Systems and methods for identifying bioactive agents utilizing unbiased machine learning

Abstract

To provide systems and methods for identifying molecules that are biologically active against a disease.SOLUTION: A method can comprise: culturing a first mammalian cell population under organoid formation conditions in the presence of a test molecule to obtain a first organoid, wherein the first mammalian cell population, when cultured under the organoid formation conditions in the absence of the test molecule, results in an organoid with a disease phenotype; imaging the first organoid following exposure to the test molecule; analyzing images of the first organoid using a neural network that has been trained to assign a probability score of disease or non-disease; and assigning the first organoid a probability score, wherein the test molecule is biologically active against the disease if the probability score of the first organoid is greater than a cutoff probability score of non-disease or lower than a cutoff probability score of disease.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] 1. Statement Regarding Federally Sponsored Research This invention was made in part through funding provided by the National Institutes of Health awarded NIHR01GM101653 and National Institutes of Health Grant No. 10102621. This invention was made with Government support under a STTR grant awarded by the United States Patent and Trademark Office (USPTO). Each person has certain rights. [Background technology]

[0002] 2.Background By differentiating stem cells in vitro, the organelles inherent to stem cells can be expressed. In theory, organoids can be used to study the self-organization ability of in vivo A tool for modeling disease in vitro and for testing potential therapeutic agents However, these large multicellular structures show great promise for use as Due to their inherent heterogeneity, it is important to fully exploit biological complexity and translate it into useful therapeutic predictions. There is a need for new quantification tools that can convert

[0003] The fact that organoids show stereotypical organization means, in principle, that they are “wild-type” (WT) or "non-disease") organoids and their counterparts, "disease" organoids. This allows for the recognition and quantification of phenotypic differences in the expression of the WT gene, thus allowing the reversion of disease phenotypes to WT. The organoids can be screened for pharmacological compounds that are

[0004] An ideal analytical scheme for downstream of such screening of organoids For each compound, two quantities relevant to translation of hit compounds to the clinic are quantified: First, the therapeutic potential must be accurately determined. Second, the toxic compounds must be quickly eliminated. Early rejection of clinical trials will prevent failure of clinical trials and reduce huge amounts of wasted investment. Ideally, potential deleterious off-target effects and cytotoxicity would be quantified. do. Summary of the Invention

[0005] 3. Overview The present disclosure relates to methods for identifying test molecules that are biologically active against a disease, and and / or an artificial nuclease that can be used in a method to quantify the potential toxicity of a test molecule. The characteristics of an exemplary neural network are described in Section 5.2 and numbered embodiments are given in Section 7.

[0006] In some embodiments, the neural network comprises a correlation between a non-disease phenotype and a disease phenotype. The weights of the artificial neurons can learn the differences between the images in the image database, e.g. Pre-trained on many images from the ImageNet database of one million images The pre-trained network can then be applied to non-diseased organoids. Trained on images of organoids and diseased organoids, the system extracts images from the various available channels. Learn the best combination of image features to generate optimal, unbiased non-disease and disease fingerprints. The training can identify the graphics processing unit. (GPU), even deep networks with many layers can be trained. Train in just a few minutes or hours depending on the size, number, and difficulty of your training After training, the trained network of the present disclosure can be used to The application of is generally very fast, e.g. thousands of images can be analyzed in a matter of seconds. do.

[0007] The present disclosure relates to methods for identifying test molecules that are biologically active against a disease, and and / or further provides a method for quantifying the potential toxicity of a test molecule.

[0008] In one aspect, the present disclosure provides a method for identifying molecules that are biologically active against a disease. Culturing a population of mammalian cells under organoid-forming conditions in the presence of a test molecule to form organoids. Obtaining mammalian cell populations by culturing mammalian cells in the absence of a bioactive molecule. When cultured under organoid-forming conditions, they give rise to organoids with disease phenotypes. Step; Imaging the organoids; Range of 0%-100% diseased or non-disease Uses a neural network that is trained to assign a probability score to and analyzing one or more images of the organoids; and assigning a probability score ranging from 0% to 100%. The probability score can range from 0 to 1, 0 to 10, or any arbitrary range. Test molecules are those that produce a probability score in organoids higher than the cutoff probability score for non-disease. or below the disease cutoff probability score, the patient is biologically active against the disease. can be considered to be a method for identifying molecules that are biologically active against disease. Exemplary methods are described in Sections 5.2.5 and 5.3, and in numbered embodiments 1-1. It is described in 22.

[0009] Advantageously, various embodiments of the neural networks and methods of the present disclosure are capable of detecting disease and W The available differences between the T and T can be used in an optimal manner. Therapeutic efficacy can be measured as defined by the degree of phenotypic reversion. A classifier that can recognize WT and disease phenotypes is required to determine how the toxicity phenotype is determined. Allows the definition of a third class corresponding to a toxicity phenotype without pre-specifying what will be visible This class responds to treatment with the test molecule but does not revert the phenotype to WT. This third class may include organoids that respond in a specific manner. It includes compounds that are considered to have a high potential for adverse effects in vivo. In embodiments, the disclosed method uses the disclosed neural network to identify test compounds. Organoids treated with the substance were assigned a score ranging from 0% to 100% (or 0–1, 0–10, if assigning a probability score of toxicity or non-toxicity (e.g., any arbitrary range) to the An exemplary method for assigning a probability score of toxicity or non-toxicity is , Sections 5.2.6 and 5.3, and numbered embodiments 1 to 122. is.

[0010] The present disclosure relates to a processor and the analysis and / or allocation steps of the disclosed method. A non-transitory storage medium (e.g., a hard disk drive) that contains processor-executable instructions for implementing the The present invention further provides a system including a computer readable medium (e.g., a hard disk, a flash drive, a CD, or a DVD). The system includes an imaging device capable of imaging the organoids (e.g., a camera). The exemplary system may further include a microscope having a microscope. It is described in 122.

[0011] The present disclosure analyzes images of organoids and assigns a diseased or non-disease probability score. The present invention further provides a method for training a neural network to Such methods are described in Section 5.2.4 and numbered embodiments 1-122.

[0012] The patent or application file contains at least one drawing executed in color. Reproduction of this patent or patent application publication with drawings is permitted upon request and payment of the necessary fee. The information will be provided by the authorities as necessary. [Brief description of the drawings]

[0013] [Figure 1] FIG. 2 illustrates the structure of an exemplary neural network of the present disclosure. [Figure 2-1] 2A is a flow chart showing exemplary steps for training a neural network according to the present disclosure and their use for analyzing the therapeutic potential and toxicity of test compounds. [Figure 2-2] 2A is a flow chart showing exemplary embodiments for training neural networks of the present disclosure and their use for analyzing the therapeutic potential and toxicity of test compounds. FIG. 2B is a flow chart showing exemplary steps for analyzing the therapeutic potential of a test molecule. [Figure 2-3] 2C is a flow chart showing exemplary steps for analyzing both the therapeutic potential and toxicity of a test molecule. [Figure 3-1]Immunofluorescence analysis of HD model organoids. Figure 3A: Lateral view: DAPI, PAX6 (green), and N-CAD (orange). Figure 3B: Top view: DAPI, neural marker PAX6 (green), neural crest marker SOX10 (red), and cranial placode marker SIX1 (yellow). [Figure 3-2] Immunofluorescence analysis of HD model organoids. (C) Lateral view showing non-neural ectoderm marker TFAP2A (blue). Epidermal cells are exclusively TFAP2A+. All scale bars represent 50 μm. [Figure 3-3] Immunofluorescence analysis of HD model organoids. (D) Schematic of human embryo during gastrulation. (E) Schematic of human HD model organoids. [Figure 4-1] Phenotypic signatures of Huntington's disease (HD) cell lines. Figure 4A: Representative images of PAX6 zones of various HD isogenic lines in HD model organoid assays. PAX6 staining allows visualization of Pax6 zones. [Figure 4-2] Phenotypic signatures of Huntington's disease (HD) cell lines (Figure 4B): Associated quantification of PAX6 area normalized by colony area. [Figure 5-1] Adaptation of HD model organoid phenotype to 96-well plates. Figure 5A: PAX6 staining of wells with WT HD model organoids. Figure 5B: PAX6 staining of wells with HD model organoids. [Figure 5-2] Figure 5C: Adaptation of HD model organoid phenotype to 96-well plates. PAX6+ area of ​​two different genetic backgrounds in multiple wells of two independent 96-well plates was quantified by threshold-based partitioning. [Figure 6-1] Figure 6: Determination of assay Z' factor in Example 1. Figure 6A: Determination of PAX6+ area by compartmentalization in WT / 20CAG control plate (left) or HD / 56CAG control plate (right). Each well is color-coded according to its average PAX6 area. [Figure 6-2]Figure 6A shows the determination of the assay Z'-factor in Example 1. Figure 6B: Calculation of well-to-well Z'-factor. Each well is color-coded with the Z'-factor and the associated value is reported. [Figure 7] Figure 7 shows machine learning assisted phenotypic analysis. Figure 7A: Training of an exemplary network to optimally recognize wild-type and disease phenotypes. Figure 7B: Quantification of phenotypic reversion. Adding drug 1 to the disease phenotype (bottom row) shows that the network is now more likely to belong to the wild-type class than in the absence of drug (top row). [Figure 8] Figure 8: Machine learning assisted phenotypic analysis. Figure 8A: Examples of images from different genetic backgrounds: WT (RUES2), Htt- / -, and Htt- / - with Htt overexpression. The network was trained to recognize WT (RUES2) and Htt- / - images. Figure 8B: The network was searched to classify images that it has never seen before, and the network assigned a probability pwt for each image. A value of 1 means that the image was classified as WT, and a value of 0 means that the image was classified as Htt- / -. WT and KO images were classified correctly. Images from rescue overexpression experiments showed a higher probability of belonging to WT. [Figure 9] Figure 1 shows the percentage of correctly and incorrectly identified 20 CAG (WT) and 56 CAG (disease) images. Total number of validation images: 754 (WT) and 692 (disease). [Figure 10] Figure 10: Rescue of Z' coefficients by neural network-assisted phenotypic analysis. Figure 10A: Comparison of Z' coefficients calculated using the regular partitioning scheme (black circles) or the machine learning toolbox (crosses). All wells showed Z'>0 with the machine learning toolbox. Figure 10B: Distribution of Z' scores obtained from machine learning-based analysis. [Figure 11]Figure 1 shows the distribution of the probability that each compound reverts the HD disease phenotype to WT. This probability is shown as pwt. Most compounds did not change the disease phenotype, but a low, broad peak close to 1 indicated that some compounds were successful in reverting the disease phenotype. [Figure 12] FIG. 12A shows an exemplary network architecture with fully connected latent vectors upstream of the final classification layer (FIG. 12A) and t-SNE clustering of latent vectors corresponding to control and compound-treated wells (FIG. 12B). [Figure 13] 1 shows an exemplary neural network training and classification of compounds. The exemplary network's final prediction, WT vs. disease, quantifies the amount of phenotypic reversion of the drug. The latent vector before final classification can be used to define both the distance to WT and disease, which can be used as a measure of toxicity. [Figure 14] Figure 1 shows the results of a phenotypic reversion screen using neurulation organoids derived from stem cells carrying a Huntington's disease mutation. Each data point represents one well with approximately 25 organoids. Green dots correspond to untreated control wells, and orange are wild-type control wells. Five compounds are close to the corners of the wild-type (WT), indicating successful reversion of the disease phenotype to WT. [Figure 15] Figure 1 shows an example of an autoencoder: Encoder and decoder neural networks are trained to make the reconstructed data match the input data as closely as possible, resulting in a low-dimensional representation of the data in a latent space. [Figure 16]Figure 1. Visualization of latent vectors. (a) Principal component analysis (PCA) of latent vectors of wild-type (WT) and disease images, with the first two principal components (PC1, PC2) displayed. (b) The same vectors overlaid with the set of compounds. (c) The same vectors as in (a), but with the variation between WT and disease removed. (d) The same vectors as in (b), but with the variation between WT and disease removed. [Figure 17] FIG. 1 shows cell viability at various drugs and concentrations as determined by the method described herein (autoencoder, "AE") as well as the traditional "MTT" assay. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] 5. Detailed Description The present disclosure provides methods for identifying test molecules that are biologically active against a disease and / or or an artificial neural network that can be used to quantify the potential toxicity of a test molecule. An exemplary neural network is described in Section 5.2. and numbered embodiments are given in Section 7.

[0015] In one aspect, the method for classifying a test organoid comprises imaging the test organoid. providing an organoid image; and assigning a classification to the test organoid. Organoids are trained using trained neural networks The method includes analyzing the image, and the classification includes a disease phenotype or a non-disease phenotype. Exemplary methods for classifying organoids are described in sections 5.2.3-5.2.5 and 5.2.6. and numbered embodiments 1 to 26.

[0016] In one aspect, a method for identifying a molecule that is biologically active against a disease comprises: Obtaining organoids by culturing mammalian cell populations under organoid-forming conditions in the presence of erythrocytes The mammalian cell population is cultured under organoid-forming conditions in the absence of the test molecule. Organoids, which when cultured in a culture medium, result in organoids with a disease phenotype; Imaging one or more images of the organoids in a range of 0% to 100%. A neural network trained to assign a disease or non-disease probability score to Analyzing the network; and assigning organoids a range of 0% to 100%. The test molecule is selected from the group consisting of a test molecule that is capable of detecting a probability score of the organoid of: Higher than the cutoff probability score for non-disease or higher than the cutoff probability score for disease If the level of activity is low, it can be considered biologically active against the disease. Exemplary methods for identifying biologically active molecules are described in Sections 5.2.5 and 5. .3, as well as numbered embodiments 27 to 121.

[0017] In some embodiments, the methods of the present disclosure involve administering 0% ~100% range (or 0-1, 0-10, or any arbitrary range) of toxicity or may or may further include assigning a probability score of non-toxicity. Exemplary methods for assigning probability scores are described in Sections 5.2.6 and 5.3. and numbered embodiments 122-136.

[0018] The present disclosure provides a process for implementing the analysis and allocation steps of the disclosed method. Non-transitory storage media (e.g., hard disk, flash drive, etc.) that contain processor-executable instructions (e.g., USB, CD, or DVD), and such non-transitory storage medium and processor The present invention further provides a system comprising: It may further include an imaging device (e.g., a microscope with a camera). The system is described in numbered embodiments 140 and 141.

[0019] The present disclosure analyzes images of organoids and assigns a diseased or non-disease probability score. The present invention further provides a method for training a neural network to Such methods are described in Section 5.2.4 and in numbered embodiments 142-159. do.

[0020] 5.1.Definition Unless otherwise defined herein, all technical terms and All scientific and technical terms have the same meaning as commonly understood by one of ordinary skill in the art. Various scientific dictionaries containing the terms used herein are well known and available to those of skill in the art. The implementation of the disclosed embodiments may involve any method similar or equivalent to that described herein. The methods and materials described in the above can be used.

[0021] The terms defined immediately below are more fully understood by reference to the specification in its entirety. The definitions are intended to illustrate certain embodiments and may provide a better understanding of the complex concepts described herein. The definitions are intended to aid in solution only. They are not intended to limit the full scope of the present disclosure. In particular, the present disclosure is not directed to the specific sequences, compositions, Understand that the present invention is not limited to algorithms, systems, methodologies, protocols, and reagents. as they may vary depending on the context in which they are used by those skilled in the art.

[0022] As used in this specification and the appended claims, the singular forms "a," "an," and " and "the" are used to refer to plural referents unless the content and context clearly dictate otherwise. Thus, for example, reference to a "device" includes two or more This includes device combinations such as:

[0023] Unless otherwise indicated, the conjunction "or" has its exact meaning as a Boolean logical operator. It is intended to be used in a variety of ways, such as selecting features in an option (where selection A is different from selection B). A or B, where they are mutually exclusive), and simultaneous selection of features (where A and B are both In some places in the text, the phrase "and" includes both A and B. The terms "and / or" are used interchangeably and "or" means mutually exclusive alternatives. It should not be construed as implying that it is being used to refer to alternatives.

[0024] Throughout this specification, amounts are defined by ranges, as well as by lower and upper limits of the ranges. Each lower limit can be combined with each upper limit to define a range. and the upper limit should each be considered a separate element.

[0025] The term "organoid" refers to the cellular self-organization, architecture, and function present in natural organs. , and form heterogeneous 3D aggregates of cells that recapitulate aspects of signaling interactions. The term "organoid" refers to a cell or tissue culture formed from a suspension cell culture. The spheroids or cell clusters differentiated on the micropatterns, as well as stem cells differentiated on the micropatterns, Contains cells.

[0026] The term "spheroid" was coined to allow for 3D growth rather than growth as a monolayer. The term "spheroid" refers to an aggregate or collection of cells cultured in a uniform manner. Note that this does not imply that the aggregates are geometric spheres. They may be highly organized with defined morphological characteristics or may be disorganized. They may be small clumps containing a single cell type or more than one cell type. Good too.

