Method and system for stem cell induction data analysis

Abstract

The application provides a stem cell induction data analysis method and system, and relates to the technical field of data analysis. The method comprises the following steps: performing local neighborhood graph topology traversal on an initial multi-dimensional feature tensor, extracting range distribution features of P data points in the neighborhood to construct a local discrete degree measurement matrix; constructing a feature sequence time evolution gradient model based on the local discrete degree measurement matrix, calculating a first-order difference slope and a second-order change curvature to obtain a change ratio correction coefficient, and performing weighted mapping fusion of the change ratio correction coefficient and the local discrete degree measurement matrix to obtain a sensitivity classification index sequence. The application can process multi-source heterogeneous data, mine local feature correlations and key regulatory factors, realize positioning of abnormal induction conditions, and improve the accuracy and pertinence of stem cell induction process analysis.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for analyzing stem cell induction data. Background Technology

[0002] Stem cell induction differentiation is a multi-factor dynamic regulatory process. There is a non-linear, multi-scale temporal coupling relationship between the physicochemical parameters of the microenvironment (such as dissolved oxygen concentration, pH, osmotic pressure, and glucose and lactate metabolic concentrations in the culture system) and the whole-genome expression profile. In actual monitoring, these two types of data are mostly recorded at different sampling frequencies and different starting time points, forming heterogeneous multi-source data streams. Existing data analysis methods typically employ one of the following strategies when dealing with this problem: first, separating and independently analyzing the time series of physicochemical parameters and the gene expression profile, comparing results only at the decision level; second, after simple splicing or linear interpolation alignment at the feature level, using global correlation coefficients (such as Pearson correlation coefficient) or overall dimensionality reduction methods (such as principal component analysis) to extract the main change patterns.

[0003] These methods may have certain limitations in application, especially when data points in a local neighborhood exhibit a nonlinear evolution trend, and this trend only occurs within a narrow time window or a limited feature subspace. Global analysis strategies may weaken or mask these subtle but biologically significant signals. Specifically, the aforementioned methods struggle to fully utilize the range distribution characteristics of multiple data points within a local neighborhood to construct a dispersion measure, and they lack the ability to adaptively model temporal evolution gradients (such as first-order difference slopes and second-order curvatures). Therefore, when calculating the change ratio correction coefficient and performing sensitivity classification, they may not be able to accurately distinguish between background fluctuations and true anomalous shifts. This leads to a major practical technical problem: how to accurately identify and locate anomalous induction conditions caused by short-term, local micro-environmental disturbances from multi-source heterogeneous time-series data.

[0004] For example, in an experiment inducing human induced pluripotent stem cells (hiPSCs) to differentiate into pancreatic β cells, between day 4 and day 5 of the culture cycle (approximately 24 hours), the dissolved oxygen concentration in a certain well slowly drifted from 21% to 20.5% due to the edge effect of the culture plate, recovering after about 6 hours. Simultaneously, the expression level of the key transcription factor PDX1 showed a transient downregulation of approximately 15% during this period. If conventional global correlation analysis of the entire culture cycle is used, this oxygen concentration change (0.5%) does not constitute a statistical outlier in the entire time series, and the overall expression trend of PDX1 still shows an upward trend, causing this localized co-variant change to be ignored. Existing single-threshold anomaly detection methods (such as those based on the 3σ principle) may not trigger an alarm, or may only label the dissolved oxygen value without correlating it with the PDX1 shift, ultimately making it difficult to reverse-map this abnormal feature back to the original parameter coordinate system (e.g., specifically, well A3 of the culture plate, from 20:00 on day 4 to 02:00 on day 5, the dissolved oxygen setpoint of 21% actually decreased to 20.5%). In this situation, researchers may not be able to obtain the precise spatiotemporal location of the anomaly in a timely manner, thus affecting the rapid correction of induced condition deviations. Summary of the Invention

[0005] This invention provides a method and system for analyzing stem cell induction data, enabling precise localization of abnormal induction conditions.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for analyzing stem cell induction data, the method comprising: Step 1: Collect time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams from the stem cell induction culture system. After feature vectorization encoding and multi-source data alignment, construct an initial multidimensional feature tensor. Step 2: Perform a topological traversal of the local neighborhood graph on the initial multidimensional feature tensor to extract the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix; construct a feature sequence temporal evolution gradient model based on the local discreteness metric matrix, calculate the first-order difference slope and the second-order change curvature to obtain the change ratio correction coefficient, and perform a weighted mapping fusion of the change ratio correction coefficient and the local discreteness metric matrix to obtain the sensitivity graded index sequence; Step 3: Based on the sensitivity grading index sequence, perform adaptive threshold segmentation and state space alignment reconstruction on the initial multidimensional feature tensor to obtain the feature alignment mapping matrix; Step 4: Perform local rank nonlinear correlation calculation on the feature alignment mapping matrix to calculate the local Spearman correlation coefficient between target feature pairs. When the local Spearman correlation coefficient crosses the preset threshold E, determine the key regulatory factor feature set. Step 5: Input the key regulatory factor feature set into the isolated forest anomaly detection network to perform recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature nodes are reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinates.

[0007] Secondly, the stem cell induction data analysis system includes: The encoding module is used to collect the time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams in the stem cell induction culture system. After feature vectorization encoding and multi-source data alignment processing, an initial multidimensional feature tensor is constructed. The fusion module performs a local neighborhood graph topological traversal on the initial multidimensional feature tensor, extracts the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix; based on the local discreteness metric matrix, it constructs a feature sequence temporal evolution gradient model, calculates the first-order difference slope and the second-order change curvature to obtain the change ratio correction coefficient, and performs a weighted mapping fusion with the change ratio correction coefficient and the local discreteness metric matrix to obtain the sensitivity graded index sequence; The reconstruction module is used to perform adaptive thresholding and state space alignment reconstruction on the initial multidimensional feature tensor according to the sensitivity-level index sequence to obtain the feature alignment mapping matrix. The calculation module is used to perform local rank nonlinear correlation calculation on the feature alignment mapping matrix, calculate the local Spearman correlation coefficient between target feature pairs, and determine the key regulatory factor feature set when the local Spearman correlation coefficient crosses the preset threshold E. The mapping module is used to input the key regulatory factor feature set into the isolated forest anomaly detection network for recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature node is reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinate.

[0008] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0009] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0010] The above-described solution of the present invention has at least the following beneficial effects: This invention achieves unified modeling of two types of heterogeneous data by standardizing, vectorizing, and aligning time-series sequences of microenvironmental physicochemical parameters with whole-genome expression profile data streams. This breaks the limitations of independent analysis or simple splicing of the two types of data, fully exploring the intrinsic correlation between them. Local distribution features are extracted through topological traversal of local neighborhood graphs, constructing a local dispersion metric matrix. Combined with a temporal evolution gradient model, the dynamic changes of features are quantified, achieving fine-grained grading of feature sensitivity. This distinguishes between background fluctuations and biologically significant true feature shifts, preventing subtle and crucial feature signals from being masked, and improving the accuracy and relevance of feature analysis. Based on the sensitivity-graded index sequence, adaptive threshold segmentation and state space alignment reconstruction of the initial multidimensional feature tensor are achieved, effectively eliminating distribution bias, dimensional differences, and temporal misalignment issues among multi-source data, optimizing the feature space structure, and improving the consistency and balance of feature data. By employing local rank-based nonlinear correlation calculus instead of traditional global correlation analysis, this method can capture the nonlinear coupling correlation of target feature pairs within a local feature subspace, efficiently mining the feature set of key regulatory factors in the stem cell induction process, and overcoming the limitation of failing to capture local and short-term feature correlations. The degree of anomaly is quantified through recursive hyperplane space partitioning and isolation path length calculation. Combined with inverse mapping transformation, the abnormal nodes in the abstract feature space are mapped back to the original parameter coordinate system, achieving precise localization of abnormal induction conditions and solving the problems of inaccurate correlation between abnormal features and original induction parameters, and difficulty in locating the conditions under which anomalies occur. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of the stem cell induction data analysis method provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of a stem cell induction data analysis system provided in an embodiment of the present invention. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] like Figure 1 As shown, embodiments of the present invention propose a method for analyzing stem cell induction data, the method comprising the following steps: Step 1: Collect time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams from the stem cell induction culture system. After feature vectorization encoding and multi-source data alignment, construct an initial multidimensional feature tensor. Step 2: Perform a topological traversal of the local neighborhood graph on the initial multidimensional feature tensor to extract the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix; construct a feature sequence temporal evolution gradient model based on the local discreteness metric matrix, calculate the first-order difference slope and the second-order change curvature to obtain the change ratio correction coefficient, and perform a weighted mapping fusion of the change ratio correction coefficient and the local discreteness metric matrix to obtain the sensitivity graded index sequence; Step 3: Based on the sensitivity grading index sequence, perform adaptive threshold segmentation and state space alignment reconstruction on the initial multidimensional feature tensor to obtain the feature alignment mapping matrix; Step 4: Perform local rank nonlinear correlation calculation on the feature alignment mapping matrix to calculate the local Spearman correlation coefficient between target feature pairs. When the local Spearman correlation coefficient crosses the preset threshold E, determine the key regulatory factor feature set. Step 5: Input the key regulatory factor feature set into the isolated forest anomaly detection network to perform recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature nodes are reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinates.

[0015] In this embodiment of the invention, through standardization, encoding, and data alignment, unified modeling of heterogeneous data on microenvironment and gene expression is achieved, and intrinsic correlations are explored. Local distribution features are extracted and combined with temporal gradient modeling to finely classify feature sensitivity and distinguish between background fluctuations and deviations in real biological features. Based on sensitivity grading, multidimensional feature tensors are adaptively segmented and spatially aligned to eliminate data bias, dimensional and temporal misalignment. Local rank correlation calculation is used to capture local nonlinear feature coupling relationships and screen key regulatory factors. Finally, the degree of abnormality is quantified and restored to the original parameter space through inverse mapping to locate the conditions for the occurrence of abnormalities induced by stem cells.

[0016] In a preferred embodiment of the present invention, step 1 includes: Step 100: Obtain the time-series sequences of microenvironmental physicochemical parameters and the whole-genome expression profile data stream from the stem cell induction culture system. Perform bandpass filtering and outlier truncation on the time-series sequences of microenvironmental physicochemical parameters, and perform logarithmic normalization and low signal-to-noise ratio feature filtering on the whole-genome expression profile data stream to obtain the cleaned microenvironmental parameter sequences and gene expression profile sequences, specifically including: Through the monitoring equipment and detection platform matched with the stem cell-induced culture system, the time series of microenvironmental physical and chemical parameters and the whole-genome expression profile data stream during the whole process of stem cell induced culture are collected synchronously, wherein the time series of microenvironmental physical and chemical parameters includes continuous time series data such as dissolved oxygen concentration, pH value, osmotic pressure, glucose metabolic concentration and lactic acid metabolic concentration in the stem cell-induced culture system, and the whole-genome expression profile data stream includes time series data of expression abundance of each gene during the stem cell induction process. Band-pass filtering processing is performed on the collected time series of microenvironmental physical and chemical parameters, and a Butterworth band-pass filter is used to implement the filtering operation. Combined with the signal characteristics of the time series of microenvironmental physical and chemical parameters in the stem cell-induced culture system, the low cut-off frequency and high cut-off frequency of the filter are set, wherein the value range of the low cut-off frequency is set to 0.01Hz to 0.1Hz, which is used to filter low-frequency interference signals irrelevant to the physiological process of stem cell induction, and the value range of the high cut-off frequency is set to 1Hz to 10Hz, which is used to filter high-frequency noise signals generated during the collection process. The specific calculation process of the filtering is as follows: a second-order band-pass filter is used to implement the filtering operation, and the expression of the transfer function is: ; wherein is the complex frequency, , = is the central angular frequency of the filter, is the low cut-off frequency, is the high cut-off frequency, is pi, is the quality factor. Substitute each data point of the time series of microenvironmental physical and chemical parameters into the transfer function, and calculate to obtain the filtered time series of microenvironmental physical and chemical parameters; after filtering is completed, outlier truncation processing is performed on the filtered time series of microenvironmental physical and chemical parameters, and the interquartile range method is used to identify outliers. First, calculate the first quartile Q1, the second quartile Q2 and the third quartile Q3 of the time series of microenvironmental physical and chemical parameters, wherein Q1 is the value at the 25th percentile after all data in the sequence are sorted from small to large, Q2 is the median, that is, the value at the 50th percentile, and Q3 is the value at the 75th percentile. Calculate the interquartile range , set the outlier judgment thresholds as and , data points in the time series of microenvironmental physical and chemical parameters that are less than or greater than are determined as outliers, and truncation processing is performed on the determined outliers, replacing the outliers with the values of the corresponding thresholds to obtain the cleaned microenvironmental parameter sequence. Meanwhile, logarithmic standardization conversion is performed on the collected whole-genome expression profile data stream, and a natural logarithm conversion method is adopted. Let the original expression abundance value of a certain gene in the whole-genome expression profile data stream be x, where x>0, and the value after logarithmic standardization conversion is Log-normalization was performed on the raw gene expression abundance values ​​to convert them into normalized values ​​that conform to a normal distribution, reducing data heteroscedasticity. After conversion, low signal-to-noise ratio (SNR) feature filtering was performed on the whole genome expression profile data stream. The SNR of each gene expression abundance sequence was calculated (SNR = sequence mean / sequence standard deviation, where the mean is the signal intensity and the standard deviation is the noise intensity). A SNR threshold of 2.0 was set, and low SNR features with an SNR less than this threshold were removed. Gene expression abundance sequences with an SNR ≥ 2.0 were retained, and the cleaned gene expression profile sequence was finally obtained.

