A method for quantifying the relationship between organoid morphological subtypes and biological functions

Abstract

This invention relates to a method for quantifying the relationship between morphological subtypes and biological functions of organoids, comprising the following steps: Step 1, organoid heterogeneity analysis: Several primary organoids are induced and cultured and morphologically differentiated to screen for organoid systems that are morphologically distinguishable among different organ types, and the morphological subtypes of these systems are observed; Step 2, external perturbation processing of organoids; Step 3, bright-field imaging and image preprocessing; Step 4, prediction of biological function parameters using a biological function identification model; Step 5, addition of organoid morphological labels to the perturbated bright-field image data; Step 6, construction of a correlation map: the scores of biological function parameters are normalized and divided into quartiles, then sequentially correlated with morphological subtypes to construct a correlation map between morphological subtypes and biological function parameter scores. This invention can reveal the relationship between morphological heterogeneity and biological function with high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of quantitative technology of the relationship between organoid morphology and function, and specifically relates to a method for quantifying the relationship between organoid morphological subtypes and biological functions. Background Technology

[0002] Patient-derived tumor organoids have demonstrated good replication of in vivo physiological relevance and serve as a platform for systematically evaluating disease mechanisms and drug efficacy. Compared to traditional two-dimensional cell lines or animal models, patient-derived tumor organoids can reduce species differences while preserving in vivo physiological relevance and complexity. In recent years, the increasingly engineered tumor organoids have significantly expanded our understanding of the mechanisms of tumorigenesis, progression, invasion, and distant metastasis. For example, molecular-level heterogeneity of tumor organoids has been systematically assessed through large-scale single-cell sequencing and gene editing technologies (CRISPR–Cas9). However, the trade-off between throughput and cost has led to a shift towards phenotype-based high-throughput screening experiments. Among these, image-based high-throughput screening, due to its endogenous single-cell resolution, can capture subtle cellular morphological changes, thereby understanding the organoid's response to external perturbations (drug perturbations and gene perturbations). In particular, this image-based large-scale phenotypic drug screening has been used to explore the morphological heterogeneity of organoids and its association with the mechanisms of action of different drugs.

[0003] However, in the current drug screening process, the identification or prediction of biological functions based on organoid morphological heterogeneity is usually carried out by constructing prediction models using deep learning methods. Although this method can identify or predict biological functions, it cannot clarify the specific quantitative relationship between organoid morphological subtypes and biological functions. Therefore, developing a method for the relationship between organoid morphological subtypes and biological functions is of great significance for revealing the relationship between morphological heterogeneity and organoid biological function parameters, discovering more organoid morphological indicators, and guiding drug development. Summary of the Invention

[0004] The purpose of this invention is to overcome the deficiencies in the prior art and to provide a method for quantifying the relationship between morphological subtypes of organoids and their biological functions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for quantifying the relationship between morphological subtypes of organoids and their biological functions includes the following steps:

[0007] Step 1: Analysis of the heterogeneity of organoids: Several primary organoids were induced and cultured and morphologically distinguished to screen organoid systems that were morphologically distinguishable from each other, and the morphological subtypes of the organoid systems were observed.

[0008] Step 2, External perturbation processing of organoids: External perturbation processing is performed on various organs in the organoid system obtained in Step 1;

[0009] Step 3: Perform bright-field Z-stack imaging on the organoids that underwent external perturbation in Step 2, and perform image preprocessing to obtain perturbed bright-field image data;

[0010] Step 4: Based on the perturbed bright-field image data, predict biological function parameters using a biological function identification model:

[0011] Step 5: Add organoid morphological labels: Add the morphological subtypes from Step 1 as labels to the corresponding perturbed bright-field image data obtained in Step 3.

[0012] Step 6: Construction of the correlation map: The scores of the biological function parameters obtained in Step 4 are normalized and divided into quartiles. Then, they are matched with the morphological subtypes in order to construct a correlation map between the morphological subtypes and the scores of the biological function parameters.

[0013] As a further technical solution, step 1, the heterogeneity analysis of organoids, specifically includes the following steps:

[0014] Step 1-1: After numbering the primary organoids according to their collection source, induce culture on several primary organoids to obtain several organoid lines. During the induction culture process, each primary organoid is cultured in a different induction medium to obtain organoids with multiple different phenotypes. That is, each organoid line contains multiple organoids with multiple different phenotypes.

[0015] Steps 1-2: Perform bright-field Z-stack imaging on the organoids obtained from induced culture, and after image preprocessing, obtain the image data of the organoids;

[0016] Steps 1-3: Based on the organoid image data obtained in Step 1-2, extract phenotypic features from all organoids and filter the extracted phenotypic features to obtain filtered phenotypic features.

[0017] Steps 1-4: Morphologically distinguish the filtered phenotypic features, determine whether the various organs in the organoid system are morphologically distinguishable, screen out organoid systems in which the various organs are morphologically distinguishable from each other, and observe the morphological subtypes of the organoid system under a microscope.

[0018] As a further technical solution, in steps 1-4, morphological differentiation is performed on the filtered phenotypic features to determine whether the various organs in the organoid system are morphologically distinguishable, and organoid systems in which the various organs are morphologically distinguishable are selected, specifically including:

[0019] Step 1-4-1: Use the UMAP algorithm to perform dimensionality reduction and clustering on the filtered phenotypic features to obtain the cluster location distribution map of all organoid phenotypic clusters and clarify the cluster location of each cluster in the cluster location distribution map of all organoid phenotypic clusters.

[0020] Step 1-4-2: Keeping the cluster position distribution of phenotypic features unchanged, extract the phenotypic composition of each organoid lineage from the phenotypic cluster position distribution map of all organoids according to the primary organoid source, and obtain several independent organoid lineage phenotypic cluster position distribution maps.

[0021] Step 1-4-3: Determine the distribution of various organs in different cluster positions within each organoid system. If the various organs in the organoid system are distributed at different cluster positions, then the various organs in the organoid system are considered morphologically distinguishable. If there are multiple organoids at one cluster position in the organoid system, then the multiple organoids at that cluster position are considered morphologically indistinguishable, that is, the various organs in the organoid system are morphologically indistinguishable from each other. Remove the organoid systems where the various organs are morphologically indistinguishable from each other, and retain the organoid systems where the various organs are morphologically distinguishable from each other for later use.

