the crucial stage of quantifying the effects of metabolite perturbation begins. The core objective of this stage is to construct a reference space that characterizes the normal cell state and, based on this space, quantify the degree of morphological deviation of each cell caused by metabolite treatment, thereby providing a quantifiable perturbation intensity index for the subsequent accurate construction of the association between metabolites and phenotypes. The construction of the reference space characterizing the normal cell state based on control group data from the corrected cell morphology data includes the following steps: Step S301: Extract the control group data that has not been treated with metabolites from the corrected cell morphology feature data, and perform singular value decomposition on the control data to obtain the principal component direction matrix that represents the core features of normal cell morphology.

[0047] Step S302: Determine the core feature dimensions and distribution range of normal cell morphology based on the principal component orientation matrix.

[0048] Step S303: A reference space representing the normal cell state is constructed by using core feature dimensions and distribution range.

[0049] The aforementioned control group data without metabolite treatment refers to the subset of morphological feature data in the corrected cell morphology feature data that identifies cell samples that have not undergone any metabolite treatment. This represents the normal cell morphology state under experimental conditions without disturbance. Singular value decomposition (SVD) is a matrix factorization method designed to decompose the control group data matrix into the product of three matrices, thereby revealing its inherent linear structure. The principal component direction matrix refers to the right singular vector matrix obtained through SVD; its column vectors define the main directions of variation in the control group data, representing the core pattern of normal cell morphological variation. Specifically, from the corrected cell morphology feature data, all cell samples belonging to the control group are selected based on the metabolite treatment identifier, forming a control group data matrix with rows equal to the number of control group cells and columns equal to the number of morphological features. This control group data matrix is ​​then centered, i.e., each feature column is subtracted from its mean in the control group. Singular value decomposition is then performed on the centered control group data matrix to obtain the left singular vector matrix, the singular value matrix, and the right singular vector matrix. The first K column vectors of the right singular vector matrix are selected to form the principal component direction matrix, where K is a preset positive integer, determined based on whether the cumulative contribution rate of singular values ​​reaches a preset threshold. This extracts the core characteristic directions of morphological variation from the normal cell population, laying the foundation for constructing a stable reference space.

[0050] The core feature dimension refers to the dimension of the linear subspace spanned by the first K principal component directions of the principal component direction matrix. This dimension K represents the minimum number of features required to capture the major variations in normal cell morphology. The distribution range refers to the statistical distribution boundary of the morphological feature projection values ​​of normal control group cells within the space defined by the core feature dimension, used to define the range of morphological fluctuations in the normal cell state. Specifically, the core feature dimension K is the number of principal component directions selected in step S301. To determine the distribution range, the centered control group data matrix is ​​projected onto the principal component direction matrix to obtain the coordinate matrix of control group cells in the core feature space. For each dimension of this coordinate matrix, i.e., each principal component direction, the statistical distribution parameters of the projection values ​​of all control group cells in that dimension are calculated. For example, the mean and standard deviation can be calculated, and the upper and lower bounds of the distribution of each dimension can be determined based on the mean plus or minus a certain number of standard deviations; or the quantiles of its empirical distribution can be directly calculated as the distribution range of that dimension. The core feature dimension and the distribution range of each dimension together define the variable region of normal cell morphology in the dimensionality-reduced space.

[0051] The reference space is a mathematically defined benchmark space used to quantify the degree of deviation in cell morphology. It is defined by the core feature dimension K and the corresponding distribution range in each dimension. This space encapsulates the typical variation patterns of normal cell morphology and their statistical boundaries. Specifically, the core feature dimension K determined in step S302 and the upper and lower bound parameters of the distribution in each principal component direction are integrated and stored to form a mathematical description of the reference space. For example, the reference space can be formally defined as a subspace spanned by K orthogonal bases, and in this subspace, the effective range of normal cell projection values ​​is defined in each dimension, such as the lower and upper limits. Here, the K orthogonal bases refer to the first K columns of the principal component direction matrix. The construction of this reference space allows the morphological characteristics of any cell, including cells treated with metabolites, to be projected into this space, and the degree of deviation from the normal state can be quantified by comparing the relative relationship between its projection position and the normal distribution range. In this way, by accurately extracting and mathematically defining the reference space of normal cell state, a stable and computable biological benchmark is provided for quantifying the effects of metabolite perturbation. Compared to directly using the original high-dimensional features to calculate distance or difference, this application uses singular value decomposition to reduce dimensionality and focus on the core patterns of normal cell morphological variations. This effectively removes noise and interference from non-critical variation directions, making the subsequently calculated leverage score more specifically reflect morphological changes caused by metabolites that exceed the normal fluctuation range. This not only enhances the accuracy of identifying metabolite-specific effects, but also, because the reference space is constructed based on control group data, makes the leverage scores comparable across different experimental batches or conditions, further improving the reproducibility and robustness of the entire model.

