Method for adaptive control of process parameters for tumor organoid culture

Abstract

This invention relates to the field of intelligent control technology, specifically disclosing an adaptive control method for parameters in the tumor organoid culture process. The method constructs a causal structure knowledge base containing multiple nodes and directed causal paths based on historical culture data. It collects multimodal sensor data streams in real time and performs time-series state analysis based on the knowledge base, outputting a current culture state vector containing proliferation activity characterization values ​​and morphological feature characterization values. When the two show a divergent trend along a preset causal path, a counterfactual reasoning program is invoked to deduce the evolutionary trajectory under conditions of executing an intervention and not executing an intervention. The two trajectories are compared; if the former shows a positive deviation of a preset magnitude, an execution command is generated. The execution command adjusts the culture environment parameters and updates the knowledge base. This invention achieves accurate decision-making under multimodal contradictory signals through causal inference and counterfactual reasoning, improving the stability and repeatability of the tumor organoid culture process.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically to an adaptive control method for parameters in the tumor organoid culture process. Background Technology

[0002] Tumor organoids, as miniature tumor models cultured in vitro in three dimensions, can highly mimic the original tumor tissue in terms of structure, function, and genetic characteristics. In recent years, they have shown broad application prospects in tumor pathogenesis research, drug screening, and personalized precision medicine. However, the in vitro culture process of tumor organoids is characterized by high heterogeneity, strong dynamic evolution, and large batch-to-batch differences. Their growth status is affected by a variety of environmental parameters, including culture medium composition, fluid shear force, pH, dissolved oxygen, and the concentration of key growth factors.

[0003] Existing methods for controlling parameters in the tumor organoid culture process lack a decision-making mechanism based on causal inference when contradictory signals appear in multimodal sensing, making it impossible to distinguish the essential differences between the biologically meaningful state of exhaustion and sensor noise or physiological fluctuations. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive control method for parameters in the tumor organoid culture process, in order to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: An adaptive control method for parameters in tumor organoid culture includes the following steps: S1. Based on the temporal correlation between multimodal sensing data and organoid state during the historical culture process, a causal structure knowledge base is constructed, which includes multiple nodes and directed causal paths between nodes. The nodes include culture environment parameters, multimodal sensing signals and organoid morphological features. S2, during the target organoid culture process, multimodal sensor data streams are collected, and time-series state analysis is performed on each multimodal sensor data stream based on the causal structure knowledge base. The current culture state vector is output in real time, which includes proliferation activity characterization value and morphological feature characterization value. S3, when it is identified that the proliferation activity characterization value and the morphological characteristic characterization value show a divergent trend along the preset causal path, the counterfactual reasoning program is invoked, and based on the causal structure knowledge base and the current culture state vector, the evolution trajectory of the morphological characteristic characterization value in the future time window is deduced under the conditions of executing the intervention action and not executing the intervention action, respectively. S4. Compare the evolutionary trajectory under the condition of performing the intervention action with the evolutionary trajectory under the condition of not performing the intervention action. If the former shows a positive deviation of a preset magnitude compared with the latter, then generate the execution instruction for the intervention action. S5 executes the execution command to adjust the culture environment parameters and updates the intervention decision data and actual observation results to the causal structure knowledge base to optimize the inference accuracy of counterfactual reasoning in subsequent culture processes.

[0006] As a further aspect of the present invention: S1 specifically includes: Acquire multiple batches of historical culture data, which include culture environment parameters, multimodal sensor signals, and corresponding organoid morphological features collected at different time points. Construct a time-series data matrix by aligning all data by time. The time series data matrix is ​​divided into sliding time windows. In each time window, the conditional independence test is performed on any two nodes. If the dependency between the two nodes under the condition of other nodes is stable in multiple consecutive time windows, then a directed causal path is established between the two nodes. An initial causal structure diagram is generated based on all established directed causal paths, and the directionality of each causal path is verified in chronological order. Causal paths that conform to temporal logic are retained, forming a causal structure knowledge base for subsequent counterfactual reasoning.

[0007] As a further aspect of the present invention: S2 specifically includes: The real-time acquired multimodal sensor data streams are split according to sensor type, and mapping associations are established with the corresponding nodes in the causal structure knowledge base. Timing alignment is performed on each sensor data stream to form a synchronized node observation sequence. Based on the directed causal path between nodes in the causal structure knowledge base, starting from the node corresponding to the proliferation activity characterization value, we trace back along the causal path to the node corresponding to the morphological feature characterization value. During the tracing process, we perform consistency checks on the observation values ​​of the nodes along the way and filter out noisy observations that are contrary to the direction of the causal path. After filtering out noise, all node observations are weighted and fused according to their topological hierarchy in the causal structure knowledge base to generate the current culture state vector, where the proliferation activity characterization value and morphological feature characterization value are output as independent components of the current culture state vector.