[0017] Step 101: Based on the cleaned microenvironment parameter sequence and gene expression profile sequence, project the discrete culture component identifiers and continuous physicochemical indicators onto a unified high-dimensional feature space, convert the gene expression abundance sequence into a normalized numerical vector, complete the vectorization encoding of the heterogeneous data source, and obtain the basic feature vector set, specifically including: Based on the cleaned microenvironment parameter sequence and cleaned gene expression profile sequence obtained in step 100, discrete culture component identifiers and continuous physicochemical indicators are first extracted from the cleaned microenvironment parameter sequence. The discrete culture component identifiers are the category identifiers of different culture components in the stem cell induction culture system, and the continuous physicochemical indicators are continuous data such as dissolved oxygen concentration, pH value, and osmotic pressure in the cleaned microenvironment parameter sequence. A feature space projection algorithm is used to project the discrete culture component identifiers and continuous physicochemical indicators to a preset high-dimensional feature space, i.e., let the category set of discrete culture component identifiers be... Each discrete culture component is identified using a one-hot encoding method. Convert to A 2D binary vector, where the corresponding Position 1 is set to 1, and all other positions are 0, resulting in a vector representation of the discrete culture component identifier; let the vector of the continuous physicochemical index be... ,in to For the numerical values ​​of different continuous physicochemical indicators, the min-max normalization method was used to standardize the vectors, resulting in normalized vector representations of the continuous physicochemical indicators. The vector representations of discrete culture component identifiers were concatenated with the normalized vector representations of the continuous physicochemical indicators to obtain high-dimensional feature vectors of microenvironment parameters. Simultaneously, the expression abundance sequence of each gene in the cleaned gene expression profile sequence was normalized, converted into a normalized numerical vector, and then processed using the z-score normalization method to obtain a normalized numerical vector of the gene expression abundance. This normalization process was performed on all genes in the cleaned gene expression profile sequence to obtain normalized numerical vectors of the expression abundance of all genes. Through these operations, the vectorization encoding of heterogeneous data sources was completed, unifying microenvironment-related features and gene expression-related features into vector forms that can be directly used for subsequent analysis, resulting in a basic feature vector set.

[0018] Step 102: Based on the timestamp sampling interval attributes of each dimension in the basic feature vector set, identify the high-frequency sampling dimension components and the low-frequency sampling dimension components, and perform time-series phase matching and sampling interval compensation on the high-frequency and low-frequency sampling dimension components to obtain time-series synchronization feature pairs, specifically including: Extract the timestamp sampling interval attribute corresponding to each feature dimension in the basic feature vector set of step 101. The timestamp sampling interval attribute is the time difference between two adjacent time points when the data of that feature dimension is collected. The sampling interval threshold is set to 10 minutes. Feature dimensions with timestamp sampling intervals less than or equal to the sampling interval threshold of 10 minutes are identified as high-frequency sampling dimension components, and feature dimensions with timestamp sampling intervals greater than 10 minutes are identified as low-frequency sampling dimension components, thus completing the identification of high-frequency and low-frequency sampling dimension components. After identification, time-series phase matching and sampling interval compensation are performed on the high-frequency and low-frequency sampling dimension components. First, a unified time-series reference axis is constructed based on the timestamp of the low-frequency sampling dimension component. The phase difference between the timestamp of the high-frequency sampling dimension component and the timestamp of the time-series reference axis is calculated. Linear interpolation is used to perform phase calibration on the high-frequency sampling dimension component. Let two adjacent timestamps on the time-series reference axis be... and The corresponding high-frequency sampling dimension components are in and Let t be the timestamp between them, where High-frequency sampling dimension components in The value at is ,exist The value at is Then the interpolated value at time t is This calculation enables time-phase matching between high-frequency sampling dimension components and low-frequency sampling dimension components.

[0019] Simultaneously, sampling interval compensation is performed on the low-frequency sampling dimension components. Cubic spline interpolation is used to increase the number of sampling points for the low-frequency sampling dimension components, ensuring that the sampling interval of the low-frequency sampling dimension components is consistent with that of the high-frequency sampling dimension components. Let the original sampling points of the low-frequency sampling dimension components be... ,in For timestamps, Construct a cubic spline interpolation function for the corresponding values. Its piecewise expression is: in each interval superior, ,in Let be the cubic spline interpolation coefficients within this interval, and satisfy the interpolation conditions. Continuity condition of the first derivative and the continuity condition of the second derivative At the same time, natural boundary conditions are adopted. and Solve for the interpolation coefficients, and In each interval The above is a cubic polynomial, with continuous first and second derivatives. During calculation, all timestamps corresponding to the high-frequency sampling dimension components (i.e., all time points with a sampling interval of 10 minutes) are first extracted. Then, each timestamp is substituted into the constructed cubic spline interpolation function to calculate the function value corresponding to each timestamp. This function value is the compensation value for the low-frequency sampling dimension component at the corresponding time point. The compensation values ​​corresponding to all timestamps are integrated in chronological order to obtain the complete compensated low-frequency sampling dimension component data. Through temporal phase matching and sampling interval compensation, the high-frequency and low-frequency sampling dimension components are synchronized in time, resulting in a temporal synchronization feature pair.

[0020] Step 103: Receive time-series synchronization feature pairs, stack them into a multidimensional structure along the time evolution dimension, microenvironment parameter dimension, and gene expression dimension, and perform cross-modal variance balancing scaling processing to construct an initial multidimensional feature tensor in the spatial coordinate system, specifically including: Upon receiving the time-series synchronization feature pairs obtained in step 102, three core construction dimensions are first determined: time evolution dimension, microenvironment parameter dimension, and gene expression dimension. The time evolution dimension represents the time progression of stem cell induction culture; the microenvironment parameter dimension represents the microenvironment-related feature dimensions within the time-series synchronization feature pairs; and the gene expression dimension represents the gene expression-related feature dimensions within the time-series synchronization feature pairs. The time-series synchronization feature pairs are then stacked along these three dimensions in a multi-dimensional structure. Specifically, the time evolution dimension is used as the first dimension, arranging the time-series synchronization feature pairs at different time points in chronological order; the microenvironment parameter dimension is used as the second dimension, arranging the microenvironment-related feature vectors at each time point in a preset order; and the gene expression dimension is used as the third dimension, arranging the gene expression-related feature vectors at each time point in a preset order, forming a three-dimensional initial structure. After stacking, cross-modal variance balancing scaling is performed on this three-dimensional initial structure to eliminate variance differences between different modalities, i.e., the variance of all feature data in the microenvironment parameter dimension is calculated separately. The variance of all feature data in the gene expression dimension Set the target variance as (Value is 1.0), calculate the scaling factor for the microenvironment parameter dimension. Scaling factor for gene expression dimension Multiply all feature data of the microenvironment parameter dimension by Multiply all feature data of the gene expression dimension by The temporal evolution dimension remains unchanged. Through the above-mentioned multidimensional structure stacking and cross-modal variance equalization scaling processing, an initial multidimensional feature tensor in the spatial coordinate system is constructed. This tensor can fully characterize the feature correlations of time, microenvironment parameters, and gene expression in the stem cell induction process.

[0021] This embodiment addresses the challenges of difficult, low-quality, and temporally asynchronous fusion of stem cell-induced multi-source heterogeneous data. First, it purifies the microenvironment and gene expression data through filtering, truncation, standardization, and noise filtering. Then, it unifies the heterogeneous data representation through vectorization encoding, achieving initial fusion. Temporal misalignment is corrected by temporal matching and interval compensation to achieve data temporal synchronization. Finally, through multidimensional stacking and variance-balanced scaling, an initial multidimensional feature tensor of fusion time, microenvironment, and gene expression information is constructed, smoothing out data differences between different modalities.

[0022] In a preferred embodiment of the present invention, step 2 includes: Step 200a: Based on the initial multidimensional feature tensor, the target data node is set as the traversal anchor point in the feature space coordinate system. Using the spatial coordinates of the target data node as the topological center, a preset number of P neighboring feature data points are selected along the spatial distribution trajectory of the initial multidimensional feature tensor, where P is a preset positive integer greater than 1. The spatial positions of the P neighboring feature data points are jointly defined by the coordinate vector of the target data node and the preset neighborhood search radius, resulting in a local neighborhood feature point set, specifically including: Based on the initial multidimensional feature tensor constructed in step 103 that characterizes the three-dimensional correlation between stem cell induction time, microenvironment, and gene expression, this initial multidimensional feature tensor is a three-dimensional tensor in a spatial coordinate system. Its three dimensions are the time evolution dimension (corresponding to the stem cell induction culture time process), the microenvironment parameter dimension (corresponding to the physicochemical parameters of the stem cell induction microenvironment), and the gene expression dimension (corresponding to the gene expression abundance during stem cell induction). Each data node in the tensor corresponds to a unique spatial coordinate vector. ,That For the time evolution dimension coordinates, For the coordinates of the microenvironment parameters, For gene expression dimension coordinates, Indexing of data nodes. In the feature space coordinate system, each data node in the initial multidimensional feature tensor is sequentially set as a traversal anchor point (i.e., a target data node). Each target data node corresponds to the combination features of a certain microenvironment parameter and gene expression at a certain moment of stem cell induction, using the spatial coordinate vector of the target data node. As the topological center, Along the spatial distribution trajectory of the initial multidimensional feature tensor (i.e., the dynamic evolution trajectory of features during stem cell induction), a predetermined number of P neighboring feature data points are selected, where P is a predetermined positive integer greater than 1, with a value of 10. These points are used to mine local feature associations around the target data node. The spatial positions of the P neighboring feature data points are defined by the coordinate vector of the target data node and a predetermined neighborhood search radius, which is 1.0. The coordinate vectors of all other data nodes in the initial multidimensional feature tensor are then calculated. Coordinate vector of target data node The Euclidean distance is used to filter out all data nodes whose Euclidean distance is less than or equal to the preset neighborhood search radius. These nodes correspond to local induction states that are spatiotemporally and spatially similar to the target node and have similar characteristics during stem cell induction. If the number of filtered data nodes is greater than or equal to P, the P data nodes with the smallest distance are selected as adjacent feature data points. If the number of filtered data nodes is less than P, all filtered data nodes are selected as adjacent feature data points, and several data nodes with the closest distance are added to make up the total number P (P - number of filtered data nodes). This ensures that the total number of nodes in the final local neighborhood feature point set is exactly P, which ultimately constitutes the local neighborhood feature point set corresponding to the target data node. The P adjacent feature data points in the set are all distributed around the target data node (corresponding to a certain local induction state). The values ​​of its time evolution dimension, microenvironment parameter dimension, and gene expression dimension are similar to the corresponding dimension values ​​of the target data node. This can centrally reflect the overall distribution characteristics of stem cell induction time process, microenvironment physicochemical parameter level, and gene expression abundance under this local induction state, as well as the local correlation between the three, and thus clearly reflect the core characteristic attributes of this local induction state. The above operations are performed sequentially on each target data node in the initial multidimensional feature tensor to obtain the set of local neighborhood feature points corresponding to each target data node.