[0052] After constructing a reference space representing the normal cell state, the crucial stage of projecting all cell data onto this reference space begins. The core objective of this stage is to quantify the deviation of each cell, including the metabolite-treated group and the control group, from the normal cell state using the established reference space. This lays the foundation for subsequent calculation of leverage scores and the construction of the association between metabolites and phenotypes. The projection of all cell data from the corrected cell morphology feature data onto the reference space to obtain the projection results includes the following steps: Step S304: Perform matrix multiplication on all cell data in the corrected cell morphology feature data and the principal component direction matrix to obtain the initial projection coordinates of each cell in the reference space.

[0053] Step S305: The initial projected coordinates are scaled so that the coordinate values ​​in each principal component direction are scaled by the reciprocal of their corresponding singular values ​​to obtain scaled projected coordinates.

[0054] Step S306: Output the scaled projection coordinates as the projection result.

[0055] The total cell data refers to the morphological feature data matrix of all cells, including those in the metabolite treatment group and the control group, obtained after preprocessing and batch effect correction. Matrix multiplication is a linear algebraic operation that multiplies the total cell data matrix by the principal component direction matrix. The initial projected coordinates are the coordinate representation of each cell in the reference space spanned by the first K principal component directions, and are a K-dimensional vector. Specifically, let the total cell data matrix be denoted as X, where rows represent cells and columns represent morphological features; let the principal component direction matrix be denoted as VK, where columns represent the first K principal component directions; then the initial projected coordinate matrix P can be calculated using the formula P = X * VK. Thus, through linear projection, high-dimensional cell morphological feature data is transformed into a low-dimensional reference space defined by the core features of normal cell states, achieving dimensionality reduction and focusing on biological significance.

[0056] Scaling refers to the proportional scaling operation performed on each column of the initial projected coordinate matrix P, i.e., the coordinate values ​​in each principal component direction. Singular values ​​refer to the singular values ​​corresponding to the first K principal component directions obtained through singular value decomposition of the control group data in step S301. These singular values ​​reflect the importance or energy of each principal component direction in describing the variation of normal cell states. Scaled projected coordinates refer to the coordinate representation obtained after scaling, which more accurately reflects the relative position of cells in the reference space. Specifically, the diagonal matrix formed by the singular values ​​corresponding to the first K principal component directions obtained from singular value decomposition is denoted as Σk. The scaled projected coordinate matrix P_scaled can be calculated using the formula P_scaled=P*(Σk)^(-1), that is, multiplying each column of the initial projected coordinate matrix P by the reciprocal of the corresponding singular value. Thus, by considering the variation contribution of each principal component direction, the projected coordinates are normalized, eliminating the bias in subsequent distance calculations caused by the different inherent variation scales of different principal component directions, ensuring the fairness and accuracy of subsequent calculations.

[0057] The projection result refers to the final coordinate representation of each cell in the normalized reference space, i.e., the scaled projection coordinate matrix P_scaled. This projection result accurately depicts the position of each cell's morphological features in the reference space defined by the core features of the normal cell state and normalized by energy. This projection result provides a direct and accurate input for the next step of calculating the leverage score of each cell. By outputting this projection result, the mapping from the high-dimensional morphological feature space to the low-dimensional biologically meaningful reference space is completed, laying the data foundation for quantifying cell state deviations. In this structured projection process, not only is all cell data effectively mapped to the reference space constructed based on the normal state, but also, by introducing singular value-based scaling, the importance differences of different principal component directions are cleverly considered, making the projection result more realistically reflect the biological deviation of cell morphology relative to the normal state. Compared to directly projecting data into the principal component space without scaling, this application effectively avoids distance calculation bias caused by energy differences in principal component directions, improves the accuracy and biological rationality of subsequent leverage score calculations, and thus indirectly enhances the quantitative accuracy of the metabolite-phenotype relationship model.