[0022] As a further technical solution, in steps 1-3, the filtering process includes the following steps: first, remove the foreground targets that are extremely large or extremely small from the extracted phenotypic features; then, perform a box-cox transformation on all extracted phenotypic features to correct the offset feature distribution; and finally, remove the phenotypic features with a median absolute deviation of 0 to obtain the filtered phenotypic features.

[0023] As a further technical solution, the types of organoids or primary organoids include one or more of the following: intestinal organoids, lung organoids, kidney organoids, liver organoids, and brain organoids.

[0024] The induction culture medium is supplemented with biologically active ingredients; the types of biologically active ingredients include one or more of small molecule inhibitors, recombinant proteins, growth factors, hormones, and vitamins; the biologically active ingredients are used to induce organoid differentiation and maturation, maintain organoid stemness, and induce one or more of the phenotypic characteristics of specific morphological subtypes.

[0025] As a further technical solution, in step 4, the biological function identification model is constructed based on various morphologically distinguishable organs in the organoid system after external perturbation processing. The biological function identification model includes a biological activity identification model and / or a biological virtual staining prediction model.

[0026] As a further technical solution, the construction of the biological activity recognition model includes the following steps:

[0027] Step a: External perturbation is performed on the various morphologically distinguishable organs in the vascular system, followed by bright-field Z-stack imaging and image preprocessing to obtain perturbed bright-field image data.

[0028] Step b, Add biological activity labels: Based on the scale or outline of individual organoids, biologists select individual organoids in the perturbed bright-field image data obtained in step a; then, based on the growth status of the organoids, biologists determine the biological activity of each type of organ and add the results of the biological activity determination as labels to the perturbed organoid bright-field image data.

[0029] Step c: Divide the organoid bright-field image data after adding labels in step b into a training set and a test set in an 8:2 ratio. Then, use the training set to train the biological activity recognition model based on the deep learning algorithm and use the test set to verify the biological activity recognition model to obtain the biological activity recognition model.

[0030] The biological activity recognition model is constructed based on a PyTorch deep target detection network.

[0031] During model training, bright-field image data from the training set and corresponding label files are input into the biological activity recognition model. The biological activity recognition model performs feature matching between the label files and the image features of the bright-field image data, and iteratively learns the organoid biological activity parameters of the label files to complete the training of the biological activity recognition model.

[0032] After training, the bright-field image data from the test set is input into the biological activity recognition model. The biological activity recognition model is used to identify image features in the bright-field image data and output organoid biological activity parameters. The correlation analysis is performed between the organoid biological activity parameters output by the biological activity recognition model and the biological activity parameters in the label file corresponding to the organoid bright-field image data to confirm the accuracy of the constructed biological activity recognition model. The biological activity recognition model is then completed.

[0033] As a further technical solution, the construction of the biological virtual staining prediction model includes the following steps:

[0034] Step A: Morphologically distinguishable organs in the organoid system are subjected to external perturbation, followed by bright-field Z-stack imaging and staining imaging. After image preprocessing, perturbed bright-field image data and perturbed staining image data are obtained.

[0035] Step B: Process the perturbed staining image data using image processing software to obtain biological function parameters, and add the biological function parameters as labels to the perturbed staining image data;

[0036] Step C: After matching the perturbed bright-field image data obtained in Step B with the stained image data with added labels in Step C, the data is divided into a training set and a test set. Then, the biological virtual staining prediction model is trained using the training set based on the deep learning algorithm, and the biological virtual staining prediction model is verified using the test set to obtain the biological virtual staining prediction model.

[0037] The biological virtual staining prediction model includes a virtual staining model (i.e., generator), a biological functional parameter prediction model (i.e., predictor), and a virtual staining discrimination model (i.e., discriminator). The virtual staining model (i.e., generator) is built based on a generative adversarial network and is used to perform virtual staining on bright-field images to generate matching virtual staining image data. The biological functional parameter prediction model (i.e., predictor) is built based on a residual network (ResNet) and is used to predict the biological functional parameters of organoids based on the virtual staining image data.

[0038] During training, bright-field image data from the training set and corresponding labeled stained image data are input into the biological virtual staining prediction model. The virtual staining model of the biological virtual staining prediction model performs feature matching between the image features of the bright-field image data and the image features of the stained image data, and iteratively learns the image features of the stained image data to complete the training of the virtual staining model. Then, the virtual stained image data output by the virtual staining model is matched with the label file of the corresponding stained image data, and the biological functional parameters of the label file are iteratively learned to complete the training of the biological functional parameter prediction model.

[0039] After training, the bright-field image data and corresponding stained image data from the test set are input into the biological virtual staining model. The virtual staining model identifies image features in the bright-field image data and outputs virtual stained image data. Then, the virtual staining discrimination model matches the virtual stained image data with the corresponding input real stained image data, scores the virtual stained images, and evaluates the accuracy of the virtual staining model in staining bright-field images. Subsequently, the biological function parameter prediction model predicts the biological function parameters of organoids based on the virtual stained image data output by the biological virtual staining model. Finally, the correlation analysis between the predicted biological function parameters and the real biological function parameters is performed to verify the accuracy of the biological function parameter prediction model. The construction of the biological virtual staining prediction model is then complete.

[0040] As a further technical solution, image preprocessing includes one or more of the following: maximum intensity projection, illumination correction, background removal, image quality control and image segmentation, fluorescence image deconvolution, and global intensity normalization.

[0041] As a further technical solution, image preprocessing includes one or more of the following: maximum intensity projection, illumination correction, background removal, image quality control and image segmentation, fluorescence image deconvolution, and global intensity normalization.

[0042] As a further technical solution, the maximum intensity projection is processed using the maximum density projection method.

[0043] As a further technical solution, illumination correction is performed using prospective methods or retrospective single-image methods.

[0044] As a further technical solution, the image segmentation is performed using Bwlabel, Watershed, and Propagate algorithms.

[0045] As a further technical solution, the external disturbance treatment includes one or more of chemical disturbance, physical disturbance, and biological disturbance;

[0046] The chemical perturbation includes one or more of drug perturbation and biologically active substance perturbation; the physical perturbation includes one or more of light stimulation, sound stimulation and electrical stimulation; and the biological perturbation includes one or more of gene knockdown or overexpression and gene editing.

[0047] As a further technical solution, in the construction of the biological virtual staining prediction model, the staining can be one or more of fluorescence staining, H&E staining, and immunohistochemical staining; the obtained perturbed staining image data includes any one of perturbed fluorescence image data, perturbed H&E staining image data, and perturbed immunohistochemical staining image data.