[0023] Step 201a involves extracting numerical extremum boundaries along each feature dimension of the local neighborhood feature point set and calculating the dimensional range vector to obtain the local range distribution set of the nodes. Specifically, this includes: For each target data node obtained in step 200a, the set of local neighborhood feature points corresponds to a local induction state during the stem cell induction process, and contains P feature data points. These data points are consistent with the initial multidimensional feature tensor, and each has three core feature dimensions (time evolution dimension, microenvironment parameter dimension, and gene expression dimension). Each dimension corresponds to a continuous value, which accurately represents the key information of stem cell induction. Specifically, the continuous value of the time evolution dimension directly corresponds to the specific duration of stem cell induction culture, in minutes. The larger the value, the longer the induction culture time, which can reflect the progress of the induction process. The continuous value of the microenvironment parameter dimension corresponds to the specific detection values ​​of various physicochemical parameters in the stem cell induction microenvironment, such as dissolved oxygen concentration (unit: mg / L), pH value (dimensionless), and osmotic pressure (unit: mOsm / kg). The magnitude of these values ​​directly determines the physicochemical state of the microenvironment. When the values ​​are in the appropriate range (dissolved oxygen 5.0 to 7.0 mg / L, pH value 7.2 to 7.4, osmotic pressure 280 mg / kg), the physicochemical state of the microenvironment is determined. When the concentration reaches 320 mOsm / kg, the microenvironment is stable and suitable, ensuring normal proliferation and differentiation of stem cells. The greater the deviation from the suitable range, the more unstable the microenvironment becomes, directly affecting stem cell metabolism, internal environment balance, and cell morphology, thus interfering with the induction process and its effectiveness. The continuous values ​​for gene expression correspond to the standardized expression abundance values ​​(dimensionless) of each gene during stem cell induction; a higher value indicates higher gene expression activity, and a lower value indicates lower expression activity. Through these three dimensions of continuous values, the specific characteristics of each feature under this local induction state can be comprehensively and accurately quantified.

[0024] To quantify the fluctuation range of various features under a specific local induction state of stem cells, it is necessary to extract the extreme value boundaries along three feature dimensions: time evolution, microenvironment parameters, and gene expression, for the set of local neighborhood feature points corresponding to each target data node. Specifically, for each feature dimension, the values ​​of all data points in that dimension are traversed, and the maximum and minimum values ​​of that dimension are selected. The difference between the two is the fluctuation range of that dimension. The larger the range, the wider the numerical distribution range and the more significant the fluctuation of the corresponding feature (time, microenvironment, gene expression). For example, a large range in the microenvironment parameter dimension means that the physicochemical parameters such as dissolved oxygen and pH value in the local microenvironment fluctuate drastically, indicating poor microenvironment stability. The smaller the range, the more concentrated the numerical distribution of the corresponding feature and the smoother the fluctuation. For example, a small range in the gene expression dimension indicates that the gene expression abundance is relatively stable under the local induction state, and the gene expression state is more stable. After extracting the extreme values ​​and calculating the ranges of each dimension, the ranges of the three feature dimensions are integrated to form a dimension-level range vector. This vector can comprehensively reflect the overall fluctuation of the three core dimensions under the local induction state. For each target data node, the above extreme value extraction and range calculation operations are repeated in the local neighborhood feature point set to obtain the dimension-level range vector corresponding to each target data node. Finally, the dimension-level range vectors of all target data nodes are summarized and integrated to form a set of node local range distributions.

[0025] Step 202a: Standardize the range vectors of each dimension in the local range distribution set of nodes using a preset dimensional adaptive scaling factor to obtain a normalized range distribution matrix. Specifically, this includes: Since the dimensional differences between microenvironmental parameter dimensions (such as dissolved oxygen and pH value) and gene expression dimensions (such as gene expression abundance) are large during stem cell induction, direct analysis will affect the accuracy of local dispersion calculation. Therefore, a preset dimensional adaptive scaling factor needs to be set to eliminate the dimensional differences of different feature dimensions. The dimensional adaptive scaling factor is calculated using the reciprocal of the global range of each feature dimension. First, the global maximum and minimum values ​​of each feature dimension (feature dimensions 1, 2, and 3) are extracted from the initial multidimensional feature tensor. Then, the global range of each feature dimension (the difference between the global maximum and minimum) is calculated. The dimensional adaptive scaling factor for each feature dimension is then set to 1 / global range. If the global range of a feature dimension is 0 (i.e., all data points in that dimension have consistent values, indicating that the feature remains stable throughout the stem cell induction process and does not affect the induction state analysis), the dimensional adaptive scaling factor for that dimension is set to 1. This results in the final dimensional adaptive scaling factor vector. .

[0026] Using this dimensional adaptive scaling factor vector, the range vectors of each dimension in the set of local range distributions of nodes obtained in step 201a are standardized and mapped. The specific mapping process is as follows: Let the range vector of a certain dimension in the set of local range distributions of nodes be... The standardized range vector obtained after standardization mapping This involves multiplying the range vector of each dimension by the corresponding dimension's adaptive scaling factor to eliminate the influence of differences in dimensions across different feature dimensions, ensuring the comparability of microenvironment parameters and gene expression fluctuation analyses. The above standardization mapping operation is performed on all dimension-level range vectors in the node local range distribution set. All standardized range vectors are then arranged in the order of the target data nodes to construct a normalized range distribution matrix. The rows of this matrix correspond to the target data nodes (corresponding to the stem cell-induced local state), and the columns correspond to the three feature dimensions. Each element in the matrix is ​​the standardized range value for the corresponding target data node and the corresponding feature dimension.

[0027] Step 203a involves performing tensor structure recombination and variance aggregation operations on the normalized range distribution matrix along the time series axis and the feature dimension axis to construct a local dispersion measure matrix characterizing the intensity of local data fluctuations. Specifically, for the normalized range distribution matrix obtained in step 202a, in order to accurately quantify the intensity of local data fluctuations at different time points and in different feature dimensions during stem cell induction (the intensity of fluctuations directly reflects the stability of the stem cell induction state), it is first determined that the rows of the matrix correspond to the target data nodes (corresponding one-to-one with the timestamps of the time evolution dimension, with each timestamp corresponding to a certain moment of stem cell induction), and the columns correspond to the three feature dimensions (time evolution dimension, microenvironment parameter dimension, and gene expression dimension). Tensor structure recombination is performed along the time series axis and the feature dimension axis. Specifically, with the time evolution axis as the first dimension, the standardized range values ​​of all target data nodes corresponding to the same time point in the normalized range distribution matrix are arranged in chronological order (corresponding to different local induction states of stem cells at the same moment). With the feature dimension axis as the second dimension, the standardized range values ​​of the three feature dimensions under each time point are arranged in a preset order (time evolution dimension, microenvironment parameter dimension, and gene expression dimension) to form a two-dimensional tensor structure. This structure can clearly present the local fluctuation distribution of stem cells at different times and with different features.

[0028] This two-dimensional tensor structure uses rows to correspond to different timestamps of stem cell induction (i.e., different induction times) and columns to correspond to three core feature dimensions: temporal evolution, microenvironment parameters, and gene expression. Each element in the structure corresponds to the standardized range value of a specific induction time and feature dimension. By observing the differences in element values, the local fluctuation differences of each feature at different induction times can be intuitively presented. Simultaneously, it can clearly distinguish the fluctuation distribution of different features at the same time and the fluctuation trend of the same feature at different induction times. Variance aggregation is performed on this two-dimensional tensor structure. That is, for the feature dimension data corresponding to each timestamp, the variance of the standardized range values ​​of all target data nodes at that timestamp in that feature dimension is calculated. The variance value can quantify the degree of local fluctuation dispersion of the corresponding feature dimension at that time. The above variance calculation operation is performed on all timestamps and all feature dimensions. The calculated variance values ​​are rearranged along the time evolution axis and the feature dimension axis to construct a local dispersion measure matrix. The rows of this matrix correspond to timestamps (different times of stem cell induction), and the columns correspond to feature dimensions. Each element in the matrix is ​​the local data fluctuation variance of the corresponding timestamp and the corresponding feature dimension, used to characterize the intensity of local data fluctuation of the corresponding feature dimension at that timestamp. The larger the value of the matrix element, the higher the degree of dispersion of the data points in the local neighborhood of the corresponding feature dimension at that timestamp, that is, the more violent the local data fluctuation, and the worse the stability of the feature at that time of stem cell induction. The smaller the value, the more concentrated the data points in the local neighborhood of the corresponding feature dimension at that timestamp, that is, the gentler the local data fluctuation, and the better the stability of the feature at that time of stem cell induction.

[0029] Step 200b: Based on the local dispersion metric matrix, construct a discrete time step difference sequence along the time evolution axis to establish a feature sequence temporal evolution gradient model characterizing the changes in the rate and direction of feature fluctuations. Specifically, this includes: to capture the dynamic evolution of each feature during stem cell induction (i.e., the changing trend of local fluctuation intensity at different times), using the local dispersion metric matrix constructed in step 203a as a basis, where each row of this matrix corresponds to a timestamp on the time evolution axis (denoted as...). , (This represents the total number of timestamps, corresponding to different moments in the entire stem cell induction process). Each column corresponds to one of the three feature dimensions (denoted as 1, 2, 3). Each element in the matrix... Indicates the first The timestamp, the first The intensity of local data fluctuations in each feature dimension (corresponding to the stability of that feature at that moment during stem cell induction). Along the time evolution axis, for each feature dimension... Extract the local data fluctuation intensity values ​​corresponding to all timestamps under this dimension, and construct the discrete time step difference sequence of this feature dimension. Let the th... Each feature dimension in timestamp The intensity of local data fluctuations is timestamp The intensity of local data fluctuations is Then the discrete-time step difference sequence of this feature dimension is: The numerical value of this sequence reflects the magnitude of the change in the local fluctuation intensity of the feature at adjacent time points, with positive and negative values ​​indicating the direction of change. The larger the absolute value, the greater the magnitude of the change in the local fluctuation intensity of the feature between two adjacent induction time points, meaning the more significant the change in the fluctuation stability of the feature during stem cell induction. For example, an absolute value greater than 0.5 indicates a drastic change in the fluctuation intensity of microenvironmental parameters or gene expression at adjacent time points, meaning the change in the local fluctuation intensity of the feature at adjacent induction time points exceeds the normal fluctuation range, corresponding to significant fluctuations in the stability of the feature induced by stem cells. An absolute value less than or equal to 0.2 indicates a smaller magnitude of the change in the local fluctuation intensity of the feature between two adjacent induction time points, meaning the change in the fluctuation stability of the feature is more gradual, corresponding to a more stable stem cell induction state. A positive value indicates that the local fluctuation intensity of the feature at the next time point is greater than that at the previous time point, meaning the fluctuation is more intense, and the stability of the feature induced by stem cells is somewhat reduced. A negative value indicates that the local fluctuation intensity of the feature at the next time point is less than that at the previous time point, meaning the fluctuation is more gradual, and the stability of the feature induced by stem cells is somewhat improved. The above operations are performed on all three feature dimensions to obtain discrete time step difference sequences corresponding to the three feature dimensions. Based on these three discrete time step difference sequences, a feature sequence temporal evolution gradient model is established to characterize the changes in the fluctuation rate and direction of the features. This model uses the time evolution axis as the horizontal axis (corresponding to the stem cell induction time) and the values ​​of the discrete time step difference sequences as the vertical axis. It can intuitively reflect the fluctuation rate (magnitude of the difference sequence values) and fluctuation direction (positive or negative of the difference sequence values) of each feature dimension (microenvironment parameters, gene expression, etc.) at different time steps.