[0058] After projecting all cell data onto a reference space constructed based on normal cell states and obtaining the scaled projected coordinates of each cell (i.e., the projection result), the crucial computational stage of quantifying the degree of morphological deviation for each cell begins. The core objective of this stage is to accurately calculate the leverage score of each cell based on the projection results through specific mathematical operations, thereby providing a quantifiable and robust perturbation strength indicator for subsequently constructing the association between metabolites and phenotypes. Figure 3 This is a flowchart illustrating the method for calculating the leverage score of each cell based on projection results, as provided in this application. Figure 3 As shown, calculating the leverage score for each cell based on the projection results includes the following steps: Step S307: Squaring the projection vector of each cell in the projection result to obtain the square vector of each cell projection vector.

[0059] Step S308: Sum the elements of the squared vector of each cell projection vector to obtain the square norm of each cell projection vector, and use the square norm of each cell projection vector as the lever score of the corresponding cell.

[0060] The projection result refers to the scaled projection coordinate matrix P_scaled output in step S306, where each row represents the coordinate representation of a cell in the reference space, i.e., a K-dimensional cell projection vector. The squaring operation is an element-wise mathematical operation, that is, squaring each coordinate value in the cell projection vector. The squared vector is a new vector obtained after the squaring operation, where each element is the square of the corresponding element in the original cell projection vector. Specifically, for the i-th cell projection vector pi=[pi1,pi2,...,piK] in the projection result matrix P_scaled, it is squared, that is, the square value (pij)^2 of each element pij is calculated, thus obtaining the squared vector pi_squared=[(pi1)^2,(pi2)^2,...,(piK)^2]. Through this step, negative values ​​in the projection coordinates are converted into positive values, and the contribution of larger coordinate values ​​(i.e., directions far from the normal center) is amplified, preparing for subsequent comprehensive measurement of the overall deviation. Here, larger coordinate values ​​refer to coordinate values ​​in directions far from the normal center.

[0061] The element-wise summation operation refers to the mathematical operation of adding all elements of a squared vector. The square norm, also known mathematically as the square of the L2 norm of a vector, refers to the sum of the squared values ​​of all elements in the cell's projection vector. The leverage score is a scalar value obtained by calculating the square norm, used to quantify the overall deviation of the corresponding cell's morphological characteristics from the reference space of the normal cell state. Specifically, for the squared vector pi_squared of the i-th cell obtained in step S307, the sum of all its elements is calculated, i.e., the leverage score leverage_i = Σ_{j=1toK}(pij)^2. This leverage score leverage_i is the final quantitative index of the i-th cell. Thus, by calculating the sum of the squares of the cell's coordinates in its projection space, the overall deviation of the cell from the center of the normal state across all core feature dimensions is effectively integrated, where the center of the normal state refers to the origin. The higher the score, the farther the cell's morphological characteristics are from the center in the reference space defined by the core features of the normal state, i.e., the greater the degree of deviation of its morphology from the normal state. This leverage score calculation method is mathematically clear, computationally efficient, and has a well-defined geometric interpretation: the squared Euclidean distance from the cell to the origin in the reduced-dimensional reference space. It can sensitively capture the comprehensive deviation of cell morphology in multidimensional space, overcoming the limitations of single-dimensional comparisons. Compared to directly using the original high-dimensional features to calculate distances or differences, this application calculates leverage scores based on a carefully constructed, energy-normalized reference space. This makes the score more specifically reflect morphological changes caused by metabolites that exceed normal fluctuations, rather than technical variations or noise. This not only enhances the accuracy and quantification reliability of identifying metabolite-specific effects, but also, due to the standardization of the calculation process, makes leverage scores comparable under different experiments or conditions, further improving the analytical consistency, interpretability, and predictive robustness of the entire metabolite-phenotype relationship model.

[0062] Step S400: Construct a correlation map between metabolites and cell phenotype clusters based on cell phenotype clusters and leverage scores.