[0027] The term "micropattern" refers to a pattern that has features on a micro-scale. For example, a micropattern may be a repeating pattern of circles or spheres with diameters on the micrometer scale. Alternatively, the micropattern may include a line width on the micrometer scale. The micropattern may comprise a repeat of lines, or the micropattern may comprise a repeat unit, e.g. , square, triangle, diamond, rhombus, or other two- or three-dimensional geometry The shape has at least one feature on the micrometer scale, e.g., height, width, length The method of the present disclosure may include shapes having different thicknesses, etc. are contemplated, which may include freeform shapes and / or geometric shapes. Micropatterns can be created by techniques including, but not limited to, lithography, stencils, and Using art-recognized micropatterning techniques, including etching and the like. It can be generated by

[0028] The term "marker" or "biomarker" generally refers to a specific detection method for a cell population in culture. Its expression or presence is detected by standard methods (or the methods disclosed herein). and can be identified with one or more cells in the cultured cell population that are of a particular type of cell. Matching molecular markers based on DNA, RNA, proteins, carbohydrates, or glycolipids A marker may be a polypeptide expressed by a cell, or may be a gene, Restriction endonuclease recognition sites or polypeptides expressed by natural cells are Even an identifiable physical location on a chromosome, such as a nucleic acid (e.g., mRNA) that The marker may be an expressed region of a gene, called a "gene expression marker." or a segment of DNA with no known coding function. The biomarker may be a cell-derived product, such as a secreted product.

[0029] "Screening" refers to a process in which one or more properties of one or more molecules are determined. For example, a typical screening process involves one or more One or more properties of one or more molecules that are members of the library are determined. The process is as follows.

[0030] A "library" is a set of at least two molecules, such as small molecule compounds, proteins, peptides, or nucleic acids. Both refer to a collection of two different molecules. For example, a library typically contains at least Large libraries typically contain at least about 10 different molecules. Contains about 100 different molecules, more typically at least about 1,000 different molecules In some applications, the library contains at least about 10,000 or more It contains different molecules.

[0031] "Selection" refers to the process of identifying one or more molecules that have one or more properties of interest. Thus, for example, screening a library to identify , reversion of a disease phenotype to a WT (non-disease) phenotype or toxicity One or more characteristics of the library members may be determined. If one or more of the following are identified as possessing a characteristic of interest (e.g., a disease phenotype), The selection is performed by isolating the library members (reversion to the T phenotype). and further testing, for example, in an animal model. The screening may be, and often is, simultaneous.

[0032] The term "extracellular matrix components" refers to components that serve as a substrate for cell attachment and growth. refers to a molecule or material, whether natural or synthetic, that functions in an extracellular environment. Examples of matrix components include, but are not limited to, collagen, laminin, fibronectin, and the like. tin, vitronectin, elastin, glycosaminoglycans, proteoglycans, and Combinations of any or all of these ingredients may be used (e.g., MATRIGEL ( (Collaborative Research, Catalog No. 40234) Can be obtained.

[0033] The term "tissue-attaching surface" refers to a technology to which cells can adhere in vitro. Surfaces that have a stain, charge, or coating. Examples of adhering surfaces include, but are not limited to: No, but stainless steel wire, VELCRO®, suture material, natural tendon, shared Plastics modified with conjugates (e.g., RGD conjugates), and textured surfaces Examples of such a tube include a silicone rubber tube having a

[0034] A "training set" is a set of images to which one or more models have been applied and based. For example, in the case of organoid phenotype models, The training set may be optionally Included are images of non-diseased and diseased organoids stained with one or more markers.

[0035] Neural Networks Network Architecture Artificial neural networks (ANNs) have been used to model mammalian systems, albeit on a much smaller scale. A processing device (algorithm or actual hardware) that models the neural structure of the brain cortex The basic unit of computation in an ANN is the neuron. In this case, the term "neuron" is used synonymously with the term "node." Large ANNs can have hundreds or thousands or millions of neurons. .

[0036] Neural networks are typically organized into layers. A layer is a set of activation functions. The pattern is input to the input layer, which is composed of many interconnected nodes that contain The input layer communicates with one or more "hidden layers" However, the actual processing is done in the hidden layer in the form of weighted "connections". When used in the literature, the term "hidden layer" is used synonymously with the term "latent layer." The hidden layer then performs a training process to determine the answer, as shown in the example network in Figure 1. The input is connected to the “output layer”.

[0037] The neural network of this disclosure is implemented using PyTorch or TensorFlow, etc. It can be coded using any machine learning framework.

[0038] 5.2.2. Convolutional Neural Networks Convolutional Neural Networks (ConvNets or CNNs) are used for image recognition and CNN is a type of ANN used for classification. CNN image classification takes an input image and processes it. This eliminates the need to process images, classify them into categories, and manually extract features. Because a computer sees an input image as an array of pixels, image resolution determines how fast the computer can The type of matrix array (in terms of h x w x d) into which the computer will process the image. In this case, h = height, w = width, and d = dimension) are determined. For example, RGB is , the array is a 6x6x3 (3 refers to RGB value) matrix image, and the grayscale image is , the array is an image of a 4x4x1 matrix.

[0039] CNNs have dozens or hundreds of layers, each of which is trained to detect different features in images. A CNN can be trained by presenting it with a set of training images. The filter can be applied to each training image at different resolutions. The output of each convolution is used as the input to the next layer. Starting from very simple functions such as degree and contour, we compound the features that uniquely define the object. The roughness can be increased.

[0040] Like other neural networks, CNNs have an input layer, an output layer, and generally It consists of many hidden layers between the input and output layers. These layers learn specific features of the data. They perform operations that modify data in order to improve the performance of the model. The three most common layers are convolutional, activation, and , or ReLU, and pooling.

[0041] When training and analyzing with deep learning CNN, each input image The image generally has a filter (kernel) layer, a pooling layer, and a fully connected layer (FC). The object is then passed through a series of convolutional layers, and a final softmax function classifies the object into a range of 0–1 (or 0% The classification will be based on a probability value (~100%, 0-10, or any arbitrary range).

[0042] Convolutional Layer Convolution is the first layer for extracting features from an input image. By learning image features using small squares of data, relationships between pixels are is stored. It takes two inputs, an image matrix and a filter or kernel. It is a mathematical operation. In the convolution of an image with various filters, the filter is applied This allows operations such as contour detection, blurring, and sharpening to be performed.

[0043] Stride The stride is the number of pixels to shift the input matrix. If the stride is 1, The filter moves one pixel at a time. If the stride is 2, the filter For example, it moves 2 pixels each time.

[0044] Padding The filter may not match the input image perfectly. In that case, the image may not match. Either pad the image with zeros (zero padding) or The latter is called validity padding and can be used to remove parts of an image that are not valid. Only a portion is kept.

[0045] 5.2.2.4. Nonlinearity (ReLU) ReLU stands for Rectified Linear Unit for nonlinear operations. The output is f(x)=max(0,x). The purpose of ReLU is to However, to learn the nonlinear relationship between the input image and the classification, we add nonlinearity to the ConvNet. The solution is to introduce tanh or sigma. There are other nonlinear functions such as id. In terms of performance, ReLU is superior to the other two. Therefore, ReLU is preferred.

[0046] Pooling Layer A pooling layer can reduce the number of parameters when the image is too large. Spatial pooling (also called subsampling or downsampling) is a method to Spatial pooling reduces the dimensionality of the image while preserving important information. , average pooling, sum pooling, etc. Max pooling is the process of extracting the largest feature from a rectified feature map. In the average pooling layer, the input is divided into rectangular pooling regions, and Downsampling is performed by calculating the average of all the elements of the feature map. The sum is called sum pooling.

[0047] 5.2.2.6.Fully connected layer A fully connected layer takes all the neurons in the previous layer and combines them into a single In the fully connected layer, the matrix is ​​flattened into a vector and the neural network is The neural network of the present disclosure can be fed to the fully connected layer in the same way as the A fully connected layer may have more than one fully connected layer. With fully connected layers, features are combined together You can create a model by combining activations such as softmax or sigmoid. A function can be used to classify the output as, for example, wild type (non-diseased) or diseased.

[0048] 5.2.3. Neural models for analyzing non-diseased and diseased organoids network Pre-trained convolutional networks are usually This is significantly more efficient than a non-disease based network, and therefore in some embodiments, Neural networks for classifying images of organoids with either a clinical or disease phenotype A pre-trained network is used to create the 2D image. In this case, we provide a convolutional NN pre-trained on the ImageNet database. There are frameworks available commercially (e.g., ResNet, DenseNet, or V For 3D images, we pre-trained the model on the Kinetics dataset. There are many commercially available NNs (e.g., ResNet or DenseNet). Training may be performed on databases other than ImageNet or Kinetics. However, ImageNet and Kinetics are very large and comprehensive image datasets. We provide a set of neural network models. Although other network architectures can be used, convolutional The network is preferred because it excels at image classification.

[0049] Residual networks (ResNets), a subclass of convolutional networks, In some embodiments, a ResNet with 18 to 152 layers is used to Various depths (18, 34, 50, 101, or The pre-trained ResNet (which has 152 layers) is a key machine learning framework. Frameworks (e.g., ResNet18, ResNet34, ResNet50, They are available in convolutional It consists of a sparse layer, a batch normalization (BatchNorm) layer, and a rectified linear (ReLU) layer. It may contain blocks of layers that are fed into a pre-trained ResNet. The final average pooling and densely connected layer ted layer) may be removed and replaced with a custom layer as described herein. It uses a convolutional layer to iterate over the input images using filters learned during the training process. We may convolve 3x3 pixel blocks. ReLU is a function f(x) = max(0, x) to all of the inputs. We introduce the nonlinearity required to learn the nonlinear relationship between images and classifications into the network. A BatchNorm layer can be used to normalize the network activations. This improves stability and allows for more efficient training.

[0050] The final layers of a NN are generally more specific to the dataset than the earlier layers, and therefore Remove the final average pooling and fully connected layers of the pre-trained network. In addition to the fully connected layers, we use an untrained adaptive average pooling layer. rage Pooling), Adaptive Maximum Pooling, Batch N orm layer, and then replace some or all of the dropout layer with non-disease phenotypes or performed a final softmax operation to classify images of organoids with disease phenotypes. Preferably, the neural network includes an adaptive average pooling layer, an adaptive max. It includes pooling layer, batch norm layer, and dropout layer, but batch Implementations that omit some of these layers, such as Norm and / or Dropout. A pooling layer reduces the spatial size of what is being represented, and A pullout layer can avoid overfitting, and a fully connected layer uses all the activations of the previous layer. has access to the classifications and puts them all together (this is the final (This is the initial setting for the layer.) In the preferred embodiment, a fully connected layer of 512 neurons is used. The final fully connected layer immediately follows this, and consists of two neural networks, one for non-disease and one for disease. A final softmax is used to sum up the activations of these neurons. It can be converted into a probability of 1.

[0051] Training Most ANNs change the weights of connections according to the input patterns presented to them. It contains some form of "learning rules."

[0052] There are many different types of learning rules used by neural networks, One of them is called the delta rule. The delta rule is a backpropagation neural network. The most common one is called backpropagational neural networks (BPNN). It is often used by the ANN class. Backpropagation is an abbreviation for backward propagation of errors.

[0053] With the delta rule, as with other types of backpropagation, "learning" is done by iterating over the forward output. Backward error propagation of activation flow and weight adjustments results in a In simpler terms, it is a supervised process that occurs when a pattern is first When presented to the network, the network makes random guesses as to what it could be. The network then calculates how far that answer is from the actual answer. Check whether the connections are high enough and adjust the connection weights appropriately.

[0054] Within each hidden layer node there is a sigmoid activation function which introduces nonlinearity.

[0055] Backpropagation follows the steepest vector of the error surface towards the "global minimum". Then, gradient descent is performed in the vector space of solutions. The global minimum is the lowest possible For most problems, the solution space is highly irregular. This can cause the network to get stuck in a "local minimum" that is not the best global solution. It has quite a few "pits" and "hills" that can be removed.

[0056] Since the nature of the error space cannot be known in advance, neural network analysis is Often many individual runs are required to determine a good solution. Learning rules include: In general, this process controls the "speed" of learning (β coefficient) and the "momentum" of learning. The speed of learning is actually dependent on the current solution. and the global minimum. Momentum is the rate at which the network converges between the error surface to overcome obstacles (local minima) and settle at or near the global minimum Assistance to:

[0057] It is also possible to overtrain a neural network, which can Ensure that the workpiece is trained to respond accurately to only one type of input. This means that an overtrained network will perform well on new inputs. Therefore, it is not useful.

[0058] The neural network of the present disclosure is trained on training images (e.g., 70%) and validation Images (e.g., 30%) divided into non-disease and disease organoid images It can be trained using a set of images (preferably multi-channel). Preferably, the training images include at least 300 non-disease images and at least 300 disease images (e.g., 300-2000 images each, 300-1500 images each , 300-1000 images each, or 300-500 images each) or more The image may be a 2D image or a 3D image. A neural network (NN) is trained on a training dataset. This can be done.

[0059] The images in the training set were classified into organelle markers, cell differentiation markers, and cell compartment markers. ols stained for one or more markers, such as markers for chromatin deficiency, ... The one or more markers may preferably include: The gross morphological characteristics of the organoids and / or the specific cells they contain The staining agent may include one or more markers that highlight the population. Exemplary staining agents include nuclear staining agents. Staining (e.g., DAPI and Hoechst nuclear stain), BrdU, calcein green, Di Membrane dyes such as I, DiO, DiD, and DiR, and fluorescently labeled antibodies against GM130 Cell Signaling's Golgi stain, ER-Tracker™ Green ER staining reagents such as ing Technology, phalloidin, and N-cadmium These include fluorescently labeled antibodies against HERIN, E-cadherin, ZO-1, and collagen. do.

[0060] Also use reagents specific to certain types of organoids or certain types of differentiated cells For example, a reagent for staining neurofilament or nestin (e.g. For example, fluorescently labeled antibodies can be used to stain brain organoids. Color reagents can be used to stain the neuronal population. By using this method, neural crest cells can be stained.

[0061] Training involves showing images to the network and determining the output probability of WT or disease. The next time the image is presented to the network, the network will By changing the weights of the network, the predictions will be closer to the values Preferably, the images are presented to the network in a random order during training. The fitting procedure is based on the major neural network frameworks. This is done using the backpropagation algorithm implemented in the network. And many times, for example, 200-800 times, 300-700 times, or 400-600 times ( For example, about 200 times, about 300 times, about 400 times, about 500 times, about 600 times, about 700 times) The images are displayed in a way that does not significantly change the content of the image, but allows the network to learn. Expand the pool of images that can be used to find the right match by applying a set of image transformations to each image. Data augmentation operations include the following: rotation, cropping, and 90-110° zooming of images. 0% scaling, and some or all of the changes in image contrast (e.g. 1, 2, Training can be done by varying the hyperparameters. Parameter (number of layers, momentum and learning rate, dropout percentage, number of epochs) ) to find the optimal set of these parameters. In an embodiment, the hyperparameters are the following: number of layers, momentum, and learning rate. Two, three, or four combinations of degree, dropout percentage, and number of epochs This includes.

[0062] Data augmentation and dropout are two strategies to avoid overfitting. Overfitting is avoided by comparing network performance on the training and validation sets. If the training accuracy is higher than the validation accuracy, then the training accuracy can be tested for underfitting. If so, the network is overfitting, and vice versa. Use the validation image set by calculating the percentage of correctly classified images. From this evaluation, a "Z'" score is obtained, which measures the accuracy of the network. A Z' score above 0.5 indicates a good screening assay. In a pre-trained network, the training The number of images can be relatively small, preferably a minimum of about 500 training images. It's fine.

[0063] Classification into WT and disease is performed by using training data to determine whether an image is WT or diseased. This is an example of supervised learning because the training data is given to the network during training. A matching strategy is a natural choice when the image classes (WT and disease) are known in advance. .

[0064] In some embodiments, unsupervised learning is used according to the methods described herein. An example of untrained learning is the use of autoencoders, see Figure 1. The encoder and decoder neural networks are then trained to reconstruct the reconstructed data into a It is trained to match the data as closely as possible, and is The autoencoder aims to reproduce the input as an output. It is used to learn efficient data coding in an unsupervised manner. An autoencoder is a type of artificial neural network. By training the network to ignore the "noise" that typically It is designed to learn representations (encodings) of a dataset for dimensionality reduction. The reconstruction side is trained together with the reduction side, and the autoencoder trains the reduced encoding. From the input, the program attempts to generate a representation that is as close as possible to the original input, i.e., its name.

[0065] A typical autoencoder includes an input layer, a hidden layer, and an output layer. Data passes through the input layer and enters the hidden layer. The number of nodes in the hidden layer is typically smaller than the number of nodes in the input layer. The number of bits is less than the number of bits required to compress or encode the data. Following encoding, the data output from the hidden layer goes to the output layer. Typically, the number of nodes in the output layer is larger than the number of nodes in the hidden layer, so the data dimensionality is reduced. increases. Therefore, decompression or decoding is performed.

