Using machine learning and / or neural networks to validate stem cells and their derivatives for use in cell therapy, drug discovery and diagnostics

Abstract

A method for non-invasively predicting properties of one or more cells and cell derivatives is provided. The method includes training a machine learning model using at least one of a plurality of training cell images representing a plurality of cells and data identifying properties for the plurality of cells. The method further includes receiving at least one test cell image representing at least one test cell to be evaluated, the at least one test cell image being non-invasively acquired and based on absorbance as an absolute measure of light, and providing the at least one test cell image to the trained machine learning model. Machine learning based on the trained machine learning model is used to predict properties of the at least one test cell.

Description

[Technical Field]

[0001] FIELD OF THE INVENTION The disclosed embodiments relate generally to the predictability of cellular function and health, and more particularly to using machine learning and / or neural networks to validate stem cells and their derivatives for use in cell therapy, drug discovery and drug toxicity testing. [Background technology]

[0002] background Many biological and clinical processes are facilitated by the addition of cells. These processes include cell therapy (e.g., stem cell therapy), drug discovery, and testing the toxic effects of compounds. Assays for cell death, proliferation, function, and health are widely used in many areas of biological and clinical research. Several types of assays are used to assess cell health, death, proliferation, and functionality, including, but not limited to, staining with dyes, antibodies, and nucleic acid probes. For toxicity assessment and cell maturity, barrier function assays, such as transepithelial permeability (TEP) and transepithelial electrical resistance (TER), provide useful benchmarks. Other selected functional assays include, but are not limited to, electron microscopy (EM); gene (DNA or RNA) expression; electrophysiological recording; or methods assessing the secretion of specific cytokines, proteins, growth factors, or enzymes; the rate and volume of fluid / small molecule transport from one side of the cell to another; the phagocytic capacity of cells; immunohistochemical (IHC) patterns; and others. Some of these assays can be used as "release" criteria to verify the functionality of cell therapy products before implantation.

[0003] However, existing release assays used in cell therapy have significant limitations. Some of the aforementioned techniques, such as IHC and EM, are not quantitative. Other release assays have been found to have a high degree of variability (e.g., cytokine release, TER), are too disruptive (e.g., gene expression and phagocytosis), or are too expensive (cytokine release and gene arrays). Furthermore, current methods, at a minimum, require opening the cell culture dish to sample the medium and, at a maximum, completely destroy the cells being measured. In some cases, these assays cannot be performed throughout drug screening and toxicity testing, and often do not provide comprehensive information about cell health and functionality. This means that current methods cannot be used to longitudinally evaluate cells as they grow, develop, and differentiate without disturbing them, for example, before administration to patients or to test for drugs or toxic compounds. More specifically, many currently used assays increase the possibility of contamination and / or render the cells under test unusable for any further evaluation / implantation, are not amenable to high throughput, and do not provide a complete picture of cell health.

[0004] Biologists can often predict whether a particular type of cell (e.g., retinal pigment epithelial (RPE) cells) is functional simply by viewing those cells under brightfield / phase contrast microscopy. This approach is non-invasive, versatile, and relatively inexpensive. Despite these advantages, human image-based analysis has drawbacks, including but not limited to sampling bias, difficulty in directly correlating visual data with function, and difficulty in drawing causal relationships.

[0005] Therefore, there is a need in the art for analytical automation that can address the above limitations of current approaches in order to deliver cell and tissue therapies to patients, as well as to discover new drugs and potentially toxic compounds in a more efficient manner. Summary of the Invention [Means for solving the problem]

[0006] Summary of the Invention Objects and advantages of the hereinafter illustrated embodiments will be set forth in and apparent from the following description. Additional advantages of the illustrated embodiments will be realized and attained by the devices, systems, and methods particularly pointed out in the description and claims, as well as the appended drawings.

[0007] According to an embodiment of the present disclosure, a method for non-invasively predicting properties of one or more cells and cell derivatives is provided. The method includes training a machine learning model using at least one of a plurality of training cell images representing a plurality of cells and data identifying properties for the plurality of cells. The method further includes receiving at least one test cell image representing at least one test cell to be evaluated, the at least one test cell image being non-invasively acquired and based on absorbance as an absolute measure of light, and providing the at least one test cell image to the trained machine learning model. A property of the at least one test cell is predicted using machine learning based on the trained machine learning model. The method further includes using the trained machine learning model to generate release criteria for a clinical cell preparation based on the predicted property of the at least one test cell.

[0008] In embodiments, the machine learning may be performed using a deep neural network, and the method may further include segmenting the image of the test image into individual cells using the deep neural network.

[0009] Additionally, in embodiments, machine learning may be performed using a deep neural network, and the method may further include classifying the test cells based on the characteristics.

[0010] In further embodiments, the prediction step may be based on the classification step and may further include determining at least one of the identity of the cell, the function of the cell, the effect of a delivered drug, the disease state, and similarity to technical repeats or previously used samples.

[0011] In embodiments, prediction of properties of test cells can be performed for either a single cell and a plurality of cells in a field of view in at least one test cell image.

[0012] In some embodiments, the method may further include visually extracting at least one feature from the test cell image. The step of training a machine learning model may be performed using the extracted at least one feature, and the step of predicting may include identifying at least one feature of the test cell using the trained machine learning model trained using the extracted at least one feature, and predicting a property of the test cell using the identified at least one feature.

[0013] Additionally, in embodiments, the test cell images are acquired using quantitative bright-field absorbance microscopy (QBAM).

[0014] Additionally, in embodiments, the method may further include receiving at least one microscopic image captured by the microscope and converting pixel intensities of the at least one microscopic image into absorbance values.

[0015] In embodiments, the method may further include at least one of calculating an absorbance confidence of the absorbance value, establishing microscope balance by benchmarking, and filtering colors when acquiring images.

[0016] Additionally, in an embodiment, the first image processing of the test image may be performed by a deep neural network.

[0017] In embodiments, the machine learning may be performed using a deep neural network, and the method may further include segmenting the at least one test image into individual cells using the deep neural network, and the features are visually extracted from the segmented individual cells.

[0018] In embodiments, predicting the property of the at least one test cell may include at least one of predicting transepithelial resistance (TER) and / or polarized secretion of vascular endothelial growth factor (VEGF), predicting a function of the at least one test cell, predicting the maturity of the at least one test cell, predicting whether the at least one test cell is derived from an identified donor, and determining whether the at least one test cell is an outlier relative to a known classification.

[0019] In certain embodiments, generating release criteria for clinical cell preparations may further include generating release criteria for drug discovery or drug toxicity by trained machine learning models.

[0020] Additionally, in embodiments, the plurality of cells and the test cells may include at least one of embryonic stem cells (ESCs), induced pluripotent stem cells (iPSCs), neural stem cells (NSCs), retinal pigment epithelial stem cells (RPESCs), mesenchymal stem cells (MSCs), hematopoietic stem cells (HSCs), and cancer stem cells (CSCs).

[0021] In embodiments, the first plurality of cells and the test cells may be derived from a plurality of at least one of ESCs, iPSCs, NSCs, RPESCs, MSCs, or HSCs, or any cells derived therefrom.

[0022] Furthermore, in embodiments, the first plurality of cells and the test cells may comprise major cell types derived from human or animal tissue.

[0023] Additionally, in embodiments, the identified and predicted properties may include at least one of physiological properties, molecular properties, cellular properties, and / or biochemical properties.

[0024] In an embodiment, the at least one extracted feature may include at least one of a cell boundary, a cell shape, and a plurality of texture metrics.

[0025] Furthermore, in embodiments, the texture metrics may include sub-cellular features.

[0026] Additionally, in embodiments, the plurality of cell images and the test cell image may include a fluorescent image, a chemiluminescent image, an emission image, or a bright field image.

[0027] In embodiments, the method may further include determining whether at least one test image is a larger image, dividing the larger image into at least two tiles, individually providing each of the tiles to a trained machine model, combining outputs of processing by the trained machine model associated with each of the tiles, and providing the combined outputs to predict properties of the test cell into a single output corresponding to the larger image.

[0028] In a further aspect of the present disclosure, a computing system is provided for performing the disclosed methods.

[0029] These and other features of the systems and methods of the present disclosure will become readily apparent to those skilled in the art from the following detailed description of the preferred embodiments taken in conjunction with the drawings.

[0030] The present disclosure will become more fully understood from the detailed description and the accompanying drawings, which illustrate one or more embodiments of the present disclosure and, together with the description, serve to explain the principles of the disclosure. Wherever possible, the same reference numerals will be used throughout the drawings to refer to the same or like elements of an embodiment. [Brief explanation of the drawings]

[0031] [Figure 1] FIG. 1 is a block diagram illustrating an example operating environment according to one embodiment of the present disclosure.

[0032] [Figure 2] 2A and 2B illustrate selected machine learning and deep neural network frameworks that can be used to predict the function, identity, disease state, and health of cells and their derivatives, according to certain embodiments of the present disclosure.

[0033] [Figure 3-1] 3A-3C illustrate the conversion of multi-dimensional data to low-dimensional data using a principal component analysis machine learning model, according to an embodiment of the present disclosure. [Figure 3-2] Same as above.

[0034] [Figure 4] FIG. 4 illustrates an example of a hierarchical clustering method for identifying similar groups, according to an embodiment of the present disclosure.

[0035] [Figure 5] FIG. 5 illustrates an example of a vector-based regression model that may be used by a selected machine learning framework, according to an embodiment of the present disclosure.

[0036] [Figure 6] FIG. 6 is a schematic diagram illustrating one embodiment of a deep neural network, according to an embodiment of the present disclosure.

[0037] [Figure 7] FIG. 7 is a schematic diagram of an exemplary convolutional neural network (CNN) model architecture according to an embodiment of the present disclosure.

[0038] [Figure 8] FIG. 8 illustrates stem cell images that may be used by selected machine learning and / or deep neural network frameworks to measure stem cell function, according to an embodiment of the present disclosure.

[0039] [Figure 9] FIG. 9 illustrates the predicted results of the mean cell resistance (TER) provided by the CNN model, according to an embodiment of the present disclosure.

[0040] [Figure 10] FIG. 10 illustrates the lack of a direct correlation between bulk absorbance measurements and TER.

[0041] [Figure 11] FIG. 11 illustrates a segmentation based similarity approach used by a CNN model, according to an embodiment of the present disclosure.

[0042] [Figure 12] FIG. 12 illustrates a comparison of principal component analysis performed with and without visual data, according to an embodiment of the present disclosure.

[0043] [Figure 13] FIG. 13 illustrates a comparison of hierarchical clustering methods performed with and without visual data, according to an embodiment of the present disclosure.

[0044] [Figure 14]FIG. 14 depicts an induced pluripotent stem cell (iPSC)-retinal pigment epithelial (RPE) cell sequence analysis according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0045] Description of Specific Embodiments The following exemplified embodiments are merely illustrative and can be embodied in various forms as will be appreciated by those skilled in the art, and therefore, the exemplified embodiments are by no means limited to those exemplified. Therefore, any structural and functional details disclosed herein should not be construed as limiting, but should be understood merely as a basis for the claims and expressions for teaching those skilled in the art to variously use the contemplated embodiments. Furthermore, the terms and phrases used herein are not intended to be limiting, but rather to provide an understandable description of the exemplified embodiments.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Although any methods and materials similar or equivalent to those described herein can also be used in the practice or testing of the exemplified embodiments, exemplary methods and materials are now described.

[0047] It should be noted that as used herein and in the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Thus, for example, a reference to "a stimulus" includes a plurality of such stimuli, a reference to "the signal" includes a reference to one or more signals and equivalents thereof known to those skilled in the art, and so forth.

[0048] It should be appreciated that the exemplary embodiments discussed below are preferably software algorithms, programs or code residing on a computer usable medium having control logic for enabling execution on a machine having a computer processor, the machine typically including a storage device configured to provide output from the execution of the computer algorithm or program.

[0049] As used herein, the term "software" is meant to be synonymous with any code or program that may reside in a host computer's processor, regardless of whether its execution is in hardware, firmware, or as a software computer product available on a disk, on a storage device, or for download from a remote machine. The embodiments described herein include such software for executing the equations, relationships, and algorithms described below. Those skilled in the art will recognize additional features and advantages of the illustrated embodiments based on the above-described embodiments. Accordingly, the illustrated embodiments should not be limited to what has been particularly shown and described, except as indicated by the appended claims.

[0050] In an exemplary embodiment, computer system components may constitute "modules" that are configured to perform and operate to perform certain operations described herein below. Accordingly, the term "module" should be understood to encompass tangible entities that are permanently configured (e.g., physically embedded) or temporarily configured (e.g., programmed) entities that are physically constructed to operate in a certain manner and to perform certain operations described herein.

[0051] FIG. 1 illustrates a general overview of one operating environment 100 according to one embodiment of the present disclosure. In particular, FIG. 1 illustrates an information processing system 102 that can be used in embodiments of the present disclosure. The processing system 102, disclosed in further detail below, can be used to verify graft function in clinical biomanufacturing, including noninvasively predicting tissue function and cellular donor identity. The processing system 102 is compatible with advances in developmental biology and regenerative medicine that aid in the creation of cell-based therapies for treating retinal degeneration, neurodegeneration, cardiac disease, and other disorders by replacing damaged or degenerated native tissue with new functional implants developed in vitro. Induced pluripotent stem cells (iPSCs) have expanded the possibilities of cell therapy, enabling the transplantation of autologous tissue. The disclosed methods supported by processing system 102 are performed non-invasively and overcome barriers to large-scale biomanufacturing by eliminating the requirement for trained users, enabling high throughput, lowering costs, and reducing the time required to make appropriate predictions (e.g., regarding tissue function and cell donor identity).

[0052] The information processing system 102 shown in Figure 1 is merely an example of a suitable system and is not intended to limit the scope of use or functionality of the embodiments of the present disclosure described below. The information processing system 102 of Figure 1 may perform and / or carry out any of the functionality described below. Any suitably configured processing system may be used as the information processing system 102 in embodiments of the present disclosure.

[0053] As shown in FIG. 1, components of information processing system 102 may include, but are not limited to, one or more processors or processing units 104, a system memory 106, and a bus 108 that couples various system components, including the system memory 106, to the processor 104.

[0054] Bus 108 may represent any one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor bus or local bus using any of a variety of bus architectures, such as, by way of example and not limitation, an Industry Standard Architecture (ISA) bus, a MicroChannel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0055] System memory 106 includes a machine learning module 109, which in one embodiment is configured to perform one or more embodiments discussed below. For example, in one embodiment, machine learning module 109 is configured to apply selected machine learning techniques to generate an output vector based on an input array of measurements using a machine learning predictive model. In another embodiment, machine learning module 109 is configured to generate an output based on an input array of images using a deep neural network model (discussed in more detail below). While FIG. 1 depicts machine learning module 109 as resident in main memory, it should be noted that machine learning module 109 may reside within processor 104, may be a separate hardware component corresponding to multiple information processing systems and / or processors, and / or may be distributed across multiple information processing systems and / or processors.

[0056] System memory 106 may also include computer-system-readable media in the form of volatile memory, such as random access memory (RAM) 110 and / or cache memory 112. Information handling system 102 may further include other removable / non-removable, volatile / non-volatile computer-system storage media. By way of example only, storage system 114 may be provided for reading from and writing to non-removable or removable, non-volatile media, such as one or more solid-state disks and / or magnetic media (commonly referred to as "hard drives"). Magnetic disk drives (e.g., "floppy disks") may be provided for reading from and writing to removable, non-volatile magnetic disks, and optical disk drives (e.g., CD-ROM, DVD-ROM, or other optical media) may be provided for reading from or writing to removable, non-volatile optical disks. In such cases, each may be connected to bus 108 by one or more data media interfaces. The memory 106 may include at least one program product having a set of program modules configured to perform the functions of embodiments of the present disclosure.

[0057] A program / utility 116 having a set of program modules 118, as well as, by way of example and not limitation, an operating system, one or more application programs, other program modules, and program data, may be stored in memory 106. Each of the operating system, one or more application programs, other program modules, and program data, or some combination thereof, may comprise an implementation of a network-based environment. The program modules 118 typically perform the functions and / or methods of embodiments of the present disclosure.

[0058] The information processing system 102 may also communicate with one or more external devices 120 (e.g., a keyboard, a pointing device, a display 122, etc.); one or more devices that enable a user to interact with the information processing system 102; and / or any devices that enable the information processing system 102 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may occur via an I / O interface 124. Nevertheless, the information processing system 102 may communicate with one or more networks (e.g., a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet)) via a network adapter 126. As depicted, the network adapter 126 communicates with other components of the information processing system 102 via a bus 108. Other hardware and / or software components may also be used with the information processing system 102. Examples include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives and data archive storage systems.