[0063] After projecting all cell data onto a reference space constructed based on normal cell states and obtaining the leverage score for each cell, the crucial stage of constructing the association between metabolites and cell phenotype clusters begins. The core objective of this stage is to systematically integrate cell phenotype cluster information with the quantified metabolite perturbation intensity to generate an interpretable and visualized association map, thus providing a structured foundation for subsequently constructing a predictable metabolite-phenotype relationship model. Constructing the association map between metabolites and cell phenotype clusters based on cell phenotype clusters and leverage scores includes the following steps: Step S401: Based on the leverage score of each cell, aggregate the leverage scores of all cells treated with the same metabolite to obtain the average perturbation intensity of each metabolite.

[0064] Step S402: Associate the average perturbation intensity of each metabolite with its corresponding cell phenotype cluster to form a relation matrix with metabolites as rows and cell phenotype clusters as columns.

[0065] Step S403: Normalize each element in the relation matrix to obtain the normalized perturbation intensity of each metabolite in each cell phenotype cluster.

[0066] Step S404: Generate a correlation map between metabolites and cell phenotypic clusters based on the normalized perturbation intensity.

[0067] The term "all cells treated with the same metabolite" refers to all cell samples with the same metabolite identifier in the corrected cell morphology characteristic data. Aggregation calculation refers to a mathematical operation that summarizes a set of values ​​to generate a representative value; in this application, the arithmetic mean is preferably calculated. The average perturbation intensity is a scalar value obtained through aggregation calculation, characterizing the average morphological deviation caused by a specific metabolite in all cells treated with it. Specifically, based on the metabolite identifiers attached to the corrected cell morphology characteristic data, all cells are grouped according to the metabolite treatment they received. For each group, i.e., all cells under the same metabolite treatment, the leverage score set calculated in step S400 is extracted. The arithmetic mean of all values ​​in this leverage score set is calculated, i.e., summed and divided by the total number of cells in the group, to obtain the average perturbation intensity of the metabolite. This assigns a global, comparable quantitative index to each metabolite, reflecting the average intensity of the cell morphological changes it causes.

[0068] The corresponding cell phenotype cluster refers to the phenotype cluster to which each cell belongs, identified through cluster analysis in step S200. Association refers to establishing a mapping relationship between metabolite identifiers, cell phenotype cluster identifiers, and quantified intensity values. The relationship matrix is ​​a two-dimensional data structure where row indices represent different metabolites, column indices represent different cell phenotype clusters, and each element in the matrix stores the association strength information of a specific metabolite within a specific cell phenotype cluster. Specifically, an initially empty matrix is ​​created with rows equal to the number of all unique metabolite identifiers and columns equal to the number of all unique cell phenotype cluster identifiers. For each metabolite, all cells processed by it are traversed, and based on the cell phenotype cluster to which each cell belongs, the leverage score of that cell is accumulated and added to the corresponding metabolite row and the corresponding cell phenotype cluster column in the relationship matrix. Then, for each element in the relationship matrix, i.e., each metabolite-cell phenotype cluster pair, the sum of its accumulated leverage scores is divided by the number of cells with that metabolite in that cell phenotype cluster to obtain the average leverage score of that metabolite in that cell phenotype cluster, and this value is used to fill the corresponding element in the relationship matrix. This led to the construction of a complete relational matrix, where each element quantifies the average morphological deviation caused by a specific metabolite in a specific cell phenotype cluster.

[0069] Normalization is a data scaling method designed to eliminate comparability issues between different cell phenotype clusters or different metabolites due to baseline differences or varying cell numbers, ensuring that the values ​​in the relationship matrix have a consistent scale. The normalized perturbation strength refers to the association strength value obtained after normalization, which can be used for fair comparisons across metabolites and cell phenotype clusters. Specifically, the relation matrix obtained in step S402 is normalized along the row direction, that is, normalized along each metabolite across all cell phenotypic clusters. For example, z-score normalization can be used, that is, for each row, the mean and standard deviation of all elements in that row are calculated, and then the mean is subtracted from each element in that row and divided by the standard deviation; or min-max normalization can be used to scale each row of elements to the range [0,1]. Preferably, this application uses z-score normalization along the row direction, so that the perturbation intensity of each metabolite in different cell phenotypic clusters can be relatively compared in the context of its own perturbation mode, highlighting its specific effect. The matrix obtained after normalization is the relation matrix containing the normalized perturbation intensity.