[0008] As a further aspect of the present invention: S3 specifically includes: Extract proliferation activity characterization values ​​and morphological feature characterization values ​​from the current culture state vector, determine the starting and ending nodes of the divergence trend between proliferation activity characterization values ​​and morphological feature characterization values ​​based on the causal structure knowledge base, and locate the unique directed causal path between the starting and ending nodes. Using the real-time observation value of the starting node in the current cultivation state vector as the initial condition, the intervention action is transformed into an assignment perturbation of the downstream adjacent nodes of the starting node along the directional causal path. The state recursion value of the perturbation propagating along the causal path to the end node is calculated in sequence to obtain the first evolutionary trajectory under the condition of executing the intervention action. Keeping the real-time observation value of the starting node unchanged and without applying any assignment perturbation, the state recursion value of each node is calculated sequentially along the same directed causal path to obtain the second evolutionary trajectory under the condition of no intervention action.

[0009] As a further aspect of the present invention: obtaining the first evolutionary trajectory under the condition of executing the intervention action specifically includes: Based on the causal structure knowledge base, the response curves between the downstream adjacent nodes of the starting node and the starting node during the historical cultivation process are determined. The mapping relationship between the observed values ​​of the starting node and the assigned values ​​of the downstream adjacent nodes is extracted from the response curves, and the intervention action is converted into an assignment perturbation quantity that conforms to the mapping relationship. The assigned perturbation is superimposed on the baseline recursive function of the downstream adjacent node in the causal structure knowledge base to generate the state value of the first node after perturbation. The state value of the current node is then used as the input parameter of the next node along the directed causal path, and the calculation is performed step by step according to the recursive function corresponding to each node. After calculating the state value of each first-level node, the state value is corrected for deviation from the actual observed value of the corresponding node in the current culture state vector. The corrected state value is used as the starting condition for the next level of recursion until the end node is calculated, thus obtaining the first evolutionary trajectory.

[0010] As a further aspect of the present invention: S4 specifically includes: The first evolutionary trajectory and the second evolutionary trajectory are paired according to the same time point within the same future time window, and the positive deviation of the first morphological value corresponding to the first evolutionary trajectory relative to the second morphological value corresponding to the second evolutionary trajectory is calculated one by one at each time point. The number of consecutive time points in which the positive deviation exceeds a preset threshold is counted. When the number of consecutive time points reaches the preset counting threshold, it is determined that the first evolutionary trajectory shows a positive deviation of a preset magnitude compared to the second evolutionary trajectory. The judgment result is converted into an execution instruction, which includes an identifier of the type of intervention action and an execution intensity parameter determined by the proliferation activity characterization value extracted from the current culture state vector.

[0011] As a further aspect of the present invention: the process for determining the execution intensity parameter is as follows: Extract the proliferation activity characterization value from the current culture state vector, input the proliferation activity characterization value into the preset intensity mapping function, and divide the execution intensity into multiple continuously increasing intensity levels according to the level of the proliferation activity characterization value; Based on the type identifier of the intervention action, the intensity response curve corresponding to the corresponding intervention action is retrieved from the causal structure knowledge base. The intensity level determined in the first step is projected onto the intensity response curve to obtain the initial value of the corresponding execution intensity parameter. Using morphological feature values ​​as correction factors, the initial value of the execution intensity parameter is reverse-compensated and corrected. When the morphological feature values ​​are lower than the historical average, the initial value of the execution intensity parameter is increased. When the morphological feature values ​​are greater than or equal to the historical average, the final execution intensity parameter is obtained.

[0012] As a further aspect of the present invention: S5 specifically includes: Collect the actual observation results within the preset time window after the intervention action is executed, compare the actual observation results with the predicted values ​​at the corresponding time points of the first evolutionary trajectory in counterfactual reasoning point by point, and calculate the deviation between the actual value and the predicted value of each node. Based on the magnitude of the deviation, the connection strength between nodes on the directed causal path is corrected by backpropagation. The larger the deviation, the greater the reduction in the connection strength from the upstream node to the downstream node, and the corrected connection strength is fixed to the corresponding path in the causal structure knowledge base. The real-time observations, assigned perturbations, and actual observation results of the starting node in this intervention decision data are stored as new time-series samples in the causal structure knowledge base to update the fitting parameters of the response curves between each node.

[0013] The beneficial effects of this invention are: (1) By constructing a causal structure knowledge base containing directed causal paths and combining it with a counterfactual reasoning procedure, this invention can identify the true intervention target based on causal inference rather than correlation analysis when the proliferation activity characterization value and the morphological characteristic characterization value show a divergent trend. This avoids decision misjudgment caused by multimodal sensor signal conflict and improves the accuracy and timeliness of intervention action triggering.

