Method and system for carrying out real-time toxicity prediction on organ-like model

Abstract

The invention discloses a method and a system for performing real-time toxicity prediction on an organ-like model, and belongs to the crossing field of biological medicine technology and artificial intelligence. The method comprises the following steps: firstly, acquiring multi-modal time sequence data of an organ-like model, wherein the data comprises multi-channel image data and non-image data; thirdly, performing fusion preprocessing on the acquired data to form a unified feature vector sequence; then, a cascaded deep learning model is utilized to process the sequence, and the model firstly extracts spatial features from fusion features of all time points through a three-dimensional convolutional neural network and then extracts time features capable of representing the toxicity dynamic process from the spatial feature sequence through a long-short-term memory neural network; the invention also can comprise a feedback control module for automatically executing intervention measures when toxicity is detected. According to the invention, the whole dynamic process of the toxic reaction can be accurately captured in real time, and highly automatic monitoring and adjustment of the organoid experiment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedicine and artificial intelligence, in particular to a method and system for real-time toxicity prediction of organoid models. BACKGROUND

[0002] As a three-dimensional cell culture model formed by stem cells self-organization in vitro, organoids can highly simulate the microstructure and physiological function of real human organs. Compared with traditional two-dimensional cell culture, it can better reproduce the interaction between cells and complex tissue structure; compared with animal models, it overcomes the inconsistency of drug reactions caused by species differences and meets the ethical requirements. Therefore, organoid technology has shown great application prospects in drug screening, disease modeling and personalized precision medicine.

[0003] As a three-dimensional cell culture model formed by stem cells self-organization in vitro, organoids can highly simulate the microstructure and physiological function of real human organs. Compared with traditional two-dimensional cell culture, it can better reproduce the interaction between cells and complex tissue structure; compared with animal models, it overcomes the inconsistency of drug reactions caused by species differences and meets the ethical requirements. Therefore, organoid technology has shown great application prospects in drug screening, disease modeling and personalized precision medicine.

[0004] Now, toxicity evaluation in vitro mostly adopts endpoint detection method or live cell imaging technology. Endpoint detection method, such as measuring ATP content, can provide quantitative indicators, but it is destructive in nature and can only provide a static snapshot at the end of the experiment, which results in the loss of the dynamic evolution process of toxicity response in time and space. Although live cell imaging technology allows long-term non-invasive dynamic observation of morphological changes of organoids, its analysis method often stops at simple interpretation of morphological changes, and it is difficult to effectively extract deep features indicating toxicity from high-dimensional and complex time-series image data.

[0005] In addition, toxicity response is a complex and multi-stage dynamic process, which may start from subtle metabolic dysfunction and then show visible morphological damage and large-scale cell death. The existing technology generally lacks a unified analysis framework that can integrate morphological, functional, metabolic and other multi-source heterogeneous information and deeply mine their internal spatio-temporal correlation. The lack of such analysis capability results in a lag in judgment relative to the actual occurrence time of toxicity events, missing the key window of observing early response and intervention.

[0006] When toxicity evaluation of organoids is needed, researchers usually make manual judgment and intervention after observing obvious damage, which is not only inefficient, but also has large evaluation result error, making it difficult to realize automatic closed-loop control of the experimental process. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides a method and system for real-time toxicity prediction of organoid models, and provides an objective, automated and dynamic real-time analysis organoid model toxicity prediction scheme.

[0008] To achieve the above object, the present application is implemented by the following technical scheme: a method and system for real-time toxicity prediction of organoid models. The method comprises the following core steps: S1 obtaining multi-modal time series data of an organoid model; S2 fusing the data; S3 extracting dynamic change features through a specific deep learning model; and S4 generating a final toxicity prediction result based on the features. This method integrates biological information of different dimensions and analyzes the time evolution law, thereby achieving accurate and early prediction of toxicity events.

[0009] In one or more embodiments, the multi-modal time series data obtained in step S1 provides a comprehensive information basis for toxicity evaluation. Specifically, the multi-channel images can include bright field images reflecting the morphological structure of the organoid, and fluorescence images reflecting the physiological state of the cells. Meanwhile, the non-image data can include metabolite concentrations in the culture medium reflecting the metabolic activity of the cells, and impedance measurements reflecting the overall structure of the organoid and the tightness of the intercellular connection. By combining morphological, physiological, metabolic and electrophysiological data, the health status of the organoid can be more comprehensively characterized.