[0070] A correlation map is a visual representation used to intuitively show the specific association patterns between metabolites and cell phenotype clusters. Generation refers to the process of creating a map based on a correlation matrix containing normalized perturbation strengths using visualization techniques. Specifically, the normalized correlation matrix is ​​used as input data. Data visualization libraries, such as Seaborn or Matplotlib in Python, are used to plot the correlation matrix as a heatmap. In this heatmap, the x-axis represents different cell phenotype clusters, the y-axis represents different metabolites, and the color intensity or hue of each cell represents the corresponding normalized perturbation strength value. For example, a continuous color map, such as from blue to red, can be used, where blue represents negative normalized perturbation strength, i.e., morphological deviation is below the average level of the metabolite; red represents positive normalized perturbation strength, i.e., morphological deviation is above the average level of the metabolite. Simultaneously, a dendrogram of row and column clustering can be used to show similar patterns between metabolites or cell phenotype clusters. The final output heatmap is the correlation map between metabolites and cell phenotype clusters. This systematic approach aggregates and correlates cellular-level quantitative perturbations (leverage scores) with metabolite and phenotypic cluster levels. By constructing a relationship matrix and performing normalization, a visualized relationship map is generated. This process not only achieves multi-scale quantification of metabolite effects (from cell to metabolite to phenotypic cluster) but also eliminates baseline differences through normalization, enabling fair comparison and visualization of the effects of different metabolites across different phenotypic clusters. This significantly enhances the ability to identify metabolite-specific phenotypic effects, allowing researchers to intuitively discover which metabolites specifically affect which cell states. It provides a clear and structured knowledge base for understanding metabolite functional mechanisms and subsequent predictive models, thereby improving the overall interpretability and practicality of the metabolite-phenotypic relationship model.

[0071] Step S500: Train a supervised classification model based on cell phenotype clusters and corrected cell morphology feature data, and construct an interpretable metabolite and phenotype prediction model based on the association map and the supervised classification model.

[0072] After systematically obtaining the cell phenotypic cluster classification results and the association map between metabolites and phenotypic clusters, this application enters the stage of supervised classification model training and final prediction model integration to achieve automated phenotypic prediction and effect analysis of samples treated with unknown metabolites. The core objective of this stage is to train a supervised classification model that can accurately predict the phenotypic cluster to which a cell belongs based on its morphological characteristics, and to effectively fuse this classifier with the constructed association map to form an end-to-end, resolvable model.

[0073] The process of training a supervised classification model based on cell phenotypic clusters and corrected cell morphological feature data includes the following steps: Step S501: Use the cluster labels of cell phenotypic clusters as supervision signals and the corrected cell morphology feature data as input features to construct a labeled training dataset.

[0074] Step S502: Use the labeled training dataset to train the preset deep neural network classification model to obtain the trained supervised classification model.

[0075] The cluster label of a cell phenotypic cluster refers to the unique identifier assigned to each cell by the cluster analysis in step S200, representing the cell state category to which the cell belongs, possessing specific morphological characteristics. The supervision signal refers to the target variable or true label used to guide model parameter optimization during model training within a supervised learning framework. The labeled training dataset is a structured dataset where each sample data point contains a corrected morphological feature vector of a cell and its corresponding cell phenotypic cluster label. Specifically, the cell phenotypic cluster identifier assigned to each cell is obtained from the output of step S200 and used as the cell's label. Each row in the corrected cell morphological feature data matrix output from step S100 is paired with the cluster label of the corresponding cell. All cell feature vector-cluster label pairs are combined to form a complete training dataset for supervised learning. This provides subsequent classifier training with biologically meaningful category labels and high-quality feature input, transforming the knowledge discovered through unsupervised clustering into supervised information usable for model training.

[0076] The pre-defined deep neural network classification model refers to a machine learning model with multi-layer nonlinear transformations. In this application, a variant of the multilayer perceptron convolutional neural network is preferred, its structure including an input layer, several hidden layers, and an output layer. Training refers to the process of adjusting the internal parameters of the deep neural network model through optimization algorithms, so that the model can accurately predict the corresponding cell phenotypic cluster labels based on the input feature vectors. A trained supervised classification model refers to a model instance whose parameters have been optimized and can be used to predict phenotypic cluster classification based on new cell morphological feature data. Specifically, the labeled training dataset constructed in step S501 is divided into a training set and a validation set according to a preset ratio. The training set data is input into the initialized deep neural network model, and the model calculates the predicted probability distribution for each cell phenotypic cluster based on the input features. The difference between the model's predicted probability distribution and the true cluster labels is calculated using the cross-entropy loss function. A backpropagation algorithm and optimizer, such as the Adam optimizer, are used to update the model's weight parameters based on the loss value, iterating until the model's performance on the validation set stabilizes or reaches a preset stopping condition. Finally, the optimal model parameters are saved to obtain the trained supervised classification model. This resulted in a strong classifier that can automatically learn from cell morphological features and identify the phenotypic cluster to which they belong, providing a core inference engine for subsequent prediction of unknown metabolite effects.