[0030] Step 201b: In the feature sequence time-series evolution gradient model, calculate the ratio of numerical increments between adjacent time steps to obtain a first-order difference slope sequence characterizing the instantaneous rate of change of the data; perform a second-order difference operation on the first-order difference slope sequence to extract slope inflection features; quantify the dynamic increase or decrease amplitude of the feature change rate based on the second-order difference operator of the discrete time series to obtain a second-order change curvature sequence characterizing the intensity of fluctuations, specifically including: In the feature sequence temporal evolution gradient model established in step 200b, to quantify the instantaneous change rate and fluctuation intensity of each feature during stem cell induction (both directly reflect the dynamic response capability of the stem cell induction state), for the discrete time step difference sequence of each feature dimension v, the ratio of numerical increments between adjacent time steps is calculated to obtain the first-order difference slope sequence characterizing the instantaneous change rate of the data. The magnitude of each element in this sequence corresponds to the speed of the instantaneous change rate of the data. The larger the value, the faster the instantaneous change rate of the data at the corresponding time step, that is, the more rapidly the fluctuation intensity of the feature changes at that moment during stem cell induction; the smaller the value, the slower the instantaneous change rate of the data at the corresponding time step, that is, the more gently the fluctuation intensity of the feature changes at that moment during stem cell induction. At the same time, the positive and negative values ​​of the sequence elements correspond to the direction of the instantaneous change of the data. Positive values ​​indicate an upward trend in the data (increased fluctuation intensity, decreased induction stability), and negative values ​​indicate a downward trend in the data (decreased fluctuation intensity, increased induction stability). Let the discrete time step difference sequence be... , No. The difference value for each time step is , No. The difference value for each time step is The time interval between adjacent time steps is (Consistent with the high-frequency sampling interval in step 102, the value is 10 min, corresponding to the sampling frequency induced by stem cells), then the... First-order difference slope at each time step The above calculation is performed on the discrete time step difference sequence of each feature dimension to obtain the first-order difference slope sequence corresponding to each feature dimension.

[0031] To capture the turning points in characteristic fluctuation rates during stem cell induction (i.e., changes in fluctuation trends, which may correspond to precursors of induction abnormalities), a second difference operation is performed on the first-order difference slope sequence for each feature dimension to extract slope turning point features. Let the first-order difference slope sequence... In the middle, the first The slope value is , No. +1 slope value Then the quadratic difference value This yields a second-order difference sequence. Based on the second-order difference operator of discrete-time series, the dynamic increase or decrease of the characteristic rate of change is quantified. The expression for the second-order difference operator is: This operator converts the values ​​of the quadratic difference sequence into a second-order change curvature sequence that characterizes the intensity of fluctuations. The larger the absolute value of the value in the sequence, the more intense the fluctuation of the feature at that time step, and the more unstable the fluctuation state of the feature at that moment during stem cell induction, and the more likely the induction anomaly will occur.

[0032] Step 202b involves fusing the first-order difference slope sequence and the second-order change curvature sequence, and jointly encoding the instantaneous rate of change and the degree of trend reversal through a nonlinear coupling mapping function to obtain the change ratio correction coefficient, which characterizes the intensity of the local data dynamic response. Specifically, this includes: To comprehensively quantify the dynamic response intensity of local data during stem cell induction (i.e., the synergistic effect of characteristic fluctuation rate and fluctuation intensity, which is directly related to the abnormal risk of the induction state), the first-order difference slope sequence of each feature dimension obtained in step 201b is first analyzed. With second-order curvature sequence Normalization is performed to eliminate dimensional differences and ensure that the two can be calculated together. The normalization method used is z-score normalization, and the normalized value of the first-order difference slope sequence is... ,in First-order difference slope sequence The mean, Its standard deviation; the normalized value of the second-order curvature sequence. ,in Second-order curvature sequence The mean, Its standard deviation. The normalized first-order difference slope sequence and the second-order change curvature sequence are fused, and a nonlinear coupling mapping function is used to jointly encode the instantaneous rate of change and the degree of trend reversal. The nonlinear coupling mapping function is selected as a combination of the Sigmoid function and linear weighting: ; in This is a weighting coefficient, with a value of 0.5, used to balance the weights of the first-order difference slope (instantaneous rate of change) and the second-order curvature (severity of fluctuation). It is a natural constant. The value represents the combined encoded value. The larger the value, the stronger the dynamic response of the local data, and the higher the risk of abnormality of this feature at that moment in stem cell induction. The above fusion encoding operation is performed on each feature dimension and each time step to obtain the change ratio correction coefficient sequence corresponding to each feature dimension. The change ratio correction coefficient sequences of all feature dimensions are integrated to obtain the change ratio correction coefficient (which is a matrix with the same dimensions as the local dispersion metric matrix) characterizing the intensity of local data dynamic response. The rows of this matrix correspond to different timestamps of stem cell induction (i.e., different induction times), and the columns correspond to the three core feature dimensions of time evolution, microenvironment parameters, and gene expression. Each element in the matrix corresponds to the correction coefficient value of a certain induction time and a certain feature dimension. The larger the coefficient value, the higher the intensity of local data dynamic response of that feature dimension at that time. That is, the more sensitive the feature is to changes in the microenvironment and time progression during stem cell induction, the more significant the synergistic effect of fluctuation rate and fluctuation intensity, and the higher the potential risk of induction anomaly. The smaller the coefficient value, the lower the intensity of local data dynamic response of that feature dimension at that time. That is, the more gradual the response of that feature to changes in external induction conditions, the more stable the fluctuation state, and the lower the potential risk of induction anomaly. Through the numerical distribution of each element in the matrix, the local data dynamic response capability of stem cell induction at different times and different feature dimensions can be comprehensively characterized.

[0033] Step 203b involves performing a multi-dimensional weighted tensor product operation on the change ratio correction coefficient and the local dispersion metric matrix to obtain the fused sensitivity quantification feature matrix. Based on a preset grading threshold range, the eigenvalues ​​in the sensitivity quantification feature matrix are discretized and mapped, and grading labels are assigned to them, resulting in a sensitivity grading index sequence carrying sensitivity level identifiers. Specifically, this includes: To distinguish between background fluctuations and biologically significant aberrations during stem cell induction and to achieve fine-grained grading of feature sensitivity, the change ratio correction coefficient matrix obtained in step 202b and the local dispersion measure matrix obtained in step 203a are first subjected to a multi-dimensional weighted tensor product operation to achieve the fusion of local fluctuation intensity and dynamic response intensity. That is, let the local dispersion measure matrix be... ( It is a set of real numbers, indicating that all elements in this feature data matrix are real numbers, meaning that the detection values ​​of time, microenvironment parameters, and gene expression during stem cell induction are all real numbers. (where 3 represents the total number of timestamps and 3 represents the number of feature dimensions), the change ratio correction coefficient matrix is... The multidimensional weighted tensor product of the two Its element calculation formula is: ,in This is the global weighting coefficient, with a value of 1.0, used to balance the weights of local dispersion and dynamic response strength. The fused sensitivity measurement feature matrix The element corresponding to the first The timestamp, the first The sensitivity metric values ​​for each feature dimension are calculated. Higher values ​​indicate higher sensitivity for that feature at that moment, making it more prone to anomalous shifts and requiring closer monitoring. The resulting fused sensitivity metric feature matrix is ​​then obtained. A preset threshold range for sensitivity is set, divided into three levels: low sensitivity, medium sensitivity, and high sensitivity. The low sensitivity range is [0, 0.3] (corresponding to stable background fluctuations during stem cell induction, with no abnormal risk), the medium sensitivity range is (0.3, 0.7] (corresponding to slight fluctuations, requiring routine monitoring), and the high sensitivity range is (0.7, 1.0] (corresponding to drastic fluctuations, indicating a risk of abnormal induction, requiring close monitoring). Based on this threshold range, a sensitivity quantification feature matrix is ​​established. Each eigenvalue in Perform discretization mapping and level label assignment, if If [0, 0.3], then assign a low sensitivity level label of 1; if If (0.3, 0.7], then the sensitivity level label in the allocation is 2; if If the value is (0.7, 1.0), then a high sensitivity level label 3 is assigned. Arrange the level labels corresponding to all feature values ​​in the order of time evolution axis and feature dimension axis to obtain a sensitivity level index sequence carrying sensitivity level identifier. This sequence can clearly characterize the local data sensitivity level of each timestamp (stem cell induction time) and each feature dimension (microenvironment, gene expression).

[0034] This embodiment solves the problems of global analysis's inability to capture local nonlinear features and distinguish between background fluctuations and real anomalies through a series of processes such as local feature mining and dispersion measurement. By using topological traversal of local neighborhood graphs and range analysis, a dispersion measurement matrix is ​​constructed to capture subtle data fluctuations within a narrow time window. Based on this matrix, a time-series evolution gradient model is constructed to quantify the rate and intensity of feature fluctuations, and correction coefficients are generated. Through weighted fusion and hierarchical threshold division, fine-grained feature sensitivity is achieved, generating a sensitivity-level index sequence. This overcomes the shortcomings of insufficient time-series gradient modeling, effectively distinguishes between background fluctuations and real feature shifts, and avoids the masking of key subtle signals.

[0035] In a preferred embodiment of the present invention, step 3 includes: Step 300: Analyze the hierarchical identifier distribution characteristics in the sensitivity grading index sequence, construct a dynamic baseline partitioning strategy model to drive data distribution bias correction, and obtain an adaptive threshold decision rule set, specifically including: The sensitivity grading index sequence carrying sensitivity level identifiers obtained in step 203b is received. This sequence is presented in matrix form, with rows corresponding to different time stamps of stem cell induction (i.e., different induction times), and columns corresponding to three core feature dimensions: temporal evolution, microenvironment parameters, and gene expression. Each element is the sensitivity level identifier (low sensitivity 1, medium sensitivity 2, high sensitivity 3) for the corresponding time and feature dimension. The distribution characteristics of the level identifiers in this sensitivity grading index sequence are analyzed, and the number of sensitivity level identifiers for the three feature dimensions at each time stamp is counted. The proportion of each sensitivity level (1, 2, 3) at each time stamp and each feature dimension is calculated. At the same time, the spatiotemporal evolution law of the sensitivity level identifiers is analyzed, that is, the trend of sensitivity level changes of the same feature dimension at different time stamps and the correlation of sensitivity levels of different feature dimensions at the same time stamp are tracked. The time interval and feature dimension where the high sensitivity level (identifier 3) is concentrated are identified. This interval and dimension correspond to the key areas where abnormal deviations are likely to occur during stem cell induction. The area where the low sensitivity level (identifier 1) is concentrated corresponds to the stable background fluctuation area.

[0036] Based on the aforementioned hierarchical identifier distribution characteristics and spatiotemporal evolution patterns, a dynamic reference surface partitioning strategy model is constructed to drive data distribution bias correction. This model takes the proportion of sensitivity levels as the core input and data distribution bias correction as the objective. By setting thresholds for the proportion of levels (0.4 for high sensitivity, 0.3 for medium sensitivity, and 0.3 for low sensitivity), different reference surface regions are divided, namely, the high sensitivity reference surface (corresponding to the region with a sensitivity level proportion ≥ 0.4), the medium sensitivity reference surface (corresponding to the region with a sensitivity level proportion < 0.4 and a sensitivity level proportion < 0.3), and the low sensitivity reference surface (corresponding to the region with a sensitivity level proportion < 0.3). Different threshold decision rules are set for different reference surface regions. For example, a strict segmentation threshold (segmentation threshold coefficient of 0.8) is used for high-sensitivity reference surface regions to accurately capture abnormal offset features; a medium segmentation threshold (segmentation threshold coefficient of 0.5) is used for medium-sensitivity reference surface regions to balance capture accuracy and computational efficiency; and a lenient segmentation threshold (segmentation threshold coefficient of 0.2) is used for low-sensitivity reference surface regions to avoid feature destruction caused by over-segmentation. All differentiated threshold decision rules are integrated to obtain an adaptive threshold decision rule set.