[0048] As a further technical solution, the biological virtual staining prediction model adopts a biological virtual fluorescence prediction model, which can be used to predict biological apoptosis parameters, etc.

[0049] As a further technical solution, the biological activity recognition model is used to identify biological activity parameters.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention pioneers a method for distinguishing organoid morphology. When quantifying the relationship between organoid morphological subtypes and biological functional parameters, and constructing biological function identification models, after inducing and culturing organoids, it is necessary to use the organoid morphology distinguishability method to determine whether several induced and cultured organoids are morphologically distinguishable from each other. Then, organoid systems that are morphologically distinguishable from each other are selected for subsequent experiments, thereby improving the accuracy of experimental results.

[0052] The biological virtual staining prediction model constructed in this invention can transform bright-field image data into corresponding staining image data, achieving unbiased prediction of organoid responses to external disturbances.

[0053] This invention combines organoid morphology exploration with deep learning algorithms to successfully construct a biological function identification model, which can quickly identify biological functions and quantify the relationship between organoid morphological subtypes and biological functions. This is of great significance for revealing the relationship between morphological heterogeneity and organoid biological function parameters, discovering more organoid morphological indicators, and guiding drug development. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of organoid morphological heterogeneity analysis based on bright-field images in one embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the culture process for inducing organoids to form specific morphological subtypes in one embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the dimension-reduced clustering map and the corresponding bright-field image of the specific morphological subtypes of organoids formed in two examples, CK02 and CK10, in one embodiment of the present invention.

[0057] Figure 4 This is a flowchart illustrating the identification of phenotypic organoids using a depth target detection network under perturbed bright-field imaging in one embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram of organoid state scores detected using a deep object detection network in one embodiment of the present invention;

[0059] Figure 6 This is a correlation map between organoid morphological subtypes and model-predicted organoid bioactivity status scores in one embodiment of the present invention.

[0060] Figure 7 This is a schematic diagram of the network architecture of a biological virtual staining prediction model in one embodiment of the present invention;

[0061] Figure 8 This is a correlation map between biological functional parameters predicted by a biological virtual staining prediction model and their true values ​​in one embodiment of the present invention.

[0062] Figure 9 This is a correlation map between organoid morphological subtypes and model-predicted organoid apoptosis state scores in one embodiment of the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the embodiments.

[0064] Example 1

[0065] The construction of a biological activity recognition model includes the following steps:

[0066] Step a: Morphologically distinguishable organs in the organoid system are subjected to external perturbation, followed by bright-field Z-stack imaging and image preprocessing to obtain perturbed bright-field image data.

[0067] Among them, the organoid system is formed by inducing and culturing primary organoids in different induction media and then conducting morphological heterogeneity analysis to determine that the various organs in the organoid system are morphologically distinguishable from each other. Only then can the various organs in the organoid system be used to construct a biological activity recognition model.

[0068] Image preprocessing for Z-stack imaging under bright field was performed using EBImage software. Image preprocessing included maximum intensity projection, illumination correction, background subtraction, image quality control, and image segmentation.

[0069] Step b, Add biological activity labels: Based on the scale or outline of individual organoids, biologists select individual organoids in the perturbed bright-field image data obtained in step a; then, based on the growth status of the organoids, biologists determine the biological activity of each type of organ and add the results of the biological activity determination as labels to the perturbed organoid bright-field image data.

[0070] Step c: Divide the bright-field organoid bright-field image data after adding labels in step b into training set and test set in a ratio of 8:2. Then, use the training set to train the biological activity recognition model based on deep learning algorithm and use the test set to verify the biological activity recognition model to obtain the biological activity recognition model.

[0071] The biological activity identification model in this embodiment is built based on a PyTorch deep object detection network, including SSD, Faster R-CNN, and YOLOv4. The performance of the biological activity identification model is evaluated based on the object detection metric [mean average precision when intersection over union (IoU) threshold was 0.5 (mAP_0.5)]. The final mAP_0.5 scores of the three networks are 0.76, 0.72, and 0.69, respectively.

[0072] During model training, bright-field image data from the training set and corresponding label files are input into the biological activity recognition model. The biological activity recognition model performs feature matching between the label files and the image features of the bright-field image data, and iteratively learns the organoid biological activity parameters of the label files to complete the training of the biological activity recognition model.

[0073] After training, the bright-field image data from the test set is input into the biological activity recognition model. The biological activity recognition model is used to identify image features in the bright-field image data and output organoid biological activity parameters. The correlation analysis is performed between the organoid biological activity parameters output by the biological activity recognition model and the biological activity parameters in the label file corresponding to the organoid bright-field image data to confirm the accuracy of the constructed biological activity recognition model. The biological activity recognition model is then completed.

[0074] Example 2

[0075] The construction of a biological virtual staining prediction model includes the following steps:

[0076] Step A: Morphologically distinguishable organs in the organoid system are subjected to external perturbation, followed by bright-field Z-stack imaging and staining imaging. After image preprocessing, perturbed bright-field image data and perturbed staining image data are obtained.

[0077] Among them, the organoid system is formed by inducing and culturing primary organoids in different induction media and then conducting morphological heterogeneity analysis to determine that the various organs in the organoid system are morphologically distinguishable from each other. Only then can the various organs in the organoid system be used to construct a biological activity recognition model.

[0078] Image preprocessing for Z-stack imaging under bright field was performed using EBImage software. Image preprocessing included maximum intensity projection, illumination correction, background subtraction, image quality control, and image segmentation.

[0079] Image preprocessing for staining imaging was performed using EBImage software. Image preprocessing included maximum intensity projection, illumination correction, background removal, global intensity normalization, and image quality control. If fluorescence staining was used, fluorescence deconvolution processing was also required.

[0080] Step B: Use image processing software to process the perturbed staining image data to obtain biological function parameters (including one or two of the following: biological activity parameters and biological apoptosis degree parameters), and add the biological function parameters as tags to the perturbed staining image data.

[0081] Step C: After matching the perturbed bright-field image data obtained in Step B with the stained image data with added labels in Step C, the data is divided into a training set and a test set. Then, the biological virtual staining prediction model is trained using the training set based on the deep learning algorithm, and the biological virtual staining prediction model is verified using the test set to obtain the biological virtual staining prediction model.