[0077] Building upon the cost-effective supervised classification model, the final stage aims to integrate all the aforementioned analytical results to construct a complete and practically applicable prediction system. The core objective of this stage is to organically combine the association graph, which quantifies the strength of the association between metabolites and phenotypes, with the supervised classification model, which predicts phenotypes at the cell level, forming a comprehensive prediction model that not only predicts the phenotype but also explains why it is that phenotype. Specifically, constructing an interpretable metabolite and phenotype prediction model based on the association graph and the supervised classification model means integrating the association graph and the trained supervised classification model to form a unified interface for metabolite and phenotype prediction models. Integration refers to the logical encapsulation and connection of the association graph and the supervised classification model in software, enabling them to work collaboratively. The unified interface for metabolite and phenotype prediction models refers to a well-defined program call interface or server endpoint that, upon receiving new cell data input, can output its predicted phenotype and related metabolite effect analysis information. Specifically, in the software system, the association graph generated in step S404 is logically associated with the trained supervised classification model obtained in step S502. This paper designs and implements a predictive wellhead interface. Internally, the interface encapsulates the following processing flow: for new input data, it first calls a supervised classification model to predict phenotypic clusters; then, based on the predicted phenotypic clusters, it queries an association graph to obtain metabolite information significantly related to those clusters. Externally, the interface provides a unified function or API that receives preprocessed cell morphology feature data and returns predicted cell phenotypic cluster labels and metabolite effect analysis results based on the association graph. This achieves an end-to-end automated process from raw feature input to final prediction and analysis output, providing users with a convenient and unified application entry point.

[0078] Preferably, after constructing an interpretable metabolite and phenotype prediction model, cell image data processed with unknown metabolites are also received. For cell image data processed with unknown metabolites, the same preprocessing and batch correction are applied and input into the supervised classification model to predict the cell phenotype clusters that are most likely to be triggered, and the inference results of cell data processed with unknown metabolites are inferred based on the association map.

[0079] Cell image data processed with unknown metabolites refers to the raw imaging data corresponding to cells that have been processed with a certain metabolite but whose corresponding metabolite is unknown. The same preprocessing and batch correction refers to strictly replicating the data preprocessing and batch effect correction process defined in step S100, including feature filtering, intra-plate normalization, global normalization, inverse normalization, and empirical Bayesian batch correction. The predicted cell phenotype cluster refers to the most likely phenotype type determined by the model interface for the input cell data. The related metabolite effect inference results refer to the list of metabolites typically associated with the predicted phenotype cluster, retrieved from the association graph, along with their normalized perturbation intensity information, used to infer the possible mode of action or intensity of the unknown metabolite. Specifically, cell images processed with unknown metabolites are acquired, and feature extraction is performed using the same image analysis software as in the training phase, such as CellProfiler, to obtain raw cell morphology feature data. This raw data is then strictly preprocessed and batch effect corrected according to the process in step S100 to ensure it is in the same feature space and distribution as the model training data. The corrected feature data is then input into the unified prediction model interface constructed above. The interface first uses a supervised classification model to calculate the probability of the cell belonging to each phenotypic cluster, and takes the cluster with the highest probability as the prediction result. Then, the interface queries the association graph to obtain information on all metabolites with high standardized perturbation strength in the predicted phenotypic cluster, serving as a reference for potential effects. Finally, the interface outputs the predicted phenotypic cluster label for the cell, as well as a list of inferred metabolite effects based on the graph. This achieves automated and quantitative prediction and analysis of unknown metabolite effects. Users can not only know what phenotypic changes have occurred in the cell, but also which metabolites typically cause these changes, greatly enhancing the model's interpretability and practical value in drug discovery or toxicity assessment. Thus, this application not only successfully transforms phenotypic knowledge discovered through unsupervised learning into supervised predictive capabilities, but also endows the model with deeper biological insights by combining classification prediction with association graph analysis. Compared to black-box models that can only output predicted categories without providing causal explanations, the metabolite and phenotypic prediction model constructed in this application achieves a unity of prediction and analysis, significantly improving the model's application value and credibility in actual biomedical research.