[0059] Various embodiments of the present disclosure relate to a computational framework that uses novel neural network and / or machine learning algorithms to analyze images of stem cells and / or multiple cell types derived from these stem cells. In at least some embodiments, the neural network and / or machine learning can be used to automatically predict one or more physiological and / or biochemical functions of the cells.

[0060] Images input into the information processing system 102 may be acquired using an external device 120, such as a brightfield microscope. The information processing system 102 may automatically process these images in real time or at selected time points using quantitative brightfield absorbance microscopy (QBAM). QBAM uses absorbance imaging based on absorbance, an absolute measure of light. The QBAM method used can be implemented on any standard brightfield microscope and provides a statistically robust method for outputting reproducible images of reliable image quality.

[0061] Specifically, applying QBAM involves converting input image pixels from relative intensity units to absorbance units, an absolute measure of light attenuation. To improve imaging reproducibility, QBAM calculates image statistics in real time as images are captured, ensuring that the absorbance value measured at each pixel has a 95% confidence level of 10 milliabsorbance units (mAU).

[0062] QBAM offers advantages over methods that use raw pixel intensities because raw pixel intensities can vary depending on microscope placement and settings (e.g., uneven lighting, bulb intensity and spectrum, camera, etc.), making image comparison difficult, even when images are captured on the same microscope. Thus, the disclosed methods can provide automation, conversion of pixel intensities to absorbance values, calculation of absorbance confidence, and establishment of microscope equilibrium through the use of benchmarking to maximize the quality of image data captured using QBAM.

[0063] In scenarios where images are processed using transmittance values, such as when using histology, the statistical calculations can be modified to generate reproducible images of tissue sections. QBAM can be generalized to any multispectral modality because the calculations are not wavelength-specific. The disclosed methods may be suitable for hyperspectral autofluorescence imaging, which can identify cell boundaries and intracellular structures in non-pigmented cells.

[0064] The information processing system 102 may use one or more processing systems and / or software applications to process the received images (including to apply machine learning and / or deep neural network processing to the images, to train models, to extract features, to perform clustering functions, and / or to perform classification functions). In embodiments, images output by the microscope are converted from one format (e.g., but not limited to, CZI) to another format (e.g., but not limited to, JPEG or TIFF) before further processing occurs. In embodiments, a script in an application, such as, but not limited to, MATLAB®, that receives the images from the microscope is adapted to manage the format output by the microscope. In embodiments, the received images and data derived from the received images are processed by one or more different programs (e.g., MATLAB®, Fiji by ImageJ). TM , C++ and / or Python®). In embodiments, data may be transferred manually from one program to another. In embodiments, scripts are adapted to allow programs to transfer data directly between each other.

[0065] Thus, the disclosed methods can be implemented in any clinical-grade biomanufacturing environment or high-throughput screening system, requiring only a bright-field microscope, a digital camera, and a desktop computer for analysis.

[0066] This description presents a set of data-driven modeling tools that can be used to evaluate the function of human stem cells (e.g., iPSCs and retinal pigment epithelial (RPE) cells). In at least some embodiments, this image analysis is used to create lot and batch release criteria for clinical preparations of iPSCs and RPE. In certain embodiments, the disclosed systems can be utilized to deliver cell and various types of tissue therapies to patients in a more efficient manner. Furthermore, the disclosed approaches can be used: 1) to distinguish between different subtypes of cells, thus offering the potential for optimizing the generation or generation of specific cell and tissue types using stem cells; 2) to distinguish between healthy and diseased cells, thereby offering the potential for discovering the underlying mechanisms of drugs or diseases that may improve the health of diseased cells; and 3) to distinguish between cells treated with drugs and toxins, thereby offering the potential for determining the efficacy of drugs on cells, the side effects of drugs and toxins, and combinations thereof. Each of these specific methods can be performed at specific time points and / or over a period of time in longitudinal studies without disturbing the cells being studied.

[0067] 2A and 2B illustrate selected machine learning and deep neural network frameworks that can be used individually or together to predict the function, identity, disease state, and health of cells and their derivatives, according to certain embodiments of the present disclosure.

[0068] 2A, the machine learning predictive model 204 is trained to correlate the input data 202 through a data-driven statistical analysis approach to generate a new, interpretable model. In one embodiment, the input data 202 may include an input array of measurements representative of at least one physiological, molecular, cellular, and / or biochemical parameter of a plurality of stem cells or a plurality of stem cell-derived cell types.

[0069] The input data 202 can be derived from extracted visual features of individual cells extracted from QBAM images, for example, using the Web Image Processing Pipeline (WIPP). The extracted visual features can be used to train a selected machine learning algorithm to predict various tissue properties, including function, donor identity from which the cells were derived, and developmental outliers (the occurrence of abnormal cells). The selected machine learning algorithm can then be used to identify critical cellular features that can contribute to predicting tissue properties.

[0070] In various embodiments, the plurality of stem cells may include at least one of embryonic stem cells (ESCs), induced pluripotent stem cells (iPSCs), neural stem cells (NSCs), retinal pigment epithelial stem cells (RPESCs), induced pluripotent stem cell-derived retinal pigment epithelial cells (iRPEs), mesenchymal stem cells (MSCs), hematopoietic stem cells (HSCs), cancer stem cells (CSCs), or any cell type derived therefrom, of human or animal origin. In some embodiments, the input data 202 may include an input array of measurements representative of at least one physiological, molecular, cellular, and / or biochemical parameter of a plurality of major cell types derived from a human tissue or any animal tissue.

[0071] In various embodiments, the machine learning predictive model 204 may include one of several types of predictive models (e.g., dimensionality reduction / rank reduction, hierarchical clustering / grouping, vector-based regression, decision trees, logistic regression, Bayesian networks, random forests, etc.). The machine learning predictive model 204 is trained on similar cellular metrics and used to provide more robust grouping of both different patients and clones of iPSC-derived RPE tissue-engineered constructs. Thus, the machine learning predictive model 204 selects and ranks individual visual parameters of the cells to generate an output vector 206 representative of, for example, the physiological and / or biochemical function of the cells.

[0072] According to another embodiment of the present disclosure, as illustrated in FIG. 2B , a neural network framework, for example using a deep neural network model 212, is trained using a training input array 208 containing labeled and / or defined images and / or data. The images in the training input array 208 depict the microstructure of various stem cells or stem cell-derived cell types. In one embodiment, each microstructure image may preferably include a cellular microstructure captured at a magnification of 2× to 200×. Additionally or alternatively, the training input array 208 may include an image array in which the ultrastructure or microstructure data of cells of interest to the researcher is highlighted with an identifier, and may also include data identifying the cell's origin, cell location, cell line, physiological data, and biochemical characteristics for the corresponding plurality of stem cells or stem cell-derived cell types.

[0073] As described in detail below, the deep neural network model 212 can consistently and autonomously analyze images, identify features within images, perform high-throughput segmentation of a given image, and correlate images with identity, safety, physiological, biochemical, or molecular outcomes.

[0074] Furthermore, the segmentations generated by the deep neural network model 212 can be overlaid back onto the original microstructure images of the training input array 208, and then image feature extraction can be performed on those microstructure images. Various quantitative cellular information, including but not limited to cell morphometry (area, perimeter, elongation, etc.), intensity (mean intensity, mode, median, etc.), or texture (entropy, standard deviation, uniformity of intensity, etc.), can be calculated for individual features within the cells, for the cells themselves, or in tissue regions where stem cells or stem cell-derived cell types are distributed.

[0075] Advantages of the deep neural network model 212 contemplated by various embodiments of the present disclosure include the combination of the predictive reliability provided by the training input array 208 with the ability to consistently determine complex relationships between cellular images and cellular identity, safety, physiological and / or biochemical processes.

[0076] The segmentations generated by the deep neural network model 212 can then be used to identify which features in the images correlate with cellular identity, safety, physiology, and / or biochemical processes by applying selected machine learning methods to the features extracted from these images. Additionally, the deep neural network model 212 can reduce the uncertainty in the training input array 208.

[0077] Key to the process provided by the deep neural network model 212 for predicting the function, identity, disease state, and health of cells and their derivatives is the ability to automatically adjust parameter groupings in a quantitative manner based on characterized images of the training input array 208. Various embodiments of the present disclosure relate to novel methods for cellular function recognition and characterization focused on deep learning, whereby recognition of various functions can be performed using only visual data. The method disclosed below should be considered a non-limiting example of how deep convolutional neural networks can be used to characterize cellular functionality. Disclosed herein are methods and systems for training the deep neural network model 212 to handle limited labeled data.

[0078] Other potential frameworks used to characterize cellular functionality when large amounts of annotated data are available typically use pixel-wise segmentation in a supervised setting. Examples of such methods / frameworks include, but are not limited to, fully convolutional networks (FCN), U-Net, pix2pix, and their extensions. However, FCN frameworks used for other tasks, such as semantic image segmentation and super-resolution, are not as deep (depth refers to the number of layers) as the disclosed method. For example, an FCN network may receive an image as input and generate the entire image as output through four hidden layers of convolutional filters. Weights are learned by minimizing the difference between the output and the clean image.

[0079] One of the challenges of applying deep learning methods to cell image analysis is the lack of large amounts of annotated data to train neural network models. However, humans can use their visual system to recognize patterns both from everyday natural images and from microscopic images. Therefore, at a very high level, the goal is to train a model using one type of image and teach it to automatically learn data patterns that can be used for the task of recognizing cell functionality.

[0080] In one embodiment, the deep neural network framework uses a two-step approach. In one example, the images in the training input array 208 include live fluorescence microscopy images, multispectral brightfield absorption images, chemiluminescence images, radioactive images, or hyperspectral fluorescence images showing RPE cells with anti-ZO-1 antibody staining (white). The tight junction protein ZO-1 represents the boundaries of RPE cells. In another example, based on an understanding of cell boundaries and visual parameters (i.e., shape, intensity, and texture metrics) in the microscopy images, the deep neural network model 212 can detect correlations between cell boundaries and visual parameters in such images in the new input array 210 (e.g., live fluorescence microscopy images, multispectral absorption brightfield images, chemiluminescence images, radioactive images, or hyperspectral fluorescence images of similar cells or cell-derived products). It should be noted that texture metrics may include multiple subcellular features. Furthermore, it should be noted that the multispectral brightfield absorption images may include images using phase contrast, differential interference contrast, or any other image with a brightfield modality.

[0081] Embodiments of the present disclosure employ a concept known in the machine learning community as transfer learning. In other words, the disclosed methods utilize the efficient and automatic retention and transfer of knowledge from one task (microscopy images) to another related task (live absorbance image analysis).

[0082] As noted above, in various embodiments, the machine learning predictive model 204 may include multiple types of predictive models. In one embodiment, the machine learning predictive model 204 may use dimensionality reduction and / or rank reduction. Figures 3A-3C illustrate the conversion of multidimensional data to low-dimensional data using a principal component analysis machine learning model, according to an embodiment of the present disclosure.

[0083] When using factor analysis as a variable reduction method, the correlation between two or more variables can be summarized by combining the two variables into a single factor. For example, two variables can be plotted on a scatter plot. A regression line outlining the linear relationship between the two variables can be fitted (e.g., by the machine learning predictive model 204 of FIG. 2A). For example, if there are two variables, a two-dimensional plot can be performed in which the two variables define a line. For three variables, a three-dimensional scatter plot can be determined, and a plane can be fitted through the data. For more than three variables, it becomes difficult to plot points on a scatter plot, but the analysis can be performed by the machine learning predictive model 204 to determine a regression summary of the relationship between the three or more variables. In a plot that captures the principal components of two or more items, a variable that approximates the regression line can be defined. Data measurements from stem cell data or stem cell-derived cell type data for new factors (i.e., represented by a regression line) can be used in future data analysis to describe the essence of the two or more items. Thus, two or more variables may be reduced to one factor, where the factor is a linear combination of the two or more variables.

[0084] Extraction of principal components (e.g., the first and second components 302 and 304 in FIG. 3A ) can be found by determining a variance-maximizing rotation of the original variable space. For example, in the scatterplot 310 shown in FIG. 3B , the regression line 312 can be the original X-axis in FIG. 3A rotated to approximate the regression line. This type of rotation is called variance-maximizing because the criterion (i.e., goal) of the rotation is to maximize the variance (i.e., scatter) of the “new” variable (factor) while minimizing the variance around the new variable. While creating a scatterplot with three or more variables is difficult, the logic of rotating the axes to maximize the variance of the new factor remains the same. In other words, the machine learning predictive model 204 continues to plot the next best-fit line 312 based on the multidimensional data 300 shown in FIG. 3A . FIG. 3C shows the final best-fit line 322, where the entire variance of the data is explained by the machine learning predictive model 204.

[0085] According to another embodiment, the machine learning predictive model 204 may use a clustering approach. FIG. 4 illustrates an example of a hierarchical clustering method for identifying similar groups according to an embodiment of the present disclosure. Clustering approaches include cluster analysis. In the context of the present disclosure, the term "cluster analysis" encompasses several different standard algorithms and methods for grouping similar types of objects into respective categories, thus integrating observed data into a meaningful structure. In this context, cluster analysis is a general data analysis process for sorting different objects into groups that maximize the relatedness between two objects if they belong to the same group, and minimize the relatedness between them otherwise. Cluster analysis can be used to find structures within data without necessarily providing an explanation for the groupings. In other words, cluster analysis can be used to find logical groupings within data.

[0086] In an alternative embodiment, the clustering approach may include hierarchical clustering. The term "hierarchical clustering" encompasses the grouping of measured results into successively less similar groups (clusters). In other words, hierarchical clustering uses some measure of similarity / distance.

[0087] FIG. 4 illustrates an example of a hierarchical cluster 402. Clustering can be used to determine the safety of a cell therapy product by identifying which cell lines have mutated cancer genes. In this case, hierarchical bootstrap clustering can be performed on a dataset of cellular features obtained from segmentation of RPE images by a convolutional neural network. Cell lines that have mutated cancerous genes are statistically different from all other non-mutated lines (see, e.g., FIG. 13). In various embodiments, the clustering approaches discussed above can be used to determine how similar groups of treatments, genes, etc. are related to each other.

[0088] According to yet another embodiment, the machine learning predictive model 204 may use vector-based regression. Reference is now made to FIG. 5, which illustrates an example of a vector-based regression model that may be used by the machine learning predictive model 204, according to an embodiment of the present disclosure. More particularly, FIG. 5 illustrates a two-dimensional data domain 500. It will be recognized that this data domain may be n-dimensional in other embodiments, and that a hyperplane in an n-dimensional space is an n-1-dimensional plane (e.g., in a three-dimensional space, the hyperplane is a two-dimensional plane). In the case of a two-dimensional data domain, the hyperplane is a one-dimensional line in the data domain.

[0089] The data domain 500 has example measurement samples 501 and 503. The training samples include samples of one class (herein referred to as “first clone” samples 501) and samples of another class (herein referred to as “second clone” samples 503). Three candidate hyperplanes are shown, labeled 502, 504, and 506. Note that the first hyperplane 502 does not separate the two classes, and the second hyperplane 504 also does not separate the two classes, but with a relatively smaller margin than the third hyperplane 506, which separates the two classes with the largest margin. Thus, the machine learning prediction model 204 outputs 506 as the classification hyperplane, i.e., the hyperplane used to classify new samples as first clones or second clones.

[0090] Measurements identified as outliers can be used in regression analysis to analyze specific parameters, relationships between parameters, etc. Dimensionality is defined as a set of samples for which the dot product of the vectors is constant over time. Vectors are typically selected to maximize the distance between the samples. In other words, candidate vectors are penalized for proximity to the collected data set. In various embodiments, this penalization, as well as the distance at which such penalization occurs, can be configurable parameters of the machine learning predictive model 204.

[0091] The bottom portion of Figure 5 illustrates hyperplane 506 in comparison with other parallel hyperplanes 512 and 514. Hyperplanes 512 and 514 separate the two classes of data so that the distance between them is as large as possible. The third hyperplane 506 is located midway between hyperplanes 512 and 514.

[0092] 3-5 are merely examples of suitable machine learning methods that may be used by the machine learning predictive model 204 and are not intended to limit the scope of use or functionality of the embodiments of the present disclosure described above. It should be noted that other machine learning methods that may be used by the machine learning predictive model 204 include, but are not limited to, partial least squares regression, local partial least squares regression, orthogonal projection to latent structure, three-pass regression filter, decision trees with recursive feature elimination, Bayesian linear and logistic models, Bayesian ridge regression, etc.

[0093] 6 illustrates an exemplary fully connected deep neural network (DNN) 600 that may be implemented by the deep neural network model 212, according to an embodiment of the present disclosure. The DNN 600 includes multiple nodes 602 organized into an input layer 604, multiple hidden layers 606, and an output layer 608. Each layer 604, 606, 608 is connected by node outputs 610. It is understood that the number of nodes 602 shown in each layer is meant to be illustrative and in no way limiting. Additionally, although the illustrated DNN 600 is shown as fully connected, the DNN 600 may have other configurations.