[0037] Step 301: Based on the adaptive threshold decision rule set, perform a multi-dimensional sliding window scan on the initial multi-dimensional feature tensor, dynamically adjust the segmentation boundary parameters to complete the adaptive threshold segmentation operation, and obtain several local feature subspace blocks, specifically including: Guided by the adaptive threshold decision rule set obtained in step 300, and using the initial multidimensional feature tensor constructed in step 103 as the processing object, this initial multidimensional feature tensor is a three-dimensional tensor (time evolution dimension × microenvironment parameter dimension × gene expression dimension), containing multi-source fusion feature data of the entire stem cell induction process. A multidimensional sliding window scan is performed on this initial multidimensional feature tensor. The sliding window settings match the adaptive threshold decision rule set. For high-sensitivity reference surface regions, the sliding window size is set to 3×3×3 (time dimension window size × microenvironment parameter dimension window size × gene expression dimension window size), and the sliding step size is set to 1 to ensure accurate capture of subtle abnormal offset features. For medium-sensitivity reference surface regions, the sliding window size is set to 5×5×5, and the sliding step size is set to 2, balancing capture accuracy and computational efficiency. For low-sensitivity reference surface regions, the sliding window size is set to 7×7×7, and the sliding step size is set to 3 to avoid redundant computation caused by over-scanning. During the sliding window scanning process, the segmentation boundary parameters are dynamically adjusted to complete the adaptive threshold segmentation operation. These parameters include the segmentation threshold and the boundary smoothing coefficient. The segmentation threshold is determined by the adaptive threshold decision rule set, and is specifically calculated as follows: , The basic segmentation threshold (with a value of 0.1) is used. Threshold coefficients for different sensitivity levels (high sensitivity) =0.8, medium sensitivity =0.5, low sensitivity =0.2); the boundary smoothing coefficient is set to 0.3 to reduce the jagged distortion of the segmentation boundary. The mean and standard deviation of the feature data within the sliding window are calculated. When the difference between a feature data point and the mean is greater than the segmentation threshold × standard deviation, the data point is determined as the segmentation boundary point. The boundary point is smoothed by combining the boundary smoothing coefficient to avoid the segmentation deviation caused by the excessive concentration of boundary points.

[0038] Through multi-dimensional sliding window scanning and dynamic boundary parameter adjustment, the initial multi-dimensional feature tensor is strictly divided into several independent local feature subspace blocks according to the sensitivity level and feature distribution pattern corresponding to the sensitivity grading index sequence. Each local feature subspace block corresponds to a specific local time interval (a series of consecutive adjacent timestamps, the duration of which is determined according to the sensitivity reference plane region; the time interval duration is 30 min in high-sensitivity regions, 50 min in medium-sensitivity regions, and 70 min in low-sensitivity regions) and a feature subspace (this feature subspace focuses on one or two related dimensions of the three core dimensions of temporal evolution, microenvironmental parameters, and gene expression, such as the subspace related to microenvironmental parameters and gene expression). Each subspace block contains only feature data of the same sensitivity level (low sensitivity 1, medium sensitivity 2, high sensitivity 3), and the feature fluctuation pattern is completely consistent with the feature distribution characteristics of that sensitivity level. That is, in the local feature subspace block of the high sensitivity level, the fluctuation amplitude of microenvironment parameters or gene expression is large (numerical range greater than 0.5), corresponding to the instability of the local region induced by stem cells; in the subspace block of the medium sensitivity level, the feature fluctuation amplitude is moderate (numerical range greater than or equal to 0.2 and less than or equal to 0.5), corresponding to the smooth transition of the induced state; in the subspace block of the low sensitivity level, the feature fluctuation amplitude is small (numerical range less than 0.2), corresponding to the stable region of the induced state, ensuring that the feature distribution of each subspace block is homogeneous.

[0039] Step 302: Based on several local feature subspace blocks, extract the principal component direction vectors of the feature distribution within each local feature subspace block and calculate the corresponding covariance matrix structure to obtain the local spatial distribution parameter set; perform eigenvector standardization operation based on the principal component direction vectors in the local spatial distribution parameter set to construct a rotation alignment basis; combine the covariance matrix structure to calculate the spatial translation offset and feature dimension scaling coefficient to obtain the affine transformation parameter set, specifically including: For the several local feature subspace blocks obtained in step 301, each local feature subspace block is a three-dimensional small tensor containing three-dimensional feature data of time, microenvironment, and gene expression within that local region. Principal component direction vectors of the feature distribution within each local feature subspace block are extracted to eliminate redundant correlations between feature dimensions and improve alignment accuracy. Principal component direction vector extraction is performed using principal component analysis (PCA). Let the feature data matrix of a certain local feature subspace block be... ,in The feature data matrix representing the local feature subspace block is used to centrally store the three-dimensional feature information of all feature data nodes within the subspace block; Represents the set of real numbers, indicating that all elements in this feature data matrix are real numbers (i.e., the detection values ​​of time, microenvironment parameters, and gene expression during stem cell induction are all real numbers). This represents the number of feature data nodes within the local feature subspace block, i.e., the total number of stem cell-induced feature data points contained within this local region. The covariance matrix of this data matrix is ​​then calculated. The covariance matrix is ​​used to measure the degree of linear correlation between the three feature dimensions.

[0040] For covariance matrix To perform eigenvalue decomposition, first solve the characteristic equation. ,in Let be the eigenvalues ​​to be solved. For the covariance matrix For a 3x3 identity matrix (i.e., a 3×3 matrix with 1s on the diagonal and 0s elsewhere), the three eigenvalues ​​of the covariance matrix can be obtained by solving its characteristic equation. ; for each feature value ( Substituting the values ​​of 1, 2, and 3 into the homogeneous linear equation system, respectively, yields the following: ,in Let be the eigenvectors to be solved. Solving this homogeneous linear system of equations will yield the corresponding eigenvalues. eigenvectors , to obtain eigenvalues and the corresponding feature vectors Select the two eigenvectors with the largest eigenvalues. and As the principal component direction vectors of this local feature subspace block, these two vectors can characterize the main directions of feature distribution within the subspace block, reflecting the core feature correlation patterns of this local region. These two principal component direction vectors can capture the core linear correlation between the time evolution dimension, microenvironment parameter dimension, and gene expression dimension within this local feature subspace block, presenting the intrinsic correlation between time progression and changes in microenvironment parameters, and fluctuations in gene expression abundance within this local induced region. For example, in microenvironment parameters, when dissolved oxygen concentration increases, the corresponding gene expression abundance will also increase synchronously (e.g., when dissolved oxygen concentration increases from 5 mg / L to 8 mg / L, the corresponding gene expression abundance will increase). From 0.3 to 0.7); when the pH value decreased from 7.2 to 7.0, the corresponding gene expression abundance decreased from 0.6 to 0.2, showing a clear synergistic change trend; the advancement of the time evolution dimension (e.g., from the 1st hour to the 3rd hour of induction) will gradually enhance the correlation strength between microenvironment parameters and gene expression, so that small changes in microenvironment parameters can be quickly transmitted to the gene expression level, thereby accurately characterizing the synergistic effect of the three core feature dimensions of time evolution, microenvironment parameters, and gene expression during stem cell induction in this local region, as well as the intrinsic correlation pattern of mutual influence and mutual regulation among the three, clearly presenting the dynamic regulatory mechanism of stem cell induction in this local region. The principal component direction vector of each local feature subspace block is integrated with the corresponding covariance matrix structure to obtain the local spatial distribution parameter set, which contains the feature distribution law and correlation characteristics of each subspace block. Based on the principal component direction vector in the local spatial distribution parameter set, feature vector normalization operation is performed to construct a rotation alignment basis, that is, let the principal component direction vector be... Standardized vector The L2 norm of the principal component direction vectors is used to construct a rotation-aligned basis by concatenating two normalized principal component direction vectors with a unit vector perpendicular to these two vectors. Simultaneously, the spatial translation offset and feature dimension scaling factor are calculated by combining the covariance matrix structure. ,in It is the local mean vector calculated from all data nodes within a single local feature subspace block. The preset global baseline mean vector (a vector composed of the global means of each feature dimension); feature dimension scaling factor. The scaling factor is calculated as the ratio of the global standard deviation of the corresponding feature dimension to the standard deviation of the corresponding feature dimension of the local feature subspace block. The scaling factor for each of the three feature dimensions is calculated and then integrated to form a feature dimension scaling factor vector. By integrating the rotation alignment basis, spatial translation offset, and feature dimension scaling factor, we obtain the affine transformation parameter set.

[0041] Step 303: Using the rotation alignment basis and affine transformation parameter set, perform coordinate system rotation transformation and linear translation scaling on several local feature subspace blocks to complete the state space alignment and reconstruction, obtaining the aligned local feature tensor, specifically including: Based on the rotation alignment basis and affine transformation parameter set obtained in step 302, coordinate system rotation transformation and linear translation scaling are performed on each of the local feature subspace blocks obtained in step 301 to achieve state space alignment and reconstruction of each local feature subspace block, eliminate feature distribution offsets between different local subspace blocks, and ensure the consistency of feature distribution patterns in each subspace block. First, coordinate system rotation transformation is performed. Let the feature vector of a data node within a local feature subspace block be... The vector after rotation transformation ,in The inverse matrix of the rotation alignment basis constructed in step 302 is used to uniformly adjust the feature distribution direction of the local feature subspace block to the global reference direction through rotation transformation, eliminating the directional offset between different subspace blocks. After the rotation transformation is completed, linear translation scaling is performed. The vector after rotation transformation is multiplied element-wise with the feature dimension scaling coefficient vector, and then the scaled vector is added with the spatial translation offset to obtain the aligned feature vector. The above rotation transformation and linear translation scaling are performed on all data nodes within the local feature subspace block to obtain the aligned local feature subspace block. The same operation is then performed on all local feature subspace blocks in sequence to complete the state space alignment and reconstruction of all local feature subspace blocks, finally obtaining the aligned local feature tensor. The feature distribution direction, scale, and position of each local subspace block in this tensor are uniformly aligned, which can reflect the correlation law of global features during stem cell induction, while preserving the subtle feature fluctuations of each local region. In this aligned local feature tensor, all local feature subspace blocks have achieved uniform alignment in direction, scale, and position. Through the synergistic change law of the three core feature dimensions of time evolution, microenvironment parameters, and gene expression within each subspace block, the common feature correlation of different local regions during the entire stem cell induction process can be intuitively presented. At the same time, it clearly reflects the overall synergistic trend of changes in microenvironment parameters and fluctuations in gene expression in the global scope, and the global regulatory role of time evolution on the relationship between the two.

[0042] Step 304: For the aligned local feature tensors, perform structural recombination and cross-block splicing mapping along the original time evolution axis and feature dimension axis to construct a feature alignment mapping matrix representing global feature consistency and distribution balance, specifically including: For the aligned local feature tensor obtained in step 303, this tensor consists of several local feature subspace blocks that have been aligned in direction, scale, and position. Each subspace block has achieved alignment and standardization in three core dimensions: temporal evolution, microenvironment parameters, and gene expression. Although unified calibration has been completed, it is still in a scattered and independent state. It is necessary to construct a complete global feature mapping system through structural recombination, cross-block splicing, and global integration to achieve the synergistic unity of local subtle features and global overall laws. First, taking the time evolution axis of stem cell induction as the core line, all aligned local feature subspace blocks are temporally sorted. Strictly following the order of induction time, each subspace block is mapped to a specific time node to ensure the continuity of the time dimension. Low-sensitivity subspace blocks correspond to the stable stage in the early stage of induction, medium-sensitivity subspace blocks correspond to the transition stage in the middle stage of induction, and high-sensitivity subspace blocks correspond to the dynamic adjustment stage in the later stage of induction, which is completely matched with the sensitivity grading standard in step 300. Meanwhile, the feature dimensions of each subspace block are uniformly standardized, and the order of the three dimensions of time evolution, microenvironment parameters, and gene expression is clarified. The microenvironment parameter dimension (including specific detection indicators such as dissolved oxygen and pH value) and the gene expression dimension (including parameters such as expression level and fluctuation amplitude) are sorted according to the preset specifications to ensure that the feature dimensions of all subspace blocks are unified and the data format is consistent, forming a recombined three-dimensional tensor structure.

[0043] A cross-block splicing operation is performed to extract feature data from all local feature sub-blocks at each time point. These data are then categorized and spliced ​​according to feature dimensions. The time evolution dimension integrates the time point data of all sub-blocks to ensure the continuity of the induction timeline. The microenvironment parameter dimension integrates specific detection data such as dissolved oxygen concentration and pH value of each sub-block (e.g., dissolved oxygen concentration maintained at 5 to 8 mg / L, pH value stable at 7.0 to 7.5). The gene expression dimension integrates the expression level data of each sub-block, clearly presenting the expression differences in different sensitivity regions (gene expression level fluctuation range of 0.6 to 0.8 in high sensitivity regions, 0.3 to 0.6 in medium sensitivity regions, and 0.1 to 0.3 in low sensitivity regions).