[0010] In one or more embodiments, the data fusion in S2 can include a preprocessing stage to improve data quality and model performance. Specifically, the multi-channel images can be enhanced using a contrast-limited adaptive histogram equalization algorithm to highlight the detailed features in the images. For the non-image data, normalization processing can be performed to eliminate the differences in dimensions and numerical ranges between different data types. The preprocessing and fusion steps convert heterogeneous data into a unified time series of fusion features, preparing for subsequent deep learning model processing.

[0011] In one or more embodiments, S3 describes the core technical principle of extracting spatiotemporal dynamic features of the present application. The deep learning model is composed of a three-dimensional convolutional neural network (3D CNN) and a long short-term memory neural network (LSTM) in cascade. First, the 3D CNN is used to process the time series of fusion features. Since the fusion features contain three-dimensional spatial information derived from images, the 3D CNN can effectively extract spatial features at each time point, which summarize the comprehensive morphological and physiological state of the organoid at a specific time.

[0012] Subsequently, the spatial feature sequence arranged in chronological order is input into an LSTM. LSTM is specifically designed to analyze the temporal dependence in sequence data, and can learn the patterns of spatial feature evolution over time, thereby effectively capturing the dynamic development process of toxic reactions and ultimately extracting temporal features that characterize the overall temporal variation.

[0013] In one or more embodiments, S4 describes the specific method for generating the final prediction result. The extracted temporal features can be input into a fully connected layer and mapped to a continuous value between 0 and 1, i.e., a toxicity prediction score, using a sigmoid activation function. This score intuitively represents the probability of toxicity. The score can then be compared with a predefined threshold to determine the toxicity status of the organoid model, and this determination serves as the final toxicity prediction result.

[0014] In one or more embodiments, the method of the present invention may further include a feedback control step. When the generated toxicity prediction result indicates a toxic state, the system can automatically trigger a preset intervention measure. This closed-loop design makes the method not only a monitoring tool, but also an automated experimental platform with proactive response capabilities.

[0015] In one or more embodiments, the automated intervention may specifically include replacing the culture medium containing the toxic substance, or eluting the test compound via a microfluidic system, in an attempt to reverse or mitigate the toxic effects, and continuously monitoring the organoid response after the intervention.

[0016] In one or more embodiments, the toxicity prediction results can be visualized through a graphical user interface, providing researchers with intuitive analytical support. For example, the visualization output may include a time-series graph showing the toxicity prediction score as the likelihood of toxicity changes over time, and a heatmap showing the spatial distribution of the importance of internal features of organoids at key moments.

[0017] A second aspect of the present invention provides a system for real-time toxicity prediction of organoid models.

[0018] The system is designed to implement all or part of the steps of the aforementioned method, including: A data acquisition device is configured to acquire multimodal time-series data of the organoid model, the multimodal time-series data including multichannel image and non-image data; A data fusion module is configured to fuse the multi-channel image and the non-image data to form a time series of fused features; A deep learning model, configured to process the time series of the fused features to extract temporal features, the deep learning model comprising a 3D CNN and an LSTM; and An output module is configured to generate toxicity prediction results for the organoid model based on the time features.

[0019] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.

[0020] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0021] This invention provides a method and system for real-time toxicity prediction of organoid models. It has the following beneficial effects: 1. More comprehensive and accurate prediction results. This invention effectively combines image data reflecting organoid morphology with non-image data reflecting its physiological function through a data fusion module. Compared to existing technologies that rely on a single data source, this invention can construct a holistic understanding of the health status of organoids from multiple dimensions, thereby obtaining more comprehensive and accurate toxicity prediction results.

[0022] 2. Enables early warning of toxic events. The LSTM used in this invention is specifically designed for analyzing time-series data, effectively capturing the dynamic patterns that evolve over time during toxic reactions. This allows the model to identify early functional changes that are not yet morphologically obvious but foreshadow subsequent damage, thus enabling early warning of toxic events and providing a valuable time window for intervention and research.

[0023] 3. Provides in-depth insights into toxicity mechanisms. This invention utilizes 3D CNN to extract spatial features of organoids. By visualizing these features, it can be intuitively shown which regions within the organoids contribute most to toxicity assessment during the toxic reaction process. This not only provides a conclusion of "toxic" or "non-toxic," but also offers important clues for researchers to understand the spatial development patterns and mechanisms of toxicity.