[0094] As an overview of the DNN 600, an image 603 to be analyzed may be input to nodes 602 in an input layer 604. Each of the nodes 602 may correspond to a mathematical function with adjustable parameters. All of the nodes 602 may be the same scalar function, perhaps differing only by different parameter values, for example. Alternatively, various nodes 602 may have different scalar functions depending on layer position, input parameters, or other discriminatory features. By way of example, the mathematical function may take the form of a sigmoid function. It is understood that other functional forms may additionally or alternatively be used. Each of the mathematical functions may be configured to receive one or more inputs and to compute or calculate a scalar output from the one or more inputs. Using the example of a sigmoid function, each node 602 may compute a sigmoid curve nonlinearity of a weighted sum of its inputs.

[0095] Thus, nodes 602 in the input layer 604 receive microscopic images 603 of cells and then generate node outputs 610, which are delivered sequentially through the hidden layers 606, with the node outputs 610 of the input layer 604 going to the nodes 602 of the first hidden layer 606, which in turn going to the nodes 602 of the second hidden layer 606, and so on. Finally, the nodes 602 of the last hidden layer 606 may be delivered to the output layer 608, which may then output a prediction 611, for example, of a particular cell characteristic (e.g., TER) or cell identity.

[0096] Prior to runtime use of the DNN 600, the DNN 600 may be trained using labeled or transcribed data. For example, during training, the DNN 600 is trained using a set of labeled / defined images in the training input array 208. As shown, the DNN 600 is considered "fully connected" because the node output 610 of each node 602 in the input layer 604 and hidden layer 606 is connected to the input of every node 602 in either the next hidden layer 606 or the output layer 608. Thus, each node 602 receives input values ​​from the immediately preceding layer 604, 606, except for the node 602 in the input layer 604, which receives input data 603. Embodiments of the present disclosure are not limited to feedforward networks and also contemplate the use of recurrent neural networks, in which at least some layers feed data back to the previous layer.

[0097] In at least one embodiment, the deep neural network model 212 may use a convolutional neural network. FIG. 7 is a schematic diagram of an example convolutional neural network model architecture according to an embodiment of the present disclosure. The architecture in FIG. 7 shows multiple feature maps, also known as activation maps. In one illustrative example, if a target image 702 to be classified (e.g., one of the images in the new input array 210) is a JPEG image having a size of 224x224, a representative array for that image may be 224x224x3 (where "3" refers to RGB values). The corresponding feature maps 704-718 may be represented by the following arrays, respectively: 224x224x64, 112x112x128, 56x56x256, 28x28x512, 14x14x512, 7x7x512, 1x1x4096, and 1x1x1000.

[0098] Furthermore, as mentioned above, a convolutional neural network includes multiple layers, one of which is a convolutional layer that performs convolution. Each convolutional layer acts as a filter to filter input data. At a high level, a CNN takes an image 702 and passes it through a series of convolutional layers, nonlinear layers, pooling (downsampling) layers, and fully connected layers to obtain an output. The output can be a single class or a set of probabilistic classes that best describes the image.

[0099] Convolution involves generating an inner product based on a filter and input data. It is conventional to immediately apply a nonlinear layer (or activation layer), such as a ReLU (rectified linear unit) layer, after each convolutional layer. The purpose of the nonlinear layer is essentially to introduce nonlinearity into the system, which has just computed linear operations (e.g., element-wise multiplication and addition) during the convolutional layer operations. After several ReLU layers, a CNN may have one or more pooling layers. Pooling layers are also referred to as downsampling layers. There are also several layer options in this category, such as max pooling. This max pooling layer essentially takes a filter (usually 2x2 in size) and a stride of the same length, applies it to the input volume, and outputs the maximum number in each subregion the filter convolves.

[0100] Each fully connected layer receives an input volume (e.g., its output is the output of the previous convolutional, ReLU, or pooling layer) and outputs an N-dimensional vector (where N is the number of classes the learning model must select). Each number in this N-dimensional vector represents the probability of a particular class. The fully connected layer processes the output of the previous layer (which corresponds to an activation map of high-level features) and determines which features are most correlated with a particular class. For example, a particular output feature from a previous convolutional layer may indicate whether a particular feature in the image is indicative of RPE cells, and such features can be used to classify a target image as "RPE cells" or "non-RPE cells."

[0101] Furthermore, the exemplary CNN architecture has a softmax layer along with the final fully connected layer, which allows for explicit modeling of bipartite graph representations (BGLs), which can be used to optimize the CNN using global back-propagation.

[0102] 7 includes multiple convolutional and ReLU layers 720, max pooling layers 722, fully connected and ReLU layers 724, and softmax layers 726. In one embodiment, the CNN network 700 may include 139 layers and 42 million parameters.

[0103] Figure 8 illustrates stem cell images that can be used with traditional machine learning and / or deep neural network frameworks to measure stem cell function, according to an embodiment of the present disclosure. In this embodiment, iPSC-RPE cells from two different cell lines are seeded into each well of a 12-well dish. Note that these RPE cells are essential for photoreceptor development and function and require functional primary cilia for full maturation. One set of cells is left untreated, while the other two sets are manipulated at the mature stage of iPSC-RPE differentiation. Another set of cells is treated with aphidicolin, a tetracyclic antibiotic that increases ciliogenesis and promotes RPE differentiation by blocking the G1 to S transition of cells, after 10 days of culture.

[0104] Another set of seeded cells is treated with HPI-4, an AAA+ ATPase dynein motor inhibitor that acts by blocking ciliary protein transport and inhibiting ciliary function, thereby inhibiting RPE differentiation and providing a preferred negative control. According to certain embodiments of the present disclosure, measurements are taken at six different time points (e.g., weeks 2-7) during the maturation stage of iPSC-RPE differentiation. Such measurements include, but are not limited to, TER, cytokine secretion profile, and phagocytic capacity. It should be noted that one of the most important functions of RPE cells is the phagocytosis of photoreceptors, resulting in outer segment shedding. Figure 8 illustrates RPE cells 802 treated with aphidicolin, untreated RPE cells 804 (control group), and RPE cells 806 treated with HPI-4. The multiple images 802-806 are used as images in a new input array 210 to be analyzed. The CNN network 700 is configured to predict TER measurements for the multiple images. In one embodiment, the plurality of input images may include approximately 15,000 images.

[0105] FIG. 9 illustrates TER prediction results 900 provided by a CNN model according to an embodiment of the present disclosure. In FIG. 9, the horizontal axis 906 represents actual TER measurements, and the vertical axis 908 represents TER values ​​predicted by the CNN network 700. Results are shown for all three sets of RPE cells grown and for all measurements taken at different time points. In this embodiment, the release criterion is configured for a TER of 400 Ohm*Cm2. In FIG. 9, region 902 contains all correctly accepted cells, and region 904 represents all cells correctly rejected by the CNN model 700. In other words, in this case, the CNN model 700 is a perfect predictor, capable of predicting mean cell resistance (TER) with 100% sensitivity and 100% specificity.

[0106] FIG. 10 illustrates the lack of a direct correlation between bulk absorbance measurements and TER. More specifically, FIG. 10 shows a plot 1002 of absorbance versus TER. In FIG. 10, the horizontal axis 1004 represents actual TER measurements, and the vertical axis 1006 represents absorbance values. Plot 1002 clearly illustrates the lack of a direct correlation between bulk absorbance measurements and TER measurements. Therefore, multiple visual parameters (e.g., spatial data, textural data, and geometric data) used by CNN model 700 are necessary for predictive capabilities.

[0107] FIG. 11 illustrates a similarity-based segmentation approach used by a CNN model, according to an embodiment of the present disclosure. In FIG. 11, a first image 1102 represents an image from the training input array 208, which includes a confocal fluorescence microscope image showing RPE cells with anti-ZO-1 antibody staining (white). The tight junction protein ZO-1 represents the boundaries of RPE cells. A second image 1104 represents the first image 1102 segmented by the CNN model 700. Additionally, in this embodiment, the image from the new input array 210 to be analyzed includes a live multispectral absorption image, represented in FIG. 11 by the third image 1106. A fourth image 1108 represents the third image 1106 segmented by the deep neural network model 212. It should be noted that while the second image 1104 can be verified by manually correcting the cell boundaries clearly visible in the first image 1102, manual correction is quite difficult, if not impossible, for the fourth image 1108. Advantageously, based on an understanding of cell boundaries and visual parameters (e.g., shape, intensity, and texture metrics), the deep neural network model 212 is able to detect cell boundaries and perform correlation of visual parameters within similar multispectral absorption images of the new input array 210.

[0108] FIG. 12 illustrates a comparison of principal component analysis performed using only molecular and physiological data with image analysis performed using only visual data, according to an embodiment of the present disclosure. Each shape-shade combination in FIG. 12 represents a different clone generated from one of three donors (here designated “Donor 2,” “Donor 3,” and “Donor 4”). PCA dimensionality reduction makes it easier to visualize the grouping of clones across the 24 different selected metrics and 27 different shape metrics utilized by the machine learning predictive model 204. FIG. 12 illustrates a comparison of plot 1202 of PCA performed on image analysis using selected metrics (e.g., TER, gene expression, growth factor release, photoreceptor outer segment phagocytosis, etc.) with plot 1204 of PCA performed on image analysis using the machine learning predictive model 204. The same donor-clone separation trends are observed between the two plots 1202 and 1204. The first clone 1206 from donor 2 clearly separates furthest from the other clones, followed by the second clone 1208 from the same donor. Furthermore, clone 1210 from donor 3 and clone 1212 from donor 4 are more closely related to each other than clones 1206 and 1208 from the other donors.

[0109] This trend is consistent across the various machine learning techniques used by the machine learning predictive model 204, as shown in FIG. 13. FIG. 13 illustrates a comparison of hierarchical clustering methods performed with and without visual data, according to an embodiment of the present disclosure. In FIG. 13, a first hierarchical clustering map 1302 represents an analysis of measurements using the selected metrics discussed above (without visual image data), and a second hierarchical map 1304 represents the visual image analysis performed by the machine learning predictive model 204. According to the illustrated embodiment, the first hierarchical clustering map 1302 and the second hierarchical clustering map 1304 are generated by applying multi-scale bootstrap resampling to hierarchical clustering of the analyzed data. It should be noted that all numbers labeled with reference numeral 1306 represent corrected probabilities of clustering, and the remaining numbers represent uncorrected probabilities.

[0110] Figure 13 illustrates that the analysis performed by the machine learning predictive model 204 not only agrees with the analysis performed using the selected metrics, but also provides valuable insight into the underlying biology of what may cause this grouping. The second hierarchical clustering map 1304 is consistent with the sequence analysis of iPSC-RPE cells for cancer genes (illustrated in Figure 14), suggesting that the first clone 1206 of donor 2 had several cancer genes mutated during reprogramming.

[0111] Figure 14 depicts an example of an embodiment of the present disclosure, in which nine clones from three different donors were tested. Only one clone, 1206, exhibited a mutation during reprogramming to iPSCs, as indicated by the highlighted region 1402.

[0112] In embodiments, because certain software applications, such as MATLAB® or Fiji, may have limitations regarding the size of input images they can handle, a large input image may be divided into overlapping tiles. The tiles may be processed separately, for example, by performing segmentation using a deep neural network or by performing particle analysis on each tile separately. The image may then be reconstructed using an appropriate software application (e.g., C++), and the tiles may then be C++ reconstructions from the individual processed tiles. The reconstructed image may not include a visual indication that tiles were used.

[0113] In summary, various embodiments of the present disclosure relate to a novel computational framework for creating lot and batch release criteria for clinical preparations of individual stem cell lines to determine the degree of similarity to previous lots or batches. Advantageously, this novel computational framework can be used for any stem cell type (e.g., but not limited to, ESCs, iPSCs, MSCs, and NSCs), any derived stem cell product (e.g., iPSC-RPE-derived cells), or any given cell line or genetically modified cell therapy product (e.g., chimeric antigen receptor (CAR) T cells) that can be imaged using brightfield multispectral imaging. Benefits provided by the deep neural network model 212 and machine learning prediction / classification model 204 contemplated by various embodiments of the present disclosure include the predictive capabilities of the models for automating culture conditions (e.g., but not limited to, cell passaging time, purity of iPSCs or other stem cell types in culture, discrimination between healthy and unhealthy iPSC cell colonies, differentiation and potency of differentiated cells, drug and toxin discrimination and potency, etc.).

[0114] In other words, these techniques allow for the convenient and efficient collection of a wide range of information, from screening for new drugs to studying the expression of novel genes, creating new diagnostic products, and monitoring cancer patients. This technology also allows for the simultaneous analysis and isolation of specific cells. Used alone or in combination with modern molecular techniques, this technology provides a useful method for correlating complex mechanisms, including the overall activity of living cells, with uniquely identifiable parameters. Furthermore, one of the key advantages offered by the disclosed computational framework is that the automated analysis performed by selected machine learning methods and modern deep neural networks virtually eliminates human bias and error. [Example]

[0115] Experimental methods and methods used We developed a platform that uses quantitative brightfield microscopy and neural networks to noninvasively predict tissue function. In experiments, we used clinical-grade iPSC-derived retinal pigment epithelium (iRPE) from age-related macular degeneration (AMD) patients and healthy donors as model systems to determine whether tissue function can be predicted from brightfield microscopy images.

[0116] The retinal pigment epithelium (RPE) is a monolayer of cells that is of clinical interest in research related to its use to treat AMD. Moreover, the appearance of RPE cells within the monolayer is known to be crucial for RPE function, and in recent clinical trials, visual inspection of RPE cells by trained technicians was used as the release criterion for biomanufacturing for transplantation. RPE cell appearance is broadly defined by the maturity of junctional complexes between adjacent RPE cells and their characteristic pigmented appearance due to melanin production. These junctional complexes correlate with tissue maturity and functionality, including barrier function (transepithelial resistance and potential (TER) and TEP) measurements) and polarized secretion of growth factors (ELISA). Therefore, cell appearance and function correlate and may predict each other.

[0117] Variability in transmitted light microscope images makes them difficult to use for automated cell analysis and segmentation. Therefore, the platform developed in this study consists of two components. The first is QBAM, which uses an automated method to capture reproducible images across different microscopes. The second component is machine learning, which uses QBAM-generated images (QBAM images) to predict multicellular function. The machine learning methods were divided into the categories of deep neural networks (DNN) and selected machine learning (SML). These methods were selected to offer speed, reproducibility, and accuracy along with a noninvasive, automated method to aid in scaling biomanufacturing processes as cell therapies are translated from the laboratory to the clinic. System overview and test case description

[0118] To achieve reproducibility of brightfield imaging across different microscopes, we developed QBAM. QBAM converts pixels from relative intensity units to absorbance units, where absorbance is an absolute measure of light attenuation. To improve imaging reproducibility, QBAM calculates real-time statistics on images as they are captured to ensure that the absorbance values ​​measured at each pixel meet a threshold confidence level (configured at 95% confidence, in this experiment, as 10 milliabsorbance units (mAU)). In this study, three different bandpass filters were used for imaging, but this method can be adjusted to accommodate any number of wavelengths. QBAM imaging is implemented as a plug-in to Micromanager (for microscopes with available hardware) or a modular Python package (for microscopes not supported by Micromanager), allowing users to obtain QBAM images with just a few button presses.

[0119] Analysis of QBAM images was selectively performed at the field of view (FOV) scale or the single-cell scale. For each of these scales, we used DNNs with different objectives. At the FOV scale, the DNN was designed to directly predict two things: the outcome of functional / maturity assays (via DNN-F) or whether two sets of QBAM images came from the same donor (via DNN-I). No image processing was performed before feeding the images into DNN-F or DNN-I.

[0120] Single-cell analysis began with a DNN (via DNN-S) that identified cell boundaries in QBAM images. Next, visual features of individual cells were extracted from the QBAM images using the Web Image Processing Pipeline (WIPP; extractable features are listed in Table 4 below). The extracted visual features were then used to train an SML algorithm to predict various tissue properties, including the function, identity, and developmental outliers (those with abnormal cellular appearances) of the donor from which the cells were derived. The SML algorithm was then used to identify critical cellular features that contributed to the prediction of tissue properties. To demonstrate the effectiveness of this imaging and analysis method, a proof-of-principle study was performed on iRPE cells derived from the following donor types: healthy, oculocutaneous albinism (OCA), and age-related macular degeneration (AMD). iRPE cells from healthy donors were imaged as they matured, while AMD and OCA donors were imaged only once at the end of their maturation time point.