[0066] The unsupervised nature of this machine learning method means that the autoencoder does not Additional information about the gene (such as whether it is derived from a wild-type or diseased cell line) It has the advantage of learning representations of data without using any algorithms, and therefore is useful for estimating compound toxicity. In addition, vector-based data representation allows for the identification of wild-type and disease phenotypes. There is also the advantage that the difference with the current type can be removed from the vector space. The difference is not relevant for toxicity determination. , in particular by using convolutional autoencoders. (a) removing the distance between wild type and disease from the latent space; (b) calculating the distance from the mean vectors of the wild type and disease phenotypes; (c) The distance calculated in step (b) and the standards for the wild-type and disease phenotypes. Comparing the deviations; and (d) determining a toxicity value Toxicity can be determined by

[0067] A flow chart illustrating an exemplary implementation of neural network training is shown in FIG. 2A. The training process 100 is performed by training a block The training image set 110 includes images of non-disease phenotypes as described above. The images may include images of organoids and disease phenotypes of organoids. Expand the image by applying an image transform (e.g., as described above). In 111, the individual images are analyzed by a neural network and then in block 11 2. Classifying organoids into diseased or non-diseased ones using a neural network In block 113, the classification is compared to the true value to determine whether the image will be processed by the neural network the next time. So that when displayed against the net, the classification is more likely to be closer to the true value. The weights of the network are updated. All images in the training set are run through the network. Once the process is complete, the process is repeated a set number of times (e.g., about 500 times). When the repetition of is completed, the training is completed, and the training A neural network is provided.

[0068] 5.2.5.Analysis Once the neural network has been "trained" to a satisfactory level, it can be run on other data. It can be used as an analytical tool for trading. Do not specify a training run, and instead allow the network to operate only in forward propagation mode. New input is presented to the input pattern, where the new input is filtered. The input is filtered and processed by the intermediate layer as if it were being trained. However, at this point the output is preserved and no backpropagation occurs. The output of the forward run is the data , which can then be used for further analysis and interpretation.

[0069] The neural network of the present disclosure can be used to image organoids exposed to a test molecule. The images of the organoids treated with the test molecule can be analyzed using a network. The network can be analyzed by the following: A probability score (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 3 For each group, 1 represents WT and 0 represents disease. The scores may be averaged (e.g., to obtain a score for the compound). It is a measure of the extent to which the phenotype is reverted to WT. Compounds are ranked by their scores. Compounds with a score >0.95 can highly clearly differentiate the disease phenotype from the WT. It can be determined that the

[0070] The accuracy of the network is preferably verified by using untreated control organoids. Untreated control organoids, preferably non-diseased organoids and diseased organoids, are examined. Both organoids could be classified as diseased or non-diseased, as could treated organoids. If the control organoids, which are larger than the cutoff volume, are correctly classified, the neural network The results can be considered accurate. For example, if the control organoids show a higher than 50% There is a probability of more than 60%, more than 70%, and more than 80% Correctly classify patients as non-diseased or diseased more than 90% of the time, or more than 95% of the time If the neural network is accurate, then the neural network can be considered accurate. In some embodiments, accuracy validation is performed simultaneously with compound screening.

[0071] 1 shows an exemplary implementation of a process for identifying molecules that are biologically active against a disease A flow chart is shown in FIG. 2B. Process 101 separates diseased organoids for testing. Starting with block 120, the child is exposed to The organoids are then imaged in block 121 to identify one or more of the organoids. The images are then processed in block 122 to generate a probability score for disease or non-disease. A trained neural network that is trained to allocate cores Finally, the neural network performs a computation on the test molecule in block 123. Exposed diseased organoids are assigned a probability score of diseased or non-disease. If the data show a reversion of the disease phenotype to a non-disease phenotype (e.g., the probability score is 5.3.3) and the test molecule is It can be identified as being biologically active against a disease.

[0072] 5.2.6.Toxicity Toxicity uses the same classification network as described above, but with the addition of the last two networks. The neurons in the fully connected layer upstream of the neurons (e.g., the neurons in the fully connected layer described in Section 5.2.1) 512 neurons). In this case, the penultimate fully connected layer is used. The fully connected upstream layer of 512 neurons is A 512-dimensional vector that contains more information than the final 2-dimensional (non-diseased vs. diseased) vector. The information contained in this high-dimensional latent vector can be used to determine how the compound In some embodiments, a measure of the degree of toxicity of the autoencoder can be defined. A classifier (e.g., a convolutional autoencoder, or a variational autoencoder, a superclassifier) or other autoencoders such as denoising autoencoders. The toxicological analysis is carried out using the 100% ethanolamine ester.

[0073] Each organoid image analyzed by the network is assigned a phenotype (WT vs. disease) score. In addition, it results in a latent vector. The latent vector of organoids exposed to the same compound can be averaged together with the WT and disease latent vectors. The vectors are optionally embedded using t-distributed stochastic n We can cluster the data into a two-dimensional space using t-SNE (traditional neighbor embedding). Other dimensionality reduction methods such as principal component analysis (PCA) can also be used. The latent vectors of the organoids treated with the offspring and the closest untreated organoid vector (W The distance from the treated organoids to the WT or disease (T or disease) indicates how far the treated organoids are from either the WT or disease. This provides a measure of how far the compound is from the toxicity of the compound. In a preferred embodiment, the upstream latent vectors are clustered before determining the distances. The distance is set to 0–1 (or 0–10, The difference between WT and disease is normalized to a range of 0% to 100%, or any arbitrary range. may be considered less relevant for more toxic test compounds. As toxicity increases, the WT-disease axis may be linearly compressed, so that each compound The resulting triangle is then arranged such that the corners represent the WT phenotype, the disease phenotype, or This refers to compounds that have a toxic phenotype. This is illustrated in Figure 13.

[0074] In one embodiment, toxicity is calculated by first (a) extracting the fully connected layers upstream of the two final nodes. and (b) detecting a first population of mammalian cells contacted with the test molecule and a second population of mammalian cells contacted with the test molecule. A latent vector associated with a first cell population that has not been subjected to the quantification and a latent vector associated with a second cell population that has not been subjected to the quantification are The difference between the potential vectors detected in the first mammalian cell population contacted with the test molecule is quantified. a latent vector associated with a first cell population that has not been contacted with the test molecule; by providing the degree of difference between the latent vectors associated with the cell populations will be done.

[0075] The degree of difference indicates the amount of latent vector associated with the first mammalian cell population contacted with the test molecule. a latent vector associated with a first cell population that has not been contacted with the test molecule; Calculated by taking the minimum of the difference between the latent vectors associated with the two cell populations will be done.

[0076] The degree of difference can be determined across the entire latent space. The space is at least 10, at least 100, at least 250, at least 400, At least 500, at least 600, at least 800, or at least 1000 dimensions Includes.

[0077] An example of a process for identifying molecules that are biologically active against a disease and assessing their toxicity A flow chart showing an exemplary implementation is shown in FIG. 2C. In process 102, blocks Blocks 120, 121, 122, and 123 are blocks of process 101 shown in FIG. 2B. The process 102 corresponds to the blocks 120, 121, 122, and 123. Further comprising culturing one or more control organoids at 130, The organoids are organoids that have not been exposed to a test molecule, preferably non-disease organoids. The control organoids include organoids and diseased organoids. The image is then analyzed by the trained neural network in block 122. In block 131, the trained neural network is A probability score of toxicity or non-toxicity is assigned to diseased organoids exposed to the test molecule. In some cases, for example, to evaluate the toxicity of a test molecule previously identified as biologically active. If so, it may be desirable to omit block 123 and perform process 102. be.

[0078] In one embodiment, toxicity is determined by normalizing the distance in the latent space by the maximum occurrence distance. The range of toxicity is 0 (no toxicity) to 1 (maximal toxicity, i.e., the number of organoids that are not diseased in the latent space). The scale of the organoids that are the furthest away from the target organoids (the non-diseased organoids that were in contact with the test molecule) is defined as This can be done.

[0079] Dose-response analysis The system and method of the present disclosure include an IC 50 and L.C. 50 Optional dose-response analysis The analysis features may further include the following four-parameter logistic function: , and can be fitted to a dose-response. y=min+(max-min) / (1+10^((x-log(IC50)) * Tilt) where x is the logarithm of the test molecule concentration and y is the therapeutic potential. 0 is the required value. The same function is applied to LC 50 It can also be used for the four formulas. The parameters are fitted using a nonlinear least squares fitting (e.g., Levenberg-Math. Levenberg-Marquardt algorithm or Trust Region Reflection It is possible to fit the data using an open source (directive) algorithm. Source Python package script or any other with curve fitting Software such as Matlab, R, or Excel can be used.

[0080] 5.2.8. Computer Use Requirements Network training and analysis can be performed on any computer. However, it is preferably implemented on a Graphical Processing Unit (GPU). It is preferable that the computations performed during training are performed in an optimal manner on the GPU. For example, GPUs can significantly speed up the training process. GPUs, such as the NVIDIA GeForce 1080 T Using a modern computer with i (e.g. 12 cores, 64GB RAM) ,Training on thousands of images can be done in under an hour. Capabilities, for example, with computers using multiple graphics cards in parallel, Training time can be reduced.

[0081] 5.3. Application of the Neural Network of the Present Disclosure The disclosed system and method provides a method for identifying disease expressions that are hidden behind large biological variability. It can be used to identify and interpret types and their reversions. The deep neural network described here correlates well with WT and diseased organoids. Different features can be identified and combined in an optimal and unbiased manner, It is ideally suited for this task.

[0082] Single or multiple microscopic images of each organoid can be taken. Preferably, there are multiple images of the image. The images may be either 2D or 3D. Each image of an organoid is referred to herein as a "channel." Multiple channels can be For example, by using different filters (e.g., 3-4 filters), The structure and expression of various proteins (e.g., differentiation markers, DAPI, membrane Combine information on WT phenotype and disease (e.g., staining, mitochondrial markers, etc.) A unique fingerprint of the patient phenotype can be obtained. The fingerprint can include: Examples of the number of cells / signal area, intensity, shape, etc. Instead of organoids, single cells and cell layers can also be used, but microparticles Organoids produced using TURN are more reproducible than single cells or cell layers. Organoids are preferred because they can be used to generate

[0083] In one embodiment, the organoids are stained with 1, 2, 3, 4, or 5 different stains. Each dye has a different color.

[0084] Using the neural network of the present disclosure, during training, WT cells, cell layers, and organoids and to learn the differences between diseased cells, cell layers, and organoids. The trained network of the present disclosure can then be subjected to a test molecule Images of organoids treated with IFN-α can be displayed, and the network shows how well the organoids are treated compared to the WT. The degree of similarity is expressed as 0 to 1 (or 0 to 10 or 0% to 100%). Images can be classified based on a score in any arbitrary range. Images with a positive result can be defined as hits that successfully rescue the phenotype. The potential toxicity of the compound can also be analyzed simultaneously.

[0085] The neural networks and methods of the present disclosure are not only useful for drug testing, It can also be used to test the effectiveness of genetic manipulations for phenotypic reversion. For example, it may silence alleles that carry disease-causing mutations, or If the affected gene causes a loss of function, the wild-type gene is overexpressed to compensate. As illustrated in the Examples, this strategy has been shown to be effective in the case of Huntington's disease. In this case, the wild-type huntingtin gene was replaced by the huntingtin knockout gene. The method described herein was used to overexpress the gene in a knockout (KO)-harboring cell line background. Using this method, the induced organoids can be tested for phenotypic reversion. Additional toxicity phenotypes are not necessarily required when testing organoids, and some In embodiments, modified organoids are scored for the probability of being wild type without assessing toxicity. This is also fine.

[0086] Advantageously, various embodiments of the disclosed systems and methods allow for the acquisition of biological information from imaging data. Identify and quantify biological properties / analytes and then use the quantified biological properties / analytes to A hierarchical analysis framework that can identify and characterize one or more pathologies based on Imaging is used as an intermediate step to assess pathology and to identify the underlying biology. This hierarchical approach to examining pathology allows for direct interpretation of pathology from the underlying imaging data. For systems and methods configured to determine and characterize This provides economic and processing advantages.

[0087] Organoid formation Organoids contain more than one cell type. Organoids are typically It begins as a single cell or small aggregates that do not exhibit any spatial organization. Over time, Self-organization occurs and various cell types emerge, which then dynamically adapt within the organoid. The specific spatial organization of organoids varies for each type of organoid. however, patterns or areas of differentiation are always present.

[0088] In various embodiments, the organoids are about or at least 10 12 Cells or Less than or equal to about or at least 10 11 A cell or less, about or at least 1 0 10 Cells or less, about or at least 10 9 Cells or less, approximately Or at least 10 8 Cells or less, about or at least 10 7 Cells or is less than, about or at least 10 6 A cell or less, about or at least 10 5 Cells or less, about or at least 10 4 Cells or less, approximately Or at least 10 3 Cells or less, about or at least 10 2 Cells or or less, about or at least 10 cells or less, or any of the above values. or a range of numbers of cells (e.g., 10 3 ~10 8 cells, or 10 4 cells ~10 10 cell , or 10 2 ~10 6 cells).

[0089] In certain aspects of the present disclosure, organoids are generated from primary cultured cells.

[0090] In other aspects, organoids are generated from immortalized cells in culture.

[0091] The primary cultured cells and immortalized cells are stem cells, e.g., totipotent stem cells or pluripotent stem cells. may be also possible.

[0092] For example, the following cells: pluripotent cells, induced pluripotent cells, adult stem cells, non-embryonic cells, or Non-embryonic stem cells can be used.

[0093] In some embodiments, the stem cells are human embryonic stem cells. In other embodiments, the stem cells are adult stem cells or non-embryonic stem cells. In some embodiments, the stem cells are induced pluripotent stem cells. The cells are of human origin.

[0094] In certain specific embodiments, the organoids comprise stem or progenitor cells. In specific embodiments, the stem or progenitor cells are selected from the group consisting of embryonic stem cells, embryonic germ cells, induced pluripotent stem cells, and the like. Stem cells, mesenchymal stem cells, bone marrow derived mesenchymal stem cells, bone marrow derived mesenchymal stromal cells, umbilical cord stem cells, Amniotic fluid hepatocytes, amniotic membrane-derived adherent cells (AMDACs), osteogenic placental adherent cells (OPACs), adipose Stem cells, limbal stem cells, dental pulp stem cells, myoblasts, endothelial progenitor cells, neural stem cells, derived from exfoliated teeth Stem cells, hair cyst stem cells, skin stem cells, parthenogenically derived stem cells stem cells, reprogrammed stem cells, amnion-derived adherent cells, or side populations Certain other specific embodiments are side population stem cells. In the present case, the organoids comprise hematopoietic stem or progenitor cells.