[0024] 4. Achieves a high degree of automation and objectivity in the evaluation process. This invention constructs an end-to-end automated process from data acquisition, feature extraction, model prediction to result output. The entire process requires no human intervention, avoiding result bias and inconsistency caused by researchers' subjective judgment, ensuring the objectivity and repeatability of toxicity assessment, and greatly freeing up manpower.

[0025] 5. Supports closed-loop feedback control and high-throughput screening. The feedback controller of this invention can automatically trigger intervention operations when the toxicity exceeds a threshold based on the real-time prediction results of the deep learning model. This closed-loop control capability not only optimizes experimental conditions, but more importantly, its highly automated nature allows the system to be easily scaled up to large-scale parallel experiments, greatly improving the efficiency of high-throughput screening of drug toxicity. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall system structure of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a diagram of the deep learning model architecture of the present invention; Figure 4 This is an example diagram of the multimodal input data of the present invention; Figure 5 This is a graphical user interface diagram illustrating the results of the present invention. Figure 6 This is a schematic diagram of the feedback control process of the present invention. Detailed Implementation

[0027] The technical solutions in 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.

[0028] Photo attached Figures 1 to 6 This invention provides a system for real-time toxicity prediction of organoid models. The system achieves an automated process from data acquisition to result output through an integrated modular design.

[0029] The system may include: an organoid culture device, a data acquisition device, a data fusion module, a deep learning model, and a result output module.

[0030] The organoid culture device provides a controlled three-dimensional culture environment for organoid models. This device can precisely maintain the specific conditions required for culture, such as temperature, humidity, and the concentration of gases like carbon dioxide, ensuring the physiological activity and stability of organoids during in vitro growth and testing.

[0031] The data acquisition device is configured to periodically and non-invasively acquire multimodal time-series data of the organoid model. The multimodal time-series data comprehensively reflects the state changes of the organoid after exposure to external stimuli. In one embodiment, the data acquisition device may include a multi-channel imaging system for acquiring multi-channel images including bright-field and fluorescence images. The bright-field images record the overall morphology, size, and boundary information of the organoid, while the fluorescence images reveal the internal physiological state of the cell through specific fluorescent probes, such as changes in mitochondrial membrane potential or the activation of key apoptosis proteins.

[0032] In addition, the data acquisition device can be configured to simultaneously collect non-image data. This non-image data may include biochemical analysis of culture medium samples to obtain metabolite concentrations, such as glucose consumption rate and lactate production rate, which are directly related to the energy metabolism status of cells. Simultaneously, it may also include a multi-electrode array or similar sensor for measuring transepithelial / endothelial resistance or other forms of impedance measurements in organoid models, which reflect the integrity of the cell barrier function.

[0033] The data fusion module is connected to the data acquisition device and its function is to receive and process the acquired heterogeneous data to prepare for subsequent deep learning analysis. This module first performs image enhancement on the input multi-channel image, for example, using a contrast-limited adaptive histogram equalization (CLAHE) algorithm to improve image contrast and detail clarity. Simultaneously, this module normalizes the input non-image data to eliminate scale differences caused by different physical units.

[0034] After preprocessing, the data fusion module integrates the enhanced image features with the normalized non-image data into a unified, high-dimensional fusion feature vector, and organizes it into a time series of fusion features in chronological order.

[0035] The deep learning model is the core of this system's analysis, receiving time series data of fused features from the data fusion module. This model is designed as a composite structure, comprising a 3D CNN and an LSTM. The 3D CNN is responsible for extracting deep spatial features from the fused feature vectors at each time point, capturing the complex changes in the three-dimensional spatial structure of organoids. Subsequently, the LSTM analyzes the time series data composed of these spatial features, learning their inherent patterns and long-term dependencies over time, ultimately outputting time features that highly summarize the dynamic process of toxicity response.

[0036] The results output module generates the final toxicity prediction result based on the temporal features output from the deep learning model. This module may contain a toxicity prediction unit that uses a fully connected layer and a sigmoid activation function to transform the temporal features into a toxicity prediction score between 0 and 1. This module further determines the specific toxicity status, such as "toxic" or "non-toxic," based on a predefined threshold. Finally, the results output module displays the toxicity prediction score, toxicity status, and related analytical charts in real time through a graphical user interface, providing researchers with intuitive and quantitative decision-making support.

[0037] In some embodiments, the system may also integrate a feedback controller that is linked to the result output module. When a toxic state is detected, the actuators in the organoid culture device can be automatically triggered to perform intervention operations, such as changing the culture medium, thereby forming a closed-loop automated experimental system.