[0014] (2) This invention updates the decision data and actual observation results of each intervention to the causal structure knowledge base in real time, corrects the connection strength between nodes by backpropagation of the deviation, and dynamically updates the fitting parameters of the response curve, thereby realizing the closed-loop iterative optimization of the control strategy, which continuously improves the deduction accuracy of counterfactual reasoning in the subsequent cultivation process, thereby enhancing the repeatability and stability of the cultivation process between different batches. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart of the adaptive control method for parameters in the tumor organoid culture process according to the present invention; Figure 2 This is a flowchart of the process of obtaining the first evolutionary trajectory in this invention; Figure 3 This is a flowchart of the process for determining the intensity parameters in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the present invention provides an adaptive control method for parameters in the tumor organoid culture process, comprising the following steps: S1. Based on the temporal correlation between multimodal sensing data and organoid state during the historical culture process, a causal structure knowledge base is constructed, which includes multiple nodes and directed causal paths between nodes. The nodes include culture environment parameters, multimodal sensing signals and organoid morphological features. S2, during the target organoid culture process, multimodal sensor data streams are collected, and time-series state analysis is performed on each multimodal sensor data stream based on the causal structure knowledge base. The current culture state vector is output in real time, which includes proliferation activity characterization value and morphological feature characterization value. S3, when it is identified that the proliferation activity characterization value and the morphological characteristic characterization value show a divergent trend along the preset causal path, the counterfactual reasoning program is invoked, and based on the causal structure knowledge base and the current culture state vector, the evolution trajectory of the morphological characteristic characterization value in the future time window is deduced under the conditions of executing the intervention action and not executing the intervention action, respectively. S4. Compare the evolutionary trajectory under the condition of performing the intervention action with the evolutionary trajectory under the condition of not performing the intervention action. If the former shows a positive deviation of a preset magnitude compared with the latter, then generate the execution instruction for the intervention action. S5 executes the execution command to adjust the culture environment parameters and updates the intervention decision data and actual observation results to the causal structure knowledge base to optimize the inference accuracy of counterfactual reasoning in subsequent culture processes.

[0019] In S1, based on the temporal correlation between multimodal sensing data and organoid states during the historical culture process, a causal structure knowledge base is constructed, containing multiple nodes and directed causal paths between nodes. Nodes include culture environment parameters, multimodal sensing signals, and organoid morphological features, specifically including: Step 1: When acquiring historical culture data from multiple batches, each batch of culture lasted 14 to 21 days. Culture environment parameters, multimodal sensor signals, and corresponding organoid morphological features were collected every 5 minutes. Culture environment parameters included culture medium perfusion rate, temperature, and carbon dioxide concentration. Multimodal sensor signals included impedance amplitude signals acquired by an impedance sensor and microscopic images acquired by an optical imaging system. Organoid morphological features were extracted from the microscopic images by performing image segmentation processing, including the number of buds, organoid diameter, and edge regularity. All collected data were aligned according to the acquisition time point to form a time-series data matrix with row vectors corresponding to time points and column vectors corresponding to different data types.

[0020] Step 2: The time-series data matrix is ​​divided into sliding time windows with a width of 6 hours and a step size of 1 hour, obtaining a subset of data points at several time points within each time window. For any two data columns representing nodes in this subset, the conditional independence between these two nodes is calculated given all other nodes. The conditional independence is determined using the partial correlation coefficient test. When the absolute value of the partial correlation coefficient is less than 0.1 and the significance probability is greater than 0.05, the two nodes are considered conditionally independent within that time window; otherwise, a dependency relationship exists. When the same pair of nodes is determined to have a dependency relationship in more than three consecutive time windows, a directed causal path is established between the two nodes. The direction is determined based on the temporal order of the observations of the two nodes within the time window, with the earlier-occurring node as the starting point of the causal path and the later-occurring node as the ending point.

[0021] Step 3: Summarize all established directed causal paths to generate an initial causal structure graph. Then, perform directionality verification on each directed causal path in the initial causal structure graph: extract all observation timestamps corresponding to the starting and ending nodes of the path in the time-series data matrix, and count the proportion of times the starting node's observation timestamp is earlier than the ending node's observation timestamp. If this proportion is greater than 95%, retain the direction of the causal path; if this proportion is less than or equal to 95%, mark the causal path as undirected and remove it from the causal structure graph. Solidify and store all the directed causal paths retained after directionality verification to form a causal structure knowledge base for counterfactual reasoning. This knowledge base is stored in a directed graph data structure, where each node corresponds to a representation value of a data type, and each directed edge corresponds to the causal transit relationship between two nodes.