[0095] In certain other specific embodiments, any of the organoids described herein are In more specific embodiments, the differentiated cells include one or more of the following: : Endothelial cells, epithelial cells, skin cells, endodermal cells, mesodermal cells, fibroblasts, bone cells, cartilage cells bone cells, natural killer cells, dendritic cells, hepatocytes, pancreatic cells, or stromal cells; · Salivary gland mucosal cells, salivary gland serous cells, von Ebner gland cells, mammary gland cells, lacrimal gland cells, ear Sweat gland cells, eccrine dark cells, eccrine clear cells, apocrine cells, molar cells Gland of Moll cells, sebaceous gland cells, Bowman's gland cells, Brunner's gland cells, seminal vesicle cells gland of Littre cells, prostatic cells, bulbar urethral gland cells, Bartholin gland cells, and glands of Littre cells. cells), endometrial cells, isolated goblet cells, gastric lining mucosal cells, gastric gland zymogen cells, gastric gland acid-secreting cells. alveolar cells, pancreatic acinar cells, Paneth cells, type II pneumocytes, Clara cells, growth hormone-producing cells, milk Thyrotropes, thyrotropes, gonadotropes, accessory Renal corticotropic cells, pituitary intermediate lobe cells, magnocellular neurosecretory cells, intestinal cells, airways Cells, thyroid epithelial cells, parafollicular cells, parathyroid cells, parathyroid chief cells, eosinophilic cells, adrenal glands Cells, Chromaffin cells, Leydig cells, Theca interna, Luteal cells, Granulosa luteal cells, Ovum alveolar luteal cells, juxtaglomerular cells, condensate cells, perivascular pole cells, mesangial cells, ·blood vessel and lymphatic vascular endothelial cells ial fenestrated cells, blood vessel and lymphatic endothelial cells ic vascular endothelial continuous cells), vascular and lymphatic endothelial splenocytes (blood vessel endothelial cells) essel and lymphatic vascular endothelial splenic cells), synovial cells, serosa cells (abdominal the lining of the tunica albuginea, pleura, and pericardial cavity), squamous cells, columnar cells, dark cells, vestibular membrane cells (of the ear endolymphatic space endothelium), basal cells of the stria vascularis, peripheral cells of the stria vascularis (endothelium of the endolymphatic space of the ear), Usius cells, Betcher cells, choroid plexus cells, soft arachnoid squamous epithelial cells, pigmented ciliated epithelial cells Cells, non-pigmented ciliated epithelial cells, corneal endothelial cells, PEG cells, Airway ciliated cells, fallopian tube ciliated cells, endometrial ciliated cells, reticulotesticular ciliated cells, efferent duct ciliated cells , ciliated ependymal cells, Epidermal keratinocytes, epidermal basal cells, fingernail and toenail keratinocytes, nail bed basal cells Cells, medullary hair stem cells, cortical hair stem cells, keratinous hair stem cells, keratinous hair root sheath cells, hair roots in the stratum corneum sheath cells, root sheath cells of Henle's layer, outer root sheath cells, hair matrix cells; ·Surface cells of stratified squamous epithelium, epithelial basal cells, urinary epithelial cells, ·Outer hair cells of the organ of Corti, inner hair cells of the organ of Corti, basal cells of the olfactory epithelium, low temperature Sensitive primary sensory neurons, thermosensitive primary sensory neurons, Merkel cells of the epidermis, olfaction Receptor neurons, pain-sensitive primary sensory neurons, photoreceptor rod cells, photoreceptor blue-sensing Receptive cone cells, photoreceptors green-sensitive cone cells, photoreceptors red-sensitive cone cells, proprioceptive Primary sensory neuron, Touch-sensitive primary sensory neuron, Type I carotid body cell, Type II carotid body cell corpuscle cells (blood pH sensor), type I hair cells of the vestibular apparatus of the ear (acceleration and gravity), Type II hair cells of the vestibular apparatus, type I taste bud cells, Cholinergic neurons, adrenergic neurons, peptidergic neurons, Inner pillar cells of the organ of Corti, outer pillar cells of the organ of Corti, inner supporting cells of the organ of Corti, External supporting cells of the organ, border cells of the organ of Corti, Hensen's cells of the organ of Corti, supporting cells of the vestibular apparatus Alveolar vesicles, taste bud supporting cells, olfactory epithelium supporting cells, Schwann cells, satellite cells, enteric glial cells, Astrocytes, neurons, oligodendrocytes, spindle neurons, Prelental epithelial cells, crystallin-containing lens fiber cells, hepatocytes, adipocytes, white Adipocytes, brown adipocytes, liver lipid cells, Renal juxtaglomerular cells, renal glomerular podocytes, renal proximal tubule brush border cells, thin loop of Henle Segment cell (loop of Henle thin segment cell), renal distal tubule cell, renal aggregate duct cell, type I pneumocyte, pancreatic duct cell, nonstriated duct cell, duct cell, Intestinal brush border cells, exocrine gland striated duct cells, on the gallbladder Epithelial cells, non-ciliated cells of the efferent duct of the testis, epithelial chief cells, epididymal basal cells, ameloblastic epithelial cells alveolar cyst, semilunar epithelial cell, interdental epithelial cell of the organ of Corti, loose connective tissue fibroblast, corneal stromal cell , tendon fibroblasts, bone marrow reticular tissue fibroblasts, non-epithelial fibroblasts, pericytes, nucleus pulposus cells, Cementoblast / cementocyte, odontoblast, odontocyte, hyaline cartilage chondrocyte alveolus, fibrocartilage chondrocyte, elastic cartilage chondrocyte, osteoblast, osteocyte, osteoclast, osteoprogenitor cell, Vitreous cells, stellate cells (ear), hepatic stellate cells (Ito cells), pancreatic stellate cells ), Red skeletal muscle cells, white skeletal muscle cells, intermediate skeletal muscle cells, nucleus of muscle spindles, sac cells of muscle spindles Nuclear chain cells of muscle spindle, satellite cells, normal cardiomyocytes, nodular Nodal heart muscle cells, Purkinje fiber cells, smooth muscle cells, iris muscle cells epithelial cells, myoepithelial cells of exocrine glands, Reticulocytes, megakaryocytes, monocytes, connective tissue macrophages, epidermal Langerhans cells, dendritic cytocytes, microglial cells, neutrophils, eosinophils, basophils, mast cells, helper T cells, supplemental Lesser T cells, cytotoxic T cells, natural killer T cells, B cells, natural killer -cells, Melanocytes, retinal pigment epithelial cells, or Oocytes / Oocytes, Spermatocytes, Spermatogonia, Sperm, Follicle cells, Sertoli cells alveoli, thymic epithelial cells, and / or renal interstitial cells.

[0096] In certain specific embodiments of any of the organoids presented herein, , Organoids are derived from liver, kidney, pancreas, thyroid, lung, intestine, colon, prostate, brain, breast, and ovary. , stomach, esophagus, tongue tissue, taste buds, inner ear, or retina.

[0097] Suitable culture conditions for mammalian cells are well known in the art or can be determined by those of skill in the art. (See, for example, Animal Cell Culture: A Practical Approach 2nd Ed., R see Dickwood, D. and Hames, B. D., eds. (Oxford University Press: New York, 1992). The culture medium may be commercially available. Non-limiting examples of media include Dulbecco's Modified Eagle's Medium (DMEM, Life Sciences). Technologies), Dulbecco's modified Eagle's medium / nutrient mixture F-12 ( DMEM / F-12, Life Technologies), Minimum Essential Medium (MEM, Sigma, St. Louis, MO), and hepatocyte medium.

[0098] The above-described medium may be supplemented with optional components at appropriate concentrations or concentrations as necessary or desirable. The cell culture medium may be supplemented with complementary ingredients or components, if desired, in amounts of The solution provides at least one component from one or more of the following categories: (1) a source of energy, usually in the form of carbohydrates such as glucose; (2) essential amino acids (3) All acids, and usually a basic set of 20 amino acids + cysteine; (4) free fatty acids or lipids, e.g. for example, linoleic acid; and (5) typically at very low concentrations, usually in the micromolar range. Trace elements are defined as inorganic compounds or naturally occurring elements required in the diet.

[0099] The medium may also be selectively supplemented with one or more components from any of the following categories: (1) salts, e.g., magnesium, calcium, and phosphate salts; (2) ) Serum, insulin, transferrin, epidermal growth factor, and fibroblast growth factor, etc. (3) protein and tissue hydrolysates, e.g., purified Peptone or peptones that can be obtained from gelatin, plant materials, or animal by-products (4) Mixtures of nucleosides, such as adenosine, thymidine, and hypoxanthine. (5) a buffer such as HEPES; (6) a gentamicin or ampicillin (7) cytoprotectants, e.g., pluronic polyols; and (8) gastrointestinal agents. Lactose.

[0100] The cultures may be supplemented with reagents to promote organoid formation. Exemplary agents used in the art to promote endothelial formation include EGF, F GF10, HGF, R-spondin, BMP4, WNT3A, retinoic acid, GSK3β inhibitors, TGF-β inhibitors, HDAC inhibitors, ROCK inhibitors, Noggin, Activin A, p38 inhibitors, and gastrin. Table 1 below lists the various types of organoids. Reagents that can be used to promote the formation of endothelial vesicles are presented.

[0101] [Table 1-1]

[0102] [Table 1-2]

[0103] Cells that are maintained in culture are those that have been transferred from a previous culture to a culture with fresh medium. In one embodiment, the organoid-producing cells ( For example, human embryonic stem cells (iPSCs) derived from human epithelial cells have been shown to be able to grow at a low rate in cell culture. At least 3 passages, at least 4 passages, at least 5 passages, at least 6 passages, at least 7 passages, at least 8 passages, at least 9 passages, at least 10 passages, at least 11 passages at least 12 passages at least 13 passages at least 14 passages at least 15 passages stable for at least 20 passages, at least 25 passages, or at least 30 passages The temperature can be maintained constant.

[0104] Organoid-producing cells contain expression vectors (constructs) such as plasmids. The expression vector constructs can be prepared by transformation, microinjection, transfection, and They can be introduced by transfection, lipofection, electroporation, or infection. An expression vector contains a coding sequence that encodes the protein to be expressed and produced. or a portion thereof. Expression vectors containing the sequences encoding the polypeptides and appropriate transcriptional and translational control elements are within the skill of the art. These methods can be produced using methods well known and practiced by those of skill in the art. The method is as follows: J. Sambrook et al., 1989, Molecular Cloning, A Laboratory Manual, Co Spring Harbor Press, Plainview, NY and FM Ausubel et al., 1989, Current Protocols in Molecular Biology, John Wiley & Sons, New York, NY Synthetic techniques, in vitro recombinant DNA techniques, and in vivo genetic recombination are used Some examples include:

[0105] In some embodiments, the organoid producing cells are grown in a microwell plate (e.g., 2 In a 4-well or 96-well plate, preferably 20 to 30 ols per well The cells are cultured under conditions that produce ganoids.

[0106] In one embodiment, organoid producing cells are cultured at a final density of 75,000 cells per well. In another embodiment, the cells are plated into wells of a 24-well plate at 100° C. Approximately 50,000 cells per well, approximately 55,000 cells per well, approximately Approximately 60,000 cells per well, approximately 65,000 cells per well, approximately 70,0 00 cells, approximately 75,000 cells per well, approximately 80,000 cells per well, Approximately 85,000 cells per well, approximately 90,000 cells per well 24-well plates at a final density of approximately 95,000 cells or 100,000 cells per well. The cells are plated into the wells of a 3×100 well plate.

[0107] In another embodiment, the cells are at least 50,000 cells per well, at least 50,000 cells per well. At least 55,000 cells per well, at least 60,000 cells per well, At least 65,000 cells per well, at least 70,000 cells per well, At least 75,000 cells per well, at least 80,000 cells per well , at least 85,000 cells per well, at least 90,000 cells per well Cells, at least 95,000 cells per well, or at least 1 0,000 cells are plated into wells of a 24-well plate at a final density of 100,000 cells.

[0108] In one embodiment, a full medium change is performed every 3 days. In another embodiment, a full medium exchange is performed at least every day for at least 2 Every day, at least every 3 days, at least every 4 days, at least every 5 days, at least every 6 days every 7 days, at least every 8 days, at least every 9 days, at least every 10 days every 11 days, at least every 12 days, at least every 13 days, or It takes place at least every 14 days.

[0109] Various types of stem cells can be used to generate organoids. For example, Human pluripotent stem cells (hPSCs), such as iPSCs (e.g., McCracken et al. al., 2014, Nature 516(7531):400-4;Spence et al., 2011, Nature 470(7332):105-9 ;Si-Tayeb et al., 2010, Hepatology 51(1):297-305;Dye et al., 2015, eLife. 4: e 05098;Lancaster et al., 2013, Nature 501(7467):373;Takasato et al., 2015, Natu re 526(7574):564-8; see Takasato et al., 2016, Nat. Protoc. 11(9):1681-92) and and human adipose-derived stem cells (hAdSCs) (e.g., Bartfeld et al., 2015, Gastroenter ology 148(1):126-136;Sato et al., 2011, Gastroenterology 141(5):1762-72;Huch e t al., 2015, Cell 160(1-2):299-312;Boj et al., 2015, Cell 160(1-2):324-38;Kart haus et al., 2014, Cell 159(1):163-175) can be used. Human embryonic stem cells such as RUES-1 or RUES-2 (rues.rockefeller.edu) Cell lines may also be used.

[0110] The composition of the extracellular matrix component solution will vary depending on the tissue being produced. Examples of suitable extracellular matrix components include, but are not limited to, collagen, laminin, and the like. fibronectin, vitronectin, elastin, glycosaminoglycan, proteo glycans, and combinations of some or all of these moieties (e.g., Ma trigel™, Collaborative Research, Catalog No. 40234). In tissues that contain cell types that are responsive to mechanical forces, the extracellular matrix The solutions of components are preferably arranged in a manner such that the cells are exposed to forces associated with the internal tension of the gel. The tissue gels or coalesces into a gel.

[0111] Organoid formation can be promoted by spatially confining stem cells during differentiation Stem cells can bind to mycoproteins such as laminin to control the geometry of cell colonies. The cells can be seeded onto a surface having patterned extracellular matrix proteins. Both the size of the micropattern and the type of medium influence the differentiation pattern results. Bosu (e.g. Deglincerti et al. 2016, Nature Protocols 11: 2223-2232; Deglincer ti et al. 2016, Current Topics in Developmental Biology, (116):99-113;Metzger e t al., 2018, Current Opinion in Genetics & Development, 52:86-91;Etoc et al., 2 (See 2016, Developmental Cell, 39(3):302-315.) Differentiating stem cells on micropatterns By applying the techniques described above, it becomes possible to generate different tissue types in a reproducible manner. For example, 3D scaffolds fabricated using hydrogels such as Matrigel™ have also been used. (See, e.g., Yin et al., 2016, Cell Stem Cell, 18(1):25-28) .

[0112] In other embodiments, the organoids are cultured, e.g., as described in the exemplary processes below. They are produced in vitro from individual cells of the tissue of interest according to the method of the present invention.

[0113] As a first step in this process, disaggregated or partially disaggregated cells are cultured in a centrifuge. The suspension is then mixed with a solution of the extracellular matrix components to create a suspension. The tissue is similar to the in vivo macroscopic morphological characteristics of tissues and is coupled to a vessel. The cells and the microorganisms are then placed in a vessel having a three-dimensional geometric shape that includes a tissue-attaching surface. The extracellular matrix components are allowed to heal or gel within the vessel, forming the vessel into a culture chamber. The cells were placed in a spheroid under conditions that allowed the cells to form organized tissues attached to the attachment surface. The vessel is surrounded by medium at .

[0114] This method is compatible with the in vitro production of a wide variety of tissues, but the At least a subset of these proteins are involved in mechanistic processes during tissue development, remodeling, or normal physiological function. The present invention is particularly suitable for tissues exposed to and affected by mechanical forces. Examples of such tissues include muscle tissue. Includes flesh, bone, skin, nerves, tendons, cartilage, connective tissue, endothelial tissue, epithelial tissue, and lung. More specific examples include skeletal and cardiac muscles (i.e., striated muscles) and smooth muscles, which are layered. or lamellar bone, as well as hyaline cartilage. When tissues contain multiple cell types, Cells of different types may be from the same or different organisms, from the same or different donors, and from the same or different The cells can be obtained from different tissues. Furthermore, the cells can be primary or immortalized. Additionally, all or a portion of the cells of the tissue may be treated with a bioactive compound (such as those described herein). A foreign DNA sequence (e.g., a foreign D encoding a receptor) that responds to NA sequence).

[0115] The composition of the extracellular matrix component solution will vary depending on the tissue being produced. Examples of suitable extracellular matrix components include, but are not limited to, collagen, laminin, and the like. fibronectin, vitronectin, elastin, glycosaminoglycan, proteo glycans, and combinations of some or all of these moieties (e.g., Ma trigel™, Collaborative Research, Catalog No. 40234). In tissues that contain cell types that are responsive to mechanical forces, the extracellular matrix The solutions of components are preferably arranged in a manner such that the cells are exposed to forces associated with the internal tension of the gel. The tissue gels or coalesces into a gel.

[0116] In vitro tissues with in vivo-like macroscopic and cellular morphological characteristics were generated. The device for producing o is a three-dimensional structure that closely resembles the macroscopic morphological characteristics of tissue in vivo. The device includes a container having a 3-dimensional geometry. Such containers must be compatible with cell and tissue culture (e.g., sterilizable). and compatibility with specific solutions of extracellular matrix components), Various morphological features that can be formed into three-dimensional geometries that closely resemble the macroscopic morphological features in vivo The tissue attachment surface (e.g., stainless steel mesh or V Elcro® and other similar products are used to generate tissue in vitro and to promote cell proliferation. The adhesive is coupled to the container so that it can be attached to and aligned between the attachment surfaces. The tissue attachment surface is compatible with cell and tissue culture (e.g., It must be sterilizable or have a suitable surface charge, texture, or coating for cell attachment. The device may be constructed from a variety of materials (including metal linings).

[0117] The tissue attachment surface may be coupled to the interior or exterior surface of the container in a variety of ways. Alternatively, the tissue-attaching surfaces may be arranged adjacent to the container and adapted to form a tissue shape. The cell culture chamber may be coupled to the cell culture chamber so as to be accessible to the cells during growth. In addition to serving as adhesion points, the attachment surface may also serve as a surface for the adhesion of certain tissue types (e.g. In some cases (e.g., muscles), this allows the tissue to generate tension between opposing attachment surfaces.

[0118] 5.3.2. Organoids with Disease Phenotypes The methods of the disclosure include identifying molecules that revert a disease phenotype in an organoid. In some embodiments, the disease phenotype is associated with a CNS disorder. For example, the disease phenotype It is used to treat neurodegenerative disorders (e.g., Huntington's disease, Alzheimer's disease, Parkinson's disease, syndrome, or ALS), autism spectrum disorder, or psychiatric illness (e.g., schizophrenia, In another embodiment, the disease expression may be associated with In yet another embodiment, the disease phenotype is associated with an infectious disease. In yet other embodiments, the disease phenotype is associated with CF.

[0119] Disease phenotypes in organoids can be used to link, for example, normal stem cell lines to disease phenotypes of interest. By genetically modifying the plant to contain a genetic defect that can be detected by As9 or TALENs). In this study, organoids with disease phenotypes were compared to normal cell lines to show that they retain the disease phenotype. In a specific embodiment, the oncogene comprises one or more mutations that affect the oncogene with a disease phenotype. Luganoids are a type of huntingtin protein that contains expanded polyglutamine ("polyQ") repeats. See, for example, WO 2017 / 147536. PolyQ repeats can be, for example, 42 to 150 glutamine residues (e.g., 4 2, 45, 58, 50, 56, 58, 67, 72, 74, or 150 glutamine residues It may be a group.