[0038] In a specific implementation of the present invention, the method aims to transform raw multidimensional biological data into accurate toxicity assessment conclusions through a series of well-defined steps.

[0039] The method begins with S1, which involves acquiring multimodal time-series data of the organoid model. This step forms the data foundation for the entire analysis process. The acquired data has two significant characteristics: firstly, it is "multimodal," meaning it comes from diverse sources and integrates biological information from different dimensions; secondly, it is "time-series," meaning the data is collected periodically at multiple consecutive time points to capture dynamic changes.

[0040] In one implementation, the multimodal data specifically includes two categories. The first category is multichannel images, such as bright-field images acquired through a microscopic imaging system that reflect the three-dimensional structure of organoids and the morphology of cell populations, and fluorescence images acquired using specific fluorescent markers that reveal key physiological processes in cells. The second category is non-image data, such as metabolite concentrations obtained by analyzing culture medium samples that reflect the overall metabolic activity of cell populations, and impedance measurements obtained by sensors that characterize the integrity of cell barrier function.

[0041] After acquiring the data, the method proceeds to S2, which involves fusing the acquired multi-channel image and non-image data. The core objective of this step is to transform the raw data, which come from diverse sources and formats, into a unified, standardized format suitable for deep learning models. Preprocessing is typically performed before fusion. For example, contrast-limited adaptive histogram equalization can be applied to multi-channel images for image enhancement to highlight subtle morphological changes caused by toxic reactions.

[0042] For non-image data, normalization is performed to eliminate biases caused by different measurement units and numerical ranges, ensuring that all types of data have comparable contribution weights in model analysis.

[0043] After preprocessing, these heterogeneous data are integrated into a fused feature vector, and the fused feature vectors of all time points constitute the time series required for subsequent analysis.

[0044] Next, the method executes S3, using a deep learning model to process the time series of fused features to extract temporal features that characterize the dynamic process of toxicity. In this embodiment, the deep learning model innovatively cascades 3D CNN and LSTM. First, leveraging the powerful spatial feature capture capability of 3D CNN, the fused features at each time point are analyzed to extract spatial features that characterize the overall organoid state at that moment.

[0045] Subsequently, these spatially arranged spatial feature sequences are input into an LSTM. The LSTM can learn long-term dependencies in sequence data, thereby accurately capturing the evolution of toxic reactions from the initial to the later stages, and ultimately extracting a highly condensed and information-rich temporal feature vector.

[0046] The final step of the method is S4, which generates a toxicity prediction result for the organoid model based on the extracted temporal features. In one specific implementation, this step processes the temporal features through a fully connected layer and uses a sigmoid activation function to output a continuous value in the range of 0 to 1; this value is the toxicity prediction score. The higher the score, the greater the likelihood that the organoid will exhibit a toxic reaction. Finally, this score is compared with a pre-set threshold to obtain the final discrete toxicity state as the toxicity prediction result.

[0047] Furthermore, the method of the present invention may also include a feedback control step. When the toxicity prediction result reaches the trigger condition, an intervention measure, such as changing the culture medium or eluting the test compound, can be automatically executed, thereby achieving closed-loop automatic control of the experimental process.

[0048] In the specific implementation of this invention, the coordinated operation of the various core technology modules is key to achieving accurate and real-time toxicity prediction. The internal implementation principles of these modules will be explained in detail below.

[0049] Regarding the acquisition and preprocessing of multimodal data, this invention integrates data from different sources to construct a comprehensive understanding of organoid states. When acquiring multi-channel images, in addition to recording bright-field images of the overall morphology, fluorescence images can be acquired using excitation light of specific wavelengths. These fluorescence images correspond to mitochondrial fluorescent probes and apoptosis markers, respectively, thereby quantifying cellular physiological activity. For non-image data, metabolite concentration measurements can be performed periodically to analyze glucose and lactate levels in culture medium samples, while impedance measurements can be performed in the frequency range of 1 kHz to 100 kHz to capture changes in cell barrier function at different frequencies.

[0050] Before feeding these heterogeneous data into a deep learning model, a crucial preprocessing step is necessary. For image data, a contrast-limited adaptive histogram equalization algorithm can be used for enhancement. This algorithm effectively enhances the local contrast of the image by equalizing the histogram of local image regions and setting a clipping limit, while avoiding artifacts caused by excessive noise amplification. For non-image data, the values ​​are scaled to a uniform range of 0 to 1 using a min-max normalization method.