[0044] After the stitching is completed, a feature alignment mapping matrix is ​​constructed. This matrix uses the time evolution axis as its vertical dimension and the three core feature dimensions as its horizontal dimension. Each row corresponds to an induced time node, and each column corresponds to a feature dimension. Each element in the matrix is ​​a standardized feature value, which can intuitively present the specific values ​​and fluctuations of different time nodes and different feature dimensions. This matrix can reflect the distribution of global features. If the standard deviation of the feature value fluctuation in the row corresponding to a certain time node is ≤0.2, it indicates that the global features are highly consistent and evenly distributed during that period. If the standard deviation is >0.2, it indicates that there is a local feature shift during that period, and the parameters need to be fine-tuned in combination with the sensitivity level of the corresponding subspace block. At the same time, this matrix can clearly quantify the fit between each local feature subspace block and the global features. The deviation between the mean of each subspace block feature and the mean of the global features is calculated. A deviation ≤0.1 is considered a high fit, a deviation between 0.1 and 0.3 is considered a moderate fit, and a deviation >0.3 is considered a low fit, requiring recalibration of the alignment parameters. Furthermore, the matrix can clearly present the correlation between local features and global features. The feature fluctuations of local subspace blocks always follow the global evolution law. The gene expression fluctuations of high-sensitivity subspace blocks have strong correlation and transmission. Abnormal fluctuations in their expression levels (such as sudden increases or decreases) will directly disrupt the distribution balance of global features, leading to a decrease in the balance of global features. On the other hand, the features of low-sensitivity subspace blocks have extremely strong stability. Their gene expression and microenvironment parameters always remain stable without significant fluctuations, ultimately achieving the synergistic unity of local subtle features and global overall laws.

[0045] This embodiment establishes dynamic benchmark partitioning and adaptive threshold rules based on sensitivity grading features to avoid the defects of fixed thresholds and lock the abnormal key regions induced by stem cells; through multi-dimensional sliding windows and dynamic parameter adjustment, it completes the accurate segmentation of local subspaces of multi-dimensional feature tensors; using principal component analysis to construct rotation alignment basis and transformation parameters, it eliminates feature redundancy and distribution offset; then, through coordinate rotation, translation and scaling, it achieves state space alignment of each local subspace, correcting feature direction, scale and position deviations; finally, through structural recombination and cross-block splicing, it constructs a feature alignment mapping matrix, taking into account both the preservation of local subtle fluctuation signals and the integration of global feature correlation rules, solving the problem of difficulty in simultaneously capturing local features and analyzing global correlations. In a preferred embodiment of the present invention, step 4 includes: Step 400: Define a sliding search window in the area where the matrix feature dimension axis intersects with the time evolution axis, extract the combination of two-dimensional feature sequences within the window coverage area to obtain the target feature pair data block; perform monotonically increasing rank mapping on each target feature pair within the target feature pair data block to convert the original continuous values ​​into an integer rank sequence, obtaining the target feature pair rank transformation sequence set, specifically including: Using the feature alignment mapping matrix constructed in step 304 as the processing object, the rows of this matrix correspond to different time stamps of stem cell induction (time evolution axis), and the columns correspond to the three core feature dimensions of time evolution, microenvironment parameters, and gene expression (feature dimension axis). Each element in the matrix is ​​the aligned feature value of the corresponding time and feature dimension. First, within the two-dimensional data region formed by the intersection of the feature dimension axis and the time evolution axis of this feature alignment mapping matrix, a sliding search window is defined, and the size of the sliding search window is set to... × ,in The number of timestamps covered by the representative window on the time evolution axis, with a value of 5 to 10 consecutive timestamps (which can be flexibly adjusted according to the temporal resolution of stem cell induction to ensure that the continuous fluctuation trend of features can be captured). The number of feature dimensions covered by the window on the feature dimension axis is represented by a value of 2 (that is, only two feature dimensions are selected each time to form a feature pair, avoiding the deviation in the determination of the relationship caused by the mixing of multiple dimensions).

[0046] The sliding search window employs a point-by-point sliding method with a sliding step size of 1. This means that after each slide, the window moves one timestamp on the time evolution axis and one feature dimension on the feature dimension axis, ensuring coverage of the entire two-dimensional data region of the feature alignment mapping matrix without missing any data nodes. For each sliding search window, the two-dimensional feature sequence combinations within its coverage area are extracted; that is, all timestamp feature values ​​corresponding to the two feature dimensions within the window are used to form target feature pair data blocks. The dimension of each target feature pair data block is... ×2, where each row corresponds to a timestamp and each column corresponds to a feature sequence of a feature dimension.

[0047] For each target feature pair within a data block, perform a monotonically increasing rank mapping operation. For a certain dimension feature sequence in the target feature pair (let this sequence be...), ,in For the first (Feature values ​​corresponding to each timestamp) First, sort all elements in the sequence in ascending order to obtain the sorted sequence. ,in Assign an integer rank to each sorted element. If elements have the same value, use the average rank method to assign the rank. Elements with the same value have the same rank. ,in The index of the first element in the sorted sequence is given; each element in the original feature sequence is replaced with its corresponding integer rank to obtain the rank sequence of the feature dimension; the rank sequences of the two feature dimensions are combined to form the target feature pair rank transformation sequence; after integrating the rank transformation sequences of all target feature pairs, the target feature pair rank transformation sequence set is obtained.

[0048] Step 401: Based on the target feature pair rank transformation sequence, calculate the squared rank difference between each dimension of the feature sequence in the target feature pair at the same sampling point. Perform nonlinear decay weighted aggregation operation through a local window smoothing weight function to solve for the local rank correlation degree value characterizing the monotonic trend correlation strength of the sequence, and obtain the intermediate result of the local rank correlation calculation, specifically including: Based on the target feature pair rank transformation sequence set obtained in step 400, for each target feature pair rank transformation sequence, the rank difference between each dimension of the feature sequence in the target feature pair at the same sampling point is extracted, and the rank difference is squared to obtain the rank difference square sequence. Let a certain target feature pair rank transformation sequence be... ,in , These are the rank sequences of the two feature dimensions, respectively. If the number of time sampling points is [number], then the number of sampling points at the same time [is] [number]. ( =1,2,..., The square of the corresponding rank difference is This yields the sequence of squared rank differences. A local window smoothing weight function is introduced to perform nonlinear decaying weighted aggregation on the rank difference squared sequence. This local window smoothing weight function uses a Gaussian weight function, and its specific expression is as follows: ,in This is the relative position index within the local window (values ​​range from -1, 0, 1, corresponding to the current sampling point, the previous sampling point, and the next sampling point). This is a smoothing coefficient, ranging from 0.8 to 1.2 (adjusted according to the severity of feature fluctuations to ensure that the core correlation trend of the features is preserved after smoothing, while eliminating random noise interference).

[0049] The weighted aggregation operation first defines a local smoothing window with a size of 3 (covering the current time sampling point and one sampling point before and after it). For each time sampling point... Extract the squared rank difference within its local smoothing window. (when When =1, only take The weights are 0.6 and 0.4 respectively; when = At that time, only take (The weights are 0.4 and 0.6 respectively). The square of each rank difference is multiplied by its corresponding weight to obtain the weighted square of the rank difference. The weighted squares of the rank differences are then summed to obtain the local rank correlation value, which characterizes the monotonicity of the sequence correlation at that time sampling point. (The summation range of the boundary sampling points is adjusted according to the boundary processing rules described above.) The smaller the local rank correlation value, the more consistent the monotonic change trends of the two feature sequences are, and the stronger the correlation. The larger the local rank correlation value, the greater the difference in the monotonic change trends of the two feature sequences are, and the weaker the correlation. After integrating the local rank correlation values ​​corresponding to all time sampling points, the intermediate result of the local rank correlation calculation is obtained.

[0050] Step 402 involves performing standardization and normalization on the intermediate results of the local rank correlation calculation, mapping the local rank correlation values ​​to local Spearman correlation coefficients to obtain the distribution set of local Spearman correlation coefficients. Specifically, this includes: performing standardization and normalization on the intermediate results of the local rank correlation calculation obtained in step 401, mapping all local rank correlation values ​​to the interval [-1, 1] to obtain the local Spearman correlation coefficients, and first calculating the global maximum value of the local rank correlation values. and global minimum ,in , , This represents the total number of local rank correlation values. A linear normalization formula is used to assign each local rank correlation value... Mapped to local Spearman correlation coefficient The specific formula is as follows ,in The function of `-1` is to set the baseline starting point of the normalization result to -1 (the lowest value of the Spearman correlation coefficient), ensuring that the mapped coefficients can cover the complete interval [-1, 1]. The function of `2` is to magnify the range of normalized values ​​by a factor of 2. The smaller the local rank correlation value (i.e., the smaller the rank difference between the two feature sequences), the closer the corresponding local Spearman correlation coefficient is to 1, indicating a stronger correlation; the larger the local rank correlation value (i.e., the greater the rank difference between the two feature sequences), the closer the corresponding local Spearman correlation coefficient is to -1, indicating a weaker correlation. After integrating all local Spearman correlation coefficients, a local Spearman correlation coefficient distribution set is obtained. This set corresponds one-to-one with the timestamps of the time evolution axis, which can intuitively reflect the correlation strength of target feature pairs at different time points.

[0051] Step 403: Each local Spearman correlation coefficient in the local Spearman correlation coefficient distribution set is compared with a preset threshold E. Corresponding feature nodes whose local Spearman correlation coefficients cross the preset threshold E are selected to obtain a high-confidence feature node index set. Based on the high-confidence feature node index set, the corresponding feature data is extracted, and topology aggregation and redundancy pruning are performed to obtain a key regulatory factor feature set, specifically including: Set preset threshold This threshold is determined based on the actual needs of stem cell induction and the statistical law of feature association strength, with a value ranging from 0.6 to 0.8. It is used to determine the confidence level of feature association. When the absolute value of the local Spearman correlation coefficient is greater than or equal to the preset threshold... When the correlation between corresponding feature nodes is high, it is determined that the relationship has high confidence; otherwise, it has low confidence. The local Spearman correlation coefficients obtained in step 402 are compared with a preset threshold. Perform a bit-by-bit comparison and judgment, if If the local Spearman correlation coefficient is high, then the feature node corresponding to that feature node is marked as a high-confidence feature node, and the index information of that feature node (including timestamp index and feature dimension index) is recorded; if If a feature node is not found to be high-confidence, it is discarded and no further processing is performed. The index information of all high-confidence feature nodes is then integrated to obtain the high-confidence feature node index set.

[0052] Based on the high-confidence feature node index set, corresponding feature data is extracted from the feature alignment mapping matrix constructed in step 304. The extraction process follows the timestamp index and feature dimension index in the index set to extract the feature values ​​corresponding to each high-confidence feature node, forming a high-confidence feature dataset. Topological aggregation processing is performed on the high-confidence feature dataset. Taking the feature dimension as the core, high-confidence feature nodes with different timestamps under the same feature dimension are aggregated to construct a feature dimension-timestamp topological association network. Nodes are high-confidence feature nodes, and edges represent the association relationships between adjacent timestamp feature nodes under the same feature dimension. Through this topological association network, continuous associated features of the same feature dimension are integrated to form aggregated features at the feature dimension level.

[0053] Redundancy pruning is performed on the aggregated feature data. A variance threshold method is used to remove redundant features. First, the variance of each aggregated feature is calculated. A redundancy threshold is set (ranging from 0.01 to 0.05). If the variance of a certain aggregated feature is less than the redundancy threshold, it indicates that the fluctuation range of that aggregated feature is extremely small, with no significant feature difference, and it is judged as a redundant feature and removed. If the variance is less than the redundancy threshold, it indicates that the variance of that aggregated feature is less than the redundancy threshold, and it is judged as a redundant feature and removed. If a redundancy threshold is reached, the aggregated feature is retained as a key feature. All retained aggregated features are then integrated to obtain a set of key regulatory factor features, which contains core features with high-confidence associations during stem cell induction.

[0054] This embodiment avoids numerical noise and dimensional scale differences by using sliding window retrieval and rank mapping, highlighting the monotonic change trend of features. It uses local window weighted aggregation to preserve local association details while suppressing single-point anomalies, accurately quantifying the temporal association strength of features. After normalization mapping to Spearman correlation coefficient, it unifies the quantification standard, adapts to the nonlinear fluctuation law of features, and improves the reliability of association judgment. Then, through threshold screening, topological aggregation, and redundancy pruning, it removes invalid and redundant features, efficiently screening out the core key regulatory factors that can characterize the intrinsic regulatory relationship between temporal evolution, microenvironment parameters, and gene expression.