[0121] iRPE from five different OCA patients and two healthy donors were imaged using QBAM to determine the sensitivity of this imaging method to biological variation and naturally low melanin levels in OCA iRPE. In addition to weekly imaging, iRPE from healthy donors were also assessed for transepithelial resistance (TER) and polarized secretion of vascular endothelial growth factor (VEGF). TER is a measure of RPE maturity and increases as tight junctions form between adjacent cells. Polarized secretion of VEGF is a measure of RPE function, with more VEGF secreted on the basal side of the cell monolayer compared to the apical side (VEGF-ratio). Finally, eight iRPE clones were obtained from three AMD donors using a clinical-grade protocol. Here, clinical-grade refers to the production of cells using xeno-free reagents and a cGMP-compliant production process. Once AMD iRPE reached maturity, QBAM imaging was performed on them. QBAM accuracy, reproducibility and sensitivity

[0122] QBAM imaging was validated using a combination of reference neutral density (ND) filters and biological samples. ND filters with known absorbance values ​​were used as references to validate QBAM imaging by comparing the absorbance measured in a UV-Vis spectrometer with that measured using QBAM imaging. The absorbance measured using QBAM imaging correlated strongly with that measured by the spectrometer across the visible spectrum. To further validate this method, the reproducibility of QBAM for ND filters was demonstrated using three additional microscopes, each equipped with different filters, objectives, and light sources. The root mean square error (RMSE) across all filters and microscopes was 66 milliabsorbance units (mAU), or approximately 4.4% at the highest absorbance value measured.

[0123] We then tested QBAM imaging on gradually maturing live iRPE from two different healthy donors. A general trend of increasing mean absorbance over time was observed. To determine the sensitivity of QBAM imaging to iRPE pigmentation, we used QBAM to image iRPE from five different patients with OCA, a condition known to cause hypopigmentation of iRPE. OCA iRPE were sequenced to identify the type of albinism (OCA1A or OCA2) and disease severity. OCA1A iRPE had severe albinism and no melanin production (OCA8 and OCA26), and therefore had the lowest image absorbance. OCA2 patients had a phenotype ranging from moderate (OCA103 and OCA9) to mild (OCA71), which correlated with the absorbance measurements performed by QBAM. Despite the low levels of pigment produced in iRPE from OCA1A patients, the absorbance values ​​were two-fold higher than the minimum sensitivity of QBAM (10 mAU). Collectively, these data demonstrate the accuracy, reproducibility, and sensitivity of QBAM imaging. Deep neural network prediction of iRPE function from absorbance images

[0124] To determine whether QBAM imaging affects cell maturation and can measure a wide range of variations in iRPE pigmentation, we imaged iRPE from each healthy donor (Healthy-1, Healthy-2) using three culture conditions: (1) control iRPE (no treatment), (2) iRPE treated with a known inducer of RPE maturation (aphidicolin), and (3) iRPE treated with a known inhibitor of RPE maturation (hedgehog pathway inhibitor-4, HPI4).

[0125] Control and aphidicolin-treated iRPE were found to mature as expected, increasing their optical density over 8 weeks of culture, whereas iRPE treated with HPI4 tended to decrease optical density over time (designated Healthy-2 and Healthy-1). Higher expression of maturation marker mRNA and protein was found in control and aphidicolin-treated iRPE than in iRPE treated with HPI4. Baseline electrical responses (TEP and TER) and their changes to physiological treatments of 1 mM potassium (K+) or 100 μM adenosine triphosphate (ATP) were significantly greater in aphidicolin-treated iRPE and significantly smaller in iRPE treated with HPI4 compared with control. Furthermore, iRPE maturation was evident from the presence of dense, native-like apical processes. From this set of experiments, we can conclude that (1) iRPE generated in clinical-grade conditions have a mature epithelial phenotype, (2) weekly QBAM imaging does not affect iRPE maturation, and (3) differences in pigmentation between mature iRPE (control and aphidicolin) and immature iRPE (HPI4) can be quantified using QBAM imaging.

[0126] Next, we evaluated the ability to predict the function and phenotype of the iRPE monolayer from QBAM images of iRPE. For iRPE from healthy donors, the mean QBAM pixel values ​​showed little correlation with TER. However, TER predictions by DNN-F were highly correlated with actual TER values ​​for the same samples, and these predictions were consistent with a 70.6 Ω cm 2 To incorporate this methodology into a biomanufacturing setting, a 400 Ω cm 2 The term TER refers to the immature iRPE monolayer (<400 Ω cm 2 ) or mature (>400 Ω·cm 2) was used as a stringent threshold for classification as immature iRPE. False positives, false negatives, and TER values ​​associated with immature iRPE were determined. Based on this TER threshold, DNN-F was 94% accurate, with a sensitivity of 100% and a specificity of 90% for classifying iRPE maturity. A similar trend was observed for VEGF-Ratio, where polarized release of VEGF did not correlate well with QBAM mean pixel values, but DNN-F predictions were highly correlated with VEGF-Ratio measurements (R 2 = 0.89), and the RMSE of VEGF-Ratio prediction was less than 1.0. The accuracy, sensitivity, and specificity of VEGF-Ratio were all 100% (where samples with a VEGF-Ratio < 3.0 were considered immature). From these experiments, we can conclude that (a) QBAM images of live cells can be used to predict TER and VEGF-Ratio with high fidelity, and (b) QBAM imaging can be used as a non-invasive means of functional validation of cells instead of measuring TER and / or VEGF-Ratio. Extraction of single-cell features from live QBAM images of iRPE monolayers

[0127] DNNs are known to have superior predictive power compared to other machine learning algorithms, but it is difficult to determine which image features the DNN uses to make predictions. To understand which iRPE cell image parameters predict monolayer function, we computed image features of individual iRPE cells in QBAM images and used them to train an SML algorithm to predict iRPE function. We created a DNN (via DNN-S) to segment individual live iRPE cells in QBAM images. We validated the DNN-S segmentation by comparing the computed cell features from 12,750 iRPE cells with the same cell features computed from manual ground truth segmentations. Comparing 44 different features in the DNN-S vs. manually corrected segmentations (see Table 1 below), the difference between their feature histograms was 7.94% ± 4.42% (mean ± standard deviation), demonstrating good pixel-by-pixel agreement (F-2 = 0.71). Table 1: Quantification and distribution of % error from manual measurements and DNN-S segmentation of 44 morphological features. [Table 1-1] [Table 1-2]

[0128] In Table 1, the % error indicates the absolute difference in the numbers at each value in the histogram. The two-sample Kolmogorov-Smirnov statistic (KSS), Kernel Maximum Mean Discrepancy (KMMD), and X2 value are three different ways to assess distributional differences; KSS is used in nonparametric distributions, KMMD makes no assumptions on the initial distribution, and X2 is a chi-square test for normally distributed features.

[0129] QBAM imaging and live-cell segmentation enable the measurement of hundreds of cellular image features, enabling noninvasive tracking of individual cells throughout iRPE maturation. Therefore, we used a trained DNN-S to segment QBAM images of live iRPE (Healthy-1 and Healthy-2 donors) treated with aphidicolin or HPI-4, or not (control). Previously published cellular image features and intensity metrics known to correlate with RPE maturation and health were then evaluated for significance. By observing the average number of neighbors each iRPE cell had in response to drug treatment, we found that HPI-4 had significantly fewer neighbors at all time points (p<0.001). Importantly, this technique allows for unprecedented hierarchical granularity in the data; it allows not only measurement of the "bulk" tissue of the entire well, but also measurements according to the field of view or at the individual cell level. We clustered treatments based on cell area and mean cell intensity, two features known to be important for RPE maturation and health. A novel metric of iRPE function identified using SML, cell minimum intensity, is further described below. These results demonstrate that (a) the accuracy of DNN-S segmentation relative to hand-drawn segmentations of individual cells and (b) the differences between iRPE treated with different molecules can be explained using distinct cellular features. Single-cell imaging features can predict iRPE maturation and function

[0130] Five different SML methods (multilayer perceptron (MLP), linear support vector machine (L-SVM), random forest (RF), principal least squares regression (PLSR), and ridge regression (RR)) were used to predict TER and VEGF-Ratio from iRPE of Healthy-2 donors using cellular features derived from cell boundary segmentation of QBAM images. MLP TER predictions showed an RMSE of 84.7 Ω cm. 2 and R2 = 0.94 was found to be the most accurate SML approach. 2 was used as a quality assurance or quality control (QA / QC) threshold to determine false positives and negatives. MLP had an accuracy of 94%, a sensitivity of 100%, and a specificity of 90%. However, a comparison of all algorithms showed that DNN-F was the most accurate predictor of TER, with the RMSE of MLP being 14.1 Ω cm higher than DNN-F. 2 For VEGF-Ratio prediction, random forest (RF) was the best SML method, but DNN-F had a 1.4-fold lower RMSE.

[0131] The advantage of predicting iRPE monolayer function from cellular features is the ability to determine single-cell characteristics that suggest tissue-level function. For each SML method, cellular features were ranked by importance. When comparing all SML models, there was similarity in the most important features for predicting TER or VEGF-Ratio, regardless of the SML method used. Interestingly, the important cellular image features for predicting TER spanned cell intensity, texture, and shape. Of the 10 most important features, three pertain to cell shape (shape), three pertain to intracellular pigment intensity (intensity), and four describe the distribution of pigment within the RPE (texture). Table 2 shows a selection of which metrics specifically represent these features, along with their 95% confidence intervals for each time point and drug treatment. Table 2. Subset of features for modeling Healthy-2 data, along with their means, 95% confidence intervals (CIs), and standard errors. [Table 2]

[0132] Taken together, the above demonstrates that live-cell segmentation, feature extraction, and SML can predict tissue TER and VEGF-Ratio with accuracy levels approaching that of DNNs analyzing QBAM images. The advantage of the SML method over DNNs is that the SML model can identify distinct cellular features indicative of iRPE monolayer function, allowing manufacturers to derive confidence intervals for cell image features relative to cell therapy release criteria. The accuracy of functional prediction is robust across multiple clinical grades of AMD patient-derived iRPE

[0133] To determine robustness across multiple donors and preparations, DNN-F and SML were used to predict the TER and VEGF-Ratio of clinical-grade iRPE from three AMD patients in eight iPSC clones. Absorbance images of iRPE samples from each AMD donor and SEM images of the corresponding iRPE apical processes confirmed the polarized phenotype of iRPE. Monolayer maturation was assessed by TER and VEGF-Ratio as well as other assays. The mean QBAM pixel value of iRPE was measured for fully mature AMD-iRPE. Similar to healthy donors, the mean absorbance was used to predict TER (R 2 = 0.015) or VEGF-Ratio (R 2 = 0.50). However, the random forest (RF) model showed a similar degree of accuracy (RMSE = 70.9 Ω cm) as for the healthy donors. 2 , R 2 DNN-F was also used to model TER, and the predicted and measured values ​​were well correlated (R 2 =0.91).

[0134] To evaluate the robustness of the approach, we trained the SML model on different combinations of AMD-iRPE monolayers. A total of 18 unique training image subsets were formed, with each image subset containing test data with images of one iRPE sample from each donor. The average TER RMSE was 86.9 Ω cm across all clonal subsets. 2 + / -14.3 Ω·cm 2 (as shown in Table 5 below), indicating that prediction error is similar when scaled to a larger donor subset (with 8 AMD-iRPE samples) compared to a single sample (Healthy-2). A 95% confidence interval was constructed from the measured and predicted values. iRPEs falling outside this region may be considered "out of specification" in a biomanufacturing environment and may be recommended for further testing.

[0135] Finally, we determined whether the features used to predict AMD-iRPE function were similar to those in Healthy-2 by assessing the cellular features most important for predicting TER and VEGF-Ratio in AMD-iRPE monolayers across all donor / sample combinations. Interestingly, only four features overlapped between these two groups: Zernike n4-l0 polynomial (shape 1), mass displacement (intensity 2), and the third inverse difference moment at 135° (texture 2) and 45° (texture 1). Overall, the model derived from clinical-grade iRPE images was able to predict iRPE phenotype across multiple donors / samples, and we were able to identify common live-cell cellular image features across multiple donors that predict iRPE monolayer function. Furthermore, the difference in feature importance between Healthy-2 and AMD-iRPE suggests that there may be both donor-specific and common features across multiple donors that predict function. Classification of developmental outliers and identity of iRPE donors using QBAM

[0136] Using QBAM images, we determined the presence of any developmental outliers based on cytoimage features in eight clinical-grade iRPE samples from three AMD patients. Developmental outliers were defined as iRPE monolayers that differed from other iRPE monolayers based on cytoimage features and warranted further analysis to determine whether the monolayers had developed properly. Principal component analysis was applied to cytoimage features from QBAM images of AMD donors / sample preparations. iRPE from a given donor clustered well with each other based on cytoimage features, with the exception of AMD1 clone A (1A) and AMD3 clone A (3A), identified in the PCA hierarchical dendrogram. Analysis of iRPE from clone 1A revealed 894 alterations in the tumor exome compared to the starting donor material. iRPE from clone 3A were found to have lower pigment levels than their "sibling" iRPE. When analyzing the features of the cell images, cell pigmentation and shape were found to be the two most dominant feature classes in distinguishing these iRPE from outliers. A full description of the features can be found in the online dataset.

[0137] For each iRPE monolayer, multiple SML models were used to predict the cell donor from QBAM images. In addition, a new DNN (DNN-I) was developed to determine whether two different iRPE images were from the same donor. The SML algorithm took features from QBAM images as input and output the donor identity. In contrast, DNN-I took two QBAM images as input and classified them as coming from the same or different donors. When trained on images of other clones from the same donor, the SML approach was able to classify the donor identity of a new clone from a given donor, but it was unable to classify "new" donors that were not present in the training data. The DNN-I strategy for binary classification of two images as "same" or "not the same" also gave it the ability to classify "new" donors that were not used during training. Linear support vector machine (L-SVM) was found to have the highest accuracy of all SML algorithms tested, with 76.4% accuracy (2.3 × chance), 64.6% sensitivity, and 82.3% specificity. Across all donor / sample combinations, DNN-I performed better, with 85.4% accuracy (2.6 × chance), 80.9% sensitivity, and 86.8% specificity. Interestingly, the cytometric features important for distinguishing AMD iRPE from each other were similar across various iRPE combinations and comprised distinct features from those used to identify tissue functional and developmental outliers. The general differences between the top 50 features used in each application were as follows: shape features were important for identifying clones as outliers (23 of the top 50 features), texture features were important for donor classification of clones (25 of the top 50 features), and shape and texture features were important for classifying iRPE function (40 of the top 50 features). Absorbance Imaging Considerations

[0138] Data input is crucial for successful analysis. Therefore, the image processing pipeline developed here begins with a reproducible and rigorous absorbance imaging method using a brightfield microscope. The QBAM technique developed here can be implemented on any standard brightfield microscope and provides high confidence in image quality and reproducibility using a real-time, automated, and statistically robust method. The advantage of using absorbance rather than raw pixel intensity is that absorbance is an absolute measure of light attenuation. Raw pixel intensity can vary depending on microscope placement and settings (e.g., uneven illumination, bulb intensity and spectrum, camera, etc.), making image comparison difficult even when images are captured on the same microscope. Converting to absorbance values ​​overcomes many challenges related to image reproducibility. The combination of automation, converting pixel intensity to absorbance values, calculating absorbance confidence, and establishing microscope equilibrium through benchmarking ensures the quality of image data captured using QBAM.

[0139] To ensure its measurements can be used in multiple contexts, we validated the robustness of QBAM in three systems: (1) synthetic standards, (2) healthy biological samples, and (3) a drug-induced model of iRPE maturation. Analysis of QBAM images showed absorbance values ​​consistent with the "known" synthetic standard, allowing us to assess pigment development in both healthy and diseased RPE. This result also highlights the robustness of measurements across multiple microscopes or imaging setups. The error between different microscopic measurements of the same sample was within 4.4% of the signal, compared with an average error of 31% for the control standard for VEGF ELISA and 100% for TER measurements in epithelium. This corresponds to a 1-2 order of magnitude reduction in variability when used in potential release assays for cell therapy products or in drug screening platforms.

[0140] QBAM is optimized for measuring the absorbance of light-absorbing cell types. Therefore, QBAM may be suitable for assays that use absorbance (e.g., viable cell counting using trypan blue staining, histological staining, or analysis of light-absorbing biological specimens such as pigmented skin cells or dopaminergic neurons expressing neuromelanin). When transmittance values ​​may be preferred (e.g., histological examination), the statistics can be modified to generate reproducible images of tissue sections. Furthermore, this method can be generalized to any multispectral modality, since none of the calculated values ​​are wavelength-specific. Finally, it is anticipated that these methods may be suitable for hyperspectral autofluorescence imaging, which can identify cell boundaries and subcellular organization in non-pigmented cells. Considerations for predicting iRPE function

[0141] Neural networks and machine learning algorithms were trained on a full range of cell phenotypes by using healthy iRPE as well as drugs known to inhibit iRPE maturation (HPI4) and promote iRPE maturation (aphidicolin). The diverse phenotypes in the training set enhanced the robustness of the algorithms. Furthermore, this method served not only as an endpoint assay for tissue health but also as a noninvasive tool for tracking tissue development over a long maturation period (approximately 35 days) in two different donors. Importantly, the accuracy of these algorithms in predicting both TER and VEGF-Ratio approached the measurement uncertainties for both TER and VEGF.