[0051] This step ensures that data with different physical dimensions are of equal importance in subsequent fusion and analysis.

[0052] The model receives the fused feature time series and analyzes it in two stages. In the first stage, a 3D CNN is responsible for extracting spatial features. This network consists of multiple convolutional layers, each using a small 3D convolutional kernel for feature extraction. The 3D CNN can effectively process fused features containing spatial dimensional information, thereby capturing the 3D structural morphology and functional distribution features of organoids at specific time points and encoding them into a high-dimensional spatial feature map S.

[0053] In the second stage, LSTM is responsible for extracting temporal features. LSTM receives a sequence of spatial features S, arranged in chronological order, generated by a 3D CNN, as input. The LSTM network consists of multiple internally complex storage units, and its unique gating mechanism enables it to learn and memorize long-term dependencies in the sequence. Through the analysis of the spatial feature sequence, LSTM can capture the entire dynamic process of a toxic reaction from latency, occurrence to development, and ultimately output a temporal feature vector T that characterizes the core laws of this process.

[0054] Toxicity prediction and feedback control are the hallmarks of this invention's automation and proactive intervention. The extracted time feature T is fed into a fully connected layer and processed by a sigmoid activation function to calculate a continuous toxicity prediction score P. The score P ranges from 0 to 1; a higher value indicates a greater likelihood of toxicity. Subsequently, the system compares the score P with a predefined threshold H. If P is greater than H, the organoid's current state is classified as "toxic"; otherwise, it is classified as "non-toxic."

[0055] A key aspect of this invention lies in its included feedback control step. When the toxicity prediction result is determined to be "toxic," the system can automatically trigger an intervention. This intervention could be sending a command to the organoid culture device to replace the culture medium containing the potentially toxic compound; or activating an integrated microfluidic system to elute the test compound with fresh culture medium. This closed-loop design transforms the system from a passive monitoring tool into an automated platform capable of proactive experimental control.

[0056] To present the analysis results intuitively to users, the results output module of this invention is equipped with a graphical user interface. This interface displays a real-time graph of the toxicity prediction score P changing over time, allowing users to clearly observe the onset and rate of development of the toxic reaction. Simultaneously, the interface can visualize the spatial characteristics S at key time points as a heatmap, highlighting the areas that contribute most to the toxicity assessment. The final toxicity status is clearly indicated by a color-coded indicator. Example 1

[0057] This embodiment provides a specific application for real-time toxicity prediction of liver organoid models. First, the liver organoid model is placed in an organoid culture device for three-dimensional culture. This device maintains a constant temperature of 37°C, a relative humidity of 95%, and a carbon dioxide concentration of 5% to simulate the in vivo physiological environment.

[0058] During the data acquisition phase, a multi-channel imaging system equipped with a high-resolution microscope and a CMOS sensor was used to periodically acquire bright-field and fluorescence images of liver organoids at 30-minute sampling intervals over a 72-hour experimental period. The fluorescence images were obtained using excitation wavelengths of 488 nm and 561 nm to capture mitochondrial fluorescence intensity and apoptosis marker fluorescence, respectively. Simultaneously, non-image data were collected, including measurements of glucose and lactate concentrations in the culture medium every 6 hours, and impedance measurements acquired in the frequency range of 1 kHz to 100 kHz.

[0059] In the data fusion and preprocessing stage, the collected multimodal data is input into the data fusion module. The module uses a contrast-limited adaptive histogram equalization algorithm with a clipping limit of 0.03 to enhance the bright field and fluorescence images. For non-image data, a min-max normalization method is used for standardization. The processed image features and non-image data are fused into a unified feature vector sequence.

[0060] In the feature extraction stage, the fused feature vector sequence is input into a deep learning model consisting of a 3D CNN and an LSTM. The 3D CNN contains 10 convolutional layers, each using a 3×3×3 kernel, with a stride of 1 to extract spatial features S. Subsequently, the spatial feature sequence S is input into an LSTM with 128 memory units to extract the final temporal features T.

[0061] In the toxicity prediction stage, a toxicity prediction score P is calculated based on the time feature T using a fully connected layer and a sigmoid activation function. For the liver organoid model, a predefined threshold H is set to 0.7. When the calculated P is greater than 0.7, the organoid is classified as toxic; if P is less than or equal to 0.7, it is classified as non-toxic. Finally, the graphical user interface of the results output module displays in real time a time series plot of the toxicity prediction score P, a heatmap of the spatial features S at key moments, and color-coded indicators representing the toxicity status. Example 2

[0062] This embodiment, based on Embodiment 1, demonstrates an implementation method that combines feedback control with a kidney organoid model. Most steps in this embodiment are the same as in Embodiment 1; the difference lies in the introduction of a feedback control mechanism.