[0022] In S2, during the target organoid culture process, multimodal sensor data streams are acquired. Based on a causal structure knowledge base, temporal state analysis is performed on each multimodal sensor data stream, and the current culture state vector is output in real time. The current culture state vector includes proliferation activity characterization values ​​and morphological feature characterization values, specifically including: Step 1: During the target organoid culture process, multimodal sensor data streams are acquired in real time. Impedance sensors acquire impedance amplitude signals once per minute, which serve as the raw data source for proliferative activity characterization values. An optical imaging system acquires microscopic images once every 6 hours; the number of buds extracted from these microscopic images after image segmentation serves as the raw data source for morphological feature characterization values. The real-time acquired impedance amplitude signals and microscopic images are streamed separately according to sensor type. Based on predefined node identifiers in a causal structure knowledge base, impedance amplitude signals are mapped to proliferative activity nodes, and microscopic images are mapped to morphological feature nodes. For the two sensor data streams with different acquisition frequencies, linear interpolation is performed on the observation values ​​of morphological feature nodes, using the acquisition time point of the impedance amplitude signal as a reference. The interpolation method involves taking the morphological feature observation values ​​from two adjacent actual observation time points, calculating a weighted average based on the time distance, and filling this average into the reference time points between the two actual observation time points, forming a time-synchronized node observation sequence.

[0023] Step 2: Based on the directed causal paths between nodes stored in the causal structure knowledge base, starting from the proliferating active node, trace back step by step along the directed edges to the morphological feature node, recording all intermediate nodes passed during the tracing process. For each node on the tracing path, perform local trend calculation on the sequence consisting of the node's observation at the current time point and the observations at the three adjacent time points: calculate the difference between the current observation and the previous observation, the difference between the previous observation and the observation two time points before that, and the difference between the observation two time points before that and the observation three time points before that. If the signs of the three differences are consistent, the current observation is determined to conform to the local change trend; if the signs of the three differences are not completely consistent, the current observation is determined to contain noise. For observations determined to be noise, replace the noise observation with the arithmetic mean of the observations at the two adjacent time points before that node to complete noise filtering.

[0024] Step 3: Based on the topological hierarchy of all nodes in the causal structure knowledge base, assign weight coefficients to each node. Nodes closer to the beginning of the causal path in the topological hierarchy are assigned a weight coefficient of 1.0. The weight coefficient decreases by 0.2 for each layer extending towards the end of the causal path; for example, the weight coefficient for the second layer node is 0.8, the weight coefficient for the third layer node is 0.6, and so on. Multiply the noise-filtered node observations by their corresponding weight coefficients and sum them to obtain the fusion total value. Then, multiply each node observation by its weight coefficient and divide by the fusion total value to obtain the normalized weight contribution value of each node. Combine the normalized weight contribution values ​​of all nodes to form the current culture state vector. The normalized weight contribution value corresponding to the proliferation activity node is output as the proliferation activity characterization value component, and the normalized weight contribution value corresponding to the morphological feature node is output as the morphological feature characterization value component.

[0025] Please see Figure 2 As shown, in S3, when it is identified that the proliferation activity characterization value and the morphological feature characterization value show a divergent trend along a preset causal path, the counterfactual reasoning program is invoked. Based on the causal structure knowledge base and the current culture state vector, the evolution trajectory of the morphological feature characterization value within the future time window is deduced under the conditions of executing the intervention action and not executing the intervention action, respectively. Specifically, this includes: Step 1: Extract proliferation activity and morphological feature representation values ​​from the current culture state vector and input them into a causal structure knowledge base for path matching. The causal structure knowledge base stores the causal relationships between all nodes in the form of a directed graph, with each node corresponding to a representation value of a data type. Traverse all directed paths in the causal structure knowledge base that start from the proliferation activity node and end at the morphological feature node. Calculate the theoretical transfer value of each intermediate node on each path. Calculate the difference between the theoretical transfer value and the actual observed value of the corresponding node in the current culture state vector. Select the path with the smallest sum of absolute differences as the unique directed causal path corresponding to the divergence trend. The starting node of this path is the proliferation activity node, the ending node is the morphological feature node, and the intermediate nodes on the path include metabolite concentration nodes and extracellular matrix deposition nodes.