[0120] Organoids with disease phenotypes can be generated by using chemical inducers, viruses, from iPSCs derived from diseased subjects, either by infection or by expression of a transgene; See, e.g., Clevers, 2018, Cell 165:1586-1597; Ho et al., 2018 , Int J Mol Sci. 19(4):936;Dutta et al., 2017, Trends Mol Med. 23(5):393-410 Please refer to.

[0121] Cells that have the same or a closely similar genotype are considered to be "isogenic." For example, normal stem cells can be modified to carry a gene for a disease form. and the resulting modified cell line can be considered isogenic to the normal cell line. As another example, a stem cell line carrying a mutated gene associated with a disease phenotype can be to provide a stem cell line having a non-disease phenotype that is isogenic to the parent stem cell line. Other mutations can include one, two, three, or more markers. and / or one or more mutations unintentionally introduced when modifying the parent cell line. Differentiation (e.g., off-target mutations introduced when using CRISPR-Cas9-mediated gene editing) The resulting cells still contain the target mutation. The cells modified from the vectors described herein will be considered isogenic to the cells modified from the vectors described herein. In some of the methods described herein, isogenic control cells or isogenic wild-type cells are used. Isogenic cell line pairs (and organoids generated from such cell line pairs) ) indicates, for example, one or a small number (e.g., 2, 3, 4, 5, or 10) of mismatches (e.g. share the same genetic background except for mismatches introduced by genetically modifying the cells This allows for the study of specific genetic discrepancies compared to wild-type cells, and allows for the study of a large number of genetic Various populations that may differ in their genetic characteristics (particularly, but not exclusively, in unknown genetic characteristics) It reduces the complexity introduced by comparing patient cells.

[0122] Table 2 lists various fluorochemicals that have been produced in the art and can be used in the methods of the present disclosure. Several types of organoids have been reported (see, for example, the review articles Dutta et al., 2017, Trends M ol Med. 23(5):393-410 and the references cited therein).

[0123] [Table 2]

[0124] 5.3.3. METHODS FOR IDENTIFYING MOLECULES WITH THERAPEUTIC CAPABILITY The disclosed neural network identifies molecules that have biological activity against a disease. It can be used to

[0125] In one embodiment, the treatment performance is determined by the final score of the network after the softmax layer. The therapeutic potential ranges from 0 (disease) to 1 (wild type), with 1 being the highest score. and are the most efficient at reverting the disease to wild-type phenotype.

[0126] For example, a molecule that is biologically active against a disease can be detected by measuring the activity of the organoids in the presence of the test molecule. formation conditions (e.g. Dutta et al., 2017, Trends Mol Med. 23(5):393-410; Clevers, 2018, Cell 165:1586-1597; or as described in Ho et al., 2018, Int J Mol Sci. 19(4):936 Culturing a population of mammalian cells in a culture medium (such as a culture medium for a mammalian cell culture medium) to obtain an organoid, The cell population exhibits a disease phenotype when cultured under organoid-forming conditions in the absence of the test molecule. and imaging the organoids after exposure to the test molecule. One or more images of the organoids were analyzed to determine disease or non-disease status ranging from 0% to 100%. The study used a neural network that was trained to assign a probability score to each patient. Analyze organoids using a probability score ranging from 0% to 100%. and the test molecule can be identified by determining whether the organoid has a probability score of being in the non-disease category. If the patient's disease is greater than the cutoff probability score or less than the disease cutoff probability score, Alternatively, the test molecule may be administered after organoid formation to a subject that is biologically active against the disease. The application and imaging may be performed after a period of exposure to the test molecule. (whether present during or after organoid formation) , especially if the test molecule is applied for the entire duration of the experiment, which can range from hours to days, weeks, or even months. The concentration of the test molecule in the culture medium may range from 10 nM to 100 The range may be in the μM range (e.g., 10 nM to 1 μM, 10 nM to 50 μM, 10 nM ~20μM, 1μM~100μM, 1μM~50μM, 1μM~20μM, 1μM~10 μM, 10μM-50μM, 20μM-50μM, or 50μM-100μM).

[0127] Cutoff probability scores for non-disease are, for example, 60%-95%, 70%-95%, 80% It can be in the range of 95%, 90%-95%, or 95%-99% (e.g., The cutoffs are 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 96% %, 97%, 98%, or 99%. The cutoff probability score for disease is: For example, 5% to 40%, 5% to 30%, 5% to 20%, 5% to 10%, or 1% to 5%. (e.g., the cutoffs may be 40%, 35%, 30%, 25%, (It may be 20%, 15%, 10%, 5%, 4%, 3%, 2%, or 1%). Thus, for example, diseased organoids exposed to a test molecule have a non-disease probability score of 96%. and the cutoff probability score for non-disease is 95%, the test molecule is As another example, the disease caused by exposure to the test molecule can be considered to be biologically active. The organoid disease probability score was 2% and the disease cutoff probability score was 5%. If so, the test molecule can be considered biologically active against the disease. As another example, if diseased organoids exposed to a test molecule had a non-disease probability score of 60%, If the cutoff probability score for non-disease is 90%, the test molecule is considered to be viable for the disease. It can be considered to be biologically inactive.

[0128] Prior to imaging the organoids with the methods of the present disclosure, the organoids are stained with organelle markers , cell differentiation markers, cell compartment markers, or combinations thereof. The cells may be stained for a marker. The marker or markers are preferably The gross morphological characteristics of organoids and / or specific cell populations contained within the organoids Exemplary staining reagents include one or more markers that highlight the clusters. Color (e.g., DAPI and Hoechst nuclear stains), BrdU, calcein green, DiI , membrane dyes such as DiO, DiD, and DiR, and fluorescently labeled antibodies against GM130. Golgi staining reagent, ER-Tracker™ Green (Cell Signali ER staining reagents such as ng Technology, phalloidin, and N-cadherin Fluorescently labeled antibodies against phosphoinositide, E-cadherin, ZO-1, and collagen .

[0129] Also use reagents specific to certain types of organoids or certain types of differentiated cells For example, a reagent for staining neurofilament or nestin (e.g. For example, fluorescently labeled antibodies can be used to stain brain organoids. Color reagents can be used to stain the neuronal population. By using this method, neural crest cells can be stained.

[0130] The test may be performed using a collection of test molecules, e.g., all or part of a small compound library ( For example, a collection of test molecules in the range of 100-100,000 or more. Compounds can be tested individually or in pools (e.g., 5- 100 pieces, 10-50 pieces, 10-20 pieces, 50-100 pieces, 5-20 pieces, or 5-10 pieces A pool of molecules can be screened. If hits are identified during matching, the molecules in the pool are individually tested to identify the specific hits. Thus, the neural network and method of the present disclosure can identify molecules. are useful for high throughput screening of test compounds.

[0131] Validation A single concentration of the test molecule can be used for initial screening of molecules (e.g., 10 After identification as a hit, the test molecule is retested at one or more additional concentrations. When a high initial concentration is used, its activity against the disease phenotype can be verified. Very high IC 50 may indicate high nonspecificity of the test molecule, To validate a hit, it may be useful to retest the hit at one or more lower concentrations. obtain.

[0132] Validation involves retesting molecules identified in the initial screen at one or more different concentrations. Alternatively or additionally, the method may include preparing a dose-response curve. In addition, validation may involve testing the molecule in an animal model of the disease.

[0133] Dose-response curves Validation is performed by IC 50 and / or L.C. 50 It may include determining IC of the test molecule 50 The value can be calculated using the network of the present disclosure by converting the test concentration into, for example, W Along the base of the T-disease-toxicity triangle (e.g., the triangle shown in Figure 14), The results can be generated by comparing the therapeutic ability measured by the network. Child LC 50 The value is calculated using the network of the present disclosure, and the test concentration is expressed as, for example, WT-disease- Along the vertical axis of a toxicity triangle (e.g., the triangle shown in Figure 14), This can be generated by comparison with more measured toxicity.

[0134] I C 50 , three or more different concentrations of the test molecule (e.g., 3 to 10 , 5–10, or 3–5) and a nonlinear least-squares fitting algorithm The probability scores of disease or non-disease at various concentrations are measured using Similarly, LC 50 Three or more different concentrations of test substances are Test multiple children (e.g., 3–10, 5–10, or 3–5) and use a nonlinear least-squares fit. A matching algorithm is used to separate the probability scores of disease or non-disease at various concentrations. It can be measured by analyzing the

[0135] Usable Capacity In general, test molecules that are closest to the WT phenotype corner of the WT-disease-virulence triangle are selected for further analysis. IC<5 μM will be selected for analysis and / or development.50 Compound having The product is favorable for further development. 50 The value is IC 50 Ideally, it should be several times larger than and in some embodiments, LC 50 IC 50 If it is greater than LC 50 IC5 At least 10 times greater than 0, or LC 50 IC 50 At least 10 If so, the test molecule is selected for further analysis or development. has a lower IC than the original test molecule 50 (e.g., less than 100 nM) and / or High LC 50 The present invention also includes preparing a derivative to identify a derivative having the formula good. EXAMPLES

[0136] 6. Working Example [Example 1] 6.1. Example 1: Generation and Characterization of HD Model Organoids HD Model Organo, a hESC-based self-organizing construct that mimics the neurulation stage of human embryos We characterized the phenotypic signature of HD in noids.

[0137] Materials and Methods Generation of human embryonic stem cell lines for use in HD model organoids.

[0138] As described in the Background section, Huntington's disease (HD) is a genetic disorder characterized by the This is caused by a mutation that leads to an expansion of the N-terminal polyQ repeat of the hepatitis C transcription protein (HTT). It is a dominant autosomal neurodegenerative disease caused by rheumatoid arthritis. Animal models accurately recapitulate the pathophysiology of human HD, likely due to species-specific differences This has hindered progress in identifying effective potential therapies for this disease. In this example, we demonstrate the function and dysfunction of HTT in healthy and >40Q cells. The human platform as a drug screening and research tool to study the whole To provide the first HD isogenic It has been described that a genetically identical human embryonic stem cell line (and an isogenic wild-type control) has been generated. is.

[0139] In this example, we applied a reverse editing strategy using CRISPR-Cas9 to transform normal hES We introduced a long stretch of polyQ tract into C. elegans, which is therefore genetically identical to its wild-type counterpart. (except for the polyQ expansion of the HTT gene), and can therefore be called isogenic. The authors report that they have generated a D strain that is stable and shows no abnormalities in HD. Human pluripotent cells are used, which can generate any cell type, including those susceptible to The advantage of this approach is that it allows comparative global transcriptome and Iasmetabolome analysis of a single genomic locus containing an extended polyQ tract Our results revealed previously undetected differences caused only by the incorporation of

[0140] 1A. Generation of isogenic human embryonic stem cell lines Briefly, using CRISPR-Cas9 technology, we cloned the previously derived RUES2 hESC lines were genetically engineered. The parent cell line, RUES2, is registered with the NIH. (NIHhESC-09-0013), The Rockefeller University and WiCell (Rockefeller It is available from the National Institute of Integrative Medicine (NII) under the designation WB33127. It encodes 22Q in one allele. The wild-type HTT locus (chromosome 4pl6) encodes 24Q in the second allele. 3) is a female (XX) strain with the 22Q allele by adding 128Q. The 150Q strain (RUES2-Q150) was thus generated. To accelerate the synthesis, we selected a gene that yielded a relatively large number of polyQs, but You may use less as long as it is 40Q or more. stomach.

[0141] The lineage tracing markers mCherry and blasticidin cassettes were also expressed in RUES2. was integrated into hESCs.

[0142] Other embryonic cell lines with wild-type HTT genes may alternatively be used and may be used as described herein. hESCs can also be genetically modified to generate engineered hESC cell lines and isogenic controls. In addition, three unaltered homozygotes with normal alleles (20 CAG codons) were A transgenic cell line was generated as a control.

[0143] The length of the insert (to create the 150Q HTT gene) was chosen to account for possible differences. To maximize the chances of detection, we modeled early-onset juvenile HD, which is the worst case scenario of the disease. (We later found that increasing the number of polyQs in this way (Although this played a role, it was not necessary.) For lineage tracing and selection purposes, The mCherry-blasticidin cassette flanked by self-cleaving peptide 2A sites The insert was placed upstream of the initiation codon (Figure 1A). The success of the gene locus modification was confirmed by PCR. The RUES2-Q150 cells were karyotypically normal and showed no significant differences in HTT cell lysate. Both HTT alleles were expressed as evidenced by estane blotting. It is predicted that expression of one HTT allele would be sufficient to confer the HD phenotype. Expression of mCherry was confirmed using fluorescent imaging.

[0144] To demonstrate that the genome editing strategy did not affect the fundamental properties of hESCs, Under pluripotent conditions (which express POU5F1(a / k / a OCT4), SOX2, NANOG One or more pluripotency markers, such as LIN28A, LIN28B, and DNMT3 This can be confirmed by testing for expression of the β-lactamase inhibitors (β-lactamase inhibitors), all of which were tested here. RUES2-Q150 cells maintained normal hESC morphological characteristics when grown at 4°C. For example, we confirmed that RUES2 expressed pluripotency markers at levels similar to wild-type RUES2 cells. In addition, the rate of cell proliferation and apoptosis was significantly increased by EdU incorporation (Fig. 1B). The degree of activation of caspase-3 was examined by immunostaining for activated caspase-3. No differences were found between the two strains. I did.

[0145] Finally, to determine the differentiation potential of these cells into HD-relevant lineages, SB431 Dual SMAD inhibition including 542 and LDN193189 was used to regulate neuronal fates Instead, we used a combination of SB (SB431542) and Noggin. We were also able to induce a neuronal fate in RUES2-Q150 cells using They express the myeloma-specific markers PAX6 and N-cadherin and have typical morphological characteristics. The rosettes were formed with SOX17, EOMES, T(BRA), CDX2, and FOX. Other differentiation markers, such as A2 and FOXG1, were also expressed to indicate differentiation. Any one of these markers would have been sufficient. Once this information is available, it will be possible to visualize the different germ layers and assess their differentiation. These results demonstrate that genome editing does not alter the fundamental properties of isogenic hESCs. It was demonstrated that this was the case.

[0146] IB. The range of different genes found in HD patients without the use of detectable and selectable markers Generation of ESCs spanning polyCAG lengths Contains the HTT gene with various polyQ lengths and a detectable marker (expression marker) mChe In addition to the HD ESC lines containing the rry and blasticidin cassettes, Applicants have Contains various polyQ stretches but lacks detectable and selectable marker cassettes We also generated cells that express polyCAG lengths in the typical range found in HD patients. A set of >40Q ES spanning 42, 48, 56, 67, and 150 CAGs C line was generated (Figure 13A). Again, the RUES2 hESC line was used as the parent line. With the exception of the penultimate codon, which was CAA, as it is in nature, All of the codons were CAG. However, as was done in 150Q modified cells, A mixture of AG and CAA codons could have been used. Cells were generated using CRISPR technology as described in Materials and Methods below. Cells in which the reQ tract is identical to that found in HD patients (essentially a CAG repeat To generate the IL-16-binding protein (IL-16), the inventors used the following cells: Coriell Insitu ND38548 from itute, GENEA020 from GENEA Biocells, and QS-001 and QS-004 from Tabrizi laboratory (UK). Using fibroblast cells to amplify mutant Huntington gene loci directly from patient samples PCR was used to construct the donor plasmid. These cells were used for this purpose. Fibroblast cells are available from a number of public and commercial sources. It would be possible to use them instead. The same goes for the rest of the cells. They are available from other sources. PolyQ tracts can also be synthesized in the laboratory.

[0147] In addition, Applicants have used a selectable marker that contained the ePigybac transposable element. This allowed the marker to be removed after identifying the selected cells. Excision-only transposases are The final product was purchased from Agen (Lexington, KY 40508). In essence, these cells do not contain a selectable marker and are homogenous only in the length of the polyQ tract. Importantly, all of these engineered cell lines are genetically distinct from wild-type control cells. phenotype was observed.

[0148] These cells express markers and selectivity profiles when compared to isogenic normal control cells. Considering that there are fewer differences than in cells containing the cassette of the possible markers, A gene that essentially lacks the insertion sequence except for the introduction of the Q tract, which is a gene that expresses Huntington's disease. Another set of embodiments for modeling is provided.

[0149] 6.1.2. Discussion 6.1.2.1. HD signatures in HD model organoids Neurulation is an important developmental transition that the central nervous system undergoes. Micropatterned culture with tocol reconstitutes the ectodermal compartment during neurulation These reconstituted embryonic parts can lead to the formation of self-organizing cell aggregates that can Transcranial fate (SOX10+, red population) and placode fate (SIX1+, yellow cells) Both were organized into neural rosettes, surrounded by a layer of epidermal cells (TFAP2+ only, blue). The developing central nervous system (PAX6+ cells, green, Fig. 3) is shown in the center. Comparison of the neurulation organoid model with its in vivo counterpart (Figures 3D-3E) It shows a high level of similarity and therefore the HD organoid model assay is compatible with human genetic disorders. It is an ideal endpoint for studying harm.