[0142] The cutoff ratio for biomanufacturing is 400 Ω cm for TER. 2 , and VEGF-Ratio was selected as 3. However, 200Ω cm 2 ~1000Ω·cm 2It is important to note that similar accuracy, sensitivity, and specificity were observed when evaluating TER values ​​ranging from 0.01 to 0.1 or VEGF-Ratio values ​​ranging from 1 to 5, and that thresholds should be configured according to manufacturer specifications. In this study, larger prediction errors were observed for iRPE derived from Healthy-2 compared with those for iRPE derived from multiple AMD donors. Two reasons for this are hypothesized: (1) Healthy-2 iRPE had a wider range of both TER and VEGF-Ratio values ​​than AMD-iRPE, which was produced in a cGMP facility with the goal of reproducibly producing mature, healthy iRPE monolayers, due to the inclusion of positive and negative controls. (2) AMD-iRPE had a larger training dataset because it included all Healthy-2 data as well as training data from AMD-iRPE. In machine learning, more data generally leads to more accurate models; however, implementation of the platform in any application requires special consideration to ensure a wider range of conditions (i.e., more donors and positive / negative controls) for training than those presented in this proof-of-principle study to ensure model robustness.

[0143] As expected, deep learning achieved lower RMSE for both TER and VEGF-Ratio predictions compared to the SML approach. However, the SML approach enabled the discovery of important cell image features. We selected these two approaches because two motivations were recognized by cellular product manufacturers, regulatory agencies, clinicians, and / or researchers. Motivation 1: In manufacturing, clinical settings, or high-throughput screening, time is often a critical factor, and a clear "go / no-go" or simple readout is desired. In these cases, the algorithm that provides the highest accuracy and is most robust to noise should be used, and deep learning is an excellent tool for this application. Motivation 2: In research, insight into the underlying mechanisms of function is frequently important. In these cases, a more readable method for determining the importance of cell image features is needed. In this case, the SML approach is desirable because its underlying architecture is simple enough to understand and allows for the importance of factors (here, cell image features) to the prediction of tissue function.

[0144] Feature extraction from QBAM images yields hundreds of features based on cell shape, intensity, and texture, both at the single-cell level and in larger cell populations. Many of these features are mathematical abstractions with no meaningful connection to cellular function. Thus, while SML models can be more interpretable than DNNs, the features that comprise these models may not be related to the underlying biology. Nevertheless, for cell manufacturing purposes, what these features are and how they relate to the underlying biology is less important than being able to identify 95% confidence intervals and ensuring that future batches / clones derived from the donor fall within this range; validating their use in SML models, regardless of their relationship to the underlying biology. Considerations for the classification and clustering of clinically relevant donor-derived iRPE

[0145] Currently, there is a critical need to develop clinically compatible, non-invasive assays to verify the identity and quality of cell therapy products immediately prior to implantation. PCA and cluster analysis can contribute to this unmet need. Using this approach or similar clustering methods, the similarity of future implants to other technical replicates (or previously successfully manufactured batches) can be assessed non-invasively for the first time. Furthermore, the classification performed using DNN-I or L-SVM can serve as a QA / QC step to detect manufacturing errors in the cell implant and to match the identity of this implant to other replicates derived from the same donor. This is particularly important for facilities that manufacture thousands of autologous therapeutics and must verify the identity of each patient dose.

[0146] Two of the most important features for identifying developmental outliers were determined to be the standard deviation of iRPE maximum intensity (intensity 8) and the Zernike n5-l3 polynomial (shape 10). The deviation in the maximum intensity parameter coincides well with the absorbance results, which showed that iRPE derived from AMD3 clone A had lower absorbance than iRPE derived from other AMD3 clones. While the Zernike polynomial was useful for detecting the shape of invasive cancer cells and for classifying tumors, it is hypothesized that this was crucial for detecting differences between AMD1 clone A (which harbored 894 tumor exome mutations) and other AMD-iRPE lines. While this cluster analysis may not be conclusive evidence that QBAM imaging alone can reveal the development of cancer gene mutations, it is proposed that in a cell therapy manufacturing environment, screening individual therapeutic replicates using cluster analysis may be able to identify outliers that may require further scrutiny. Also, because this assay is non-invasive, this information may provide surgeons with additional confidence about the quality of the actual graft delivered to the patient, something not previously possible.

[0147] In conclusion, the study presented here demonstrates that QBAM imaging can be used to noninvasively assess pigmentation development in healthy and diseased iRPE. A DNN can analyze these images and accurately predict cellular TER and VEGF-Ratio in 10 different iRPE preparations. In addition, QBAM images contain sufficient information to enable the DNN to accurately segment the RPE boundary of live RPE. Once segmented, hundreds of features can be calculated for each cell, and these features can be used to predict cell function, identify outlier samples, and confirm donor identity. All of this information can be obtained for tissue to be transplanted into patients in just a few minutes using automated brightfield microscopy, without the need for expert input from a clinician. Therefore, QBAM has the potential to be applied in biomanufacturing settings, where thousands of manufactured RPE units can be noninvasively tested and qualified by technicians for clinical use. Experimental model and subject details Origin and culture of iRPE cells The origin of human cells

[0148] All human studies were performed under NIH Institutional Review Board-approved protocol #11-E1-0245. A total of 15 iRPE cell lines derived from 10 different donors were used in this study. These iRPE lines were derived from three types of patients: healthy, AMD, and OCA. iRPE derived from healthy patients (designated Healthy-1 and Healthy-2) were derived from iPSC lines BEST4C and LORDY9, respectively. iRPE derived from AMD patients are referred to according to donor and clone number. For example, AMD1A refers to cells derived from AMD donor #1 and clone A. Different clones for each donor were replicates, and each clone was fully reproduced from iPSC generation to iRPE differentiation. AMD clones were previously reported in Sharma et al., Patient-Specific Clinical-Grade iPS Cell-Derived Retinal Pigment Epithelium Patch Rescues Retinal Degeneration in Rodent and Pig Eyes, Sci. Transl. Med. Under Review (2018). A summary of the clone numbers per donor is as follows: AMD1 had clone A and clone B, AMD2 had clone A, clone B, and clone C, and AMD3 had clone A, clone B, and clone C. iRPE obtained from OCA patients (also referred to as albinism patients) were derived from five different patients (each a single clone), represented by OCA8, OCA26, OCA103, OCA9, and OCA71. Details of the age and sex of each donor, as well as vital source information, are provided in Table 3. Table 3. Important sources [Table 3]

[0149] Cell culture conditions and media All cells were cultured in a cell culture incubator at 37°C and 5% CO2. Depending on the developmental stage of the cells, different cell culture media were used. The different cell culture media used include:

[0150] Neuroectoderm induction medium (NEIM) was prepared using DMEM / F-12 (Thermo Fisher, 11330-032) supplemented with 1% (weight / volume) N-2 (Thermo Fisher, A13707-01), 1x B-27 (Thermo Fisher, 17504-044), 10 μmol / L LDN-193189 (Stemgent, 04-0074-10), 10 nmol / L SB431452 (R&D Systems, 1614), 0.5 μmol / L CKI-7 dihydrochloride (Sigma Aldrich, C0742-5mg), and 1 ng / ml IGF-1 (R&D Systems, 291-GMP-5.5ug).

[0151] DMEM / F-12 (Thermo Fisher, 11330-032), KOSR (Thermo Fisher, 11330-032), supplemented with 1% N-2 (Thermo Fisher, A13707-01), 1x B-27 (Thermo Fisher, 17504-044), 100 μmol / L LDN-193189 (Stemgent, 04-0074-10), 100 nmol / L SB431452 (R&D Systems, 1614), 5 μmol / L CKI-7 dihydrochloride (Sigma Aldrich, C0742-5mg), 10 ng / ml IGF-1 (R&D Systems, 291-GMP-5.5ug), and 1 μmol / L PD0325901 (Stemgent, 04-0006) were used. RPE induction medium (RPEIM), which is a medium containing 100% RPE cells per 1000 μg of ...

[0152] RPE commitment medium (RPECM) was prepared using DMEM / F-12 (KOSR, Thermo Fisher Scientific, A1099201) supplemented with 1% N-2 (Thermo Fisher Scientific, A13707-01), 1x B-27 (Thermo Fisher Scientific, 17504-044), nicotinamide 10 mmol / L (Sigma-Aldrich, PHR1033-1G), and activin A 100 ng / ml (R&D Systems, AFL338).

[0153] RPE growth medium (RPEGM) was MEMα (Thermo Fisher, 12571-063) supplemented with 1% N-2 (Thermo Fisher, A13707-01), 1% (wt / vol) glutamine (GlutaMAX Supplement, Thermo Fisher, 35050-061), 1% (wt / vol) non-essential amino acids (MEM Non-Essential Amino Acid Solution (100x), Thermo Fisher, 11140-050), 0.25 mg / ml taurine (Sigma Aldrich, PHR1109-1G), 20 ng / ml hydrocortisone (50 μmol / L solution, Sigma Aldrich, H-6909), 13 pg / ml triiodothyronine (Sigma Aldrich, T5516), and 5% (vol / vol) FBS (fetal bovine serum, GE Healthcare / Hyclone, SH30071.03).

[0154] MEMα (Thermo Fisher, A13707-01) supplemented with 1% (mass / volume) N-2 (Thermo Fisher, 35050-061), 1% (mass / volume) glutamine (Thermo Fisher, 11140-050), 0.25 mg / ml taurine (Sigma Aldrich, PHR1109-1G), 10 ng / ml hydrocortisone (Sigma Aldrich, H-6909), 13 pg / ml triiodothyronine (Sigma Aldrich, T5516), 5% (volume / volume) FBS (GE Healthcare / Hyclone, SH30071.03), and 50 μM PGE2 (prostaglandin E2, R&D Systems, 2296). RPE maturation medium (RPEMM), which is a medium containing 100% soluble erythrocyte membrane protein (100% soluble erythrocyte membrane protein). Transformation and differentiation of human iRPE

[0155] All iRPE were generated using a previously outlined clinical-grade protocol (Sharma et al., 2018). Briefly, iPSC clones were derived from CD34+ PBMCs using a previously published protocol (Mack et al., "Generation of Induced Pluripotent Stem Cells from CD34+ Cells across Blood Drawn from Multiple Donors with Non-Integrating Episomal Vectors," WB (2011)). iPSC cells were then seeded onto tissue culture plates (Thermo Sci., 140675) or T75 flasks (Corning, 430641U) coated with 5 μg / ml vitronectin (ThermoFisher, A14701S) and containing E8 medium (Essential 8 Medium, ThermoFisher, A1517001). After 2 days, the cell culture medium was changed to RPEIM for 10 days, then to RPECM for another 10 days. On day 22, the cell culture medium was switched to RPEGM. On day 27, cells were detached using Versene solution (0.2 g EDTA (ethylenediaminetetraacetic acid) per liter of phosphate-buffered saline, Thermo Fisher Scientific, 15040-066) and replated into new culture plates (Thermo Sci., 140675) or T75 flasks (Corning, 430641U) containing RPEGM. On day 42, cells were detached using CTS TrypLE Select Enzyme (Thermo Fisher Scientific, A12859-01) and transferred to biodegradable nanofiber scaffolds [Stellenbosch Nanofiber Company, FiberScaff-RPE TM The iRPE were replated at 500,000 cells / ml onto 3D cell culture scaffolds (AMD strain) or 12-well 0.4 μm polycarbonate transwell plates (Corning, 3401) (Healthy and OCA strains) and cultured in RPEMM. The iRPE seeded on day 42 were considered time 0, and all timing for drug treatment, imaging, and harvesting for assays in the experiments outlined in this document was counted from this day. Learn more about how Quantitative bright-field absorbance microscopy Basic principles

[0156] The basic principle behind QBAM imaging is absorbance measurement, which is an absolute measurement of light attenuation as described in Equation 1.

number

[0157] where A is the absorbance value at wavelength A, I(λ) is the intensity of light passing through the sample, and I0(λ) is the intensity of light in the absence of sample. One of the reasons why absorbance is of interest in this study is because the Beer-Lambert law establishes a relationship between absorbance and chemical concentration, as shown in Equation 2:

number

[0158] C is the chemical concentration in the sample, ε is the molar attenuation coefficient, and ι is the path length of the ray bundle. In the case of retinal pigment epithelial cells (RPE), melanin is produced by healthy RPE as they mature. Therefore, a doubling of absorbance may indicate a doubling of the amount of melanin in the cells. By converting pixel values ​​in RPE images to absorbance values, the images become melanin concentration maps that can be tracked noninvasively over time. In addition to the relationship to melanin concentration, using absorbance imaging had several advantages, the most important of which was reproducibility. Calculating absorbance values ​​requires an internal reference standard that allows values ​​to be compared across microscopes with different configurations. Using QBAM imaging, each pixel value was converted to an absorbance value by dividing the intensity of each pixel in the sample image (I(λ)) by the corresponding pixel in an image captured without the sample (I0(λ)). As with most measurements, there are several factors that must be considered when making a measurement to ensure reproducibility. QBAM imaging mitigates some of the sources of error through various procedures, including benchmarking, internal calibration and real-time statistics. Pixel-level absorbance calculation

[0159] For QBAM imaging, three different images were required to calculate the absorbance at each pixel: i) an image captured with the optical shutter closed at exposure time ε (I Dark (λ)), ii) an image captured with no sample in the field of view and the optical shutter open at exposure time ε (I Bright (λ)), and iii) the image of the sample (I(λ)). In terms of Equation 1, (I Bright where I(λ) was the blank reference image (I(λ)), but I(λ) was the reference image (I(λ)) to account for the camera bias and readout currents, as described in Equation 3. Dark It was necessary to include an additional term for (λ):

number

[0160] The subscripts i and j indicate the pixel location in the ith row and jth column of each image. Note that calculating transmittance (the term in this log function) is a priori background correction method reported by Young, 2001 (Young, Shading correction: compensation for illumination and sensor inhomogeneities. Curr. Protoc. Cytom. Chapter 2, Unit 2.11, (2001)). This means that calculating pixel-level absorbance essentially corrects for non-uniform illumination and is one factor that makes QBAM imaging robust across microscope configurations. Pixel-level confidence interval calculation

[0161] According to Equation 3, there were three different measurements taken to calculate absorbance, each with its own potential source of error. To improve the reproducibility of pixel-level absorbance measurements, the QBAM imaging method uses statistics to calculate confidence intervals for the absorbance value at each pixel. The goal of these statistical methods was to capture enough image data to ensure that the absorbance value at each pixel had a 95% confidence interval of 0.01 absorbance units (10 mAU). This means that the lower end of the dynamic range of QBAM is 10 mAU.

[0162] The standard deviation of the absorbance values ​​was calculated according to Equation 4:

number

[0163] In Equation 4, σ A(λ)i,j is the standard deviation of the absorbance value at position (i,j). σ I(λ)i,j and σ IBright(λ)i,j are the pixel intensity values ​​I(λ) for the pixel at location (i,j) in the sample image, respectively. i,j and the pixel intensity value I for the pixel at location (i,j) in the bright reference image.Bright(λ)i,j Equation 4 was derived using uncertainty propagation, but two things should be noted about this derivation. First, I(λ) i,j and I Bright (λ) i,j It was assumed that the dark reference image I was not correlated. This was not verified experimentally, but if these variables were correlated, an additional term would be subtracted from Equation 4, which would mean that the standard deviation of the absorbance was overestimated by this equation. Second, the standard deviation of the dark reference image I accounted for less than 1% of the standard deviation of the absorbance. Dark The error introduced by (λ) was ignored.