[0082] like Figure 7 As shown, the biological virtual staining prediction model includes a virtual staining model (i.e., generator), a biological functional parameter prediction model (i.e., predictor), and a virtual staining discrimination model (i.e., discriminator). The virtual staining model (i.e., generator) is constructed based on a generative adversarial network and is used to perform virtual staining on bright-field images to form matching virtual staining image data. The biological functional parameter prediction model (i.e., predictor) is constructed based on a residual network (ResNet) and is used to predict the biological functional parameters of organoids (including biological activity parameters, biological apoptosis parameters, etc.) based on the virtual staining image data.

[0083] During training, bright-field image data from the training set and corresponding labeled stained image data are input into the biological virtual staining prediction model. The virtual staining model performs feature matching between the image features of the bright-field image data and the image features of the stained image data, and iteratively learns the image features of the stained image data to complete the training of the virtual staining model in the biological virtual staining model. Then, the virtual stained image data output by the virtual staining model is matched with the label file of the corresponding stained image data, and the biological functional parameters of the label file are iteratively learned to complete the training of the biological functional parameter prediction model in the biological virtual staining model.

[0084] After training, the bright-field image data and corresponding stained image data from the test set are input into the biological virtual staining model. The virtual staining model identifies image features in the bright-field image data and outputs virtual stained image data. Then, the virtual staining discrimination model matches the virtual stained image data with the corresponding real stained image data, scores the virtual stained images, and evaluates the accuracy of the virtual staining model in staining bright-field images. Next, the biological function parameter prediction model predicts the biological function parameters of organoids based on the virtual stained image data output by the biological virtual staining model. Finally, the predicted biological function parameters are correlated with the real biological function parameters (derived from the label files added to the corresponding stained image data) to verify the accuracy of the biological function parameter prediction model. The construction of the biological virtual staining prediction model is then complete.

[0085] Example 3

[0086] A method for quantifying the relationship between morphological subtypes of organoids and their biological functions includes the following steps:

[0087] Step 1: Heterogeneity analysis of organoids: Several primary organoids were induced and cultured, and morphologically differentiated to screen for organoid systems that were morphologically distinguishable from each other, and the morphological subtypes of the organoid systems were observed; specifically including:

[0088] Step 1-1: After numbering the primary organoids according to their source (the source refers to collection from different individuals), several primary organoids are induced and cultured to obtain several organoid lines. During the induction and culture process, such as... Figure 2 As shown, each primary organoid was cultured using a different induction medium to obtain organoids with multiple different phenotypes, that is, each organoid line contains multiple organoids with multiple different phenotypes.

[0089] The types of organoids or primary organoids include intestinal organoids, lung organoids, kidney organoids, liver organoids, brain organoids, etc.

[0090] The induction culture medium contains biologically active ingredients that induce organoid differentiation and maturation, maintain organoid stemness, and induce the formation of specific morphological subtype phenotypic characteristics; the types of biologically active ingredients include small molecule inhibitors, recombinant proteins, growth factors, hormones, vitamins, etc.

[0091] The 20 primary organoids used in this embodiment are numbered CK01, CK02, ..., CK20. The various organs obtained by induction culture from the same primary organoids constitute an organoid system. The type of organoid used in this embodiment is colorectal cancer organoids. The biologically active ingredients used for induction are Wnt3a recombinant protein and Trametinib small molecule inhibitor.

[0092] Steps 1-2, as follows Figure 1 As shown, bright-field Z-stack imaging was performed on organoids obtained from induced culture, and image preprocessing was performed to obtain image data of the organoids.

[0093] In this embodiment, during bright-field Z-stack imaging, 16 images are sampled in the Z direction with a step size of 10 micrometers.

[0094] The image preprocessing described in steps 1-2 is performed using EBImage software. Image preprocessing includes maximum intensity projection, illumination correction, background removal, image quality control, and image segmentation.

[0095] Steps 1-3: Based on the organoid image data obtained in Step 1-2, extract phenotypic features from all organoids and filter the extracted phenotypic features to obtain filtered phenotypic features.

[0096] The phenotypic features include shape, intensity, texture, etc.

[0097] The filtering process includes the following steps: First, remove the foreground objects that are extremely large or extremely small from the extracted phenotypic features. Then, perform a box-cox transformation on all extracted phenotypic features to correct the offset feature distribution. After that, remove the phenotypic features with a median absolute deviation of 0 to obtain the filtered phenotypic features.

[0098] In this embodiment, a total of 401 phenotypic features were extracted; after filtering, 163 phenotypic features remained.

[0099] Steps 1-4: Morphologically differentiate the filtered phenotypic features to determine whether the various organs in the organoid system are morphologically distinguishable. Select organoid systems where each organ is morphologically distinguishable from the others, and determine the morphological subtypes of the organoids through microscopic observation. Specifically, this includes:

[0100] First, the UMAP algorithm is used to perform dimensionality reduction and clustering on the filtered phenotypic features to obtain the distribution map of cluster locations for all organoid phenotypic features (see...). Figure 1 This method clarifies the cluster position of each cluster in the distribution map of all organoid phenotypic clusters; different organoids are distributed in different cluster positions in the distribution map of all organoid clusters; in this embodiment, the 163 phenotypic features after filtering are clustered into 13 clusters, and each of the 13 clusters occupies one cluster position (i.e., a cluster) in the distribution map of all organoid clusters.

[0101] Then, keeping the cluster position distribution of phenotypic features unchanged (i.e., the cluster position of each cluster), according to the primary organoid origin (primary organoid number), the phenotypic composition of each organoid lineage is extracted from the phenotypic cluster position distribution map of all organoids (phenotypic dimensionality reduction clustering map of all organoids), resulting in several independent organoid lineage phenotypic cluster position distribution maps; taking the organoids CK02 and CK10 as examples, the obtained independent organoid lineage phenotypic cluster position distribution maps are shown below. Figure 3 Distribution map of phenotypic cluster locations of independent organoid lines;

[0102] Then, the distribution of each type of organoid in each organoid system at the 13 cluster locations obtained above is determined. If each type of organoid is distributed at different cluster locations, it is considered that the morphology of each type of organoid in the organoid system is distinguishable, and the organoid system can be used for subsequent experiments. If there are multiple organoids at one cluster location in the organoid system, it is considered that the morphology of the multiple organoids at that cluster location is indistinguishable, that is, the morphology of each type of organoid in the organoid system is indistinguishable. Several organoids induced by the primary organoid are discarded, and the organoid systems in which the morphology of each type of organoid is distinguishable are retained for later use. Based on the morphological differentiation determination, 8 organoid systems out of the 20 groups of organoids cultured in this embodiment can be used for subsequent experiments. The 8 organoid systems are CK02, CK04, CK06, CK07, CK08, CK10, CK11, and CK13.