[0063] In this embodiment, when the results output module determines that the toxicity score P exceeds a preset threshold of 0.7, the system automatically triggers an intervention command. This command is sent to the microfluidic system integrated in the organoid culture device, which immediately executes a preset intervention operation, such as replacing the existing culture medium with fresh culture medium or eluting the test compound. The entire intervention process is completed within minutes of the toxicity status being confirmed. Subsequently, the system continues to monitor the organoid's response to the intervention, thereby achieving a complete closed-loop automated process of "monitoring-judgment-intervention-re-monitoring".

[0064] To ensure the real-time nature of the entire process, the deep learning model in this embodiment is deployed on an edge computing server. This server is equipped with a high-performance graphics processing unit (GPU), whose computing power is sufficient to ensure that the latency for processing and predicting the data collected at each time point is less than 10 minutes. Example 3

[0065] This embodiment describes the application of the present invention in the field of personalized drug screening. In this scenario, the organoid model used is a tumor organoid derived from a specific patient, with the aim of evaluating the toxic side effects of various candidate drugs on that patient.

[0066] To quickly detect drug-induced toxic reactions, the data acquisition sampling interval was shortened to 15 minutes. The rest of the system's processing flow is basically the same as in Example 1.

[0067] The key feature of this embodiment lies in the expanded functionality of the results output module. In addition to displaying the toxicity prediction score and status of individual drugs, the module's graphical user interface also provides a comparative analysis view. This view can display the toxicity prediction score curves of multiple candidate drugs side-by-side at the same time scale, thus providing researchers or clinicians with an intuitive basis for comparison to assist them in selecting the safest and most effective treatment for a specific patient.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time toxicity prediction of organoid models, characterized in that, Includes the following steps: S1: Obtain the multimodal time series data of the organoid model, wherein the multimodal time series data includes multi-channel image and non-image data; S2: The multi-channel image and the non-image data are fused to form a time series of fused features; S3: Process the time series of the fused features using a deep learning model that includes a 3D convolutional neural network and a long short-term memory neural network to extract time features; S4: Based on the time features, generate the toxicity prediction results of the organoid model.

2. The method according to claim 1, characterized in that, In S1, the multi-channel image includes bright-field image and fluorescence image; the non-image data includes the concentration of metabolites in the culture medium and the impedance measurement of the organoid model.

3. The method according to claim 1, characterized in that, In step S2, before fusing the multi-channel image and the non-image data, the method further includes: The multi-channel image is enhanced using a contrast-limited adaptive histogram equalization algorithm; and the non-image data is normalized.

4. The method according to claim 1, characterized in that, In step S3, spatial features are extracted from the time series of the fused features using the three-dimensional convolutional neural network.

5. The method according to claim 4, characterized in that, The spatial features are further analyzed using the long short-term memory neural network to extract the temporal features.

6. The method according to claim 1, characterized in that, S2 includes: using a fully connected layer and a sigmoid activation function to calculate a toxicity prediction score based on the time features; and determining the toxicity status of the organoid model as the toxicity prediction result based on the toxicity prediction score according to a predefined threshold.

7. The method according to claim 1, characterized in that, The method further includes a feedback control step: when the toxicity prediction result indicates toxicity, automatic intervention is triggered.

8. The method according to claim 7, characterized in that, The automated intervention includes changing the culture medium or eluting the test compound.

9. The method according to claim 1, characterized in that, The toxicity prediction results are visualized through a graphical user interface, and the visualization output includes a time series plot of the toxicity prediction score and a heatmap of spatial features.

10. A system for real-time toxicity prediction of organoid models, characterized in that, include: A data acquisition device is used to acquire multimodal time series data of the organoid model, wherein the multimodal time series data includes multi-channel image and non-image data; The data fusion module is used to fuse the multi-channel image and the non-image data to form a time series of fused features; A deep learning model, configured to process the time series of the fused features to extract temporal features, the deep learning model including a three-dimensional convolutional neural network and a long short-term memory neural network; And a result output module, used to generate toxicity prediction results for the organoid model based on the time features.

Images