[0164] To ensure reproducibility of absorbance measurements, pixel calculations were regulated so that pixel absorbance values ​​had a 95% confidence interval of 0.01 absorbance units, assuming a normal distribution as shown in Equation 5.

number

[0165] where n is I(λ) i,j is the number of images captured for a sample, n is simply the number of captured images for a sample, and I Bright Note that the right hand side of Equation 5 represents an overestimation of the confidence interval since it does not include the number of captured images of Reducing chromatic aberration with color filters

[0166] When comparing images between different transmitted light microscopes, one reason the images appear different is the light spectrum. Everything about how the light spectrum is generated, manipulated using optical components, and collected in the microscope can vary from microscope to microscope. For example, different light sources emit different light spectra that can change with age and temperature, different objective lenses correct for chromatic aberration differently, and different cameras have different sensitivities to different wavelengths of light. While absorbance measurements can mitigate issues related to the spectral emission of light sources and the spectral sensitivity of cameras for a single microscope setup, color filters must be reproducible between imaging sessions and between microscopes. For this study, images of fixed RPE (AMD and OCA cells) were captured on a Zeiss AxioImager M2 microscope using three different color filters: 405 nm (ET405 / 10x, Chroma, Bellows Falls, VT), 548 nm (ET548 / 10x, Chroma, Bellows Falls, VT), and 640 nm (ET640 / 20x, Chroma, Bellows Falls, VT). Images of live RPE (Healthy-2 cells) were captured on a Zeiss AxioObserver Z1 microscope using three different color filters: 461 nm (FF01-461 / 5-25, Semrock, Rochester, NY), 541 nm (FF01-541 / 3-25, Semrock, Rochester, NY), and 671 nm (FF01-671 / 3-25, Semrock, Rochester, NY). Microscope benchmarking and blank image capture

[0167] The goal of benchmarking is to determine whether the pixel intensity of a microscope image responds directly to changes in illumination. In the case of QBAM imaging, doubling the exposure time should double the pixel intensity, so this is accomplished by finding how pixel intensity changes when changing the camera's exposure time.

[0168] The benchmarking protocol calculates the following quantities: pixel intensity as a function of exposure time, variance of pixel intensity as a function of mean pixel intensity, optimal exposure time, and minimum acceptable pixel intensity. These quantities and their calculations are described in detail below. This benchmarking protocol is used for each image filter because image characteristics can vary with wavelength.

[0169] The first step in the benchmarking protocol involved capturing images at various exposure times with the camera shutter closed and open. Prior to capturing images, the user defined the number of images, n, to be captured at each exposure time. The camera shutter was then closed and n images were captured at 1 ms, 2 ms, 4 ms, etc., up to 1024 ms, with all images saved for quality control. The camera shutter was then opened and n images were captured at 1 ms, 2 ms, and 4 ms. For each set of images captured at an exposure time, the mean and standard deviation of pixel intensity were calculated for each pixel. The exposure time was then doubled, n images were captured, and the mean and standard deviation were calculated for each pixel. This process was repeated until the image was overexposed. Overexposure at a particular exposure time was determined as follows:

number

[0170] In the formula, σ 2 I(λ,ε)i,j is the wavelength λ and exposure time 2 ε is the variance of intensity at pixel (i,j) captured in ms, and I*J is the total number of pixels in the image. Since overexposure is defined by reaching the maximum pixel intensity value of the camera, the standard deviation of overexposed pixels is 0. Therefore, for an exposure time of 2 ε The image captured at ε-1Equation 6 is satisfied when there is less variance in the image than in the captured image. However, this is a crude test for overexposure, as the standard deviation can fall faster than the maximum pixel intensity value (when full well capacity is reached) without meeting the overexposure criterion of Equation 6. For charge-coupled devices (CCDs), pixel intensities are known to follow a Poisson distribution, so a function was created to estimate the standard deviation based on pixel intensity, as shown in Equation 7:

number

[0171] In the formula, σ λ (p) is the standard deviation of pixel intensity versus pixel intensity p at wavelength λ, and α and β were parameters calculated from linear regression of mean pixel intensity versus standard deviation in images captured at 1 ms, 2 ms, and 4 ms.

[0172] A linear regression was then performed to determine how pixel intensity varied with exposure time. Camera exposure times that resulted in overexposed images were excluded from the linear regression, and if an image captured at a given exposure time had an error of more than 5% according to Equation 7, the image was classified as overexposed. According to the linear regression of pixel intensity against exposure time, the ideal exposure time was calculated in Equation 8 as follows:

number

[0173] In the formula, ε λ is the ideal exposure time, and p D is the maximum camera pixel bit depth, and σ λ (p D) was the standard deviation estimate (from Equation 7), and a and b were the slope and intercept of the linear regression of pixel intensity against exposure time. The ideal exposure time is the time at which the average exposure time should be three standard deviations below the maximum possible pixel intensity, thereby minimizing the number of individually overexposed pixels.

[0174] Once the ideal exposure time is known, I Bright The number of images to capture and average to obtain was calculated using Equation 9, which uses the standard deviation estimation function of Equation 7:

number

[0175] In the formula, n Bright is the pixel intensity p ε is the number of images required to have a 95% confidence interval of E for . For this study, E was set to 2 pixel intensity units, which corresponds to approximately 0.05% error for the 12-bit camera used in this study. By setting E = 2, n in Equation 9 Bright ≒σ λ (p ε ) 2 However, this approximation should not be used for cameras with different bit depths.

[0176] Minimum pixel intensity for accurate calculation of absorbance ( p min) The absorbance is a logarithmic function, and I Bright Pixel intensities that are low relative to σ can have significantly larger errors than pixel intensities that are higher. During live cell imaging, if any pixel in an image has a value less than the minimum pixel intensity, it is unlikely that the absorbance value for that pixel will have a 95% CI less than 0.01. Equation 10 below uses Equation 7 and Equation 4 to estimate the minimum pixel intensity value that satisfies Equation 5:

number

[0177] By using the minimum pixel intensity value during live cell imaging, adjust the method of capturing images to satisfy Equation 5. Assessment of microscopic equilibrium

[0178] A microscope used for QBAM imaging is equilibrated so that it captures images of the same field of view with little variation. Several factors can change how images taken of the same field of view can vary from one image capture to the next (e.g., changes in light bulb intensity as the light bulb heats up, or changes in camera sensitivity as the camera heats up with use). When images are captured repeatedly, the microscope should reach an equilibrium state where successive images have minimal variation between them. Using the benchmarking method described above, we developed an equilibrium metric that is measured when the microscope has reached equilibrium. The equilibrium metric is shown in Equation 11:

number

[0179] In the formula, E q is the balance metric, and a t is the slope obtained from the linear regression of pixel intensity against exposure time as described in the benchmarking section,

number

[0180] After ensuring the microscope is equilibrated, benchmark the microscope and capture a reference image (I Bright and I Dark ), and the sample is imaged. For each field of view of the sample, n images were captured and averaged. A bright reference image was always captured by focusing the microscope on the same medium in which the sample was prepared. For fixed cell samples mounted on microscope slides, a bright reference image was captured by focusing on a blank section of the microscope glass. For live cell imaging, a bright reference image was captured in a well containing the same volume of medium in which the cells were cultured. The average of any pixel value was calculated during benchmarking to obtain the minimum pixel value p min If the ideal exposure time was less than p, then the ideal exposure time was doubled and an additional n images were captured and averaged. This process of doubling the exposure time and capturing additional images was repeated until all pixels in the averaged image were p min All images captured at each exposure time were saved and used to convert from pixels to absorbance values ​​to increase the reliability of the measurements.

[0181] To calculate the absorbance for each field, Equation 3 is written as I(λ) i,j I(λ) was used with the average value of n images used in i,j <p min In the case of , the image captured at the longer exposure time was used to calculate the absorbance at that pixel by dividing the pixel intensity by the factor of increase in exposure time. Because pixel intensities follow a Poisson distribution, dividing the pixel intensity by the factor of increase in exposure time reduces the standard deviation by the same amount, resulting in a tighter confidence interval. iRPE culture, assays and imaging immunostaining

[0182] Tissues were prepared for immunohistochemistry by placing wells in 4% (wt / vol) paraformaldehyde (Electron Microscopy Science, 157-4-100) for 20 minutes. The immunohistochemistry blocking solution (IBS) consisted of 500 ml of 1x DPBS (Dulbecco's Phosphate Buffered Saline, Life Technologies, 14190250), 5% (wt / vol) bovine serum albumin (Sigma Aldrich, A3311), 0.5% (wt / vol) Triton® X-100 (Sigma Aldrich, X100-100ml), and 0.5% (wt / vol) TWEEN® 20 (Sigma Aldrich, P2287-100ml). Fixed cells were washed three times with IBS and permeabilized with IBS for 2 hours at room temperature. The cells were then stained with the following primary antibodies for 1 hour at room temperature: RPE65 (anti-RPE65 monoclonal antibody [1:300, Abcam, ab13826]), ezrin (monoclonal anti-ezrin antibody, 1:100, Sigma-Aldrich, E8897), or GT335 (anti-polyglutamyl-modified monoclonal antibody, 1:1000, Adipogen, AG-20B-0020). After primary staining, the samples were washed with IBS, and goat anti-mouse IgG Alexa Fluor 594 secondary antibody (1:300, Thermo Fisher, A-11032) was added and incubated overnight at 4°C. All antibodies were diluted in IBS solution. After overnight incubation, the samples were washed with IBS and anti-ZO-1 Alexa Fluor 488 mouse monoclonal antibody (1:200, Thermo Fisher, 339188) was added and incubated for 1 hour at room temperature. Additionally, nuclei were stained with Hoechst 33342 dye (1:1000, Thermo Fisher Scientific, H3570) for 15 minutes at room temperature. After staining, cells were washed with D-PBS (Dulbecco's Phosphate-Buffered Saline, Life Technologies, 14190250) and mounted on slides. All images were captured using a Zeiss AxioImager M2 microscope or a Zeiss Axio Scan Z1 slide scanner.Z-stacks were acquired over 50 μm in 1.5 μm increments along the z-direction, and maximum intensity projections were used for all analyses. ELISA quantification of VEGF

[0183] Vascular endothelial growth factor (VEGF) was measured from supernatants collected weekly from each well. At least five replicate wells were selected for each treatment at each time point, with three technical replicates measured per well. ELISA was performed using a singleplex Luminex® multiplex ELISA according to the manufacturer's protocol (R&D Systems). Electrophysiological measurements

[0184] Electrical resistance measurements used to predict iRPE monolayers were measured using EVOM2 and EndOhm chambers (World Precision Instruments, EVOM2 and ENDOHM-12, respectively). Each strain was measured for its resistance in 12 wells per treatment per week. Measurements of intracellular transepithelial potential and resistance in the presence of 1 mM potassium (Sigma Aldrich, P5405-1KG) and 100 μM ATP (Sigma Aldrich, A9187-500MG) were performed as previously reported (Miyagishima et al., In Pursuit of Authenticity: Induced Pluripotent Stem Cell-Derived Retinal Pigment Epithelium for Clinical Applications, STEM CELLS Transl. Med. 5, 1562-1574 (2016)). Briefly, RPE monolayer cultures derived from iPSC lines were mounted in modified "Ussing" chambers as previously described (Maminishkis et al., The P2Y2 Receptor Agonist INS37217 Stimulates RPE Fluid Transport In Vitro and Retinal Reattachment, Rat. Invest. Ophthalmol. Vis. Sci. 43, 3555-3566 (2002); Peterson et al., Extracellular ATP Activates Calcium Signaling, Ion, and Fluid Transport in Retinal Pigment Epithelium, J. Neurosci. 17, 2324-2337 (1997)). Transepithelial potential (TEP) was measured using a calomel electrode in series with Ringer's solution and an agar bridge. Signals from intracellular microelectrodes were converted to basolateral membrane potentials (V) relative to the basal bath. b ) and measure the apical membrane potential (V a ) into the equation: V a =V b -TEP was calculated by passing a current pulse of 2 to 4 mA through the tissue, and the resulting TEP, V aand V b The total transepithelial resistance (R t ) and the ratio of apical to basolateral membrane resistance (R A / R B ) was obtained. Gene expression in iRPE

[0185] Total RNA was isolated using NucleoSpin RNA (Machery-Nagel, 740955) according to the manufacturer's protocol. RNA was quantified using an ND-1000 spectrophotometer (Nanodrop Technologies) and the manufacturer's protocol. cDNA synthesis was performed using the Script cDNA Synthesis Kit (Bio-Rad, 1708891) and the manufacturer's protocol. A custom gene array containing primers for the following genes was purchased from Bio-Rad: MITF, PAX6, BEST1, CLDN19, PMEL, TYR, OCA2, RPE65, RLBP1, and the housekeeping genes RPL13A and B2M. Sybr green-based QPCR was performed on a ViiA 7 Real-Time PCR System (Thermo Fisher Scientific) using the EXPRESS One-Step SYBR GreenER Kit with premixed ROX (Thermo Fisher, 1179001K) according to the manufacturer's protocol. Each sample was run in at least two independent wells with triplicate technical replicates (2x HPI4, 3x aphidicolin, and 5x control). Wells at HPI4 had to be excluded due to extremely low RNA extraction yields, resulting in insufficient amplified cDNA for measurement. The mean CT of housekeeping genes for each well was used to determine the plotted -ΔCT values. Data were analyzed using R software (R Development Core Team, 2011). Electron microscopy of iRPE

[0186] Mature iRPE monolayers were fixed overnight in 10% (wt / vol) glutaraldehyde (Electron Microscopy Sciences, 16120). Samples were then washed three times with D-PBS and immersed in 25% (wt / vol) ethanol for 10 minutes. The samples were then removed and placed sequentially in 50%, 75%, 90%, 95%, and 100% ethanol solutions for 10 minutes each. After immersion in 100% ethanol, the samples were removed, placed in a critical point dryer (Leica EM CPD300), and processed for red blood cells according to the manufacturer's protocol. After CPD treatment, the samples were sputter-coated with 10 nm of gold and imaged using a Zeiss EVO 10 scanning electron microscope at a voltage of 10 kV and an amperage of 10 μA. Drug treatment and imaging

[0187] After differentiation, iRPE from Healthy-1 and Healthy-2 donors were seeded into 12-well 0.4 μm polycarbonate transwell plates (Corning, 3401) at a seeding density of 500,000 cells / ml in 0.5 ml. Only six of the 12 wells were used to minimize time spent outside the incubator during imaging. QBAM imaging and drug treatment began 2 or 1 week after seeding of Healthy-1 and Healthy-2 iRPE, respectively. For each culture plate, two of the six wells were treated with 3.0 μmol / L aphidicolin (Sigma Aldrich, A0781-10MG) or 30.0 μM HPI4 (hedgehog pathway inhibitor 4, Sigma Aldrich, H4541-25MG), or no supplements to RPEMM medium. RPEMM was changed three times per week, and imaging was performed on the same day each week immediately after medium change. Room-temperature medium minimized condensation, so to prevent fogging of the culture plate lids during imaging, the culture medium in the wells was replaced with room-temperature RPEMM (containing drugs, if necessary) immediately before imaging. To capture bright reference images for QBAM imaging, drug-free RPEMM was added to wells without cells. For each well, a 4x3 grid of overlapping images was captured for image stitching using a Zeiss AxioImager M2 microscope with a 10x objective (10-15% image overlap). For Healthy-2 and Healthy-1, a total of six plates were prepared, each with six wells, resulting in a total of 36 wells imaged and analyzed. iRPE derived from AMD samples was previously reported by Sharma et al. (Sharma et al., 2018), and iRPE derived from OCA donors were cultured as described above, but imaged only after they matured and were fixed with PFA. A total of 5 well replicates of OCA8, 6 replicates of OCA9, 4 replicates of OCA26, 8 replicates of OCA71 and 4 replicates of OCA103 were imaged.For AMD data, one repeat of AMD1 clone A, five repeats of AMD1 clone B, one repeat of AMD2 clone A, five repeats of AMD2 clone B, one repeat of AMD2 clone C, one repeat of AMD3 clone A, five repeats of AMD3 clone B, and one repeat of AMD3 clone C were imaged.

[0188] For iRPE imaged from the Healthy-2 donor, images were captured using three color filters, while Healthy-1 was imaged using a different filter set. The filters used to image Healthy-1 allowed a broader bandwidth of light to pass, thereby benefiting from shorter exposure times and, consequently, the time the imaged cells spent outside the incubator. To ensure that iRPE maturation was not adversely affected by imaging, Healthy-1 was first imaged using broadband filters. Once it was established that iRPE maturation in Healthy-1 was not adversely affected by QBAM imaging, narrower optical bandwidth filters were used to image Healthy-2 cells. All images were acquired using a Zeiss AxioImager M2 microscope. Deep neural network prediction of assays Network Architecture and Training

[0189] A deep convolutional neural network (DNN-F) was designed to predict transepithelial resistance (TER) and VEGF-Ratio. The basic structure of this network consisted of a series of preliminary convolutional layers followed by an inception layer deeper in the network (Szegedy et al., Going Deeper with Convolutions, ArXiv14094842 Cs (2014)). The network takes a 1024 × 1024 × 3 image as input and produces two values ​​as output: TER and VEGF-Ratio. The network was first trained to predict TER values, and then a VEGF-Ratio prediction layer was added to the end of the network. The three-channel image used as input consisted of three QBAM images captured at different wavelengths (488 nm, 561 nm, and 633 nm). The network had approximately 11 million parameters.

[0190] Prior to training, we improved the predictive ability of DNN-F by scaling and weighting cell measurements. TER measurements were divided by 1000 and VEGF-Ratio measurements by 10 to roughly scale their values ​​to a range of 0–1 for training. Initially, all images were given equal weights, but later in training, each image was weighted based on its associated TER value. Weights were generated by binning TER values ​​into bins with a range of 100 ohms, with the smallest bin containing TER values ​​ranging from 0 to 135 ohms. Weights were then assigned by dividing the total number of TER measurements by the number of measurements in each bin. As training progressed, weights were gradually decreased until all weights had a value of 1.