[0026] Step 2: Extract the historical correspondence between the starting node and its downstream adjacent nodes based on the causal structure knowledge base. From the historical training data, select all time points where the starting node's observed value fluctuates within 10% of the current observed value. Extract the values ​​of the downstream adjacent nodes corresponding to these time points and use the arithmetic mean of these downstream adjacent node values ​​as the baseline value. Define the intervention action as adjusting the downstream adjacent node's value to 1.5 times the baseline value. Calculate the assignment perturbation based on the ratio between the current observed value of the starting node and the historical observed values ​​of the starting node. The formula for calculating the assignment perturbation is as follows: ;in, This represents the assigned perturbation amount. This represents the preset standard intervention perturbation amount, which is the average change in the values ​​assigned to downstream adjacent nodes when similar interventions were performed during historical training processes. This represents the real-time observation value of the starting node. This represents the average observed value of the starting node under the corresponding intervention conditions during the historical cultivation process. The calculated perturbation is used as a specific quantitative expression of the intervention action along the causal path.

[0027] Step 3: Retrieve the baseline recursive function corresponding to the downstream adjacent nodes from the causal structure knowledge base. This baseline recursive function is a linear function obtained by least squares fitting based on the scatter distribution between the observed values ​​of the starting node and the assigned values ​​of the downstream adjacent nodes in the historical training data. The assigned perturbation is then superimposed on the calculation result of the baseline recursive function to generate the perturbed state value of the first node. The formula for calculating the perturbed state value is as follows: ;in, This represents the state value of the first node after the disturbance. This represents the baseline state value calculated by substituting the real-time observations from the starting node into the baseline recursive function. This represents the assigned perturbation amount. and Indicates the weighting coefficient. The value is 0.6. The value is 0.4. and The sum is 1. The value of this weighting coefficient is determined based on the goodness of fit of the baseline recursive function in the historical culture data; the higher the goodness of fit, the greater the weighting coefficient. The larger the value, the lower the value. The larger the value, the better.

[0028] Step 4: Using the state value of the first node after the disturbance as the state value of the current node, and following the directed causal path, use the state value of the current node as the input parameter for the next node. For each subsequent node on the path, retrieve the corresponding baseline recursive function from the causal structure knowledge base. This baseline recursive function is fitted based on the historical correspondence between the observation values ​​of the preceding nodes and the observation values ​​of the current node. Substitute the state value of the preceding node into the baseline recursive function to calculate the recursive state value of the current node. After completing the state value calculation for each level of node, compare the recursive state value of that node with the actual observation value of the corresponding node in the current culture state vector, calculate the difference between the two, and add 30% of this difference as a correction to the recursive state value to obtain the corrected state value. Use the corrected state value as the input parameter for the next level of recursion. Repeat the above process until the final node is calculated. Arrange the final state values ​​of the final node sequentially according to the time points within the future time window to form the first evolutionary trajectory under the condition of executing the intervention action.

[0029] Step 5: Keep the real-time observations of the starting node unchanged and do not apply any perturbations; that is, set the perturbation value to 0. Following the step-by-step recursive calculation method described in Step 4, starting from the starting node, calculate the state recursive values ​​of each node sequentially along the same directed causal path. Perform bias correction processing on each node until the endpoint is reached. Arrange the state recursive values ​​of the endpoint at each time point sequentially to form the second evolutionary trajectory under the condition of no intervention.

[0030] Please see Figure 3 As shown, in S4, the evolutionary trajectory under the condition of performing an intervention action is compared with the evolutionary trajectory under the condition of not performing an intervention action. If the former deviates positively by a preset magnitude compared to the latter, an execution instruction for the intervention action is generated, specifically including: Step 1: Align the first and second evolutionary trajectories using the same future time window, set to 12 hours from the current time. Take one time point every hour, resulting in 13 time points. For each time point, extract the first morphological value of the first evolutionary trajectory and the second morphological value of the second evolutionary trajectory. The positive deviation is calculated as follows: subtract the second morphological value from the first morphological value, divide the difference by the second morphological value, and multiply the result by 100% to obtain the positive deviation magnitude at that time point. If the first morphological value is less than the second morphological value, the positive deviation magnitude is counted as zero.

[0031] Step 2: Set the positive deviation threshold to 15% and the consecutive time point counting threshold to 3 time points. Starting from the first time point of the future time window, sequentially determine whether the positive deviation at each time point is greater than 15%. When a time point with a positive deviation greater than 15% is detected, start counting and continue to judge adjacent time points; if the positive deviation at adjacent time points is still greater than 15%, the count is incremented; if the positive deviation at adjacent time points is less than or equal to 15%, the count is reset to zero, and the detection starts again. When 3 consecutive time points are counted, it is determined that the first evolutionary trajectory shows a positive deviation of the preset magnitude compared to the second evolutionary trajectory. If 3 consecutive time points are not counted by the end of the future time window, it is determined that no positive deviation of the preset magnitude is shown.