[0150] The emergence of numerous homogenous and stereotyped self-organizing organoids is linked to neuroepithelial morphogenesis. Because of their popularity, we have chosen to use them to model Huntington's disease (HD). We therefore investigated whether a series of HD isogenic Using hESC lines, HD model organoids have one major feature: As shown in Figures 4A-4B, the CAG expansion was associated with the PAX6 area. Interestingly, the HTT- / - strain showed the most dramatic increase in the X6+ area. The current model is shown.

[0151] 6.1.2.2. Observation of HD model organoid phenotype in 96-well plates HD model organoid formation was adapted to micropatterned 96-well plates HD-specific organoid phenotypes were reproducibly observed in 96-well plates and On average, 28 HD model organoids were observed per group (Figure 5). The HD model organoids formed with hESCs (Figure 5B) were significantly more abundant than those with WT hESCs. The PAX6+ region was found to be large (Figure 5A). The individual PAX6+ regions of each colony were sectioned using the WT background and H. D An example of quantification of background in multiple wells is shown. In agreement with previous results, HD-specific organoids The PAX6+ area was significantly greater in the HD background than in the WT. Therefore, we performed the HD model organoid formation protocol for 9 We successfully adapted it to a 6-well plate and used it in high-throughput screening operations. This provided the ability to screen compounds for phenotypic reversion.

[0152] 6.1.2.3. Screening robustness: Calculation of Z' coefficient When performing high-throughput screening (HTS), there is a high rate of false positives and false negatives. To avoid inaccuracies, a robust assay is required. This involves the use of positive and negative controls. This suggests strong separation of the controls. A commonly used tool is the Z' coefficient. In the ideal case, Z' converges towards a value of 1. A good screen assay is To be robust enough to constitute a good screening platform, 0.5, but in the case of complex phenotype screening, Z' It is widely recognized that values ​​>0 are acceptable. Negative values ​​indicate that the assay is This means that it cannot be the basis for a military offensive.

[0153] To measure the Z' factor, two different genetic backgrounds in a control 96-well plate The average PAX6+ area of ​​the WT / 20CAG control plate was quantified. Overall, the WT / 20CAG control plate showed It was observed that the D / 56CAG plate had a larger PAX6-positive area than the D / 56CAG plate (Figure 6 A). The Z' factor associated with the assay for each well was then measured. Row r and column For all wells placed in c, the following were measured:

[0154]

number

[0155] When this calculation was performed (Figure 6B), 1) the separation between the positive and negative controls was not good enough; 2) there is a significant edge effect and the plate It was found that the best Z' scores were obtained at the center of the However, positive values ​​were not achieved, with the average Z' across the plate being -17. Therefore, it was decided to use a different analytical scheme to obtain positive Z' coefficients.

[0156] [Example 2] 6.2. Example 2: Training a neural network to recognize HD organoids Training It should be noted that the quantification shown in Figure 5C shows that the data are highly variable and that the positive and negative controls are not significantly different. Although the separation from the sex controls is small, the distinction between the WT and HD phenotypes is clear to the human eye. This seemed self-evident (Figures 5A-5B). This is because the human brain is AX6 compares unique features such as area, but also a number of others: shape of the region, phase By integrating the intensity of images and a host of other features that we may not even be aware of, To reconstruct this process computationally, machine learning techniques are being used. was used.

[0157] Deep neural networks have been gaining popularity for unbiased image classification over the past few years. It has been used frequently, for example, to Thanks to the neural network built, Google Image Search now uses images that were previously It is now possible to classify photos of cats that the person has never seen before as "cats." can be used for phenotypic analysis, improving the robustness of screening techniques. We hypothesized that it would be possible to achieve this using a computer equipped with a GPU and publicly available We used a capable deep neural network architecture as a starting point.

[0158] The strategy used can be described in two steps. The first training phase (Figure 1) In 7A), the network was trained on images belonging to the WT background and the HD background obtained under the control condition. At the end of training, the network will It learns to recognize the present type and presents previously unseen images from either WT or HD backgrounds. In the second phase, the Using the trained network, we test the network on images from the screening. The results of our screening strategy will be analyzed by searching for the The network identified diseased HD organoids treated with SO as belonging to the HD class. Hit compounds revert the phenotype to WT, and the images are then would be classified as WT (Fig. 7B ).

[0159] In a first proof-of-concept study, we demonstrated that the most dramatic HD phenotypes were obtained in an Htt- / - background. As shown by earlier studies by the authors (Fig. 4), we used WT and Htt- We trained the network to efficiently recognize / - HD model organoids. We attempted to replicate this in vivo using 400 WT HD model organoids (Figure 8A). Images and trained with 400 HD HTt- / - model organoid images We performed a study to identify previously unseen WT or Htt- / - HD model organoids. When we presented the ids to the network, it was able to assign them to the correct class with 97% accuracy. The results are shown in Figure 8B. judges whether the new search image belongs to the WT class image or the HD class image. In a further step, this tool can be used to analyze phenotypic reversion. To demonstrate that this is possible, the Htt gene was transfected into separate cell lines in an Htt- / - background. This allowed the rescue of protein expression. on the basis that it rescues the HD model organoid phenotype. The images of HD model organoids in an Htt- / - background with Htt overexpression were collected. When the WT-class was searched for in the network (Fig. 8B), they were classified as belonging to the WT class with a higher probability. Thus, phenotypic rescue was quantitatively demonstrated.

[0160] [Example 3] 6.3. Example 3: High Throughput of Compound Libraries for Reversion of HD Phenotype screening The network will be deployed as a high-throughput platform for discovering compounds that reverse the HD phenotype. The HD model organoids were cultured in isogenic medium as described above. Transgenic WT and HD hESCs were cultured on a micropatterned 96-well plate Multiple positive control plates with WT cells and untreated 56CAG HD cells were prepared in duplicate. Multiple negative control plates were prepared with strains. Compounds were then added to the 56CAG HD strain The mixture was added to wells of a 96-well plate containing

[0161] Whether the robustness of the assay, quantified by the Z' factor, can be increased First, HD model organoids in the control wells (WT and HD) were randomly The training set and validation set were split into 70% and 30%, respectively. Thus, there are 67 training wells each and 29 validation wells each. Each well contained approximately 25 organoids. Then, 1750 WT We trained the neural network using 1691 HD images and 1691 HD images. The trained network was trained as described in the Methods section. The accuracy of classifying individual images is shown in Figure 9.

[0162] We then used our machine learning approach to quantify the Z' coefficient. In this study, we tested whether it was possible to enhance the robustness of our assay using 100% ethanolamine. Preliminary experiments with multi-well plates were reexamined using neural network techniques, and The Z' coefficients of the various wells of the well plate were calculated. A significant increase was observed (Figure 10A), compared with a significant negative increase using the previous partitioning-based method. The Z' factor per well was positive in this case. (Figure 10B), most of the Z' coefficients are higher than 0.5 and all are positive. Only one well showed a slightly negative Z', but this well was not in line with the actual screen. The wells at the corners of the plate served as control wells during the scan, as well as the wells at the corners of the plate This is not so important because it corresponds to the first and last rows of this particular well. Others have positive Z' values ​​regardless of whether the wells are at the edge of the plate or in the center. Thus, the innovative machine learning toolbox described herein can be used to identify hit molecules. This allows the analysis of HD-specific phenotype screens under ideal conditions for identifying Cut.

[0163] A screening run of 800 molecules was then performed. To train the neural network, one pair of WT (RUES2) controls was used. One control plate and one HD / 56CAG control plate were included. All other plates At this rate, HD / 56CAG organoids (HD model organoids) were formed. Compounds were applied to these wells. Wells with less than 10 intact organoids were discarded. In these cases, the compounds were assumed to have a toxic effect and were excluded from the analysis. The network reverts phenotypes based on the tools described in the previous section. The ability of the compounds was measured for each well. The network was then applied to each HD model organoid. Score p between 0 (disease) and 1 (WT). WT This means that the compound is in the disease HD table. This score is then averaged across each well. The histogram of the scores obtained per well is shown in FIG.

[0164] As expected, most compounds did not alter the HD / 56CAG phenotype and were consistent with wild-type class II. Their associated probability p wt was close to zero. This is because HD model organoids This means that the network has recognized that the device is not affected and still belongs to the HD class. However, a small number of compounds showed phenotypic reversion, with a probability of belonging to the wild-type class p wt Even more interestingly, the p of the 13 compounds was wt > It was 0.95.

[0165] [Example 4] 6.4. Example 4: Toxicity Analysis Since the number of hits was relatively large in the screening of Example 3, the present inventors Partly due to the fact that the work is tied to only two classes, WT and HD. We investigated whether false positive compounds could appear in the hit list. We found that it had a detrimental effect on HD model organoids, resembling the WT phenotype in appearance. A third class of toxic compounds that result in neither the HD phenotype nor the HD phenotype was extracted from the network. Instead, we decided to use these compounds to treat HD model organoids in a completely different way. Therefore, we performed a binary analysis of WT and HD. We used our previous network, which shows the choice as the output, but We extracted the fully connected layer upstream of the NN. This layer has 512 neurons. The activations can be interpreted as 512-dimensional latent vectors, which allows us to use the same latent space It is now possible to express each phenotype and quantify how different two phenotypes are from each other. The network with this latent vector is shown in Figure 12A.

[0166] Using this network architecture, we model the effect of each compound in terms of two related components: The therapeutic potential was determined by the level of toxicity of the compound and its therapeutic potential. Toxicity was measured by the output of two WT and diseased neurons. This was measured by assessing the differences between Ronnie and the two control phenotypes WT and disease. The extent of this difference was due to the difference in the potential vectors associated with the compound-treated colonies and the two controls. This was calculated by taking the minimum difference between the associated latent vectors. It was possible to perform this on the entire 12-dimensional latent space. However, Use t-SNE or Principal Component Analysis (PCA) to encode the vectors In order to extract only the most relevant information, we first reduce the dimensionality of the vectors. It was advantageous to do so. The difference in the vectors in the reduced space was then calculated. The t-SNE clustering of vectors corresponding to wells and treatments is shown in Figure 12B. It is being done.

[0167] For optimal display of the results, the effect of each compound was plotted as a triangular graph as shown in Figure 13. Plotted on a phenotypic map.

[0168] When the results of screening using this methodology are plotted (Figure 14), high treatment potency (clustered in the WT corner of the plot, showing high phenotypic reversion) as well as low virulence Five compounds were found that exhibited toxicity (still far from the toxicity angle of the triangle).

[0169] [Example 5] 6.5 Example 5: Analysis of toxicity with autoencoders A method based on convolutional autoencoders was used to assess toxicity. A neural network without it encodes data and compresses it into a low-dimensional latent representation. (Figure 15).

[0170] The unsupervised nature of this machine learning method means that the autoencoder learns The data were presented without any additional information (e.g., derived from wild-type or diseased cell lines) that could be used to This has the advantage of learning from the actual data and therefore not biasing estimates of compound toxicity. In addition, vector-based data expression involves the use of vector space to detect differences between wild-type and disease phenotypes. This has the advantage that the differences are not relevant to determining toxicity. Toxicity was determined in the following way. First, the potential difference between wild type and disease was examined. Then, the distance of the wild-type phenotype and the disease phenotype from the mean vector is The standard deviation of the wild-type phenotype and the disease phenotype are calculated and compared. This distance is defined as toxicity. (Figure 16).

[0171] The accuracy of this method can be verified by applying various concentrations of drugs with known toxic effects and comparing the results with Conventional colorimetric "MTT" assay of cell viability performed using 56CAG RUES2 cells. The survival rate was tested by comparing the survival rate with the mean survival rate (here, survival rate is defined as the reciprocal of toxicity). The comparison showed consistent results, validating the autoencoder method for determining toxicity. It can be confirmed that this is the case (Figure 17).