[0191] For training, mean squared error (MSE) regression was used as the objective function. Standard stochastic gradient descent was used with a constant learning rate. No image preprocessing was performed other than random cropping before feeding the images to the network. Each image was 1040 × 1388 pixels. Images in the training data were randomly cropped just before feeding them to the network, while test data were cropped at the same position each time. Test and training data were created by assigning one culture plate as test data, and the remaining plates were used as training data (two replicates of each treatment per plate; aphidicolin, HPI4, control). The best network was determined by finding the network with the lowest MSE in the test set. Fluorescence Image Segmentation

[0192] A deep convolutional neural network was designed to segment RPE fluorescently labeled for the tight junction protein 1 (ZO-1), highlighting cell boundaries and enabling accurate cell segmentation. Because cell boundaries in QBAM images have lower contrast than in RPE fluorescent images, making manual segmentation difficult, our goal was to obtain a highly accurate segmentation method to generate ground-truth cell boundary labels for QBAM. The approach was to: 1) train a DNN (DNN-Z) to segment cell boundaries in ZO-1 fluorescent images using corresponding images (where the cell boundaries were drawn by an expert technician); 2) collect QBAM images and fluorescent images of RPE fluorescently stained for ZO-1; 3) use DNN-Z to segment cell boundaries using the ZO-1 fluorescent images; and 4) train a new DNN (DNN-S) to segment cells in QBAM images using the ZO-1 segmentation. Human and mouse data

[0193] Images of retinal pigment epithelial (RPE) cells labeled for the cellular tight junction protein zonula occludens-1 (ZO-1) were collected from various human donors under various conditions. Human RPE were derived from induced pluripotent stem cells (iRPE). iRPE were obtained from healthy human donors or human donors with one of two different disease phenotypes: oculocutaneous albinism (OCA) or age-related macular degeneration (AMD). iRPE images were derived from a total of 10 different human donors. All iRPE were allowed to mature for at least 6 weeks, fixed in paraformaldehyde, labeled for ZO-1, mounted on microscope slides, and then imaged. Images were captured at magnifications ranging from 10x to 40x on three different microscopes. The human data were supplemented with images of whole adult mouse retinas labeled for cell boundaries and labeled with ZO-1, kindly provided by the laboratory of John Nickerson. Reference segmentation images and post-processing

[0194] Among many segmentation methods, the FogBank algorithm (Chalfoun et al., FogBank: a single cell segmentation across multiple cell lines and image modalities, BMC Bioinformatics 15, 431 (2014)) was selected to segment each fluorescently labeled image. The FogBank algorithm defines RPE cell regions using thresholding derived from intensity distributions in combination with a geodesic distance map of the contours. The FogBank segmentation results were reviewed and corrected by a human subject to obtain ground truth segmentation data. After generating the ground truth data, the images were divided into 256 × 256 tiles, resulting in 4,064 images of 256 × 256 pixels in size. Architecture and Training of Deep Convolutional Neural Networks for RPE Segmentation

[0195] A supervised DNN-based segmentation algorithm was designed in MATLAB® (R2017a) using the open-source machine learning framework MatConvNet (Vedaldi et al., MatConvNet - Convolutional Neural Networks for MATLAB. ArXiv14124564 Cs (2014)). The same architecture was used to segment fluorescent images labeled for ZO-1 (cellular tight junction staining) and cells in QBAM images. The basic layer structure was the inception layer used in GoogLeNet (Szegedy et al., 2014), and its higher-order structure followed the U-Net architecture (Ronneberger et al., U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, (Springer, Cham), pp. 234-241 (2015)). The network took a 256x256 pixel image as input and output a 200x200 pixel image, with cell boundaries having positive pixel values ​​and cell bodies having negative pixel values. The network had over 15 million parameters.

[0196] To train the network, we use the ADADELTA optimization (ε = 10 -6 and ρ = 0.9). Batch normalization was used with 10 images in each batch. The F1 score was used to determine the best network model, and correctly labeled cell boundaries were considered true positives. A dropout layer was used with a 50% dropout rate before the final convolution used to classify pixels. The objective function was a modified logistic-logarithmic loss function as shown in Equation 12:

number

[0197] where x is the predicted pixel class, c is the actual pixel class, c=+1 represents the cell boundary, c=-1 represents the cell body, and B=max{0,-cx}. Image preprocessing

[0198] Fluorescence images, but not QBAM images, were preprocessed to normalize pixel values ​​based on regional statistics. This normalization process 1) allows for image scaling for faster training of the DNN, since scaling image values ​​reduces training time; 2) helps remove some of the local background fluorescence or improves contrast in poorly stained areas; and 3) allows images of different bit depths or contrasts to be processed by the same network. Because pixel values ​​are absolute scales, typically between 0 and 1, QBAM images were not normalized before feeding them into the network. The integral image was used to calculate the local mean and standard deviation within a 127 x 127 pixel box centered on each pixel in the image. Each pixel was normalized based on the local pixel mean and standard deviation according to Equation 13:

number

[0199] In the formula, p i,j is the pixel value at row i and column j, and μ i,j and σ i,j is pixel p i,j are the mean and standard deviation of a 127 × 127 pixel region centered on z i,j is the normalized pixel value. Pixel weight assignment for DNN training

[0200] To improve the accuracy of the DNN, the training weights for each pixel in an image were adjusted based on the normalized pixel value according to the classification of each pixel. This adjustment helped the DNN correctly label cell boundaries in signal-poor areas. The cell boundary pixels were then labeled with μ i,j and σi,j But, p i,j, was normalized as described in Equation 1, except that it is the mean and standard deviation of only the boundary pixels in a 127 × 127 pixel region around z. i,j All boundary pixels with a weight of 1 are assigned a weight of -z i,j This resulted in all bright pixels having a weight of 1, and dark pixels were assigned a weight inversely proportional to their pixel intensity. Weights for cell bodies were assigned based on their proximity to the cell boundary. A distance transform for each cell body was performed to the cell boundary, and the square root of that distance was then used as the pixel training weight. Cell body weights were truncated so that all weights greater than 10 were assigned a value of 10.

[0201] Training weights were applied to images during training using two different methods. The first method applied the weights conventionally, multiplying the loss during backpropagation by the weights. The second method used the weights as pixel class labels, c, in the logistic regression function (Equation 1). In both the first and second methods, c>0 for cell boundaries and c<0 for cell bodies. In the first method, c=±1, while in the second method, c was equal to the pixel weight. The second method was found to be better for several reasons. First, the second method yielded higher F1 scores on the test dataset. Second, no deviations in F1 scores (which may indicate overfitting) were observed between the test and training datasets while training the DNN. QBAM image segmentation Determining cell boundaries in QBAM images for DNN training

[0202] A DNN was developed to segment QBAM images (DNN-S). A subset of samples was imaged using QBAM imaging in addition to fluorescence imaging of ZO-1, specifically patients with AMD. Fluorescence images were captured along with transmitted light images of the iRPE. Images of the same regions were then captured using QBAM imaging. The QBAM images were registered to the fluorescence images of the iRPE by finding 256 × 256 pixel regions that had at least 97% correlation between the transmitted light images captured during fluorescence and QBAM imaging. Correlation was assessed using generalized normalized cross-correlation using fast Fourier transforms. The fluorescence images were segmented using the DNN, and the DNN segmentation was used as cell boundary labels for the QBAM images. DNN ensemble segmentation of QBAM images

[0203] Unlike the DNN used to segment RPE fluorescence images, which achieved an F1 score of over 80%, the DNN segmenting QBAM images achieved an F1 score of ≈60%. This was due in part to the lack of overlap between ZO-1 boundary fluorescence and the boundaries observed in QBAM images of some regions of the RPE monolayer. To improve the segmentation accuracy of DNN-S, we used an ensemble approach, training seven different networks and using the consensus of all networks to determine the RPE boundary in QBAM images, achieving an F1 score of 66%.

[0204] No preprocessing of the QBAM images was performed before processing them with DNN-S. The structure of the DNN-S was the same as that used to segment the fluorescence images, except that the input image had three channels instead of one. The three channels were three QBAM images captured of the same field of view, but each QBAM image was captured at a different wavelength (488 nm, 561 nm, and 633 nm). Segmentation preprocessing for feature extraction File selection for Healthy-2 and AMD

[0205] From the directory structure containing all red, green, and blue brightfield images, calibration and raw images, absorbance images from the blue channel, and ZO-1 fluorescently labeled images, for the Healthy-2 dataset (81,646 files in 1,564 folders), 2,580 absorbance files with a size of 1,388 pixels by 1,040 pixels were selected for segmentation. For the AMD dataset, 20 absorbance images with a size of 2,700 pixels by 3,000 pixels were selected. Invert colors and skeletonize segmentation results

[0206] To remove artifacts and fill small gaps, a morphological binary operation called fill was applied to the segmented images, including cell borders. The segmentation was then thinned so that the cell borders were one pixel thick. Next, the images were inverted so that the foreground corresponded to the cell interior and the background corresponded to the cell boundary. Finally, the binary-inverted images were labeled so that each cell region was labeled with a unique id. Removal of peripheral cells in the image

[0207] To calculate accurate cell-level features from absorbance images on reference segmentation masks, each mask was preprocessed to remove segments that reached the image boundary. This preprocessing step eliminated feature values ​​that may correspond to partial cells. The same preprocessing routine was applied to the computer-segmented images when they were evaluated for segmentation accuracy. Image Segmentation

[0208] For the AMD dataset, the large image stitched FOV poses scalability challenges for RAM during DNN-based model training. The size of the FOV varies depending on the experimental collection. Therefore, the collected images were preprocessed by stitching them into a large mosaic image and then dividing the mosaic image into 512x512 image tiles with 0% overlap. Image feature extraction Feature extraction of albinism iRPE

[0209] From a directory structure containing 1,978 files in 439 folders, including all red, green, and blue bright-field images, calibration and raw images, blue channel absorbance images, and ZO-1 fluorescently labeled images, 381 blue channel absorbance files were selected for image-level feature extraction. These files corresponded to a 3 × 3 field of view (FOV) with 10% overlap. Due to the fact that some absorbance values ​​were below zero (calibration artifact), a binary mask was created for each image to exclude these negative values ​​from feature extraction. A feature vector was calculated for each absorbance image (FOV) on each corresponding mask using the Web Image Processing Pipeline (WIPP). The feature vector included 14 intensity-based and 5 texture-based image features, originally implemented in MATLAB® (R2017a). The intensity-based features corresponded to central moment and entropy properties. Texture-based features were derived from gray-level co-occurrence matrices (GLCMs) and included contrast, uniformity, correlation, energy, and entropy. Scatter plots of the features were generated using scripts in R and the ggplot package, and selected plots were included in the main manuscript. Feature extraction of AMD and healthy iRPE

[0210] To extract as many cellular-level features as possible, we leveraged several widely used image libraries for feature extraction integrated into WIPP. Although feature values ​​for the same feature definition may vary, they provide insight into the amount of variation introduced not only by image acquisition but also by image processing. Intensity, shape, and texture features were calculated using feature extractors in WIPP (Bajcsy et al., Web Microanalysis of Big Image Data, Springer International Publishing (2018b)), specifically, extractors from MATLAB (2017a) and CellProfiler (Kamentsky et al., Improved structure, function, and compatibility for CellProfiler: modular high-throughput image analysis software, Bioinformatics 27, 1179-1180, 2011 (2011)), applied to each cell region. Features corresponding to cell regions located at the image border and very large connected regions were discarded. Table 4 below summarizes these features: Table 4: List of cell feature class types extracted per cell region and the software that derived the features. [Table 4-1] [Table 4-2]

[0211] Table 4 lists 40 cell features, but many features have subsets such as the angle or number of moments, resulting in a total of 315 cell features measured. Assay prediction machine learning model

[0212] Machine learning-based regression models were constructed on cell-level features extracted from RPE cells, and the trained models could be used to predict TER and VEGF-Ratio measurements. All of these models were tested and trained on the same dataset for comparability. For TER prediction for healthy donor lines treated with aphidicolin or HPI4, the data were divided into test and training data by assigning five culture plates as training data and the sixth as test data (two replicates of each treatment per plate). A similar procedure was performed for VEGF-Ratio prediction, except that a subset of wells from plate 5 was selected and removed from the training set to serve as the test set (one well from each treatment). This was done because ELISA was not performed on the supernatants of all wells, and therefore only a subset of wells was available for this analysis. To eliminate bias in sample selection, the selection of these plates / wells for training / testing was performed blindly before data acquisition or analysis.

[0213] For AMD cell lines, SML models were trained from scratch on different combinations of iRPE monolayers. A total of 18 unique training image subsets were formed, with each image subset containing test data with images of one clone from each donor. The average performance across all 18 subsets was used to identify key features regardless of the machine learning method used or the clone combination used for testing. Table 5 details all combinations that were not included in the training data. Table 5: List of 18 different clonal subsets tested [Table 5-1] [Table 5-2] Feature Preprocessing

[0214] RPE cell features were extracted for each cell. These features included intensity, shape, and texture characteristics. Once extracted, the mean, standard deviation, skewness, and kurtosis of each feature were calculated for each cell set within each image segmentation (described in the "Image Segmentation" section above). A subset of features was found to be highly correlated and have various amplitude ranges. Feature preprocessing was performed to extract highly correlated features and normalize their dynamic range. This preprocessing was applied to all features and consisted of (1) z-normalization and (2) a correlation-based feature selection (CFS) method, which extracted all features with a correlation greater than 99.5%. This was performed independently for all datasets within this manuscript. Therefore, highly correlated features in the combined dataset of AMD-iRPE and Healthy-2 lines were analyzed for cocorrelation separately from the dataset used for AMD-iRPE sample identity (which was analyzed for cocorrelation separately from the Healthy-2-only dataset). This resulted in different total numbers of features used in each of these models. Normalization and pruning of highly correlated variables were also performed on the training / test data, independently of the other data. To assess the ability of DNN-S to accurately segment individual cell features and thus provide the most granular representation of the data, for certain data, z-score normalization and feature correlation were performed on individual cell metrics rather than on the mean, standard deviation, skewness, and kurtosis of cells within the region of interest. Therefore, the number of metrics for this analysis was significantly smaller than for the other models.