[0055] In a preferred embodiment of the present invention, step 5 includes: Step 500: Initialize the tree structure hyperparameters and random subsampling strategy of the isolated forest anomaly detection network, load the key regulatory factor feature set into the multidimensional feature input layer, and construct the initial root node set to be segmented, specifically including: The key regulatory factor feature set obtained in step 403 is used as the processing object. This set contains core features with high-confidence correlations during stem cell induction and is the core data foundation for locating abnormal induction conditions. First, the construction of the isolated forest anomaly detection network is defined. This network is based on an unsupervised anomaly detection algorithm and is constructed by combining the high-dimensionality and temporal characteristics of key regulatory factor features during stem cell induction. The core is an ensemble learning network composed of multiple isolated trees. Its construction logic fits the core requirements of anomaly detection, that is, abnormal feature nodes are more easily isolated than normal nodes because they deviate from the normal feature distribution. By integrating multiple isolated trees, the segmentation bias of a single decision tree is avoided. At the same time, it adapts to the distribution characteristics of key regulatory factor features and solves the problems of poor performance and low computational efficiency in adapting to high-dimensional data. The construction of this network follows the core process of the isolated forest algorithm, that is, multiple isolated trees are constructed by random sampling and recursive segmentation, and then integrated to form a forest structure, which is used to capture abnormal patterns in the key regulatory factor feature set.

[0056] Based on the isolated forest anomaly detection network constructed above, its tree structure hyperparameters and random subsampling strategy are initialized, clarifying the specific values ​​and setting basis of each hyperparameter. Specifically, the tree structure hyperparameters are set as follows: the number of decision trees (forest size) is set to 100, and the preset maximum tree depth of a single decision tree is set to... ,in The depth setting represents the total number of feature nodes in the feature set of key regulatory factors. This depth setting can avoid overfitting or underfitting of the decision tree and ensure that abnormal feature patterns can be accurately captured. The minimum number of samples for a leaf node is set to 2. That is, when the number of samples for a certain node is less than 2, the segmentation stops to ensure that the leaf node has a certain degree of isolation purity.

[0057] The random subsampling strategy is specifically set as follows: each time a single decision tree is constructed, samples are randomly drawn from the key regulatory factor feature set, with a sampling ratio of 0.7 and a sampling quantity not exceeding 256 (if the total number of feature nodes in the key regulatory factor feature set is less than 256, all nodes are sampled). This sampling strategy reduces computation while preserving the core distribution information of the features, avoiding the influence of single-sample bias on anomaly detection results. After the hyperparameters and random subsampling strategy are initialized, the key regulatory factor feature set is loaded into the multidimensional feature input layer of the isolated forest anomaly detection network. This input layer is a fundamental component of the isolated forest anomaly detection network, and its dimension is consistent with the feature dimension of the key regulatory factor feature set, corresponding to the three core feature dimensions of time evolution, microenvironment parameters, and gene expression, and is used to receive key regulatory factor feature data. After loading, the entire key regulatory factor feature set is used as the initial root node set to be segmented. Each initial root node corresponds to one key regulatory factor feature node, and the dimension of the initial root node set is consistent with the dimension of the key regulatory factor feature set. Each root node contains all feature information of that feature node (timestamp index, feature value, feature dimension index).

[0058] Step 501: Based on the initial set of root nodes to be segmented, construct a random segmentation hyperplane in the multidimensional feature space, perform a recursive hyperplane space partitioning operation, and iteratively binary-cut the initial set of root nodes to be segmented until the leaf node isolation purity condition is met or the preset maximum tree depth is reached, resulting in a multi-level decision tree forest structure, specifically including: For each decision tree's initial root node, a feature dimension is randomly selected within the multidimensional feature space (temporal evolution, microenvironmental parameters, and gene expression three-dimensional space). ( ∈{1,2,3}, corresponding to the three core feature dimensions respectively, and then a segmentation threshold is randomly generated within the value range of this feature dimension. Segmentation threshold The value range is between the minimum and maximum values ​​of all feature values ​​under this feature dimension. Specifically, it is a random real number within this interval. For example, if the minimum value of all feature values ​​under this feature dimension is 0.1 and the maximum value is 0.9, then the segmentation threshold is... Any real number between 0.1 and 0.9 can be randomly selected to ensure that the segmentation threshold falls within the actual numerical range of that feature dimension, effectively dividing the feature space. Based on the selected feature dimension... and segmentation threshold Construct a randomly partitioning hyperplane, the equation of which is: = ,in Given all feature values ​​under this feature dimension, the hyperplane divides the feature space corresponding to the current node into two subspaces.

[0059] In the current node, the feature dimensions The value is less than or equal to the segmentation threshold. The feature nodes are divided into left child nodes; feature dimensions The value is greater than the segmentation threshold The feature node is assigned to the right child node, completing one node segmentation. For the resulting left and right child nodes, the above random hyperplane construction and bisection operation is repeated recursively until one of the following two termination conditions is met: First, the isolation purity of the leaf node corresponding to the current node reaches a preset standard, where isolation purity = ... / ,in This is the number of abnormal candidate nodes in the node whose feature values ​​deviate from those of other feature nodes within the same node by a preset deviation threshold. The preset deviation threshold is set to 1.5 times the standard deviation of all feature values ​​within the node. The first condition is the total number of feature nodes for that node. If the isolation purity is greater than or equal to 0.8, the node is deemed to meet the isolation purity condition, and the segmentation stops. The second condition is that the depth of the current decision tree reaches the maximum tree depth preset in step 500. At this point, regardless of whether the isolation purity meets the standard, the segmentation is stopped. Following the above process, 100 decision trees are constructed. After all decision trees are integrated, a multi-level decision tree forest structure is formed. In this structure, each decision tree contains a root node, intermediate splitting nodes, and leaf nodes. Each leaf node corresponds to a set of key regulatory factor feature nodes with similar feature distributions.

[0060] Step 502: Traverse the distribution paths of each feature node in the key regulatory factor feature set within the multi-level decision tree forest structure, calculate the number of splitting edges traversed from the root node to the terminal leaf node, solve for the average isolation path length across the tree, and perform unbiased estimation correction to obtain the anomaly score vector sequence, specifically including: For each feature node in the key regulatory factor feature set, 100 decision trees in the forest are traversed sequentially. In each decision tree, starting from the root node, the direction (left child node or right child node) of the feature node is determined based on the feature values ​​of each dimension, until the terminal leaf node is reached. The number of split edges that the feature node passes through from the root node to the leaf node in the decision tree is recorded. This number of split edges is called the split edge number parameter. Each decision tree corresponds to a unique split edge number parameter, and 100 decision trees correspond to 100 split edge number parameters, which correspond to the number of split edges that the feature node passes through in each decision tree.

[0061] The average cross-tree isolation path length of the feature node is calculated throughout the entire forest. This average reflects the average isolation difficulty of the feature node in the forest. Abnormal nodes, due to their feature distribution deviating from the normal range, typically have shorter average isolation path lengths than normal nodes. To eliminate the influence of sample size on the average isolation path length, the average cross-tree isolation path length is calculated. Perform unbiased estimation corrections by introducing a correction function. ,in For the first harmonic number of order, , Here, is Euler's constant, with a value of 0.5772. Total number of feature nodes. Corrected average isolation path length. ,when When =1, the correction function =0, at this time the corrected average isolation path length is directly set to =1.

[0062] The anomaly score of this feature node is calculated based on the corrected average isolation path length. The abnormal score ranges from (0,1). The closer the score is to 1, the more likely the feature node is to be an abnormal node; the closer the score is to 0, the more likely the feature node is to be a normal node. Arrange the abnormal scores of all feature nodes in the order of their feature node indices to obtain the abnormal score vector sequence.

[0063] Step 503 involves comparing each dimension component of the abnormal score vector sequence with a preset threshold B, filtering out target feature nodes whose abnormal scores exceed the preset threshold B, and obtaining an abnormal feature node index set. Specifically, this includes setting a preset threshold B for abnormal judgment. This threshold is determined based on the statistical regularity of abnormal features during stem cell induction and the actual application requirements, with a value range of 0.7 to 0.8. Its core function is to distinguish between normal feature nodes and abnormal feature nodes. The higher the value of threshold B, the higher the confidence of the abnormal judgment and the fewer abnormal nodes are filtered out; the lower the value of threshold B, the wider the range of abnormal judgment, which can avoid missing potential abnormal nodes. Each component of the abnormal score vector sequence obtained in step 502 (i.e., the abnormal score of each feature node) is compared bit by bit with a preset threshold B. If the abnormal score of a feature node is greater than or equal to B, the feature node is determined to be an abnormal feature node, and its index information (including feature dimension index, timestamp index, and position index in the key regulatory factor feature set) is recorded. If the abnormal score is less than B, the feature node is determined to be a normal feature node, and no further processing is performed; the index information of the node is discarded. The index information of all the identified abnormal feature nodes is integrated and sorted in chronological order by timestamp to obtain an abnormal feature node index set. This index set locates all abnormal nodes in the key regulatory factor feature set.

[0064] Step 504: The inverse mapping transformation matrix constructed during the multi-source data alignment processing stage is invoked to project the target feature nodes in the abnormal feature node index set from the abstract feature space back to the original microenvironment physicochemical parameter time-series coordinate system. The corresponding time-series timestamps, culture medium component concentration gradients, and environmental control parameters are analyzed to obtain the abnormal induction condition location coordinates, specifically including: The inverse mapping transformation matrix constructed in the multi-source data alignment processing stage (steps 300 to 304) is invoked. This inverse mapping transformation matrix is ​​the inverse of the feature alignment mapping matrix constructed in step 304, denoted as... Its dimension is consistent with the affine transformation matrix. Its function is to project the feature nodes in the abstract feature space (the standardized feature space where the key regulatory factors reside) backwards to the original microenvironment physicochemical parameter time-series coordinate system (i.e., the coordinate system at the time of original data acquisition, including the time evolution axis and the microenvironment physicochemical parameter axis), thus achieving a correspondence between abstract features and original actual parameters. Based on the abnormal feature node index set obtained in step 503, the feature vector of each abnormal feature node in the abstract feature space is extracted. ( (For the index of abnormal feature nodes), input the feature vector into the inverse mapping transformation matrix, and perform the back projection operation, i.e. ,in Let this be the coordinate vector of the anomalous feature node in the time-series coordinate system of the original microenvironment physicochemical parameters. The translation offset set in step 303 It is a feature dimension scaling coefficient vector. Through this operation, the abnormal feature nodes are accurately mapped from the abstract feature space to the original coordinate system.

[0065] The coordinate vector obtained by backprojection is analyzed to extract three key types of information: first, the specific timestamp aligned with the time sequence, clarifying the stem cell induction stage where the anomaly occurred; second, the concentration gradient of culture medium components, including the specific values ​​and changes in microenvironmental physicochemical parameters such as dissolved oxygen concentration and pH value; and third, the culture environment control parameters (such as culture temperature and stirring speed) at the corresponding timestamp. These three types of information are integrated into anomaly induction condition location coordinates in the format of (time sequence timestamp, dissolved oxygen concentration, pH value, culture temperature, stirring speed), with each anomaly feature node corresponding to one coordinate. Summarizing all coordinates provides a complete picture of the specific time of the anomaly occurrence, microenvironmental conditions, and control parameters, achieving precise location of the anomaly induction conditions.

[0066] This embodiment first optimizes the hyperparameter settings and sampling strategy of isolated forests to balance detection accuracy and computational efficiency, adapting to the feature distribution of key regulatory factors. Then, relying on a multi-level decision tree forest structure, it improves the robustness of anomaly identification and balances detection accuracy and computational efficiency. After unbiased correction of the average isolation path length, it eliminates the interference of sample quantity and accurately quantifies the degree of feature anomaly. By screening abnormal nodes through threshold comparison, it can adapt to different induction scenarios and improve the accuracy of anomaly screening. Finally, it uses an inverse mapping transformation matrix to realize the reverse projection of the feature space to the original physicochemical time-series coordinate system, which can not only identify abnormal nodes, but also accurately locate the time, microenvironment parameters and culture control conditions corresponding to the anomalies.