[0080] Figure 4 This is a schematic diagram illustrating the overall process of constructing a model of the relationship between metabolites and phenotypes, as provided in this application. Figure 4 As shown, building a model of the relationship between metabolites and phenotypes requires sequentially performing data preprocessing and batch effect correction, dimensionality reduction and clustering to identify cell phenotype clusters, calculating leverage scores to quantify metabolite perturbation effects, constructing a metabolite-phenotype relationship map, and building a classifier for prediction.

[0081] Figure 5This is a schematic diagram illustrating the system connections for constructing a model of the relationship between metabolites and phenotypes, as provided in this application. Figure 5 As shown, a system for constructing a model of the relationship between metabolites and phenotypes includes: a preprocessing module, a dimensionality reduction and clustering module, a data computation module, and a model construction module.

[0082] The system comprises the following modules: a preprocessing module, which acquires raw cell morphology data, performs preprocessing and batch effect correction on the raw data to obtain corrected cell morphology data; a dimensionality reduction and clustering module, which performs dimensionality reduction on the corrected cell morphology data to obtain low-dimensional cell morphology representations, and performs cluster analysis on these low-dimensional representations to identify multiple cell phenotype clusters; a data calculation module, which constructs a reference space representing the normal cell state based on control group data from the corrected cell morphology data, projects all cell data from the corrected cell morphology data onto the reference space to obtain projection results, and calculates the leverage score for each cell based on the projection results; and a model building module, which constructs a correlation map between metabolites and cell phenotype clusters based on cell phenotype clusters and leverage scores, trains a supervised classification model based on cell phenotype clusters and corrected cell morphology data, and constructs an interpretable metabolite and phenotype prediction model based on the correlation map and the supervised classification model.

[0083] The other functions performed by the preprocessing module, dimensionality reduction clustering module, data calculation module, and model building module, as well as the technical details of each function, are the same or similar to the corresponding features in the method for constructing a metabolite-phenotype relationship model described above, so they will not be repeated here.

[0084] This application also provides a computer storage medium storing a computer program that, when run on a computer, enables the computer to perform the steps in the method for constructing a metabolite-phenotype relationship model described above.

[0085] It should be understood that although the steps in the flowcharts in the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order requirement for the execution of these steps, and they can be performed in other orders.

[0086] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a model of the relationship between metabolites and phenotypes, characterized in that, The method includes: Obtain raw cell morphology feature data, and perform preprocessing and batch effect correction on the raw cell morphology feature data to obtain corrected cell morphology feature data; The corrected cell morphology feature data is subjected to dimensionality reduction processing to obtain a low-dimensional cell morphology feature representation, and cluster analysis is performed on the low-dimensional cell morphology feature representation to identify multiple cell phenotype clusters; Based on the control group data in the corrected cell morphology feature data, a reference space representing the normal cell state is constructed. All cell data in the corrected cell morphology feature data are projected onto the reference space to obtain the projection result. Based on the projection result, the leverage score of each cell is calculated. A correlation map between metabolites and cell phenotype clusters is constructed based on the cell phenotype clusters and the leverage scores; A supervised classification model is trained based on the cell phenotype clusters and the corrected cell morphology feature data. An interpretable metabolite and phenotype prediction model is then constructed based on the association map and the supervised classification model.

2. The method according to claim 1, characterized in that, The preprocessing and batch correction of the original cell morphology feature data to obtain the corrected cell morphology feature data includes: The original cell morphology feature data is filtered to remove features with variance below a preset threshold to obtain filtered cell morphology feature data. The filtered cell morphology data are subjected to in-plate normalization to eliminate in-plate position effects in order to obtain in-plate cell morphology data. The cell morphology feature data within the plate are globally standardized to ensure that each feature has zero mean and unit variance, thus obtaining global cell morphology feature data. The global cell morphology feature data is subjected to inverse normal transformation to make the distribution of each feature normal to obtain normal cell morphology feature data; The empirical Bayesian correction algorithm is applied to perform batch effect correction on the normal cell morphology feature data to obtain the corrected cell morphology feature data.