[0172] 7. Specific embodiments The present disclosure is illustrated by the following specific embodiments. 1. A method for classifying a test organoid, comprising: imaging the test organoids to prepare an organoid image; Trained to assign classifications to test organoids Analyzing the organoid images using a neural network, , comprising a disease phenotype or a non-disease phenotype, The method includes: 2. Cells of the organoid having a disease phenotype are characterized by one or more phenotypes that confer the disease phenotype. 2. The method of embodiment 1, comprising a mutation of: 3. The cells of the organoids with non-disease classification and disease classification are isogenic; 3. The method of embodiment 1 or 2. 4. Organoids are multicellular aggregates, stem cells, pluripotent cells, induced pluripotent cells, human embryonic Any of embodiments 1 to 3, comprising stem cells, adult stem cells, totipotent stem cells, or non-embryonic cells. The method described in one. 5. The method of any one of embodiments 1-4, wherein the disease comprises Huntington's disease. 6. Imaging includes staining the organoids for one or more markers 6. The method according to any one of embodiments 1 to 5. 7. One or more markers are organelle markers, cell differentiation markers, cell compartment markers, 7. The method of embodiment 6, comprising a marker, a marker for detecting a marker, or a combination thereof. 8. The method of embodiment 6 or 7, wherein the one or more markers include a nuclear marker. Law. 9. Any one of embodiments 6 to 8, wherein the one or more markers include a membrane marker. The method according to 10. Any of embodiments 6 to 9, wherein the one or more markers include a cell differentiation marker. The method according to any one of the preceding claims. 11. At least two images of each organoid are analyzed, each for different markers and and corresponding to different color channels. 12. At least three images of each organoid are analyzed, with each image showing a different marker and corresponding to different color channels. 13. The method according to any one of embodiments 1 to 12, wherein the organoid image comprises a two-dimensional image. Method of posting. 14. The method according to any one of embodiments 1 to 13, wherein the organoid image includes a three-dimensional image. Method of posting. 15. The trained neural network identifies organoids belonging to non-disease classifications. The system was trained on multiple images of organoids belonging to disease categories. A neural network is a network that has multiple layers and two nodes. 15. The method of any one of the preceding claims, comprising a final layer comprising a coating. 16. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 10 nodes. 17. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 100 nodes. 18. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 250 nodes. 19. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 400 nodes. 20. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 500 nodes. 21. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 600 nodes. 22. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 800 nodes. 23. The neural network includes a fully connected layer upstream of the final two nodes, 16. The method of embodiment 15, wherein the combined layer comprises at least 1000 nodes. 24. The neural network includes a fully connected layer upstream of the final two nodes, 24. The method of any one of embodiments 15 to 23, wherein the combined layer includes 512 nodes. 25. The fully connected layer upstream of the final two nodes is the penultimate layer. 25. The method according to any one of aspects 1 to 24. 26. Analysis Extracting the fully connected layers upstream of the two final nodes, and Calculating the degree of difference 26. The method of any one of embodiments 15 to 25, further comprising: 27. A method for identifying a test molecule that is biologically active against a disease, comprising: (a) culturing a first population of mammalian cells under organoid-forming conditions to form a first organoid; obtaining organoids, the organoid forming conditions including exposure to a test molecule, of mammalian cell populations cultured under organoid-forming conditions in the absence of biologically active molecules. which, when generated, results in an organoid having a disease phenotype; (b) imaging the first organoid after exposure to the test molecule; (c) Train subjects to assign a disease or non-disease probability score ranging from 0% to 100%. Using a neural network trained on the first organoids, analyzing the plurality of images; (d) Assigning a probability score to the first organoid, ranging from 0% to 100%. Including, The probability score of the first organoid is higher than the cutoff probability score for non-disease or or disease cutoff probability score, the test molecule is biologically active against the disease. The way it is. 28. The first population of mammalian cells harbors one or more mutations that confer a disease phenotype. 28. The method of embodiment 27, comprising: 29. The method of embodiment 27 or 28, further comprising verifying the activity of the test molecule against the disease. Or the method according to embodiment 28. 30. The method of embodiment 29, wherein the validation comprises determining the IC50 of the test molecule. . 31. The IC50 of a test molecule is determined using at least three different concentrations of the test molecule. Repeating the method according to embodiment 27, and performing a nonlinear least squares fitting algorithm. This involves using a genomic algorithm to analyze the probability scores of disease or non-disease at different concentrations. 31. The method of embodiment 30. 32. If the IC50 of the test molecule is less than 5 μM, the test molecule is discarded for further analysis or 32. The method of embodiment 31, further comprising the step of selecting for development. 33. Prepare derivatives of test molecules to find derivatives with lower IC50 33. The method of embodiment 32, comprising the steps of: 34. The method of embodiment 33, wherein the lower IC50 is lower than 100 nM. 35. The neural network generates a probability score of toxicity or non-toxicity ranging from 0% to 100%. 35. Any of embodiments 27 to 34, further trained to allocate cores. The method described in one. 36. First organoids are given a probability score of toxicity or non-toxicity ranging from 0% to 100%. 36. The method of embodiment 35, further comprising the step of assigning: 37. The method of embodiment 35 or embodiment 36, further comprising determining the LC50 of the test molecule. 37. The method of claim 36. 38. The determination of the LC50 of a test molecule is performed using at least three different concentrations of the test molecule. Repeating the method according to embodiment 36, and performing a nonlinear least squares fitting algorithm. This involves using a statistical algorithm to analyze the probability scores of toxicity or non-toxicity at different concentrations. 38. The method of embodiment 37. 39. Determining IC50 and LC50 of test molecules and further analysis The method further comprises the step of selecting a test molecule having an LC50 greater than the IC50 for the purpose of determining whether the test molecule has an LC50 greater than the IC50. 39. The method according to any one of embodiments 27 to 38. 40. The method of embodiment 39, wherein the LC50 is at least 10 times greater than the IC50. . 41. The method of embodiment 39, wherein the LC50 is at least 100 times greater than the IC50. Law. 42. The method further comprises the step of assaying the activity of the test molecule in an animal model of the disease. 42. The method according to any one of embodiments 27 to 41. 43. The method of embodiment 2 further comprising the step of verifying the accuracy of the neural network. 7. The method according to any one of claims 7 to 42. 44. Verification of the accuracy of neural networks (e) Propagating a second mammalian cell population of the same cell type as the first cell population into an organoid-shaped The second organoids are cultured under synthetic conditions to yield organoids that do not have a disease phenotype. obtaining a step; (f) imaging the second organoid; and (g) Assigning a probability score to the second organoid ranging from 0% to 100%. And, The neural network then determines the probability score of the second organoid, which is the cutoff for disease-free status. If it is higher than the probability score or lower than the disease cutoff probability score, it is accurate. considered as a step; or (h) Propagating a second mammalian cell population of the same cell type as the first cell population into an organoid-shaped Cultivate under synthetic conditions to produce a third organoid that will result in an organoid with a disease phenotype. Steps to obtain; (i) imaging the third organoid; and (j) Assigning a probability score to the third organoid, ranging from 0% to 100%. And, The neural network then determines the probability score of the third organoid, which is the cutoff for disease-free status. If it is below the probability score or above the disease cutoff probability score, it is accurate. The step is considered 44. The method of embodiment 43, comprising: 45. A first population of mammalian cells harboring one or more mutations that confer a disease phenotype. the second mammalian cell population has one or more mutations other than those of the first mammalian cell. 45. The method of embodiment 44, which is isogenic to the population. 46. ​​The method of embodiment 44 or embodiment 45, wherein step (e) is performed simultaneously with step (a). 46. ​​The method according to embodiment 45. 47. A first population of mammalian cells is cultured in the presence of a test molecule for at least 3 days; The method according to any one of embodiments 27 to 46. 48. The concentration of the test molecule in the culture is in the range of 10 nM to 100 μM, according to embodiment 27. The method according to any one of claims 1 to 47. 49. The method according to embodiment 48, wherein the concentration of the test molecule is in the range of 1 μM to 20 μM. . 50. Organoid formation conditions are achieved by culturing mammalian cell populations on micropatterns. The method of any one of embodiments 27 to 49, comprising: 51. Any of embodiments 27 to 50, wherein the first mammalian cell population is a population of stem cells. or one of the methods described above. 52. The method of embodiment 51, wherein the stem cells are totipotent stem cells. 53. The method of embodiment 51, wherein the stem cells are human embryonic stem cells. 54. The method of embodiment 51, wherein the stem cells are pluripotent stem cells. 55. The organoid formation conditions include one or more differentiation factors in the culture medium. 55. The method according to any one of embodiments 51 to 54. 56. The method of embodiment 55, wherein the differentiation factor is a neuronal differentiation factor. 57. The method of embodiment 56, wherein the neural differentiation factor is BMP4. 58. A first mammalian cell population is a hanchi having an expanded polyglutamine repeat. 58. The method of any one of embodiments 27 to 57, wherein the nucleic acid sequence encodes an inhibitor protein. 59. Extended polyglutamine repeats contain 42-150 glutamine residues , The method according to embodiment 58. 60. An embodiment in which the expanded polyglutamine repeat contains 42 glutamine residues. 59. The method according to claim 59. 61. An embodiment in which the expanded polyglutamine repeat contains 45 glutamine residues. 59. The method according to claim 59. 62. An embodiment in which the expanded polyglutamine repeat contains 48 glutamine residues. 59. The method according to claim 59. 63. An embodiment in which the expanded polyglutamine repeat contains 50 glutamine residues. 59. The method according to claim 59. 64. An embodiment in which the expanded polyglutamine repeat contains 56 glutamine residues. 59. The method according to claim 59. 65. An embodiment in which the expanded polyglutamine repeat contains 58 glutamine residues. 59. The method according to claim 59. 66. An embodiment in which the expanded polyglutamine repeat contains 67 glutamine residues. 59. The method according to claim 59. 67. An embodiment in which the expanded polyglutamine repeat contains 72 glutamine residues. 59. The method according to claim 59. 68. An embodiment in which the expanded polyglutamine repeat contains 74 glutamine residues. 59. The method according to claim 59. 69. The extended polyglutamine repeat contains 150 glutamine residues. 60. The method of claim 59. 70. The method of embodiment 2, wherein the first mammalian cell population is cultured in a microwell plate. 7. The method according to any one of claims 7 to 69. 71. The embodiment in which the microwell plate is a 24-well microwell plate. 71. The method according to claim 70. 72. First mammalian cell population produces 20-30 organoids per well 72. The method of embodiment 70 or embodiment 71, wherein the cell is cultured under conditions that 73. The method further comprises staining the organoids for one or more markers. 73. The method according to any one of embodiments 27 to 72. 74. One or more markers are organelle markers, cell differentiation markers, cell division markers, 74. The method of embodiment 73, comprising a marker, a marker for detecting a tumor, or a combination thereof. 75. The method of embodiment 74, wherein the one or more markers include a nuclear marker. 76. The method of embodiment 74 or embodiment 75, wherein the one or more markers include a membrane marker. 75. The method according to claim 75. 77. The method of any one of embodiments 74 to 76, wherein the one or more markers include a cell differentiation marker. The method according to any one of the preceding claims. 78. At least two images of each organoid are analyzed, each with different markers and and corresponding to different color channels. 79. At least three images of each organoid were analyzed, each with a different marker and and corresponding to different color channels. 80. The method of any one of embodiments 27 to 79, wherein the image comprises a two-dimensional image. 81. The method of any one of embodiments 27 to 79, wherein the image comprises a three-dimensional image. 82. Any of embodiments 27-79, wherein the probability score is given on a scale of 0 to 1. The method described in one. 83. Any one of embodiments 27 to 81, wherein the cutoff probability score is 95%. The method described. 84. The method of embodiment 27-8, in which a cutoff probability score of at least 85% indicates non-disease. 1. The method according to any one of claims 1 to 1. 85. The method of embodiment 27-8, in which a cutoff probability score of at least 90% indicates non-disease. 1. The method according to any one of claims 1 to 1. 86. The method of embodiment 27-8, in which a cutoff probability score of at least 95% indicates non-disease. 1. The method according to any one of claims 1 to 1. 87. The method of embodiment 27-8, in which a cutoff probability score of at least 99% indicates non-disease. 1. The method according to any one of claims 1 to 1. 88. A collection of test molecules is prepared to identify molecules that are biologically active against a disease. A method for screening a collection of antibodies, the method comprising the steps of: A method comprising the step of carrying out the method according to any one of aspects 27 to 87. 89. The collection represents all or part of a small molecule compound library. 89. The method of embodiment 88. 90. The method of embodiment 88 or 89, wherein the collection comprises 100 to 100,000 test molecules. Or the method according to embodiment 89. 91. Any one of embodiments 88 to 90, in which the test molecules are screened individually. The method according to 92. Any of embodiments 88-90, in which the test molecules are screened in pools. The method described in one. 93. A score higher than the cutoff probability score for non-disease or the cutoff probability score for disease For each member of the pool that results in a probability score lower than The method according to embodiment 92, further comprising carrying out any one of the methods according to any one of the preceding claims. Law. 94. An embodiment in which the neural network is a convolutional neural network. 27 to 93. 95. The method of embodiment 94, wherein the neural network is a residual network. Law. 96. The method of embodiment 94 or embodiment 95, wherein the neural network includes 18 to 152 layers. The method of embodiment 95. 97. The layers are convolutional layers, batch normalization layers, and rectified linear layers. 97. The method of embodiment 96, comprising a ReLU layer. 98. A convolutional layer is capable of convolving 3x3 pixel blocks of the input image. , The method according to any one of embodiments 94 to 97. 99. The ReLU layer is capable of applying the function f(x) = max(0,x) to the input. The method according to any one of embodiments 72 to 76. 100. The BatchNorm layer normalizes the activations of a neural network. The method according to any one of embodiments 94 to 99, wherein the method is capable of: 101. Any of embodiments 94 to 10, in which the neural network further includes a pooling layer. 10. The method according to any one of claims 1 to 0. 102. Any of embodiments 94 to 102, wherein the neural network further includes a dropout layer. 101. The method according to any one of claims 101 to 101. 103. The method of any one of embodiments 94 to 102, wherein the neural network further includes a fully connected layer. The method according to any one of the preceding claims. 104. A neural network is capable of performing a softmax operation. , The method according to any one of embodiments 94 to 103. 105. The neural network is pre-trained, according to any one of embodiments 94 to 1. 04. A method according to any one of claims 04 to 04. 106. Network (a) a first training set of input images of organoids having a disease phenotype; and Receive a second training set of input images of organoids with a disease phenotype and a non-disease phenotype. Steps; (b) comparing the disease or non-disease probability scores of the input image with their true values; P; and (c) During the backpropagation process through the neural network, Step of updating the node weights of The method is trained to assign a disease or non-disease probability score by a method comprising: 106. The method of embodiment 105, wherein 107. A first training set of input images and a second training set of input images 107. The method of embodiment 106, wherein each of the sets contains at least 500 images. 108. The network receives input images in a random order, as in embodiment 106 or The method of embodiment 107. 109. Steps (a) to (c) are to be repeated at least 200 to 800 times. and performing steps (a) to (c) includes one epoch. 8. A method according to any one of claims 1 to 8. 110. The method of embodiment 109, comprising about 500 epochs. 111. An embodiment further comprising the step of dilating the input image before each subsequent epoch. 106 to 110. The method according to any one of claims 106 to 110. 112. Enhancement can rotate, crop, scale, and or changing the contrast of the image, or changing two, three or all four 112. The method of embodiment 111, comprising a combination of operations. 113. Repeat steps (a) to (c) at least once using different hyperparameters. The method according to any one of embodiments 106 to 112, comprising repeating five times. 114. Repeat steps (a) to (c) at least once using different hyperparameters. The method according to any one of embodiments 106 to 112, comprising repeating 10 times. 115. Repeat steps (a) to (c) at least once using different hyperparameters. The method according to any one of embodiments 106 to 112, comprising repeating 20 times. 116. Hyperparameters include the number and type of layers, momentum and learning rate, Dropout percentage, number of epochs, or two, three, or all four handicap The method according to any one of embodiments 113 to 115, comprising a combination of hyperparameters. Law. 117. A neural network includes a fully connected layer with 512 nodes. The method according to any one of aspects 89 to 97. 118. The final fully connected layer has two nodes reflecting disease and non-disease phenotypes. The method according to any one of embodiments 89 to 98, 119. (a) culturing a third population of mammalian cells of the same cell type as the first population of cells, the third population of mammalian cells being a cell type that has a disease phenotype; culturing under organoid forming conditions resulting in organoids that (b) culturing a fourth mammalian cell population of the same cell type as the first cell population, the fourth mammalian cell population being a cell type that possesses a disease phenotype; culturing under organoid-forming conditions that do not result in organoids; (c) Imaging of organoids formed from the third and fourth mammalian cell populations. ; and (d) One or more images of organoids formed from the third mammalian cell population. and one or more images of organoids formed from the fourth mammalian cell population. The latent vectors of disease and non-disease phenotypes were analyzed using neural networks. Steps to get The method of any one of embodiments 103 to 118, further comprising: 120. The first and third mammalian cell populations contain one or more a fourth population of mammalian cells that is identical to the first population of mammalian cells except for the one or more mutations; and the third population of mammalian cells is isogenic to the second population of mammalian cells. 121. Culturing a first mammalian cell population, culturing a third mammalian cell population, and culturing a fourth mammalian cell population. The mammalian cell populations of embodiment 119 or embodiment 120 are cultured simultaneously. The method according to 122.Neural networks, (a) extracting the fully connected layers upstream of the two final nodes; and (b) a latent vector associated with a first population of mammalian cells contacted with the test molecule; and A latent vector associated with a first cell population that has not been contacted with the test molecule and a second cell population that has not been contacted with the test molecule. The difference between the latent vectors associated with the population is quantified to determine whether the first a mammalian cell population of said test molecule and a latent antibody associated with a first cell population that has not been contacted with said test molecule. providing a degree of difference between the vector and a potential vector associated with a second cell population. Steps to take A method comprising: determining a probability score of toxicity or non-toxicity for an organoid exposed to a test molecule; The method of any one of embodiments 35 to 121, wherein the method further comprises: 123. The degree of difference is associated with a first mammalian cell population contacted with the test molecule. A latent vector associated with a first cell population that has not been contacted with the test molecule. by taking the minimum of the difference between the latent vector associated with the second cell population and the latent vector associated with the first cell population. The method of embodiment 122, wherein the calculated value is: 124. The method of embodiment 123, wherein the degree of difference is determined across the entire latent space. method. 125. The method of embodiment 124, wherein the latent space includes at least 10 dimensions. 126. The method of embodiment 124, wherein the latent space includes at least 100 dimensions. . 127. The method of embodiment 124, wherein the latent space includes at least 250 dimensions. . 128. The method of embodiment 124, wherein the latent space includes at least 400 dimensions. . 129. The method of embodiment 124, wherein the latent space includes at least 500 dimensions. . 130. The method of embodiment 124, wherein the latent space includes at least 600 dimensions. . 131. The method of embodiment 124, wherein the latent space includes at least 800 dimensions. . 132. The method of embodiment 124, wherein the latent space includes at least 1000 dimensions. Law. 133. The method according to any one of embodiments 123 to 132, in which the latent space is reduced. Law. 134. The dimensionality reduction method is t-SNE or principal component analysis. 134. The method of embodiment 133, wherein the method is a PCA (Protein Crystal Analysis) method. 135. The probability scores for disease or non-disease and toxicity or non-toxicity are calculated using the The method according to any one of embodiments 133 to 134, further comprising plotting the data in a graph format. Method of posting. 136. The graph format is an equilateral triangle with the corners representing non-disease phenotypes, disease phenotypes, and 136. The method of embodiment 135, wherein said phenotype is a virulence phenotype. 137. (a) a processor; and (b) processor-executable instructions for implementing the analysis step of any one of claims 1 to 19. Non-transitory storage media including A system including: 138. (a) a processor; and (b) A processor-implemented method for implementing step (c) of any one of embodiments 27 to 84. A non-transitory storage medium containing executable instructions. A system including: 139. (a) a processor; and (b) a non-transitory method comprising: storage medium A system including: 140. To receive organoid images and use neural networks , one or more images of the organoid are analyzed using a neural network, Processor implementation to assign a disease or non-disease probability score ranging from % to 100% A non-transitory storage medium containing executable instructions. 141. The method according to any one of embodiments 94 to 118 and 122 to 136, 141. A non-transitory storage medium as described in embodiment 140, comprising a neural network for use. 142. Analyze one or more images of organoids to detect disease ranging from 0% to 100%. Or train a neural network to assign a probability score of non-disease. 1. A method for detecting a cellular function of a blood vessel comprising: (a) a first training set of input images of organoids having a disease phenotype; and Receive a second training set of input images of organoids with a disease phenotype and a non-disease phenotype. Steps; (b) comparing the disease or non-disease probability scores of the input image with their true values; P; and (c) During the backpropagation process through the neural network, Step of updating the node weights of The method includes: 143. A first training set of input images and a second training set of input images 143. The method of embodiment 142, wherein each of the sets contains at least 500 images. 144. The network receives input images in a random order, as in embodiment 142 or The method of embodiment 143. 145. The method further includes repeating steps (a) to (c) at least 200 times. , A method according to any one of embodiments 142 to 144. 146. The method further includes repeating steps (a) to (c) at least 400 times. , A method according to any one of embodiments 142 to 144. 147. The method further includes repeating steps (a) to (c) at least 500 times. , A method according to any one of embodiments 142 to 144. 148. The method further includes repeating steps (a) to (c) at least 600 times. , A method according to any one of embodiments 142 to 144. 149. The method further includes repeating steps (a) to (c) at least 800 times. , A method according to any one of embodiments 142 to 144. 150. The method further includes repeating steps (a) to (c) at least 1,000 times. 145. The method according to any one of embodiments 142 to 144. 151. The method of embodiment 145, comprising about 500 epochs. 152. In an embodiment further comprising the step of dilating the input image before each subsequent epoch. 145 to 151. 153. Extensions can rotate, crop, scale, and or changing the contrast of the image, or changing two, three or all four 153. The method of embodiment 152, comprising a combination of operations. 154. Repeat steps (a) to (c) at least once using different hyperparameters. The method of any one of embodiments 142 to 153, comprising repeating five times. 155. Repeat steps (a) to (c) at least once using different hyperparameters. The method according to any one of embodiments 142 to 153, comprising repeating 10 times. 156. Repeat steps (a) to (c) at least once using different hyperparameters. The method of any one of embodiments 142 to 153, comprising repeating 20 times. 157. Hyperparameters include the number of layers, momentum and learning rate, and dropout. Percentage, number of epochs, or two, three, or all four hyperparameters 155. The method of embodiment 154, comprising a combination of meters. 158. Any of embodiments 142 to 157, including a fully connected layer having 512 nodes. or one of the methods described above. 159. The final fully connected layer has two nodes that reflect the diseased and non-disease phenotypes. The method according to any one of embodiments 142 to 158, 160. (a) An imaging device capable of imaging organoids; (b)(i) receiving an organoid image from an imaging device; and (ii) using a neural network to generate a sequence of one or more images of the organoid; The results are analyzed using a neural network to determine whether the patient is diseased or not, ranging from 0% to 100%. To assign a probability score of The configured processor A system including: 161. The neural network according to any one of embodiments 94 to 118 and 122 to 136 According to embodiment 160, a neural network is used in the method according to any one of the preceding claims. The system described. 162. (a) An imaging device capable of imaging organoids; (b)(i) receiving an organoid image from an imaging device; and (ii) training a neural network to identify one or more of the organoids; The number of images is analyzed and 0% is detected according to the method according to any one of embodiments 142 to 159. Assign a probability score of disease or non-disease ranging from ~100% The configured processor A system including: 163. (a) An imaging device capable of imaging organoids; (b) i. receiving an organoid image from an imaging device; ii. Train the neural network to identify one or more of the organoids. The image is analyzed, and 0% to 159% of the image is analyzed according to the method of any one of embodiments 142 to 159. to assign a probability score of disease or non-disease ranging from 100%; and iii. Using a neural network to generate a sequence of one or more images of the organoid The results are analyzed using a neural network to determine whether the patient is diseased or not, ranging from 0% to 100%. To assign a probability score of The configured processor A system including: 164. A method for determining the toxicity of a test molecule, comprising: a) Carrying out or obtaining a method according to any one of embodiments 27 to 136 Collecting or obtaining data, e.g., resulting in Data (e.g., image data) is represented in a latent space; (b) removing the distance between wild type and disease from the latent space; (c) calculating the distance from the mean vectors of the wild type and disease phenotypes; (d) The distance calculated in step (c) and the standards for the wild-type and disease phenotypes. Comparing the deviations; and (e) determining a toxicity value Including, Optionally, steps (a)-(e) are performed using an autoencoder, e.g. a convolutional autoencoder. A method implemented by an encoder. 165. The determination of toxicity value involves assigning a probability score of toxicity or non-toxicity ranging from 0% to 100%. 165. The method of embodiment 164, comprising assigning 166. First organoids were assigned a probability score of toxicity or non-toxicity ranging from 0% to 100%. 166. The method of embodiment 165, further comprising the step of assigning a 167. The method of embodiment 165 or further comprising determining the LC50 of the test molecule. The method of embodiment 166. 168. The determination of the LC50 of a test molecule is performed using at least three different concentrations of the test molecule. repeating the method of embodiment 166 using a nonlinear least squares fitting method; Using an algorithm to analyze the probability scores of toxicity or non-toxicity at different concentrations 168. The method of embodiment 167, comprising: 169. Determining IC50 and LC50 of test molecules and further analysis and further selecting test molecules having an LC50 greater than the IC50 for the The method of any one of embodiments 164 to 168, comprising: 170. The compound according to embodiment 169, wherein the LC50 is at least 10 times greater than the IC50 method. 171. The compound according to embodiment 169, wherein the LC50 is at least 100 times greater than the IC50. How to. 172. The neural network includes an autoencoder, and optionally 27. Any one of embodiments 1 to 26, wherein the encoder is a convolutional autoencoder. The method according to 173. The method according to any one of embodiments 1 to 26, wherein the classification further comprises a virulence phenotype. method. 174. The method of embodiment 173, wherein the virulence phenotype is assigned by an autoencoder. The method described. 175. The virulence phenotype is determined by carrying out any one of (a) to (e) of embodiments 164 to 171. The method of embodiment 174, wherein the autoencoder is determined by performing the autoencoder. 176. The method of any one of steps (a) to (e) of embodiments 164 to 171 is performed. 137. The method of claim 136, further comprising: A system described in any one of 139 and 160 to 164. 177. Steps (a) to (e) of the method according to any one of embodiments 164 to 171 140. The method of claim 140, further comprising: 141. A non-transitory storage medium according to claim 141.