[0103] Finally, for the remaining morphologically distinguishable organoids, their morphological subtypes were observed under a microscope and numbered, such as phenotype I, phenotype II, phenotype III, ..., with different numbers representing different morphological subtypes. In this embodiment, organoids CK02 and CK10 are used as examples. Figure 3 The bright-field image data shows that the morphological subtype of cells in organoid CK02 is vesicular, while the morphological subtype of cells in organoid CK10 is solid. When numbering, phenotype I refers to vesicular and phenotype II refers to solid.

[0104] Step 2, External perturbation processing of organoids: External perturbation processing is performed on various organs in the organoid system obtained in Step 1;

[0105] The external disturbance treatment includes various types such as chemical disturbance, physical disturbance, and biological disturbance;

[0106] The chemical perturbations include: drug perturbations, biologically active substance perturbations, etc.; physical perturbations include light stimulation, sound stimulation, electrical stimulation, etc.; biological perturbations include gene knockdown or overexpression, gene editing, etc.

[0107] Step 3: Perform bright-field Z-stack imaging on the organoids that underwent external perturbation in Step 2, and perform image preprocessing to obtain perturbed bright-field image data;

[0108] like Figure 4 The Z-stack imaging image is shown 72 hours after the administration of the traditional Chinese medicine. In this embodiment, 16 images were sampled in the Z direction with a stride of 10 micrometers.

[0109] The image preprocessing described in step 3 is performed using EBImage software. The preprocessing items for brightfield imaging images include maximum intensity projection, illumination correction, background removal, and image quality control.

[0110] Step 4: Construction of biological function identification model and identification of biological function parameters: such as Figure 5 As shown, a biological function identification model is constructed based on a deep learning algorithm, and the constructed biological function identification model is used to predict the biological function parameters of organoids.

[0111] In this embodiment, the biological function identification model adopts the biological activity identification model to identify the biological activity of organoids after perturbation treatment.

[0112] The construction of a biological activity recognition model includes the following steps:

[0113] Step (1), add biological activity labels: Based on the scale or outline of a single organoid, biologists select a single organoid in the perturbed bright-field image data obtained in step 3; then, based on the growth status of the organoids, biologists determine the biological activity of each type of organ and add the results of the biological activity determination as labels to the perturbed organoid bright-field image data.

[0114] In this embodiment, organoids treated with high concentrations of oxaliplatin were determined by biologists to be "failures," i.e., positive controls. Therefore, positive controls were added as tags to the bright-field image data of these organoids. On the other hand, organoids treated with DMSO were determined by biologists to be "successes," i.e., negative controls. Therefore, negative controls were added as tags to the bright-field image data of these organoids.

[0115] Step (2): Divide the bright field organoid bright field image data after adding labels in step (1) into training set and test set in a ratio of 8:2. Then, use the training set to train the biological activity recognition model based on the deep learning algorithm and use the test set to verify the biological activity recognition model to obtain the biological activity recognition model.

[0116] The biological activity identification model in this embodiment uses the PyTorch deep object detection network framework, including SSD, Faster R-CNN, and YOLOv4. The performance of the biological activity identification model is evaluated based on the object detection metric (mean average precision when intersection over union (IoU) threshold was 0.5 (mAP_0.5)). The final mAP_0.5 scores of the three networks are 0.76, 0.72, and 0.69, respectively.

[0117] During model training, bright-field image data from the training set and corresponding label files are input into the biological activity recognition model. The biological activity recognition model performs feature matching between the label files and the image features of the bright-field image data, and iteratively learns the organoid biological activity parameters of the label files to complete the training of the biological activity recognition model.

[0118] After training, the bright-field image data from the test set is input into the biological activity recognition model. The biological activity recognition model identifies image features in the bright-field image data and outputs organoid biological activity parameters. Correlation analysis is performed between the organoid biological activity parameters output by the biological activity recognition model and the biological activity parameters in the corresponding label file of the organoid bright-field image data. When the correlation coefficient is greater than 0.8, the biological activity recognition model is considered to have high accuracy and can be used to identify organoid biological activity parameters through bright-field image data. The biological activity recognition model is then complete.

[0119] Step 5: Adding organoid morphological labels: Add morphological subtype labels to the perturbed bright-field image data obtained in Step 3;

[0120] Add the morphological subtypes from step 1 as labels to the corresponding perturbed bright-field image data obtained in step 3;

[0121] Step 6, Construction of the Association Map: The scores of the biological functional parameters obtained in Step 4 are normalized and divided into quartiles, and then mapped sequentially to morphological subtypes to construct an association map. Figure 6 The diagram shows the correlation between morphological subtypes and scores of biological functional parameters.

[0122] Example 4

[0123] A method for quantifying the relationship between morphological subtypes of organoids and their biological functions includes the following steps:

[0124] Step 1: Heterogeneity analysis of organoids: Several primary organoids were induced and cultured, and morphologically differentiated to screen for organoid systems that were morphologically distinguishable from each other, and the morphological subtypes of the organoid systems were observed; specifically including:

[0125] Step 1-1: After numbering the primary organoids according to their source (the source refers to collection from different individuals), several primary organoids are induced and cultured to obtain several organoid lines. During the induction and culture process, such as... Figure 2 As shown, each primary organoid was cultured using a different induction medium to obtain organoids with multiple different phenotypes, that is, each organoid line contains multiple organoids with multiple different phenotypes.

[0126] The types of organoids or primary organoids include: intestinal organoids, lung organoids, kidney organoids, liver organoids, brain organoids, etc.

[0127] The induction culture medium contains biologically active ingredients that induce organoid differentiation and maturation, maintain organoid stemness, and induce the formation of specific morphological subtype phenotypic characteristics; the types of biologically active ingredients include small molecule inhibitors, recombinant proteins, growth factors, hormones, vitamins, etc.

[0128] This embodiment uses 20 groups of primary organoids (collected from 20 different individuals), numbered CK01, CK02, ..., CK20. Organoids obtained from the same group of primary organoids through induction and culture constitute an organoid system. The organoid type used in this embodiment is colorectal cancer organoids. The biologically active ingredients used for induction are Wnt3a recombinant protein and Trametinib small molecule inhibitor.