[0215] Among the many ML models for classification and prediction, we considered a subset of ML models that predict continuous variables, such as TER or VEGF-Ratio measurements. This subset of ML models includes multilayer perceptrons (MLPs), linear support vector machines (L-SVMs), random forests (RFs), partial least squares regression (PLSR), and ridge regression (RR). For the donor classification task, we chose identical models (except for ridge regression, which does not have a classification format). Therefore, for classification, we replaced ridge regression with a Naive Bayes (NB) classifier. For all models, feature weights were scaled based on their absolute magnitude (0–1), and then averaged across all methods to determine which features had the highest relative weights compared to prediction. Features with high averages suggest relatively important features for prediction, as these important features were consistently identified (i.e., ML model-independent). All models were trained using 10–30 k-fold cross-validation and then tested on "dropout" data not in the training or validation sets within the k-fold cross-validation. Multilayer Perceptron (MLP)

[0216] The MLP model approximates the nonlinear relationship between independent and dependent variables. In the multilayer perceptron model, cell-level features were considered as independent variables, and TER / VEGF values ​​were considered as dependent variables. An MLP model with four layers (two hidden layers) was used, with 70 neurons in the first hidden layer and 40 neurons in the second hidden layer. The number of hidden layers and the number of neurons in each hidden layer were empirically chosen to avoid overfitting the model to the input data and to avoid being too simple to robustly fit the input data. To identify features important for TER / VEGF prediction, Garson's algorithm (Garson, Interpreting neural-network connection weights, AI Expert 6, 46+ (1991)) was used to calculate the relationship weights. Linear Support Vector Machine (L-SVM)

[0217] Support vector machines construct a hyperplane in a high-dimensional space to maximize separation between data. In the case of a linear support vector machine (L-SVM), the plane is linearly correlated with each dimension / feature in that high-dimensional space, and the L-SVM weights can be directly translated into weights / importance for features. The SVM function has a term that optimizes the penalization of the hyperplane for both the "closeness" of the fit to the data (cost) and the distance at which penalization occurs (epsilon). Therefore, while training the model, we iterated the cost and epsilon over a range of 0.001 to 1.2 and 0.001 to 4096, respectively. All model optimization was performed in R using the Liblinear package. Random Forest (RF)

[0218] Random forests use a "forest" or ensemble of decision trees to predict a factor. Decision trees can be thought of as a flow diagram for decision making. Decision trees are constructed from class-labeled training sequences, where each node represents a test on an attribute, each branch represents the outcome of a test, and each terminal node classifies or regresses. Random forests combine multiple deep decision trees trained on different parts of the same training set with the goal of reducing variance in the decision tree output. Thus, the number of trees and the branching depth of each tree can be optimized. Here, 125 trees were evaluated over a range of 5 to 1500 branches to measure optimal model performance. All model optimizations were performed in R using the Caret package. Relative importance of variables was calculated as shown in Liaw, A. et al., Classification and Regression by randomForest, R News 2, 18-22 (2002). Partial Least Squares Regression (PLSR)

[0219] PLSR creates a linear regression model by projecting predicted variables (TER and VEGF-Ratio) and observable variables (cell image features) into a low-dimensional space. The PLSR model finds multidimensional directions in "X" space (predicted variables) that explain the greatest multidimensional variance directions in "Y" space (observable variables). In this way, high-dimensional data can be used to predict real-world outcomes. The only variable to optimize for PLSR is the number of dimensional components required to predict the desired outcome. For the model optimized in this report, the number of components was varied from 1 to 50 to determine which had the highest predictive power. Model optimization was performed in R using the Caret and PLS packages. Variable importance was defined as the sum of the absolute values ​​of each feature coefficient within each dimensional component multiplied by the percent of the total variance each component explained in the raw data. Ridge Regression (RR)

[0220] RR is a specialized form of ordinary least squares regression, in which the predictive model is optimized to minimize the sum of squared residuals between predicted and actual outcomes, and includes a gamma function to eliminate collinear, redundant, or confounding variables through optimization using L2 regularization. As with ordinary least squares regression, variable importance can be determined from the absolute value of the feature coefficient weights. The model was optimized to reduce the number of variables to fewer than 25 total features. Model optimization was performed using the Caret and foba packages in R. DNN model for classifying donor identity

[0221] We created a DNN to determine whether two iRPE QBAM images were from the same donor. The inputs were two 1024 × 1024 × 3 pixel images. The basic layer structure consisted of low-rank expansion as reported by Jaderberg et al., "Speeding up Convolutional Neural Networks with Low Rank Expansions," ArXiv 14053866 Cs (2014). The low-rank expansion layer consisted of a 1 × 3 convolutional layer followed by a 3 × 1 convolutional layer. The first layer had eight neurons, and each subsequent layer doubled the number of neurons compared to the previous layer. Each convolutional operation was followed by a leaky ReLU layer with a 0.1 leakage value. Each low-rank expansion layer was batch normalized and then followed by a 3 × 3 max-pooling layer with a stride of 2. A residual layer was added before each max-pooling operation except for the first layer, where the residual layer consisted of a 1 × 1 convolution operation to scale the input layer to the same size as the output layer. The final layer was a 3 × 3 average pooling layer followed by a fully connected or dense convolutional layer (size = 64 × 64 × 256). The best network was determined based on the F1 score. The simplicity and small size of this network allowed us to train it on the same 18 training / test datasets described in the following section. Machine learning models for classifying donor identity and outliers

[0222] Machine learning-based classification models were constructed on cell-level features extracted from RPE cells, and the trained models could be used to predict clonal outliers or donor identity. All models except the ridge regression model were evaluated for their ability to correctly classify donor identity. A naive Bayesian model was used instead of ridge regression. For clonal outlier prediction, all data were subjected to PCA and clustering of iRPE lines using a hierarchical clustering method.

[0223] For AMD cell lines, we trained the SML model from scratch on different combinations of iRPE monolayers. A total of 18 unique training image subsets were formed, with each image subset containing test data with images of one clone from each donor. Table 5 details all combinations that were not included in the training data. This approach used images of donor-derived clones to attempt to predict the "parent" donor strains of clones the network had not seen before. Principal component analysis (PCA)

[0224] PCA and hierarchical clustering were used to identify developmental outliers. PCA finds the maximum variance hyperplane that best describes a set of correlated variables by using an orthogonal transformation to convert a set of correlated variables into a reduced set of linear, uncorrelated variables called principal components. In large dimensional spaces (e.g., multiple features), PCA is useful for reducing the dimensionality of the data to determine whether the cytometric features as a whole can classify different donors / clones relative to each other. Images were grouped by clone and day level, and the mean, standard deviation, skewness, and kurtosis of each feature were calculated for each clone / day combination. These aggregate features were then subjected to PCA. Principal components 1 and 2 accounted for over 75% of the total variance in the data and were therefore considered preferred indicators of variability across samples. Feature importance was defined as the sum of the absolute values ​​of each feature coefficient within each dimensional component multiplied by the percentage of the total variance each component explained in the raw data. PCA was performed using base R. Hierarchical Clustering

[0225] Hierarchical clustering was performed on the PCA output, where the mean, standard deviation, skewness, and kurtosis of individual features were calculated for each clone. For hierarchical clustering, the Euclidean distance between all clones was calculated across all principal components, and the complete linkage distance was used for clustering. Split heights were selected at three branches representing the three donors used in this study. Clustering was performed using base R functions. Naive Bayes

[0226] The naive Bayes model describes the probability of a cell type being of donor origin based on the prior probability of cells originating from that donor with a given set of features. The naive Bayes model assumes strong (naive) independence between features. That is, the naive Bayes classifier assumes that each feature in a cell image independently contributes to the probability that a cell is of a given donor origin, regardless of possible correlations between features. The model was optimized for the amount of Laplacian correction required, ranging from 0.0 to 0.1, and for whether the features required normal or kernel density. Feature importance was obtained from the absolute value of the feature coefficient in the best-fit model. Model optimization was performed using the Caret and klaR packages in R. Quantitative and statistical analysis

[0227] All significance between groups shown for the albinism strain was performed using linear mixed-effects models controlling for repeated measurements over time from a single well and multiple images taken per well. These models were evaluated using the multicomp and nlme packages in R. 2 Values, confidence intervals and Kolmogorov-Smirnov statistics were calculated in base R.

[0228] In certain illustrative embodiments described above, it should be recognized that the various non-limiting embodiments described herein may be used separately, combined, or selectively combined for particular applications. Moreover, some of the various features of the above non-limiting embodiments may be used without the corresponding use of other features described. Thus, the foregoing description should be considered merely illustrative of the principles, teachings, and exemplary embodiments of the present disclosure, and not limiting thereof.

[0229] It is to be understood that the above-described arrangements are merely illustrative of the application of the principles of the illustrated embodiments. Numerous modifications and alternative arrangements can be devised by those skilled in the art without departing from the scope of the illustrated embodiments, and the appended claims are intended to cover all such modifications and arrangements. The present invention provides, for example, the following items. (Item 1) 1. A method for non-invasively predicting a property of one or more cells and cell derivatives, said method comprising: training a machine learning model using at least one of a plurality of training cell images representing a plurality of cells and data identifying characteristics for the plurality of cells; receiving at least one test cell image representative of at least one test cell to be evaluated, the at least one test cell image being non-invasively acquired and based on absorbance as an absolute measure of light; providing the at least one test cell image to the trained machine learning model; predicting a property of the at least one test cell using machine learning based on the trained machine learning model; and generating release criteria for a clinical cell preparation based on the predicted properties of the at least one test cell using the trained machine learning model. A method comprising: (Item 2) 2. The method of claim 1, wherein the machine learning is performed using a deep neural network, and the method further comprises segmenting the images of the at least one test image into individual cells by the deep neural network. (Item 3) 2. The method of claim 1, wherein the machine learning is performed using a deep neural network, and the method further comprises classifying the at least one test cell based on the characteristic. (Item 4) 4. The method of claim 3, wherein the predicting step further comprises determining at least one of cell identity, cell function, effect of a delivered drug, disease state, and similarity to technical repeats or previously used samples based on the classifying step. (Item 5) Item 10. The method of item 1, wherein the step of predicting the properties of the at least one test cell can be performed on either a single cell or a plurality of cells in a field of view in the at least one test cell image. (Item 6) 2. The method of claim 1, further comprising visually extracting at least one feature from the at least one test cell image, wherein training the machine learning model is performed using the extracted at least one feature, and predicting comprises identifying at least one feature of the at least one test cell using the trained machine learning model trained using the extracted at least one feature, and predicting the property of the at least one test cell using the identified at least one feature. (Item 7) 2. The method of claim 1, wherein the at least one test cell image is obtained using quantitative bright-field absorbance microscopy (QBAM). (Item 8) The method comprises: receiving at least one microscopic image captured by a microscope; and converting pixel intensities of said at least one microscopic image into absorbance values. 8. The method of claim 7, further comprising: (Item 9) The method comprises: calculating an absorbance confidence of the absorbance value; Establishing equilibrium of the microscope by benchmarking; and color filtering when acquiring said image; 8. The method of claim 7, further comprising at least one of: (Item 10) 4. The method of claim 3, wherein the first image processing of the at least one test image is performed by the deep neural network. (Item 11) Item 10. The method according to item 10, wherein the machine learning is performed using a deep neural network, and the method further comprises a step of segmenting the at least one test image into individual cells by the deep neural network, and the features are visually extracted from the segmented individual cells. (Item 12) 2. The method of claim 1, wherein predicting the property of the at least one test cell comprises at least one of predicting transepithelial resistance (TER) and / or polarized secretion of vascular endothelial growth factor (VEGF), predicting a function of the at least one test cell, predicting a maturity of the at least one test cell, predicting whether the at least one test cell is derived from an identified donor, and determining whether the at least one test cell is an outlier relative to a known classification. (Item 13) 2. The method of claim 1, wherein the step of generating release criteria for clinical cell preparations further comprises generating release criteria for drug discovery or drug toxicity using the trained machine learning model. (Item 14) 2. The method of claim 1, wherein the plurality of cells and the at least one test cell comprise at least one of embryonic stem cells (ESCs), induced pluripotent stem cells (iPSCs), neural stem cells (NSCs), retinal pigment epithelial stem cells (RPESCs), mesenchymal stem cells (MSCs), hematopoietic stem cells (HSCs), and cancer stem cells (CSCs). (Item 15) 16. The method of claim 15, wherein the first plurality of cells and the at least one test cell are derived from a plurality of at least one of ESCs, iPSCs, NSCs, RPESCs, MSCs, or HSCs, or any cells derived therefrom. (Item 16) 16. The method of claim 15, wherein the first plurality of cells and the at least one test cell comprise a major cell type derived from human or animal tissue. (Item 17) 2. The method of claim 1, wherein the identified and predicted properties include at least one of physiological properties, molecular properties, cellular properties, and / or biochemical properties. (Item 18) 7. The method of claim 6, wherein the at least one extracted feature comprises at least one of a cell boundary, a cell shape, and a plurality of texture metrics. (Item 19) Item 19. The method of item 18, wherein the plurality of texture metrics comprises a plurality of sub-cellular features. (Item 20) Item 10. The method of item 1, wherein the plurality of cell images and the at least one test cell image comprise a fluorescent image, a chemiluminescent image, an emission image, or a bright field image. (Item 21) determining whether a test image of the at least one test image is a large image; dividing the large image into at least two tiles; individually providing each of the tiles to the trained machine model; combining the outputs of the trained machine model processing associated with each of the tiles; and providing the combined outputs to predict a property of the at least one test cell into a single output corresponding to the larger image. Item 1, the method of claim 1 further comprising: (Item 22) 1. A computing system for non-invasively predicting properties of one or more cells and cell derivatives, said computing system comprising: a memory configured to store instructions; a processor arranged in communication with said memory wherein the processor, upon execution of the instructions, configured to train a machine learning model using at least one of a plurality of training cell images representing a plurality of cells and data identifying characteristics for the plurality of cells; configured to receive at least one test cell image representative of at least one test cell to be evaluated, the at least one test cell image being non-invasively acquired and based on absorbance as an absolute measure of light; configured to provide the at least one test cell image to the trained machine learning model; predicting a property of the at least one test cell using machine learning based on the trained machine learning model; and configuring the trained machine learning model to generate release criteria for a clinical cell preparation based on the predicted property of the at least one test cell. Computing system.

Claims

1. 1. A method for non-invasively predicting retinal pigment epithelial (RPE) cell maturity of one or more cells, the method comprising: obtaining a machine learning model trained with a plurality of training cell images acquired using a fluorescence microscope, the training cell images comprising a plurality of cells and labeled data identifying a cellular maturity level for the plurality of cells, the labeled data comprising a transepithelial electrical resistance of the plurality of cells; receiving at least one test cell image comprising one or more cells to be evaluated, the at least one test cell image being obtained by a fluorescent microscope; providing the at least one test cell image to the trained machine learning model; and predicting the maturity of the one or more cells of the at least one test cell image using machine learning based on the trained machine learning model. A method comprising:

2. The method described in claim 1, wherein the predicted maturity of one or more cells is used as a release criterion for transplantation into clinical cell preparations.

3. 2. The method of claim 1, wherein the machine learning is performed using a deep neural network, the method further comprising segmenting images of the at least one test cell image into individual cells by the deep neural network.

4. The method of claim 1 , wherein the machine learning is performed using a deep neural network.

5. 2. The method of claim 1, wherein predicting the maturity of the at least one test cell can be performed on either a single cell or a field of view of multiple cells in the at least one test cell image.

6. 2. The method of claim 1, further comprising the step of visually extracting at least one feature from the at least one test cell image, wherein the machine learning model is trained using the extracted at least one feature, and the predicting step comprises identifying at least one feature of the at least one test cell using the trained machine learning model trained using the extracted at least one feature, and predicting the maturity of the at least one test cell using the identified at least one feature.

7. The at least one test cell image is acquired using quantitative bright field absorbance microscopy (QBAM), where acquiring the at least one test cell image using QBAM includes receiving at least one microscopic image captured by a microscope; and converting pixel intensities of said at least one microscopic image into absorbance values. The method of claim 1 , comprising:

8. The method further comprising: calculating an absorbance confidence of the absorbance value; Establishing equilibrium of the microscope by benchmarking; and color filtering when acquiring said at least one microscopic image. The method of claim 7 , further comprising at least one of:

9. 2. The method of claim 1, wherein the machine learning is performed using a deep neural network, the method further comprising a step of segmenting an image of the at least one test cell image into individual cells by the deep neural network, and at least one feature is visually extracted from the segmented individual cells.

10. 2. The method of claim 1, wherein predicting the maturity of the at least one test cell comprises at least one of predicting transepithelial resistance (TER) and / or polarized secretion of vascular endothelial growth factor (VEGF).

11. 3. The method of claim 2, wherein the release criteria for clinical cell preparations are for drug discovery or drug toxicity.

12. 2. The method of claim 1, wherein the plurality of cells comprises at least one of embryonic stem cells (ESCs), induced pluripotent stem cells (iPSCs), neural stem cells (NSCs), retinal pigment epithelial stem cells (RPESCs), mesenchymal stem cells (MSCs), hematopoietic stem cells (HSCs), and cancer stem cells (CSCs).

13. 2. The method of claim 1, wherein the plurality of cells is derived from a plurality of at least one of ESCs, iPSCs, NSCs, RPESCs, MSCs or HSCs or any cells derived therefrom.

14. The method of claim 13 , wherein the plurality of cells comprises major cell types derived from human or animal tissue.

15. The method of claim 1 , wherein the predicted maturity comprises at least one of a physiological characteristic, a molecular characteristic, a cellular characteristic and / or a biochemical characteristic.

16. The method of claim 6 , wherein the extracted at least one feature comprises at least one of a cell boundary, a cell shape, and a plurality of texture metrics.

17. The method of claim 16 , wherein the plurality of texture metrics comprises a plurality of sub-cellular features.

18. The method of claim 1 , wherein the plurality of training cell images and the at least one test cell image further comprise a fluorescent image, a chemiluminescent image, a radiographic image, or a bright field image.

19. dividing the test cell image into at least two tiles; providing each of the tiles individually to the trained machine learning model; combining outputs of the trained machine learning models associated with each of the tiles; and providing the combined outputs to predict the maturity of the at least one test cell into a single output corresponding to the test cell image. The method of claim 1 further comprising:

20. 1. A computing system for non-invasively predicting retinal pigment epithelium (RPE) cell maturity of one or more cells, the computing system comprising: A memory configured to store instructions; and A processor configured in communication with the memory wherein the processor, upon execution of the instructions, obtaining a machine learning model trained using a plurality of training cell images acquired using a fluorescence microscope, the training cell images including labeled data identifying a plurality of cells and a cellular maturity level for the plurality of cells, the labeled data including a transepithelial electrical resistance of the plurality of cells; configured to receive at least one test cell image including one or more cells to be evaluated, the at least one test cell image being obtained using a fluorescent microscope; providing the at least one test cell image to the trained machine learning model; and predicting a maturity of the one or more cells of the at least one test cell image using machine learning based on the trained machine learning model. Computing system.

21. The computing system of claim 20, wherein the predicted maturity of the one or more cells is used by the processor as a release criterion for transplantation into a clinical cell preparation.

Images