[0032] Step 3: Extract the proliferation activity characterization value from the current culture state vector. The value ranges from 0 to 100. The intensity mapping function divides this range into five intensity levels: Level 1 corresponds to a proliferation activity characterization value greater than 0 and less than or equal to 20; Level 2 corresponds to a value greater than 20 and less than or equal to 40; Level 3 corresponds to a value greater than 40 and less than or equal to 60; Level 4 corresponds to a value greater than 60 and less than or equal to 80; and Level 5 corresponds to a value greater than 80 and less than or equal to 100. Each intensity level has a preset baseline intensity parameter, with the baseline intensity parameters for Levels 1 to 5 being 0.2, 0.4, 0.6, 0.8, and 1.0, respectively.

[0033] Based on the type identifier of the intervention action, the corresponding intensity response curve is retrieved from the causal structure knowledge base. This intensity response curve is an increasing curve describing the mapping relationship between the baseline intensity parameter and the actual execution intensity parameter. The five key points on the curve correspond to the baseline intensity parameter and the corresponding initial value of the execution intensity parameter for the five intensity levels, respectively. The initial value of the execution intensity parameter for the first intensity level is 0.3, for the second intensity level it is 0.5, for the third intensity level it is 0.7, for the fourth intensity level it is 0.9, and for the fifth intensity level it is 1.1. The initial value of the execution intensity parameter corresponding to the intensity level determined in the first step is used as the initial value of the current execution intensity parameter.

[0034] Morphological feature values ​​at the same time point during the previous three culture processes are extracted from historical culture data, and their arithmetic mean is calculated as the historical average. The morphological feature values ​​in the current culture state vector are compared with the historical average: when the morphological feature value is lower than the historical average, the initial value of the execution intensity parameter is multiplied by 1.2 to obtain the final execution intensity parameter; when the morphological feature value is greater than or equal to the historical average, the initial value of the execution intensity parameter remains unchanged and is used as the final execution intensity parameter.

[0035] Step 4: Combine the intervention action type identifier with the final determined execution intensity parameter to form an execution command. The execution command data format is structured data containing two fields: the first field is the intervention action type code, and the second field is the execution intensity parameter value. This execution command is then output to the culture environment control unit for adjusting the corresponding culture environment parameters.

[0036] In S5, execution instructions are used to adjust culture environment parameters and update the intervention decision data and actual observation results to the causal structure knowledge base to optimize the inference accuracy of counterfactual reasoning in subsequent culture processes. Specifically, this includes: Step 1: After the intervention is performed, a preset time window is set for the next 24 hours. Within this time window, actual observation values ​​for each node are collected every 2 hours. The actual observation values ​​for proliferating active nodes are obtained by taking the arithmetic mean of 11 data points collected by an impedance sensor at a frequency of once per minute, and then taking the data 5 minutes before and after the 2-hour time point. The actual observation values ​​for morphological feature nodes are obtained by extracting the data from microscopic images acquired by an optical imaging system at that time point after image segmentation. The actual observation values ​​at each time point are compared point-by-point with the predicted values ​​at the corresponding time points of the first evolutionary trajectory in counterfactual reasoning. The deviation is calculated as follows: subtract the predicted value from the actual observation value, divide the difference by the predicted value, and then multiply the result by 100% to obtain the percentage deviation for that node at that time point. If the predicted value is zero, the deviation is directly taken as the actual observation value.

[0037] Step 2: For each directed edge in the directed causal path, the initial connection strength is set to 1.0. Starting from the endpoint node, a backpropagation correction is performed step-by-step towards the starting node. For the endpoint node, the arithmetic mean of the deviations at all time points within the next 24-hour window is taken as the node's mean deviation. The mean node deviation is divided by 100 to obtain the attenuation coefficient. When the attenuation coefficient is greater than 0.5, 0.5 is used as the attenuation coefficient. The difference between the initial connection strength multiplied by 1 and the attenuation coefficient is used to obtain the corrected connection strength. The corrected connection strength is used as the new connection strength of the directed edge and is used to calculate the mean deviation of the upstream nodes: the mean deviation of the upstream nodes is equal to the mean deviation of the current node multiplied by the corrected connection strength. The above process is repeated to correct the connection strength of each upstream node in turn until the correction reaches the downstream adjacent node of the starting node. All corrected connection strengths are permanently stored in the corresponding directed edges in the causal structure knowledge base.