[0129] Steps 1-2, as follows Figure 1 As shown, bright-field Z-stack imaging was performed on organoids obtained from induced culture, and image preprocessing was performed to obtain image data of the organoids.

[0130] In this embodiment, during bright-field Z-stack imaging, 16 images are sampled in the Z direction with a step size of 10 micrometers.

[0131] The image preprocessing described in steps 1-2 is performed using EBImage software. Image preprocessing includes maximum intensity projection, illumination correction, background removal, image quality control, and image segmentation.

[0132] Steps 1-3: Based on the organoid image data obtained in Step 1-2, extract phenotypic features from all organoids and filter the extracted phenotypic features to obtain filtered phenotypic features.

[0133] The phenotypic features include shape, intensity, texture, etc.

[0134] The filtering process includes the following steps: First, remove the foreground objects that are extremely large or extremely small from the extracted phenotypic features. Then, perform a box-cox transformation on all extracted phenotypic features to correct the offset feature distribution. After that, remove the phenotypic features with a median absolute deviation of 0 to obtain the filtered phenotypic features.

[0135] In this embodiment, a total of 401 phenotypic features were extracted; after filtering, 163 phenotypic features remained.

[0136] Steps 1-4: Morphologically differentiate the filtered phenotypic features to determine whether the various organs in the organoid system are morphologically distinguishable. Select organoid systems where each organ is morphologically distinguishable from the others, and determine the morphological subtypes of the selected organoid systems. Specifically, this includes:

[0137] First, the UMAP algorithm is used to perform dimensionality reduction and clustering on the filtered phenotypic features to obtain the distribution map of cluster locations for all organoid phenotypic features (see...). Figure 1 This method clarifies the cluster position of each cluster in the distribution map of all organoid phenotypic clusters; different organoids are distributed in different cluster positions in the distribution map of all organoid clusters; in this embodiment, the 163 phenotypic features after filtering are clustered into 13 clusters, and each of the 13 clusters occupies one cluster position (i.e., a cluster) in the distribution map of all organoid clusters.

[0138] Then, keeping the cluster position distribution of phenotypic features unchanged (i.e., the cluster position of each cluster), according to the primary organoid origin (primary organoid number), the phenotypic composition of each organoid lineage is extracted from the phenotypic cluster position distribution map of all organoids (phenotypic dimensionality reduction clustering map of all organoids), resulting in several independent organoid lineage phenotypic cluster position distribution maps; taking the organoids CK02 and CK10 as examples, the obtained independent organoid lineage phenotypic cluster position distribution maps are shown below. Figure 3 Distribution map of phenotypic cluster locations of independent organoid lines;

[0139] Then, the distribution of each type of organoid in each organoid system at the 13 cluster locations obtained above is determined. If each type of organoid is distributed at different cluster locations, it is considered that the morphology of each type of organoid in the organoid system is distinguishable, and the organoid system can be used for subsequent experiments. If there are multiple organoids at one cluster location in the organoid system, it is considered that the morphology of the multiple organoids at that cluster location is indistinguishable, that is, the morphology of the organoids in the organoid system is indistinguishable, and the organoid system needs to be discarded. The organoid systems in which the morphology of each type of organoid is distinguishable are retained for later use. Based on the morphological differentiation determination, 8 organoid systems out of the 20 organoid systems cultured in this embodiment can be used for subsequent experiments. The 8 organoid systems are CK02, CK04, CK06, CK07, CK08, CK10, CK11, and CK13.

[0140] Finally, the morphological subtypes of the preserved organoid lines were observed under a microscope and numbered, such as phenotype I, phenotype II, phenotype III, ..., with different numbers representing different morphological subtypes. In this embodiment, the CK02 organoid lineage and the CK10 organoid lineage are used as examples. Figure 3 As can be seen from the bright-field image data, the cell morphology subtype of the CK02 organoid system in this embodiment is vesicular, and the cell morphology subtype of the CK10 organoid system is solid. Therefore, in this embodiment, phenotype I refers to vesicular and phenotype II refers to solid.

[0141] Step 2, External perturbation processing of organoids: External perturbation processing is performed on the morphologically distinguishable organs in the organoid system obtained in Step 1.

[0142] The external disturbance treatment includes various types such as chemical disturbance, physical disturbance, and biological disturbance;

[0143] The chemical perturbations include: drug perturbations, biologically active substance perturbations, etc.; physical perturbations include light stimulation, sound stimulation, electrical stimulation, etc.; biological perturbations include gene knockdown or overexpression, gene editing, etc.

[0144] Step 3: Perform bright-field Z-stack imaging and fluorescence staining imaging on the organoids that underwent external perturbation in Step 2, and perform image preprocessing to obtain perturbed bright-field image data;

[0145] In this embodiment, 16 images are sampled in the Z direction with a stride of 10 micrometers during the imaging process.

[0146] The image preprocessing described in step 3 is performed using EBImage software. The preprocessing items for brightfield imaging images include maximum intensity projection, illumination correction, background removal, and image quality control.

[0147] Step 4: Construction of biological function identification model and identification of biological function parameters: Construct a biological function identification model based on deep learning algorithm, and use the constructed biological function identification model to predict the biological function parameters of organoids.

[0148] In this embodiment, the biological function identification model adopts a biological virtual staining prediction model, which in turn adopts a biological virtual fluorescence staining prediction model.

[0149] The biological functional parameters predicted by the biological virtual fluorescence staining prediction model are biological apoptosis parameters.

[0150] The construction of a biological virtual fluorescence staining prediction model includes the following steps:

[0151] Step 1) Perform fluorescence Z-stack imaging and image preprocessing on the organoids that underwent external perturbation in Step 2 to obtain perturbed fluorescence image data;

[0152] In this embodiment, 16 images were sampled in the Z-direction during fluorescence Z-stack imaging, with a stride of 10 micrometers.

[0153] In this embodiment, the image preprocessing of fluorescence Z-stack imaging was performed using EBImage software. The image preprocessing included maximum intensity projection, illumination correction, fluorescence deconvolution, background removal, global intensity normalization, and image quality control.