[0067] like Figure 2 As shown, embodiments of the present invention also provide a stem cell induction data analysis system, including: The encoding module is used to collect the time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams in the stem cell induction culture system. After feature vectorization encoding and multi-source data alignment processing, an initial multidimensional feature tensor is constructed. The fusion module performs a local neighborhood graph topological traversal on the initial multidimensional feature tensor, extracts the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix; based on the local discreteness metric matrix, it constructs a feature sequence temporal evolution gradient model, calculates the first-order difference slope and the second-order change curvature to obtain the change ratio correction coefficient, and performs a weighted mapping fusion with the change ratio correction coefficient and the local discreteness metric matrix to obtain the sensitivity graded index sequence; The reconstruction module is used to perform adaptive thresholding and state space alignment reconstruction on the initial multidimensional feature tensor according to the sensitivity-level index sequence to obtain the feature alignment mapping matrix. The calculation module is used to perform local rank nonlinear correlation calculation on the feature alignment mapping matrix, calculate the local Spearman correlation coefficient between target feature pairs, and determine the key regulatory factor feature set when the local Spearman correlation coefficient crosses the preset threshold E. The mapping module is used to input the key regulatory factor feature set into the isolated forest anomaly detection network for recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature node is reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinate.

[0068] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0069] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0070] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0071] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for analyzing stem cell induction data, characterized in that, The method includes: Step 1: Collect time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams from the stem cell induction culture system. After feature vectorization encoding and multi-source data alignment, construct an initial multidimensional feature tensor. Step 2: Perform a topological traversal of the local neighborhood graph on the initial multidimensional feature tensor to extract the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix; construct a feature sequence temporal evolution gradient model based on the local discreteness metric matrix, calculate the first-order difference slope and the second-order change curvature to obtain the change ratio correction coefficient, and perform a weighted mapping fusion of the change ratio correction coefficient and the local discreteness metric matrix to obtain the sensitivity graded index sequence; Step 3: Based on the sensitivity grading index sequence, perform adaptive threshold segmentation and state space alignment reconstruction on the initial multidimensional feature tensor to obtain the feature alignment mapping matrix; Step 4: Perform local rank nonlinear correlation calculation on the feature alignment mapping matrix to calculate the local Spearman correlation coefficient between target feature pairs. When the local Spearman correlation coefficient crosses the preset threshold E, determine the key regulatory factor feature set. Step 5: Input the key regulatory factor feature set into the isolated forest anomaly detection network to perform recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature nodes are reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinates.

2. The stem cell induction data analysis method according to claim 1, characterized in that, Step 1 includes: The time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams in the stem cell induction culture system were obtained. Bandpass filtering and outlier truncation were performed on the time-series sequences of microenvironmental physicochemical parameters, and logarithmic normalization and low signal-to-noise ratio feature filtering were performed on the whole-genome expression profile data streams to obtain the cleaned microenvironmental parameter sequences and gene expression profile sequences. Based on the microenvironment parameter sequence and gene expression profile sequence after cleaning, discrete culture component identifiers and continuous physicochemical indicators are projected onto a unified high-dimensional feature space, gene expression abundance sequences are converted into normalized numerical vectors, and vectorization encoding of heterogeneous data sources is completed to obtain a basic feature vector set. Based on the timestamp sampling interval attributes of each dimension in the basic feature vector set, the high-frequency sampling dimension component and the low-frequency sampling dimension component are identified, and the high-frequency sampling dimension component and the low-frequency sampling dimension component are subjected to time phase matching and sampling interval compensation to obtain the time synchronization feature pair. Receive time-synchronized feature pairs, stack them into multidimensional structures along the time evolution dimension, microenvironment parameter dimension and gene expression dimension, and perform cross-modal variance equalization scaling processing to construct the initial multidimensional feature tensor in the spatial coordinate system.

3. The stem cell induction data analysis method according to claim 2, characterized in that, A local neighborhood graph topological traversal is performed on the initial multidimensional feature tensor to extract the range distribution features of P data points within the neighborhood to construct a local dispersion measure matrix, including: Based on the initial multidimensional feature tensor, the target data node is set as the traversal anchor point in the feature space coordinate system. The spatial coordinates of the target data node are used as the topological center. A preset number of P adjacent feature data points are selected along the spatial distribution trajectory of the initial multidimensional feature tensor, where P is a preset positive integer greater than 1. The spatial position of the P adjacent feature data points is defined by the coordinate vector of the target data node and the preset neighborhood search radius, thus obtaining a set of local neighborhood feature points. For a set of local neighborhood feature points, numerical extreme value boundaries are extracted along each feature dimension and the dimension-level range vector is calculated to obtain the set of local range distributions of nodes. The range vectors of each dimension in the local range distribution set of nodes are standardized and mapped using a preset dimensional adaptive scaling factor to obtain a normalized range distribution matrix; For the normalized range distribution matrix, tensor structure reconstruction and variance aggregation operations are performed along the time series axis and the feature dimension axis to construct a local discreteness metric matrix that characterizes the intensity of local data fluctuations.

4. The stem cell induction data analysis method according to claim 3, characterized in that, A temporal evolution gradient model of the feature sequence is constructed based on the local discreteness metric matrix. The first-order difference slope and the second-order change curvature are calculated to obtain the change ratio correction coefficient. The change ratio correction coefficient is then fused with the local discreteness metric matrix through a weighted mapping to obtain the sensitivity-level index sequence, including: Based on the local discreteness metric matrix, a discrete time step difference sequence is constructed along the time evolution axis to establish a feature sequence time evolution gradient model that characterizes the changes in characteristic fluctuation rate and direction. In the feature sequence time-series evolution gradient model, the ratio of numerical increments between adjacent time steps is calculated to obtain the first-order difference slope sequence that characterizes the instantaneous rate of change of the data; a second difference operation is performed on the first-order difference slope sequence to extract the slope inflection feature; the dynamic increase or decrease of the feature change rate is quantified based on the second-order difference operator of the discrete time series to obtain the second-order change curvature sequence that characterizes the intensity of fluctuation. By fusing the first-order difference slope sequence and the second-order change curvature sequence, and using a nonlinear coupling mapping function to jointly encode the instantaneous rate of change and the degree of trend reversal, a change ratio correction coefficient characterizing the intensity of the dynamic response of local data is obtained. The change ratio correction coefficient and the local dispersion measure matrix are subjected to a multi-dimensional weighted tensor product operation to obtain the fused sensitivity measure feature matrix. Based on the preset grading threshold range, the feature values ​​in the sensitivity quantification feature matrix are discretized and mapped and grade labels are assigned to obtain a sensitivity grading index sequence carrying sensitivity level identifiers.

5. The stem cell induction data analysis method according to claim 4, characterized in that, Step 3 includes: The distribution characteristics of hierarchical identifiers in the sensitivity grading index sequence are analyzed, a dynamic benchmark partitioning strategy model driving data distribution bias correction is constructed, and an adaptive threshold decision rule set is obtained. Based on the adaptive threshold decision rule set, a multi-dimensional sliding window scan is performed on the initial multi-dimensional feature tensor, and the segmentation boundary parameters are dynamically adjusted to complete the adaptive threshold segmentation operation, resulting in several local feature subspace blocks. Based on several local feature subspace blocks, the principal component direction vectors of the feature distribution within each local feature subspace block are extracted and the corresponding covariance matrix structure is calculated to obtain the local spatial distribution parameter set; based on the principal component direction vectors in the local spatial distribution parameter set, feature vector standardization is performed to construct a rotation alignment basis, and the spatial translation offset and feature dimension scaling coefficient are calculated in combination with the covariance matrix structure to obtain the affine transformation parameter set; By using the rotation alignment basis and affine transformation parameter set to perform coordinate system rotation transformation and linear translation and scaling on several local feature subspace blocks, the state space alignment and reconstruction is completed, and the aligned local feature tensor is obtained. For the aligned local feature tensors, structural reorganization and cross-block splicing mapping are performed along the original time evolution axis and feature dimension axis to construct a feature alignment mapping matrix that represents the consistency and distribution balance of global features.

6. The stem cell induction data analysis method according to claim 5, characterized in that, Step 4 includes: A sliding retrieval window is defined in the area where the matrix feature dimension axis and the time evolution axis intersect. The combination of two-dimensional feature sequences within the window coverage area is extracted to obtain the target feature pair data block. Monotonically increasing rank mapping is performed on each target feature pair in the target feature pair data block to convert the original continuous values ​​into an integer rank sequence, thus obtaining the target feature pair rank transformation sequence set. Based on the target feature pair rank transformation sequence, the square of the rank difference between each dimension feature sequence in the target feature pair at the same sampling point is calculated. The nonlinear decay weighted aggregation operation is performed through the local window smoothing weight function to solve the local rank correlation degree value that characterizes the monotonic trend correlation strength of the sequence, and the intermediate result of the local rank correlation calculation is obtained. The intermediate results of the local rank correlation calculation are standardized and normalized to map the local rank correlation degree values ​​to the local Spearman correlation coefficients, thus obtaining the local Spearman correlation coefficient distribution set. Each local Spearman correlation coefficient in the local Spearman correlation coefficient distribution set is compared with a preset threshold E. Corresponding feature nodes whose local Spearman correlation coefficients cross the preset threshold E are selected to obtain a high-confidence feature node index set. Based on the high-confidence feature node index set, the corresponding feature data is extracted, and topology aggregation and redundancy pruning are performed to obtain the key regulatory factor feature set.

7. The stem cell induction data analysis method according to claim 6, characterized in that, Step 5 includes: The tree structure hyperparameters and random subsampling strategy of the isolated forest anomaly detection network are initialized, and the key regulatory factor feature set is loaded into the multidimensional feature input layer to construct the initial root node set to be segmented. Based on the initial set of root nodes to be segmented, a random segmentation hyperplane is constructed in the multidimensional feature space. The recursive hyperplane space partitioning operation is performed, and the initial set of root nodes to be segmented is iteratively bisected until the leaf node isolation purity condition is met or the preset maximum tree depth is reached, thus obtaining a multi-level decision tree forest structure. Traverse the distribution path of each feature node in the key regulatory factor feature set in the multi-level decision tree forest structure, calculate the number of split edges traversed from the root node to the terminal leaf node, solve the average isolation path length across the tree and perform unbiased estimation correction to obtain the anomaly score vector sequence. Each dimension component of the abnormal score vector sequence is compared with a preset threshold B to determine the abnormal feature nodes whose abnormal scores exceed the preset threshold B, thus obtaining the abnormal feature node index set. The inverse mapping transformation matrix constructed during the multi-source data alignment processing stage is invoked to project the target feature nodes in the abnormal feature node index set from the abstract feature space to the original microenvironment physicochemical parameter time series coordinate system. The corresponding time series timestamps, culture medium component concentration gradients and environmental control parameters are analyzed to obtain the abnormal induction condition location coordinates.

8. A stem cell induction data analysis system, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The encoding module is used to collect time-series sequences of microenvironmental physicochemical parameters and whole-genome expression profile data streams in stem cell induction culture systems. An initial multidimensional feature tensor is constructed by feature vectorization encoding and multi-source data alignment. The fusion module is used to perform a local neighborhood graph topological traversal on the initial multidimensional feature tensor and extract the range distribution features of P data points in the neighborhood to construct a local discreteness metric matrix. A feature sequence time-series evolution gradient model is constructed based on the local discreteness metric matrix. The first-order difference slope and the second-order change curvature are calculated to obtain the change ratio correction coefficient. The change ratio correction coefficient and the local discreteness metric matrix are then weighted and fused to obtain the sensitivity-level index sequence. The reconstruction module is used to perform adaptive thresholding and state space alignment reconstruction on the initial multidimensional feature tensor according to the sensitivity-level index sequence to obtain the feature alignment mapping matrix. The calculation module is used to perform local rank nonlinear correlation calculation on the feature alignment mapping matrix, calculate the local Spearman correlation coefficient between target feature pairs, and determine the key regulatory factor feature set when the local Spearman correlation coefficient crosses the preset threshold E. The mapping module is used to input the key regulatory factor feature set into the isolated forest anomaly detection network for recursive hyperplane spatial partitioning and isolation path length calculation to obtain anomaly score vector sequence. When the anomaly score vector sequence exceeds the preset threshold B, the anomaly feature node is reverse-mapped to the original parameter coordinate system to obtain the anomaly-induced condition location coordinate.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images