[0038] Step 3: The real-time observations of the starting node, the assigned perturbation values, and the actual observations at each time point within the next 24-hour window from the data used in this intervention decision are compiled into a new time-series sample. The historical sample set upon which the response curve between the starting node and its downstream neighboring nodes is based is retrieved from the causal structure knowledge base. This historical sample set contains the most recent 10 batches of culture data, each batch including the starting node observation and the corresponding downstream neighboring node assignment. The new time-series sample is added to this historical sample set, and the oldest sample in the set is removed, keeping the total number of samples below 10. Based on the updated sample set, the fitting parameters of the response curve are recalculated: using the least squares method, with the starting node observation as the independent variable and the downstream neighboring node assignment as the dependent variable, the coefficients of the first-order term and the constant term of the fitted line are calculated. The newly calculated coefficients of the first-order term and the constant term replace the corresponding response curve parameters in the causal structure knowledge base, completing the update of the response curve.

[0039] The working principle of this invention is as follows: First, based on the temporal correlation between multimodal sensor data and organoid states during the historical culture process, a causal structure knowledge base is constructed, including nodes such as culture environment parameters, multimodal sensor signals, and organoid morphological features, as well as their directed causal paths. Then, during the target organoid culture process, multimodal sensor data streams are collected in real time, and temporal state analysis is performed based on this causal structure knowledge base, outputting a current culture state vector containing proliferation activity characterization values ​​and morphological feature characterization values. When it is identified that the proliferation activity characterization values ​​and morphological feature characterization values ​​show a divergent trend along a preset causal path, counterfactual reasoning is invoked. The program, based on a causal structure knowledge base and the current cultivation state vector, deduces the evolution trajectory of morphological feature representation values ​​within a future time window under conditions of implementing an intervention and not implementing an intervention. The evolution trajectories under the two conditions are compared. If the evolution trajectory under the intervention condition deviates positively by a preset magnitude compared to the evolution trajectory under the non-intervention condition, an execution instruction for the intervention is generated. Finally, this instruction is executed to adjust the cultivation environment parameters, and the intervention decision data and actual observation results are updated to the causal structure knowledge base to optimize the deduction accuracy of counterfactual reasoning in subsequent cultivation processes.

[0040] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An adaptive control method for parameters in tumor organoid culture, characterized in that, Includes the following steps: S1. Based on the temporal correlation between multimodal sensing data and organoid state during the historical culture process, a causal structure knowledge base is constructed, which includes multiple nodes and directed causal paths between nodes. The nodes include culture environment parameters, multimodal sensing signals and organoid morphological features. S2, during the target organoid culture process, multimodal sensor data streams are collected, and time-series state analysis is performed on each multimodal sensor data stream based on the causal structure knowledge base. The current culture state vector is output in real time, which includes proliferation activity characterization value and morphological feature characterization value. S3, when it is identified that the proliferation activity characterization value and the morphological characteristic characterization value show a divergent trend along the preset causal path, the counterfactual reasoning program is invoked, and based on the causal structure knowledge base and the current culture state vector, the evolution trajectory of the morphological characteristic characterization value in the future time window is deduced under the conditions of executing the intervention action and not executing the intervention action, respectively. S4. Compare the evolutionary trajectory under the condition of performing the intervention action with the evolutionary trajectory under the condition of not performing the intervention action. If the former shows a positive deviation of a preset magnitude compared with the latter, then generate the execution instruction for the intervention action. S5 executes the execution command to adjust the culture environment parameters and updates the intervention decision data and actual observation results to the causal structure knowledge base to optimize the inference accuracy of counterfactual reasoning in subsequent culture processes.

2. The adaptive control method for parameters in tumor organoid culture according to claim 1, characterized in that, S1 specifically includes: Acquire multiple batches of historical culture data, which include culture environment parameters, multimodal sensor signals, and corresponding organoid morphological features collected at different time points. Construct a time-series data matrix by aligning all data by time. The time series data matrix is ​​divided into sliding time windows. In each time window, the conditional independence test is performed on any two nodes. If the dependency between the two nodes under the condition of other nodes is stable in multiple consecutive time windows, then a directed causal path is established between the two nodes. An initial causal structure diagram is generated based on all established directed causal paths, and the directionality of each causal path is verified in chronological order. Causal paths that conform to temporal logic are retained, forming a causal structure knowledge base for subsequent counterfactual reasoning.

3. The adaptive control method for parameters in tumor organoid culture according to claim 1, characterized in that, S2 specifically includes: The real-time acquired multimodal sensor data streams are split according to sensor type, and mapping associations are established with the corresponding nodes in the causal structure knowledge base. Timing alignment is performed on each sensor data stream to form a synchronized node observation sequence. Based on the directed causal path between nodes in the causal structure knowledge base, starting from the node corresponding to the proliferation activity characterization value, we trace back along the causal path to the node corresponding to the morphological feature characterization value. During the tracing process, we perform consistency checks on the observation values ​​of the nodes along the way and filter out noisy observations that are contrary to the direction of the causal path. After filtering out noise, all node observations are weighted and fused according to their topological hierarchy in the causal structure knowledge base to generate the current culture state vector, where the proliferation activity characterization value and morphological feature characterization value are output as independent components of the current culture state vector.