[0154] Step 2) Based on the perturbed fluorescence image data, biological apoptosis parameters are calculated using EBImage software, and then the biological apoptosis parameters are added as tags to the perturbed fluorescence image data;

[0155] In this embodiment, the biological apoptosis parameter used is the fluorescence intensity of the apoptosis index Caspase-3;

[0156] Step 3) After matching the perturbed bright-field image data obtained in Step 3 with the labeled fluorescence image data added in Step 2), the data is divided into a training set and a test set. Then, the biological virtual fluorescence staining prediction model is trained using the training set based on the deep learning algorithm, and the biological virtual fluorescence staining prediction model is verified using the test set to obtain the biological virtual fluorescence staining prediction model.

[0157] like Figure 7 As shown, the biological virtual fluorescence staining prediction model in this embodiment includes a virtual staining model (i.e., generator), a biological functional parameter prediction model (i.e., predictor), and a virtual staining discrimination model (i.e., discriminator). The virtual staining model (i.e., generator) is constructed based on a generative adversarial network and is used to perform virtual staining on bright-field images to form matching virtual staining image data. The biological functional parameter prediction model (i.e., predictor) is constructed based on a residual network (ResNet) and is used to predict biological apoptosis parameters of organoids based on the virtual fluorescence staining image data.

[0158] During training, bright-field image data from the training set and corresponding labeled fluorescence image data are input into the biological virtual fluorescence staining prediction model. The virtual staining model performs feature matching between the image features of the bright-field image data and the image features of the fluorescence image data, and iteratively learns the image features of the fluorescence image data to complete the training of the virtual staining model in the biological virtual fluorescence staining prediction model. Then, the virtual fluorescence staining image data output by the virtual staining model is matched with the label file of the corresponding fluorescence image data, and the biological apoptosis parameters of the label file are iteratively learned to complete the training of the biological functional parameter prediction model in the biological virtual fluorescence staining model.

[0159] After training, the bright-field image data and corresponding fluorescence image data from the test set are input into the biological virtual fluorescence staining model. The virtual staining model identifies image features in the bright-field image data and outputs virtual fluorescence staining image data. Then, the virtual staining discrimination model matches the virtual fluorescence staining image data with the corresponding input real fluorescence image data, scoring the virtual fluorescence staining images. A score higher than 0.8 indicates that the virtual staining model can accurately stain the bright-field images. Subsequently, the biological functional parameter prediction model predicts the biological apoptosis parameters of the organoids based on the virtual fluorescence staining image data output by the biological virtual staining model. Finally, as shown... Figure 8As shown, correlation analysis was performed between the predicted biological apoptosis parameters and the actual biological apoptosis parameters (derived from the tag files added to the corresponding fluorescence image data). When the correlation coefficient was higher than 0.8, it proved that the biological functional parameter prediction model could accurately predict biological apoptosis parameters through fluorescence image data, and the biological virtual fluorescence staining prediction model was completed. In this embodiment, the correlation coefficient between the predicted biological apoptosis parameters and the actual biological apoptosis parameters was 0.91 after correlation analysis.

[0160] Step 5: Adding organoid morphological labels: Add morphological subtype labels to the perturbed bright-field image data obtained in Step 3;

[0161] Add the morphological subtypes from step 1 as labels to the corresponding perturbed bright-field image data obtained in step 3;

[0162] Step 6: Construction of the correlation map: The scores of the biological functional parameters (apoptosis parameters in this example) obtained in Step 4 are normalized and divided into quartiles, and then matched sequentially with morphological subtypes to construct a correlation map as follows: Figure 9 The correlation map shown is between morphological subtypes and scores of biological functional parameters (in this example, biological apoptosis parameters).

[0163] The embodiments described above are merely preferred embodiments of the present invention, and not an exhaustive list of all possible implementations of the present invention. Any obvious modifications made by those skilled in the art without departing from the principles and spirit of the present invention should be considered to be included within the scope of protection of the claims of the present invention.

Claims

1. A method for quantifying the relationship between morphological subtypes of organoids and their biological functions, characterized in that, Includes the following steps: Step 1: Analysis of the heterogeneity of organoids: Several primary organoids were induced and cultured and morphologically distinguished to screen organoid systems that were morphologically distinguishable from each other, and the morphological subtypes of the organoid systems were observed. Step 2, External perturbation processing of organoids: External perturbation processing is performed on various organs in the organoid system obtained in Step 1; Step 3: Perform bright-field Z-stack imaging on the organoids that underwent external perturbation in Step 2, and perform image preprocessing to obtain perturbed bright-field image data; Step 4: Based on the perturbed bright-field image data, predict biological function parameters using a biological function identification model: Step 5: Add organoid morphological labels: Add the morphological subtypes from Step 1 as labels to the corresponding perturbed bright-field image data obtained in Step 3. Step 6: Construction of the correlation map: The scores of the biological function parameters obtained in Step 4 are normalized and divided into quartiles. Then, they are matched with the morphological subtypes in order to construct a correlation map between the morphological subtypes and the scores of the biological function parameters.

2. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 1, characterized in that, Step 1, the heterogeneity analysis of organoids, specifically includes the following steps: Step 1-1: After numbering the primary organoids according to their collection source, induce culture on several primary organoids to obtain several organoid lines. During the induction culture process, each primary organoid is cultured in a different induction medium to obtain organoids with multiple different phenotypes. That is, each organoid line contains multiple organoids with multiple different phenotypes. Steps 1-2: Perform bright-field Z-stack imaging on the organoids obtained from induced culture, and after image preprocessing, obtain the image data of the organoids; Steps 1-3: Based on the organoid image data obtained in Step 1-2, extract phenotypic features from all organoids and filter the extracted phenotypic features to obtain filtered phenotypic features. Steps 1-4: Morphologically distinguish the filtered phenotypic features, determine whether the various organs in the organoid system are morphologically distinguishable, screen out organoid systems in which the various organs are morphologically distinguishable from each other, and observe the morphological subtypes of the organoid system under a microscope.

3. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 2, characterized in that, In steps 1-4, morphological differentiation is performed on the filtered phenotypic features to determine whether the various organs in the organoid system are morphologically distinguishable, and organoid systems in which the various organs are morphologically distinguishable are selected. Specifically, this includes: Step 1-4-1: Use the UMAP algorithm to perform dimensionality reduction and clustering on the filtered phenotypic features to obtain the cluster location distribution map of all organoid phenotypic clusters and clarify the cluster location of each cluster in the cluster location distribution map of all organoid phenotypic clusters. Step 1-4-2: Keeping the cluster position distribution of phenotypic features unchanged, extract the phenotypic composition of each organoid lineage from the phenotypic cluster position distribution map of all organoids according to the primary organoid source, and obtain several independent organoid lineage phenotypic cluster position distribution maps. Step 1-4-3: Determine the distribution of various organs in different cluster positions within each organoid system. If the various organs in the organoid system are distributed at different cluster positions, then the various organs in the organoid system are considered morphologically distinguishable. If there are multiple organoids at one cluster position in the organoid system, then the multiple organoids at that cluster position are considered morphologically indistinguishable, that is, the various organs in the organoid system are morphologically indistinguishable from each other. Remove the organoid systems where the various organs are morphologically indistinguishable from each other, and retain the organoid systems where the various organs are morphologically distinguishable from each other for later use.

4. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 2, characterized in that, In steps 1-3, the filtering process includes the following steps: first, remove the foreground targets that are extremely large or extremely small from the extracted phenotypic features; then, perform a box-cox transformation on all extracted phenotypic features to correct the offset feature distribution; and finally, remove the phenotypic features with a median absolute deviation of 0 to obtain the filtered phenotypic features.

5. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 2, characterized in that, The types of organoids or primary organoids include one or more of the following: intestinal organoids, lung organoids, kidney organoids, liver organoids, and brain organoids. The induction culture medium is supplemented with biologically active ingredients; the types of biologically active ingredients include one or more of small molecule inhibitors, recombinant proteins, growth factors, hormones, and vitamins; the biologically active ingredients are used to induce organoid differentiation and maturation, maintain organoid stemness, and induce one or more of the phenotypic characteristics of specific morphological subtypes.

6. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 1, characterized in that, In step 4, the biological function identification model is constructed based on the various morphologically distinguishable organs in the organoid system after external perturbation processing. The biological function identification model includes a biological activity identification model and / or a biological virtual staining prediction model.

7. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 6, characterized in that, The construction of the biological activity recognition model includes the following steps: Step a: External perturbation is performed on the various morphologically distinguishable organs in the vascular system, followed by bright-field Z-stack imaging and image preprocessing to obtain perturbed bright-field image data. Step b, Add biological activity labels: Select individual organoids in the perturbed bright-field image data obtained in step a according to their scale or outline; then, determine the biological activity of each organoid according to its growth status, and add the results of the biological activity determination as labels to the perturbed organoid bright-field image data. Step c: Divide the bright-field image data of organoids after adding labels in step b into a training set and a test set. Then, based on the deep learning algorithm, train the biological activity recognition model using the training set and verify the biological activity recognition model using the test set to obtain the biological activity recognition model. The biological activity recognition model is constructed based on a PyTorch deep target detection network. During model training, bright-field image data from the training set and corresponding label files are input into the biological activity recognition model. The biological activity recognition model performs feature matching between the label files and the image features of the bright-field image data, and iteratively learns the organoid biological activity parameters of the label files to complete the training of the biological activity recognition model. After training, the bright-field image data from the test set is input into the biological activity recognition model. The biological activity recognition model is used to identify image features in the bright-field image data and output organoid biological activity parameters. The correlation analysis is performed between the organoid biological activity parameters output by the biological activity recognition model and the biological activity parameters in the label file corresponding to the organoid bright-field image data to confirm the accuracy of the constructed biological activity recognition model. The biological activity recognition model is then completed.

8. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 6, characterized in that, The construction of the biological virtual staining prediction model includes the following steps: Step A: Morphologically distinguishable organs in the organoid system are subjected to external perturbation, followed by bright-field Z-stack imaging and staining imaging. After image preprocessing, perturbed bright-field image data and perturbed staining image data are obtained. Step B: Process the perturbed staining image data using image processing software to obtain biological function parameters, and add the biological function parameters as labels to the perturbed staining image data. Step C: After matching the perturbed bright-field image data obtained in Step B with the stained image data with added labels in Step C, the data is divided into a training set and a test set. Then, the biological virtual staining prediction model is trained using the training set based on the deep learning algorithm, and the biological virtual staining prediction model is verified using the test set to obtain the biological virtual staining prediction model. The biological virtual staining prediction model includes a virtual staining model, a biological functional parameter prediction model, and a virtual staining discrimination model. The virtual staining model is based on an adversarial network and is used to perform virtual staining on bright-field images to form matching virtual staining image data. The biological functional parameter prediction model is based on a residual network and is used to predict the biological functional parameters of organoids based on the virtual staining image data. During training, bright-field image data from the training set and corresponding labeled stained image data are input into the biological virtual staining prediction model. The virtual staining model of the biological virtual staining prediction model performs feature matching between the image features of the bright-field image data and the image features of the stained image data, and iteratively learns the image features of the stained image data to complete the training of the virtual staining model. Then, the virtual stained image data output by the virtual staining model is matched with the label file of the corresponding stained image data, and the biological functional parameters of the label file are iteratively learned to complete the training of the biological functional parameter prediction model. After training, the bright-field image data and corresponding stained image data from the test set are input into the biological virtual staining model. The virtual staining model identifies image features in the bright-field image data and outputs virtual fluorescent stained image data. Then, the virtual staining discrimination model matches the virtual stained image data with the corresponding input real stained image data, scores the virtual stained images, and evaluates the accuracy of the virtual staining model in staining bright-field images. Subsequently, the biological function parameter prediction model predicts the biological function parameters of organoids based on the virtual stained image data output by the biological virtual staining model. Finally, the correlation analysis between the predicted biological function parameters and the real biological function parameters is performed to verify the accuracy of the biological function parameter prediction model. The construction of the biological virtual staining prediction model is then complete.

9. A method for quantifying the relationship between morphological subtypes of organoids and their biological functions according to claim 7 or 8, characterized in that, Image preprocessing includes one or more of the following: maximum intensity projection, illumination correction, background removal, image quality control and image segmentation, fluorescence image deconvolution, and global intensity normalization.

10. The method for quantifying the relationship between morphological subtypes and biological functions of organoids according to claim 6, characterized in that, The external disturbance treatment includes one or more of chemical disturbance, physical disturbance, and biological disturbance; The chemical perturbation includes one or more of drug perturbation and biologically active substance perturbation; the physical perturbation includes one or more of light stimulation, sound stimulation and electrical stimulation; and the biological perturbation includes one or more of gene knockdown or overexpression and gene editing.

Images