3. The method according to claim 1, characterized in that, The step of dimensionality reduction processing of the corrected cell morphology data to obtain a low-dimensional cell morphology feature representation includes: Principal component analysis was performed on the corrected cell morphology data to extract the first N principal components as a preliminary dimensionality-reduced representation of cell morphology features. The preliminary dimensionality-reduced cell morphology feature representation is input into a nonlinear dimensionality reduction model. By constructing a similarity graph between cells and performing diffusion mapping, a low-dimensional cell morphology feature representation is obtained.

4. The method according to claim 1, characterized in that, The reference space for characterizing normal cell states, constructed based on the control group data in the corrected cell morphology feature data, includes: Extract the control group data that has not been treated with metabolites from the corrected cell morphology feature data, and perform singular value decomposition on the control group data to obtain the principal component direction matrix that characterizes the core features of normal cell morphology. The core feature dimensions and distribution range of normal cell morphology are determined based on the principal component orientation matrix. A reference space representing the normal cell state is constructed by using the core feature dimensions and distribution range.

5. The method according to claim 1, characterized in that, The cluster analysis of the low-dimensional cell morphological feature representation identifies multiple cell phenotypic clusters, including: Based on the low-dimensional cell morphology features, a cell similarity map is constructed, wherein the cell similarity map is used to represent the adjacency relationship between cells; A community detection algorithm is performed on the cell similarity map to identify multiple cell phenotype clusters.

6. The method according to claim 1, characterized in that, The calculation of the leverage score for each cell based on the projection results includes: The squared vector of each cell projection vector in the projection result is obtained by squaring the projection vector of each cell. The square norm of each cell projection vector is obtained by summing the elements of the squared vector of each cell projection vector, and the square norm of each cell projection vector is used as the leverage score of the corresponding cell.

7. The method according to claim 6, characterized in that, The construction of the association map between metabolites and cell phenotype clusters based on the cell phenotype clusters and the leverage score includes: The leverage scores of all cells treated with the same metabolite are aggregated and calculated based on the leverage score of each cell to obtain the average perturbation intensity of each metabolite. The average perturbation intensity of each metabolite is correlated with its corresponding cell phenotype cluster to form a relation matrix with metabolites as rows and cell phenotype clusters as columns; Normalization is performed on each element in the relation matrix to obtain the normalized perturbation intensity of each metabolite in each cell phenotype cluster; Based on the standardized perturbation intensity, a correlation map between metabolites and cell phenotype clusters is generated.

8. The method according to claim 1, characterized in that, The method further includes: The system receives cell image data processed with unknown metabolites. After the same preprocessing and batch correction, the cell image data processed with unknown metabolites is input into the supervised classification model to predict the cell phenotypic clusters that it will trigger with the highest probability. Based on the association map, the system infers the inference results of the cell image data processed with unknown metabolites.

9. A system for constructing a model of the relationship between metabolites and phenotypes, characterized in that, The system includes: a preprocessing module, a dimensionality reduction and clustering module, a data computation module, and a model building module; wherein, The preprocessing module is used to acquire raw cell morphology feature data, and to preprocess and perform batch effect correction on the raw cell morphology feature data to obtain corrected cell morphology feature data. The dimensionality reduction clustering module is used to perform dimensionality reduction processing on the corrected cell morphology feature data to obtain a low-dimensional cell morphology feature representation, and to perform cluster analysis on the low-dimensional cell morphology feature representation to identify multiple cell phenotype clusters. The data calculation module is used to construct a reference space representing the normal cell state based on the control group data in the corrected cell morphology feature data, project all cell data in the corrected cell morphology feature data onto the reference space to obtain the projection result, and calculate the leverage score of each cell based on the projection result. The model building module is used to construct a correlation map between metabolites and cell phenotype clusters based on the cell phenotype clusters and the leverage score, train a supervised classification model based on the cell phenotype clusters and the corrected cell morphology feature data, and construct an interpretable metabolite and phenotype prediction model based on the correlation map and the supervised classification model.

10. A computer-readable storage medium having a computer program stored thereon that can run on a processor, characterized in that, When the computer program is executed by the processor, it implements a method for constructing a metabolite-phenotype relationship model as described in any one of claims 1 to 8.

Images