4. The adaptive control method for parameters in tumor organoid culture according to claim 1, characterized in that, S3 specifically includes: Extract proliferation activity characterization values ​​and morphological feature characterization values ​​from the current culture state vector, determine the starting and ending nodes of the divergence trend between proliferation activity characterization values ​​and morphological feature characterization values ​​based on the causal structure knowledge base, and locate the unique directed causal path between the starting and ending nodes. Using the real-time observation value of the starting node in the current cultivation state vector as the initial condition, the intervention action is transformed into an assignment perturbation of the downstream adjacent nodes of the starting node along the directional causal path. The state recursion value of the perturbation propagating along the causal path to the end node is calculated in sequence to obtain the first evolutionary trajectory under the condition of executing the intervention action. Keeping the real-time observation value of the starting node unchanged and without applying any assignment perturbation, the state recursion value of each node is calculated sequentially along the same directed causal path to obtain the second evolutionary trajectory under the condition of no intervention action.

5. The adaptive control method for parameters in tumor organoid culture according to claim 4, characterized in that, The first evolutionary trajectory obtained under the condition of executing the intervention action specifically includes: Based on the causal structure knowledge base, the response curves between the downstream adjacent nodes of the starting node and the starting node during the historical cultivation process are determined. The mapping relationship between the observed values ​​of the starting node and the assigned values ​​of the downstream adjacent nodes is extracted from the response curves, and the intervention action is converted into an assignment perturbation quantity that conforms to the mapping relationship. The assigned perturbation is superimposed on the baseline recursive function of the downstream adjacent node in the causal structure knowledge base to generate the state value of the first node after perturbation. The state value of the current node is then used as the input parameter of the next node along the directed causal path, and the calculation is performed step by step according to the recursive function corresponding to each node. After calculating the state value of each first-level node, the state value is corrected for deviation from the actual observed value of the corresponding node in the current culture state vector. The corrected state value is used as the starting condition for the next level of recursion until the end node is calculated, thus obtaining the first evolutionary trajectory.

6. The adaptive control method for parameters in tumor organoid culture according to claim 1, characterized in that, S4 specifically includes: The first evolutionary trajectory and the second evolutionary trajectory are paired according to the same time point within the same future time window, and the positive deviation of the first morphological value corresponding to the first evolutionary trajectory relative to the second morphological value corresponding to the second evolutionary trajectory is calculated one by one at each time point. The number of consecutive time points in which the positive deviation exceeds a preset threshold is counted. When the number of consecutive time points reaches the preset counting threshold, it is determined that the first evolutionary trajectory shows a positive deviation of a preset magnitude compared to the second evolutionary trajectory. The judgment result is converted into an execution instruction, which includes an identifier of the type of intervention action and an execution intensity parameter determined by the proliferation activity characterization value extracted from the current culture state vector.

7. The adaptive control method for parameters in tumor organoid culture according to claim 6, characterized in that, The process for determining the execution intensity parameter is as follows: Extract the proliferation activity characterization value from the current culture state vector, input the proliferation activity characterization value into the preset intensity mapping function, and divide the execution intensity into multiple continuously increasing intensity levels according to the level of the proliferation activity characterization value; Based on the type identifier of the intervention action, the intensity response curve corresponding to the corresponding intervention action is retrieved from the causal structure knowledge base. The intensity level determined in the first step is projected onto the intensity response curve to obtain the initial value of the corresponding execution intensity parameter. Using morphological feature values ​​as correction factors, the initial value of the execution intensity parameter is reverse-compensated and corrected. When the morphological feature values ​​are lower than the historical average, the initial value of the execution intensity parameter is increased. When the morphological feature values ​​are greater than or equal to the historical average, the final execution intensity parameter is obtained.

8. The adaptive control method for parameters in tumor organoid culture according to claim 1, characterized in that, S5 specifically includes: Collect the actual observation results within the preset time window after the intervention action is executed, compare the actual observation results with the predicted values ​​at the corresponding time points of the first evolutionary trajectory in counterfactual reasoning point by point, and calculate the deviation between the actual value and the predicted value of each node. Based on the magnitude of the deviation, the connection strength between nodes on the directed causal path is corrected by backpropagation. The larger the deviation, the greater the reduction in the connection strength from the upstream node to the downstream node, and the corrected connection strength is fixed to the corresponding path in the causal structure knowledge base. The real-time observations, assigned perturbations, and actual observation results of the starting node in this intervention decision data are stored as new time-series samples in the causal structure knowledge base to update the fitting parameters of the response curves between each node.

Images