[0173] All publications, patents, patent applications, and other documents cited in this application are the property of their respective owners. Any publications, patents, patent applications, or other documents incorporated by reference for any purpose. They may be included in their entirety for all purposes to the same extent as if each document were individually indicated. No. 6,399,433, which is incorporated herein by reference.

[0174] Although various specific embodiments have been shown and described, it is understood that the spirit and scope of the present invention is not to be construed as limiting the scope of the present invention. It will be understood that various modifications can be made without departing from the scope of the invention.

Claims

1. 1. A method for classifying a test organoid, comprising: imaging the test organoid to prepare an organoid image; A training method for assigning a classification to the test organoids. and analyzing the organoid image using the neural network. wherein said classification comprises a disease phenotype or a non-disease phenotype. The method includes:

2. The cells of the organoid having a disease phenotype are selected from one or more of the following that confer the disease phenotype: The method of claim 1 , comprising multiple mutations.

3. The cells of the organoids having a non-disease classification and a disease classification are isogenic. The method according to claim 1 or 2.

4. The organoids may be multicellular aggregates, stem cells, pluripotent cells, induced pluripotent cells, human embryonic stem cells, Any one of claims 1 to 3, comprising stem cells, adult stem cells, totipotent stem cells, or non-embryonic cells. The method according to claim 5.

5. The method of any one of claims 1 to 4, wherein the disease comprises Huntington's disease.

6. Imaging includes staining the organoids for one or more markers. The method according to any one of claims 1 to 5.

7. The one or more markers may be a marker for a cell organelle, a marker for cell differentiation, a marker for a cell compartment, 7. The method of claim 6, comprising a marker, a marker, or a combination thereof.

8. The one or more markers may be a nuclear marker, a membrane marker, a differentiation marker, or the like. The method according to any one of claims 6 to 7, comprising a combination of these.

9. At least two images of each organoid were analyzed, each for different markers and The method according to any one of claims 6 to 8, wherein the first and second color channels correspond to:

10. The organoid image according to any one of claims 1 to 9, wherein the organoid image comprises a two-dimensional image or a three-dimensional image.

3. The method according to claim 1 .

11. The trained neural network is used to identify organoids belonging to a non-disease classification. The neural network was trained on multiple images of organoids and disease classifications. A neural network, the neural network having a plurality of layers and two The method of any one of claims 1 to 10, including a final layer that includes the nodes.

12. The neural network includes a fully connected layer upstream of the final two nodes, The coupling layer may include at least 10 nodes, at least 100 nodes, at least 250 nodes, at least 400 nodes, at least 500 nodes, at least 6 00 nodes, at least 800 nodes, or at least 1000 nodes The method of claim 11 , comprising:

13. 1. A method for identifying a test molecule that is biologically active against a disease, comprising: (a) culturing a first population of mammalian cells under organoid-forming conditions to produce a first organoid; obtaining organoids, wherein the organoid forming conditions include exposure to a test molecule; The first population of mammalian cells is cultured in the organoid form in the absence of a biologically active molecule. when cultured under nutrient conditions, results in an organoid having a disease phenotype; (b) imaging the first organoid after exposure to the test molecule; (c) train the subjects to assign a disease or non-disease probability score ranging from 0% to 100%. and detecting one of the first organoids using a neural network that has been trained on the first organoid. or analyzing a plurality of images; (d) assigning a probability score to the first organoid ranging from 0% to 100%. Top Including, The probability score of the first organoid is higher than a cutoff probability score for non-disease. If the probability score is equal to or lower than the cutoff probability score for the disease, the test molecule is effective against the disease. and optionally said first mammalian cell population is biologically active as a signaling agent for said disease expression. The method includes one or more mutations that confer a type.

14. The neural network generates a probability score of toxicity or non-toxicity ranging from 0% to 100%. and the method is further trained to assign a unique address to the Each organoid was assigned a probability score of toxicity or non-toxicity ranging from 0% to 100%. The method of claim 13 further comprising the step of:

15. Validating the accuracy of the neural network, and optionally Verifying the accuracy of the neural network (e) culturing a second population of mammalian cells of the same cell type as the first population of cells by organotyping; A second organoid is cultured under organoid-forming conditions to result in an organoid that does not have the disease phenotype. Obtaining the ganoids; (f) imaging the second organoid; and (g) assigning the second organoid a probability score ranging from 0% to 100%. Tep, the probability score of the second organoid is higher than the cutoff probability score for non-disease; or if the disease probability score is lower than the cutoff probability score, the neural network is accurate. is considered to be; or (h) culturing a second mammalian cell population of the same cell type as the first cell population to form an organoid. A third organoid is cultured under formation conditions to result in an organoid having the disease phenotype. obtaining an id; (i) imaging the third organoid; and (j) assigning the third organoid a probability score ranging from 0% to 100%. Tep, the probability score of the third organoid is lower than the cutoff probability score for non-disease; or higher than the cutoff probability score for disease, the neural network is accurate The step The method of claim 13, comprising:

16. the first population of mammalian cells containing one or more mutations that confer the disease phenotype; wherein the second population of mammalian cells has, other than the one or more mutations, 16. The method of claim 15, wherein the mammalian cell population is isogenic to the mammalian cell population of claim 15.

17. 17. The method of claim 15 or claim 16, wherein step (e) is carried out simultaneously with step (a). The method according to

18. a. the first population of mammalian cells is cultured in the presence of the test molecule for at least three days; Or b. The concentration of the test molecule in the incubation is between 10 nM and 100 μM, optionally 1 μM in the range of 20 μM or c. The organoid formation conditions are achieved by culturing a population of mammalian cells on a micropattern. or d. The first population of mammalian cells is selected from the group consisting of stem cells, optionally totipotent stem cells, human embryonic stem cells, A population of stem cells or pluripotent stem cells, and optionally, said organoid forming conditions and optionally comprising one or more differentiation factors in the culture medium, the differentiation factors being neurodifferentiation factors. and optionally, said neural differentiation factor is BMP4; or e. Any combination of a. to d. The method according to any one of claims 13 to 17.

19. The first mammalian cell population is a population of huntingtin polypeptides having expanded polyglutamine repeats. and optionally the expanded polyglutamine repeat is The method according to any one of claims 13 to 18, comprising from 2 to 150 glutamine residues.

20. The first population of mammalian cells is cultured in a microwell plate, and optionally, The microwell plate according to claim 13 to claim 16, wherein the microwell plate is a 24-well microwell plate.

20. The method of any one of claims 19 to 20.

21. The first population of mammalian cells produces 20-30 organoids per well. The method of claim 20, wherein the cell is cultured under conditions comprising:

22. Further comprising staining the organoids for one or more markers. Optionally, the one or more markers are organelle markers, cell differentiation markers, 22. The method according to claim 13, comprising administering to said patient a marker comprising a marker for detecting a cellular component of a tumor, a cell compartment marker, or a combination thereof. The method described above.

23. At least two images of each organoid were analyzed, each for different markers and The method according to any one of claims 13 to 22, wherein the first and second color channels correspond to:

24. The probability score is given on a scale of 0 to 1, and / or the cutoff probability The method according to any one of claims 13 to 23, wherein the score is 95%.

25. At least 85%, at least 90%, at least 95%, or at least 99% The method according to any one of claims 13 to 24, wherein the cut-off probability score is indicative of non-disease. method.

26. Screening a collection of test molecules to identify molecules that are biologically active against a disease A method for screening comprising the steps of:

26. The method according to any one of claims 13 to 25, optionally comprising carrying out a test Methods in which molecules are screened individually or in pools 。

27. the neural network is a convolutional neural network, and optionally , the neural network is a residual network optionally including 18 to 152 layers. Optionally, the layers include a convolution layer, a batch normalization layer, and and a rectified linear unit (ReLU) layer, and optionally, the convolution layer includes an input image A 3×3 pixel block of an image can be convolved, and optionally, the ReLU layer However, it is possible to apply a function f(x)=max(0,x) to the input, and optionally Alternatively, the BatchNorm layer can normalize the neural network activations. The method according to any one of claims 13 to 26, wherein

28. The neural network is pre-trained, and optionally, the network K, (a) a first training set of input images of organoids having a disease phenotype; and and receiving a second training set of input images of organoids having a non-disease phenotype. Steps: (b) comparing the disease or non-disease probability scores of the input images with their true values; Tep; and (c) during a backpropagation process through the neural network, Steps for updating the weights of the nodes in the network The method is trained to assign a disease or non-disease probability score by a method comprising:

28. The method of claim 27 .

29. a. The first training set of input images and the second training set of input images each of the image sets contains at least 500 images; and / or b. The network receives the input images in a random order.

30. The method of claim 28.

30. Further comprising repeating steps (a)-(c) at least 200-800 times. , performing steps (a)-(c) comprises one epoch, and optionally the method further comprises:

30. The method of claim 28 or 29, comprising 500 epochs.

31. (a) contacting a third mammalian cell population of the same cell type as said first cell population with said disease indication; culturing under organoid forming conditions resulting in organoids having a phenotype; (b) contacting a fourth population of mammalian cells of the same cell type as said first population of cells with said disease indication. culturing under organoid forming conditions resulting in atypical organoids; (c) isolating the organoids formed from the third and fourth mammalian cell populations. imaging; and (d) one or more of the organoids formed from the third population of mammalian cells. and one of the organoids formed from the fourth mammalian cell population. The plurality of images is analyzed using the neural network to identify the disease phenotype and Obtaining a latent vector of a non-disease phenotype The method of any one of claims 1 to 30, further comprising:

32. The neural network, (a) extracting the fully connected layers upstream of the two final nodes; and (b) a latent vector associated with the first population of mammalian cells contacted with a test molecule. a latent vector associated with the first cell population that has not been contacted with a test molecule; The difference between the latent vector associated with the second cell population is quantified to identify the test molecule and the first mammalian cell population that has been contacted and the first cell population that has not been contacted with a test molecule. a latent vector associated with the second cell population; providing a measure of difference between the A probability score of toxicity or non-toxicity is calculated for organoids exposed to a test molecule by a method comprising: The method according to any one of claims 14 to 31, further comprising assigning a

33. The degree of difference is associated with the first mammalian cell population contacted with the test molecule. The latency vector and the latency vector associated with the first cell population that has not been contacted with a test molecule. the minimum difference between the current vector and the potential vector associated with the second cell population. Optionally, the measure of difference is calculated by taking the smallest value across the entire latent space. The method of claim 32, wherein the

34. (a) a processor; and (b) a processor-executable program for implementing the analyzing step of any one of claims 1 to 12. Non-transitory storage medium containing functional instructions A system including:

35. (a) a processor; and (b) a processor-executable program for implementing step (c) of any one of claims 13 to 25. Non-transitory storage medium containing functional instructions A system including:

36. To receive organoid images and use a neural network to analyze organoids One or more images of the nose are analyzed using a neural network to determine the degree of noise reduction from 0% to 10%. Processor executable instructions for assigning a disease or non-disease probability score in the range of 0%. Non-transitory storage media, including:

37. One or more images of the organoids are analyzed to determine whether the organoid is diseased or not, ranging from 0% to 100%. To train a neural network to assign disease probability scores 1. A method comprising: (a) a first training set of input images of organoids having a disease phenotype; and and receiving a second training set of input images of organoids having a non-disease phenotype. Steps: (b) comparing the disease or non-disease probability scores of the input images with their true values; Tep; and (c) during a backpropagation process through the neural network, Steps for updating the weights of the nodes in the network The method includes:

38. (a) an imaging device capable of imaging the organoid; (b)(i) receiving an organoid image from the imaging device; and (ii) using a neural network to identify one or more of the organoids; The images are analyzed using a neural network to determine whether the image is diseased or not, ranging from 0% to 100%. To assign a probability score for a disease The configured processor A system including:

39. (a) an imaging device capable of imaging the organoid; (b)(i) receiving an organoid image from the imaging device; and (ii) training a neural network to identify one or more of the organoids; and analyzing the number of images to determine the disease or To assign a probability score for non-disease The configured processor A system including:

40. a. an imaging device capable of imaging the organoids; b. i. receiving an organoid image from the imaging device; ii. Training the neural network to recognize one or more of the organoids. and analyzing the image to determine whether the disease or non-disease ranges from 0% to 100% according to the method of claim 37. to assign a disease probability score, and iii. Using the neural network, generate a sequence for one or more of the organoids. The system analyzes a number of images using a neural network to identify disease or assigns a probability score for non-disease The configured processor A system including:

41. 1. A method for determining toxicity of a test molecule, comprising: a) the method according to claim 13; (b) removing the distance between wild type and disease from the latent space; (c) calculating the distance from the mean vectors of the wild type and disease phenotypes; (d) the distance calculated in step (c) and said wild type phenotype and said disease phenotype. comparing the standard deviations of the patterns; and (e) determining the toxicity value The method includes:

Images